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/vmacache.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/rbtree_augmented.h>
42#include <linux/notifier.h>
43#include <linux/memory.h>
44#include <linux/printk.h>
45#include <linux/userfaultfd_k.h>
46#include <linux/moduleparam.h>
47#include <linux/pkeys.h>
48#include <linux/oom.h>
49#include <linux/sched/mm.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#include "internal.h"
57
58#ifndef arch_mmap_check
59#define arch_mmap_check(addr, len, flags) (0)
60#endif
61
62#ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
63const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
64const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
65int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
66#endif
67#ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
68const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
69const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
70int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
71#endif
72
73static bool ignore_rlimit_data;
74core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
75
76static void unmap_region(struct mm_struct *mm,
77 struct vm_area_struct *vma, struct vm_area_struct *prev,
78 unsigned long start, unsigned long end);
79
80/* description of effects of mapping type and prot in current implementation.
81 * this is due to the limited x86 page protection hardware. The expected
82 * behavior is in parens:
83 *
84 * map_type prot
85 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
86 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
87 * w: (no) no w: (no) no w: (yes) yes w: (no) no
88 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
89 *
90 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
91 * w: (no) no w: (no) no w: (copy) copy w: (no) no
92 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
93 *
94 * On arm64, PROT_EXEC has the following behaviour for both MAP_SHARED and
95 * MAP_PRIVATE:
96 * r: (no) no
97 * w: (no) no
98 * x: (yes) yes
99 */
100pgprot_t protection_map[16] __ro_after_init = {
101 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
102 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
103};
104
105#ifndef CONFIG_ARCH_HAS_FILTER_PGPROT
106static inline pgprot_t arch_filter_pgprot(pgprot_t prot)
107{
108 return prot;
109}
110#endif
111
112pgprot_t vm_get_page_prot(unsigned long vm_flags)
113{
114 pgprot_t ret = __pgprot(pgprot_val(protection_map[vm_flags &
115 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
116 pgprot_val(arch_vm_get_page_prot(vm_flags)));
117
118 return arch_filter_pgprot(ret);
119}
120EXPORT_SYMBOL(vm_get_page_prot);
121
122static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
123{
124 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
125}
126
127/* Update vma->vm_page_prot to reflect vma->vm_flags. */
128void vma_set_page_prot(struct vm_area_struct *vma)
129{
130 unsigned long vm_flags = vma->vm_flags;
131 pgprot_t vm_page_prot;
132
133 vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
134 if (vma_wants_writenotify(vma, vm_page_prot)) {
135 vm_flags &= ~VM_SHARED;
136 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
137 }
138 /* remove_protection_ptes reads vma->vm_page_prot without mmap_sem */
139 WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
140}
141
142/*
143 * Requires inode->i_mapping->i_mmap_rwsem
144 */
145static void __remove_shared_vm_struct(struct vm_area_struct *vma,
146 struct file *file, struct address_space *mapping)
147{
148 if (vma->vm_flags & VM_DENYWRITE)
149 atomic_inc(&file_inode(file)->i_writecount);
150 if (vma->vm_flags & VM_SHARED)
151 mapping_unmap_writable(mapping);
152
153 flush_dcache_mmap_lock(mapping);
154 vma_interval_tree_remove(vma, &mapping->i_mmap);
155 flush_dcache_mmap_unlock(mapping);
156}
157
158/*
159 * Unlink a file-based vm structure from its interval tree, to hide
160 * vma from rmap and vmtruncate before freeing its page tables.
161 */
162void unlink_file_vma(struct vm_area_struct *vma)
163{
164 struct file *file = vma->vm_file;
165
166 if (file) {
167 struct address_space *mapping = file->f_mapping;
168 i_mmap_lock_write(mapping);
169 __remove_shared_vm_struct(vma, file, mapping);
170 i_mmap_unlock_write(mapping);
171 }
172}
173
174/*
175 * Close a vm structure and free it, returning the next.
176 */
177static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
178{
179 struct vm_area_struct *next = vma->vm_next;
180
181 might_sleep();
182 if (vma->vm_ops && vma->vm_ops->close)
183 vma->vm_ops->close(vma);
184 if (vma->vm_file)
185 fput(vma->vm_file);
186 mpol_put(vma_policy(vma));
187 vm_area_free(vma);
188 return next;
189}
190
191static int do_brk_flags(unsigned long addr, unsigned long request, unsigned long flags,
192 struct list_head *uf);
193SYSCALL_DEFINE1(brk, unsigned long, brk)
194{
195 unsigned long retval;
196 unsigned long newbrk, oldbrk, origbrk;
197 struct mm_struct *mm = current->mm;
198 struct vm_area_struct *next;
199 unsigned long min_brk;
200 bool populate;
201 bool downgraded = false;
202 LIST_HEAD(uf);
203
204 brk = untagged_addr(brk);
205
206 if (down_write_killable(&mm->mmap_sem))
207 return -EINTR;
208
209 origbrk = mm->brk;
210
211#ifdef CONFIG_COMPAT_BRK
212 /*
213 * CONFIG_COMPAT_BRK can still be overridden by setting
214 * randomize_va_space to 2, which will still cause mm->start_brk
215 * to be arbitrarily shifted
216 */
217 if (current->brk_randomized)
218 min_brk = mm->start_brk;
219 else
220 min_brk = mm->end_data;
221#else
222 min_brk = mm->start_brk;
223#endif
224 if (brk < min_brk)
225 goto out;
226
227 /*
228 * Check against rlimit here. If this check is done later after the test
229 * of oldbrk with newbrk then it can escape the test and let the data
230 * segment grow beyond its set limit the in case where the limit is
231 * not page aligned -Ram Gupta
232 */
233 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
234 mm->end_data, mm->start_data))
235 goto out;
236
237 newbrk = PAGE_ALIGN(brk);
238 oldbrk = PAGE_ALIGN(mm->brk);
239 if (oldbrk == newbrk) {
240 mm->brk = brk;
241 goto success;
242 }
243
244 /*
245 * Always allow shrinking brk.
246 * __do_munmap() may downgrade mmap_sem to read.
247 */
248 if (brk <= mm->brk) {
249 int ret;
250
251 /*
252 * mm->brk must to be protected by write mmap_sem so update it
253 * before downgrading mmap_sem. When __do_munmap() fails,
254 * mm->brk will be restored from origbrk.
255 */
256 mm->brk = brk;
257 ret = __do_munmap(mm, newbrk, oldbrk-newbrk, &uf, true);
258 if (ret < 0) {
259 mm->brk = origbrk;
260 goto out;
261 } else if (ret == 1) {
262 downgraded = true;
263 }
264 goto success;
265 }
266
267 /* Check against existing mmap mappings. */
268 next = find_vma(mm, oldbrk);
269 if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
270 goto out;
271
272 /* Ok, looks good - let it rip. */
273 if (do_brk_flags(oldbrk, newbrk-oldbrk, 0, &uf) < 0)
274 goto out;
275 mm->brk = brk;
276
277success:
278 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
279 if (downgraded)
280 up_read(&mm->mmap_sem);
281 else
282 up_write(&mm->mmap_sem);
283 userfaultfd_unmap_complete(mm, &uf);
284 if (populate)
285 mm_populate(oldbrk, newbrk - oldbrk);
286 return brk;
287
288out:
289 retval = origbrk;
290 up_write(&mm->mmap_sem);
291 return retval;
292}
293
294static inline unsigned long vma_compute_gap(struct vm_area_struct *vma)
295{
296 unsigned long gap, prev_end;
297
298 /*
299 * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
300 * allow two stack_guard_gaps between them here, and when choosing
301 * an unmapped area; whereas when expanding we only require one.
302 * That's a little inconsistent, but keeps the code here simpler.
303 */
304 gap = vm_start_gap(vma);
305 if (vma->vm_prev) {
306 prev_end = vm_end_gap(vma->vm_prev);
307 if (gap > prev_end)
308 gap -= prev_end;
309 else
310 gap = 0;
311 }
312 return gap;
313}
314
315#ifdef CONFIG_DEBUG_VM_RB
316static unsigned long vma_compute_subtree_gap(struct vm_area_struct *vma)
317{
318 unsigned long max = vma_compute_gap(vma), subtree_gap;
319 if (vma->vm_rb.rb_left) {
320 subtree_gap = rb_entry(vma->vm_rb.rb_left,
321 struct vm_area_struct, vm_rb)->rb_subtree_gap;
322 if (subtree_gap > max)
323 max = subtree_gap;
324 }
325 if (vma->vm_rb.rb_right) {
326 subtree_gap = rb_entry(vma->vm_rb.rb_right,
327 struct vm_area_struct, vm_rb)->rb_subtree_gap;
328 if (subtree_gap > max)
329 max = subtree_gap;
330 }
331 return max;
332}
333
334static int browse_rb(struct mm_struct *mm)
335{
336 struct rb_root *root = &mm->mm_rb;
337 int i = 0, j, bug = 0;
338 struct rb_node *nd, *pn = NULL;
339 unsigned long prev = 0, pend = 0;
340
341 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
342 struct vm_area_struct *vma;
343 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
344 if (vma->vm_start < prev) {
345 pr_emerg("vm_start %lx < prev %lx\n",
346 vma->vm_start, prev);
347 bug = 1;
348 }
349 if (vma->vm_start < pend) {
350 pr_emerg("vm_start %lx < pend %lx\n",
351 vma->vm_start, pend);
352 bug = 1;
353 }
354 if (vma->vm_start > vma->vm_end) {
355 pr_emerg("vm_start %lx > vm_end %lx\n",
356 vma->vm_start, vma->vm_end);
357 bug = 1;
358 }
359 spin_lock(&mm->page_table_lock);
360 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
361 pr_emerg("free gap %lx, correct %lx\n",
362 vma->rb_subtree_gap,
363 vma_compute_subtree_gap(vma));
364 bug = 1;
365 }
366 spin_unlock(&mm->page_table_lock);
367 i++;
368 pn = nd;
369 prev = vma->vm_start;
370 pend = vma->vm_end;
371 }
372 j = 0;
373 for (nd = pn; nd; nd = rb_prev(nd))
374 j++;
375 if (i != j) {
376 pr_emerg("backwards %d, forwards %d\n", j, i);
377 bug = 1;
378 }
379 return bug ? -1 : i;
380}
381
382static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
383{
384 struct rb_node *nd;
385
386 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
387 struct vm_area_struct *vma;
388 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
389 VM_BUG_ON_VMA(vma != ignore &&
390 vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
391 vma);
392 }
393}
394
395static void validate_mm(struct mm_struct *mm)
396{
397 int bug = 0;
398 int i = 0;
399 unsigned long highest_address = 0;
400 struct vm_area_struct *vma = mm->mmap;
401
402 while (vma) {
403 struct anon_vma *anon_vma = vma->anon_vma;
404 struct anon_vma_chain *avc;
405
406 if (anon_vma) {
407 anon_vma_lock_read(anon_vma);
408 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
409 anon_vma_interval_tree_verify(avc);
410 anon_vma_unlock_read(anon_vma);
411 }
412
413 highest_address = vm_end_gap(vma);
414 vma = vma->vm_next;
415 i++;
416 }
417 if (i != mm->map_count) {
418 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
419 bug = 1;
420 }
421 if (highest_address != mm->highest_vm_end) {
422 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
423 mm->highest_vm_end, highest_address);
424 bug = 1;
425 }
426 i = browse_rb(mm);
427 if (i != mm->map_count) {
428 if (i != -1)
429 pr_emerg("map_count %d rb %d\n", mm->map_count, i);
430 bug = 1;
431 }
432 VM_BUG_ON_MM(bug, mm);
433}
434#else
435#define validate_mm_rb(root, ignore) do { } while (0)
436#define validate_mm(mm) do { } while (0)
437#endif
438
439RB_DECLARE_CALLBACKS_MAX(static, vma_gap_callbacks,
440 struct vm_area_struct, vm_rb,
441 unsigned long, rb_subtree_gap, vma_compute_gap)
442
443/*
444 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
445 * vma->vm_prev->vm_end values changed, without modifying the vma's position
446 * in the rbtree.
447 */
448static void vma_gap_update(struct vm_area_struct *vma)
449{
450 /*
451 * As it turns out, RB_DECLARE_CALLBACKS_MAX() already created
452 * a callback function that does exactly what we want.
453 */
454 vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
455}
456
457static inline void vma_rb_insert(struct vm_area_struct *vma,
458 struct rb_root *root)
459{
460 /* All rb_subtree_gap values must be consistent prior to insertion */
461 validate_mm_rb(root, NULL);
462
463 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
464}
465
466static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
467{
468 /*
469 * Note rb_erase_augmented is a fairly large inline function,
470 * so make sure we instantiate it only once with our desired
471 * augmented rbtree callbacks.
472 */
473 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
474}
475
476static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
477 struct rb_root *root,
478 struct vm_area_struct *ignore)
479{
480 /*
481 * All rb_subtree_gap values must be consistent prior to erase,
482 * with the possible exception of the "next" vma being erased if
483 * next->vm_start was reduced.
484 */
485 validate_mm_rb(root, ignore);
486
487 __vma_rb_erase(vma, root);
488}
489
490static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
491 struct rb_root *root)
492{
493 /*
494 * All rb_subtree_gap values must be consistent prior to erase,
495 * with the possible exception of the vma being erased.
496 */
497 validate_mm_rb(root, vma);
498
499 __vma_rb_erase(vma, root);
500}
501
502/*
503 * vma has some anon_vma assigned, and is already inserted on that
504 * anon_vma's interval trees.
505 *
506 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
507 * vma must be removed from the anon_vma's interval trees using
508 * anon_vma_interval_tree_pre_update_vma().
509 *
510 * After the update, the vma will be reinserted using
511 * anon_vma_interval_tree_post_update_vma().
512 *
513 * The entire update must be protected by exclusive mmap_sem and by
514 * the root anon_vma's mutex.
515 */
516static inline void
517anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
518{
519 struct anon_vma_chain *avc;
520
521 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
522 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
523}
524
525static inline void
526anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
527{
528 struct anon_vma_chain *avc;
529
530 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
531 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
532}
533
534static int find_vma_links(struct mm_struct *mm, unsigned long addr,
535 unsigned long end, struct vm_area_struct **pprev,
536 struct rb_node ***rb_link, struct rb_node **rb_parent)
537{
538 struct rb_node **__rb_link, *__rb_parent, *rb_prev;
539
540 __rb_link = &mm->mm_rb.rb_node;
541 rb_prev = __rb_parent = NULL;
542
543 while (*__rb_link) {
544 struct vm_area_struct *vma_tmp;
545
546 __rb_parent = *__rb_link;
547 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
548
549 if (vma_tmp->vm_end > addr) {
550 /* Fail if an existing vma overlaps the area */
551 if (vma_tmp->vm_start < end)
552 return -ENOMEM;
553 __rb_link = &__rb_parent->rb_left;
554 } else {
555 rb_prev = __rb_parent;
556 __rb_link = &__rb_parent->rb_right;
557 }
558 }
559
560 *pprev = NULL;
561 if (rb_prev)
562 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
563 *rb_link = __rb_link;
564 *rb_parent = __rb_parent;
565 return 0;
566}
567
568static unsigned long count_vma_pages_range(struct mm_struct *mm,
569 unsigned long addr, unsigned long end)
570{
571 unsigned long nr_pages = 0;
572 struct vm_area_struct *vma;
573
574 /* Find first overlaping mapping */
575 vma = find_vma_intersection(mm, addr, end);
576 if (!vma)
577 return 0;
578
579 nr_pages = (min(end, vma->vm_end) -
580 max(addr, vma->vm_start)) >> PAGE_SHIFT;
581
582 /* Iterate over the rest of the overlaps */
583 for (vma = vma->vm_next; vma; vma = vma->vm_next) {
584 unsigned long overlap_len;
585
586 if (vma->vm_start > end)
587 break;
588
589 overlap_len = min(end, vma->vm_end) - vma->vm_start;
590 nr_pages += overlap_len >> PAGE_SHIFT;
591 }
592
593 return nr_pages;
594}
595
596void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
597 struct rb_node **rb_link, struct rb_node *rb_parent)
598{
599 /* Update tracking information for the gap following the new vma. */
600 if (vma->vm_next)
601 vma_gap_update(vma->vm_next);
602 else
603 mm->highest_vm_end = vm_end_gap(vma);
604
605 /*
606 * vma->vm_prev wasn't known when we followed the rbtree to find the
607 * correct insertion point for that vma. As a result, we could not
608 * update the vma vm_rb parents rb_subtree_gap values on the way down.
609 * So, we first insert the vma with a zero rb_subtree_gap value
610 * (to be consistent with what we did on the way down), and then
611 * immediately update the gap to the correct value. Finally we
612 * rebalance the rbtree after all augmented values have been set.
613 */
614 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
615 vma->rb_subtree_gap = 0;
616 vma_gap_update(vma);
617 vma_rb_insert(vma, &mm->mm_rb);
618}
619
620static void __vma_link_file(struct vm_area_struct *vma)
621{
622 struct file *file;
623
624 file = vma->vm_file;
625 if (file) {
626 struct address_space *mapping = file->f_mapping;
627
628 if (vma->vm_flags & VM_DENYWRITE)
629 atomic_dec(&file_inode(file)->i_writecount);
630 if (vma->vm_flags & VM_SHARED)
631 atomic_inc(&mapping->i_mmap_writable);
632
633 flush_dcache_mmap_lock(mapping);
634 vma_interval_tree_insert(vma, &mapping->i_mmap);
635 flush_dcache_mmap_unlock(mapping);
636 }
637}
638
639static void
640__vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
641 struct vm_area_struct *prev, struct rb_node **rb_link,
642 struct rb_node *rb_parent)
643{
644 __vma_link_list(mm, vma, prev, rb_parent);
645 __vma_link_rb(mm, vma, rb_link, rb_parent);
646}
647
648static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
649 struct vm_area_struct *prev, struct rb_node **rb_link,
650 struct rb_node *rb_parent)
651{
652 struct address_space *mapping = NULL;
653
654 if (vma->vm_file) {
655 mapping = vma->vm_file->f_mapping;
656 i_mmap_lock_write(mapping);
657 }
658
659 __vma_link(mm, vma, prev, rb_link, rb_parent);
660 __vma_link_file(vma);
661
662 if (mapping)
663 i_mmap_unlock_write(mapping);
664
665 mm->map_count++;
666 validate_mm(mm);
667}
668
669/*
670 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
671 * mm's list and rbtree. It has already been inserted into the interval tree.
672 */
673static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
674{
675 struct vm_area_struct *prev;
676 struct rb_node **rb_link, *rb_parent;
677
678 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
679 &prev, &rb_link, &rb_parent))
680 BUG();
681 __vma_link(mm, vma, prev, rb_link, rb_parent);
682 mm->map_count++;
683}
684
685static __always_inline void __vma_unlink_common(struct mm_struct *mm,
686 struct vm_area_struct *vma,
687 struct vm_area_struct *prev,
688 bool has_prev,
689 struct vm_area_struct *ignore)
690{
691 struct vm_area_struct *next;
692
693 vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
694 next = vma->vm_next;
695 if (has_prev)
696 prev->vm_next = next;
697 else {
698 prev = vma->vm_prev;
699 if (prev)
700 prev->vm_next = next;
701 else
702 mm->mmap = next;
703 }
704 if (next)
705 next->vm_prev = prev;
706
707 /* Kill the cache */
708 vmacache_invalidate(mm);
709}
710
711static inline void __vma_unlink_prev(struct mm_struct *mm,
712 struct vm_area_struct *vma,
713 struct vm_area_struct *prev)
714{
715 __vma_unlink_common(mm, vma, prev, true, vma);
716}
717
718/*
719 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
720 * is already present in an i_mmap tree without adjusting the tree.
721 * The following helper function should be used when such adjustments
722 * are necessary. The "insert" vma (if any) is to be inserted
723 * before we drop the necessary locks.
724 */
725int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
726 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
727 struct vm_area_struct *expand)
728{
729 struct mm_struct *mm = vma->vm_mm;
730 struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
731 struct address_space *mapping = NULL;
732 struct rb_root_cached *root = NULL;
733 struct anon_vma *anon_vma = NULL;
734 struct file *file = vma->vm_file;
735 bool start_changed = false, end_changed = false;
736 long adjust_next = 0;
737 int remove_next = 0;
738
739 if (next && !insert) {
740 struct vm_area_struct *exporter = NULL, *importer = NULL;
741
742 if (end >= next->vm_end) {
743 /*
744 * vma expands, overlapping all the next, and
745 * perhaps the one after too (mprotect case 6).
746 * The only other cases that gets here are
747 * case 1, case 7 and case 8.
748 */
749 if (next == expand) {
750 /*
751 * The only case where we don't expand "vma"
752 * and we expand "next" instead is case 8.
753 */
754 VM_WARN_ON(end != next->vm_end);
755 /*
756 * remove_next == 3 means we're
757 * removing "vma" and that to do so we
758 * swapped "vma" and "next".
759 */
760 remove_next = 3;
761 VM_WARN_ON(file != next->vm_file);
762 swap(vma, next);
763 } else {
764 VM_WARN_ON(expand != vma);
765 /*
766 * case 1, 6, 7, remove_next == 2 is case 6,
767 * remove_next == 1 is case 1 or 7.
768 */
769 remove_next = 1 + (end > next->vm_end);
770 VM_WARN_ON(remove_next == 2 &&
771 end != next->vm_next->vm_end);
772 VM_WARN_ON(remove_next == 1 &&
773 end != next->vm_end);
774 /* trim end to next, for case 6 first pass */
775 end = next->vm_end;
776 }
777
778 exporter = next;
779 importer = vma;
780
781 /*
782 * If next doesn't have anon_vma, import from vma after
783 * next, if the vma overlaps with it.
784 */
785 if (remove_next == 2 && !next->anon_vma)
786 exporter = next->vm_next;
787
788 } else if (end > next->vm_start) {
789 /*
790 * vma expands, overlapping part of the next:
791 * mprotect case 5 shifting the boundary up.
792 */
793 adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
794 exporter = next;
795 importer = vma;
796 VM_WARN_ON(expand != importer);
797 } else if (end < vma->vm_end) {
798 /*
799 * vma shrinks, and !insert tells it's not
800 * split_vma inserting another: so it must be
801 * mprotect case 4 shifting the boundary down.
802 */
803 adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT);
804 exporter = vma;
805 importer = next;
806 VM_WARN_ON(expand != importer);
807 }
808
809 /*
810 * Easily overlooked: when mprotect shifts the boundary,
811 * make sure the expanding vma has anon_vma set if the
812 * shrinking vma had, to cover any anon pages imported.
813 */
814 if (exporter && exporter->anon_vma && !importer->anon_vma) {
815 int error;
816
817 importer->anon_vma = exporter->anon_vma;
818 error = anon_vma_clone(importer, exporter);
819 if (error)
820 return error;
821 }
822 }
823again:
824 vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
825
826 if (file) {
827 mapping = file->f_mapping;
828 root = &mapping->i_mmap;
829 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
830
831 if (adjust_next)
832 uprobe_munmap(next, next->vm_start, next->vm_end);
833
834 i_mmap_lock_write(mapping);
835 if (insert) {
836 /*
837 * Put into interval tree now, so instantiated pages
838 * are visible to arm/parisc __flush_dcache_page
839 * throughout; but we cannot insert into address
840 * space until vma start or end is updated.
841 */
842 __vma_link_file(insert);
843 }
844 }
845
846 anon_vma = vma->anon_vma;
847 if (!anon_vma && adjust_next)
848 anon_vma = next->anon_vma;
849 if (anon_vma) {
850 VM_WARN_ON(adjust_next && next->anon_vma &&
851 anon_vma != next->anon_vma);
852 anon_vma_lock_write(anon_vma);
853 anon_vma_interval_tree_pre_update_vma(vma);
854 if (adjust_next)
855 anon_vma_interval_tree_pre_update_vma(next);
856 }
857
858 if (root) {
859 flush_dcache_mmap_lock(mapping);
860 vma_interval_tree_remove(vma, root);
861 if (adjust_next)
862 vma_interval_tree_remove(next, root);
863 }
864
865 if (start != vma->vm_start) {
866 vma->vm_start = start;
867 start_changed = true;
868 }
869 if (end != vma->vm_end) {
870 vma->vm_end = end;
871 end_changed = true;
872 }
873 vma->vm_pgoff = pgoff;
874 if (adjust_next) {
875 next->vm_start += adjust_next << PAGE_SHIFT;
876 next->vm_pgoff += adjust_next;
877 }
878
879 if (root) {
880 if (adjust_next)
881 vma_interval_tree_insert(next, root);
882 vma_interval_tree_insert(vma, root);
883 flush_dcache_mmap_unlock(mapping);
884 }
885
886 if (remove_next) {
887 /*
888 * vma_merge has merged next into vma, and needs
889 * us to remove next before dropping the locks.
890 */
891 if (remove_next != 3)
892 __vma_unlink_prev(mm, next, vma);
893 else
894 /*
895 * vma is not before next if they've been
896 * swapped.
897 *
898 * pre-swap() next->vm_start was reduced so
899 * tell validate_mm_rb to ignore pre-swap()
900 * "next" (which is stored in post-swap()
901 * "vma").
902 */
903 __vma_unlink_common(mm, next, NULL, false, vma);
904 if (file)
905 __remove_shared_vm_struct(next, file, mapping);
906 } else if (insert) {
907 /*
908 * split_vma has split insert from vma, and needs
909 * us to insert it before dropping the locks
910 * (it may either follow vma or precede it).
911 */
912 __insert_vm_struct(mm, insert);
913 } else {
914 if (start_changed)
915 vma_gap_update(vma);
916 if (end_changed) {
917 if (!next)
918 mm->highest_vm_end = vm_end_gap(vma);
919 else if (!adjust_next)
920 vma_gap_update(next);
921 }
922 }
923
924 if (anon_vma) {
925 anon_vma_interval_tree_post_update_vma(vma);
926 if (adjust_next)
927 anon_vma_interval_tree_post_update_vma(next);
928 anon_vma_unlock_write(anon_vma);
929 }
930 if (mapping)
931 i_mmap_unlock_write(mapping);
932
933 if (root) {
934 uprobe_mmap(vma);
935
936 if (adjust_next)
937 uprobe_mmap(next);
938 }
939
940 if (remove_next) {
941 if (file) {
942 uprobe_munmap(next, next->vm_start, next->vm_end);
943 fput(file);
944 }
945 if (next->anon_vma)
946 anon_vma_merge(vma, next);
947 mm->map_count--;
948 mpol_put(vma_policy(next));
949 vm_area_free(next);
950 /*
951 * In mprotect's case 6 (see comments on vma_merge),
952 * we must remove another next too. It would clutter
953 * up the code too much to do both in one go.
954 */
955 if (remove_next != 3) {
956 /*
957 * If "next" was removed and vma->vm_end was
958 * expanded (up) over it, in turn
959 * "next->vm_prev->vm_end" changed and the
960 * "vma->vm_next" gap must be updated.
961 */
962 next = vma->vm_next;
963 } else {
964 /*
965 * For the scope of the comment "next" and
966 * "vma" considered pre-swap(): if "vma" was
967 * removed, next->vm_start was expanded (down)
968 * over it and the "next" gap must be updated.
969 * Because of the swap() the post-swap() "vma"
970 * actually points to pre-swap() "next"
971 * (post-swap() "next" as opposed is now a
972 * dangling pointer).
973 */
974 next = vma;
975 }
976 if (remove_next == 2) {
977 remove_next = 1;
978 end = next->vm_end;
979 goto again;
980 }
981 else if (next)
982 vma_gap_update(next);
983 else {
984 /*
985 * If remove_next == 2 we obviously can't
986 * reach this path.
987 *
988 * If remove_next == 3 we can't reach this
989 * path because pre-swap() next is always not
990 * NULL. pre-swap() "next" is not being
991 * removed and its next->vm_end is not altered
992 * (and furthermore "end" already matches
993 * next->vm_end in remove_next == 3).
994 *
995 * We reach this only in the remove_next == 1
996 * case if the "next" vma that was removed was
997 * the highest vma of the mm. However in such
998 * case next->vm_end == "end" and the extended
999 * "vma" has vma->vm_end == next->vm_end so
1000 * mm->highest_vm_end doesn't need any update
1001 * in remove_next == 1 case.
1002 */
1003 VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
1004 }
1005 }
1006 if (insert && file)
1007 uprobe_mmap(insert);
1008
1009 validate_mm(mm);
1010
1011 return 0;
1012}
1013
1014/*
1015 * If the vma has a ->close operation then the driver probably needs to release
1016 * per-vma resources, so we don't attempt to merge those.
1017 */
1018static inline int is_mergeable_vma(struct vm_area_struct *vma,
1019 struct file *file, unsigned long vm_flags,
1020 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1021{
1022 /*
1023 * VM_SOFTDIRTY should not prevent from VMA merging, if we
1024 * match the flags but dirty bit -- the caller should mark
1025 * merged VMA as dirty. If dirty bit won't be excluded from
1026 * comparison, we increase pressure on the memory system forcing
1027 * the kernel to generate new VMAs when old one could be
1028 * extended instead.
1029 */
1030 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
1031 return 0;
1032 if (vma->vm_file != file)
1033 return 0;
1034 if (vma->vm_ops && vma->vm_ops->close)
1035 return 0;
1036 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
1037 return 0;
1038 return 1;
1039}
1040
1041static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
1042 struct anon_vma *anon_vma2,
1043 struct vm_area_struct *vma)
1044{
1045 /*
1046 * The list_is_singular() test is to avoid merging VMA cloned from
1047 * parents. This can improve scalability caused by anon_vma lock.
1048 */
1049 if ((!anon_vma1 || !anon_vma2) && (!vma ||
1050 list_is_singular(&vma->anon_vma_chain)))
1051 return 1;
1052 return anon_vma1 == anon_vma2;
1053}
1054
1055/*
1056 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1057 * in front of (at a lower virtual address and file offset than) the vma.
1058 *
1059 * We cannot merge two vmas if they have differently assigned (non-NULL)
1060 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1061 *
1062 * We don't check here for the merged mmap wrapping around the end of pagecache
1063 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
1064 * wrap, nor mmaps which cover the final page at index -1UL.
1065 */
1066static int
1067can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1068 struct anon_vma *anon_vma, struct file *file,
1069 pgoff_t vm_pgoff,
1070 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1071{
1072 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1073 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1074 if (vma->vm_pgoff == vm_pgoff)
1075 return 1;
1076 }
1077 return 0;
1078}
1079
1080/*
1081 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1082 * beyond (at a higher virtual address and file offset than) the vma.
1083 *
1084 * We cannot merge two vmas if they have differently assigned (non-NULL)
1085 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1086 */
1087static int
1088can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1089 struct anon_vma *anon_vma, struct file *file,
1090 pgoff_t vm_pgoff,
1091 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1092{
1093 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1094 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1095 pgoff_t vm_pglen;
1096 vm_pglen = vma_pages(vma);
1097 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1098 return 1;
1099 }
1100 return 0;
1101}
1102
1103/*
1104 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1105 * whether that can be merged with its predecessor or its successor.
1106 * Or both (it neatly fills a hole).
1107 *
1108 * In most cases - when called for mmap, brk or mremap - [addr,end) is
1109 * certain not to be mapped by the time vma_merge is called; but when
1110 * called for mprotect, it is certain to be already mapped (either at
1111 * an offset within prev, or at the start of next), and the flags of
1112 * this area are about to be changed to vm_flags - and the no-change
1113 * case has already been eliminated.
1114 *
1115 * The following mprotect cases have to be considered, where AAAA is
1116 * the area passed down from mprotect_fixup, never extending beyond one
1117 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1118 *
1119 * AAAA AAAA AAAA AAAA
1120 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
1121 * cannot merge might become might become might become
1122 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
1123 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
1124 * mremap move: PPPPXXXXXXXX 8
1125 * AAAA
1126 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
1127 * might become case 1 below case 2 below case 3 below
1128 *
1129 * It is important for case 8 that the vma NNNN overlapping the
1130 * region AAAA is never going to extended over XXXX. Instead XXXX must
1131 * be extended in region AAAA and NNNN must be removed. This way in
1132 * all cases where vma_merge succeeds, the moment vma_adjust drops the
1133 * rmap_locks, the properties of the merged vma will be already
1134 * correct for the whole merged range. Some of those properties like
1135 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1136 * be correct for the whole merged range immediately after the
1137 * rmap_locks are released. Otherwise if XXXX would be removed and
1138 * NNNN would be extended over the XXXX range, remove_migration_ptes
1139 * or other rmap walkers (if working on addresses beyond the "end"
1140 * parameter) may establish ptes with the wrong permissions of NNNN
1141 * instead of the right permissions of XXXX.
1142 */
1143struct vm_area_struct *vma_merge(struct mm_struct *mm,
1144 struct vm_area_struct *prev, unsigned long addr,
1145 unsigned long end, unsigned long vm_flags,
1146 struct anon_vma *anon_vma, struct file *file,
1147 pgoff_t pgoff, struct mempolicy *policy,
1148 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1149{
1150 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1151 struct vm_area_struct *area, *next;
1152 int err;
1153
1154 /*
1155 * We later require that vma->vm_flags == vm_flags,
1156 * so this tests vma->vm_flags & VM_SPECIAL, too.
1157 */
1158 if (vm_flags & VM_SPECIAL)
1159 return NULL;
1160
1161 if (prev)
1162 next = prev->vm_next;
1163 else
1164 next = mm->mmap;
1165 area = next;
1166 if (area && area->vm_end == end) /* cases 6, 7, 8 */
1167 next = next->vm_next;
1168
1169 /* verify some invariant that must be enforced by the caller */
1170 VM_WARN_ON(prev && addr <= prev->vm_start);
1171 VM_WARN_ON(area && end > area->vm_end);
1172 VM_WARN_ON(addr >= end);
1173
1174 /*
1175 * Can it merge with the predecessor?
1176 */
1177 if (prev && prev->vm_end == addr &&
1178 mpol_equal(vma_policy(prev), policy) &&
1179 can_vma_merge_after(prev, vm_flags,
1180 anon_vma, file, pgoff,
1181 vm_userfaultfd_ctx)) {
1182 /*
1183 * OK, it can. Can we now merge in the successor as well?
1184 */
1185 if (next && end == next->vm_start &&
1186 mpol_equal(policy, vma_policy(next)) &&
1187 can_vma_merge_before(next, vm_flags,
1188 anon_vma, file,
1189 pgoff+pglen,
1190 vm_userfaultfd_ctx) &&
1191 is_mergeable_anon_vma(prev->anon_vma,
1192 next->anon_vma, NULL)) {
1193 /* cases 1, 6 */
1194 err = __vma_adjust(prev, prev->vm_start,
1195 next->vm_end, prev->vm_pgoff, NULL,
1196 prev);
1197 } else /* cases 2, 5, 7 */
1198 err = __vma_adjust(prev, prev->vm_start,
1199 end, prev->vm_pgoff, NULL, prev);
1200 if (err)
1201 return NULL;
1202 khugepaged_enter_vma_merge(prev, vm_flags);
1203 return prev;
1204 }
1205
1206 /*
1207 * Can this new request be merged in front of next?
1208 */
1209 if (next && end == next->vm_start &&
1210 mpol_equal(policy, vma_policy(next)) &&
1211 can_vma_merge_before(next, vm_flags,
1212 anon_vma, file, pgoff+pglen,
1213 vm_userfaultfd_ctx)) {
1214 if (prev && addr < prev->vm_end) /* case 4 */
1215 err = __vma_adjust(prev, prev->vm_start,
1216 addr, prev->vm_pgoff, NULL, next);
1217 else { /* cases 3, 8 */
1218 err = __vma_adjust(area, addr, next->vm_end,
1219 next->vm_pgoff - pglen, NULL, next);
1220 /*
1221 * In case 3 area is already equal to next and
1222 * this is a noop, but in case 8 "area" has
1223 * been removed and next was expanded over it.
1224 */
1225 area = next;
1226 }
1227 if (err)
1228 return NULL;
1229 khugepaged_enter_vma_merge(area, vm_flags);
1230 return area;
1231 }
1232
1233 return NULL;
1234}
1235
1236/*
1237 * Rough compatbility check to quickly see if it's even worth looking
1238 * at sharing an anon_vma.
1239 *
1240 * They need to have the same vm_file, and the flags can only differ
1241 * in things that mprotect may change.
1242 *
1243 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1244 * we can merge the two vma's. For example, we refuse to merge a vma if
1245 * there is a vm_ops->close() function, because that indicates that the
1246 * driver is doing some kind of reference counting. But that doesn't
1247 * really matter for the anon_vma sharing case.
1248 */
1249static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1250{
1251 return a->vm_end == b->vm_start &&
1252 mpol_equal(vma_policy(a), vma_policy(b)) &&
1253 a->vm_file == b->vm_file &&
1254 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) &&
1255 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1256}
1257
1258/*
1259 * Do some basic sanity checking to see if we can re-use the anon_vma
1260 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1261 * the same as 'old', the other will be the new one that is trying
1262 * to share the anon_vma.
1263 *
1264 * NOTE! This runs with mm_sem held for reading, so it is possible that
1265 * the anon_vma of 'old' is concurrently in the process of being set up
1266 * by another page fault trying to merge _that_. But that's ok: if it
1267 * is being set up, that automatically means that it will be a singleton
1268 * acceptable for merging, so we can do all of this optimistically. But
1269 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1270 *
1271 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1272 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1273 * is to return an anon_vma that is "complex" due to having gone through
1274 * a fork).
1275 *
1276 * We also make sure that the two vma's are compatible (adjacent,
1277 * and with the same memory policies). That's all stable, even with just
1278 * a read lock on the mm_sem.
1279 */
1280static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1281{
1282 if (anon_vma_compatible(a, b)) {
1283 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1284
1285 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1286 return anon_vma;
1287 }
1288 return NULL;
1289}
1290
1291/*
1292 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1293 * neighbouring vmas for a suitable anon_vma, before it goes off
1294 * to allocate a new anon_vma. It checks because a repetitive
1295 * sequence of mprotects and faults may otherwise lead to distinct
1296 * anon_vmas being allocated, preventing vma merge in subsequent
1297 * mprotect.
1298 */
1299struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1300{
1301 struct anon_vma *anon_vma;
1302 struct vm_area_struct *near;
1303
1304 near = vma->vm_next;
1305 if (!near)
1306 goto try_prev;
1307
1308 anon_vma = reusable_anon_vma(near, vma, near);
1309 if (anon_vma)
1310 return anon_vma;
1311try_prev:
1312 near = vma->vm_prev;
1313 if (!near)
1314 goto none;
1315
1316 anon_vma = reusable_anon_vma(near, near, vma);
1317 if (anon_vma)
1318 return anon_vma;
1319none:
1320 /*
1321 * There's no absolute need to look only at touching neighbours:
1322 * we could search further afield for "compatible" anon_vmas.
1323 * But it would probably just be a waste of time searching,
1324 * or lead to too many vmas hanging off the same anon_vma.
1325 * We're trying to allow mprotect remerging later on,
1326 * not trying to minimize memory used for anon_vmas.
1327 */
1328 return NULL;
1329}
1330
1331/*
1332 * If a hint addr is less than mmap_min_addr change hint to be as
1333 * low as possible but still greater than mmap_min_addr
1334 */
1335static inline unsigned long round_hint_to_min(unsigned long hint)
1336{
1337 hint &= PAGE_MASK;
1338 if (((void *)hint != NULL) &&
1339 (hint < mmap_min_addr))
1340 return PAGE_ALIGN(mmap_min_addr);
1341 return hint;
1342}
1343
1344static inline int mlock_future_check(struct mm_struct *mm,
1345 unsigned long flags,
1346 unsigned long len)
1347{
1348 unsigned long locked, lock_limit;
1349
1350 /* mlock MCL_FUTURE? */
1351 if (flags & VM_LOCKED) {
1352 locked = len >> PAGE_SHIFT;
1353 locked += mm->locked_vm;
1354 lock_limit = rlimit(RLIMIT_MEMLOCK);
1355 lock_limit >>= PAGE_SHIFT;
1356 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1357 return -EAGAIN;
1358 }
1359 return 0;
1360}
1361
1362static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1363{
1364 if (S_ISREG(inode->i_mode))
1365 return MAX_LFS_FILESIZE;
1366
1367 if (S_ISBLK(inode->i_mode))
1368 return MAX_LFS_FILESIZE;
1369
1370 if (S_ISSOCK(inode->i_mode))
1371 return MAX_LFS_FILESIZE;
1372
1373 /* Special "we do even unsigned file positions" case */
1374 if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1375 return 0;
1376
1377 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1378 return ULONG_MAX;
1379}
1380
1381static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1382 unsigned long pgoff, unsigned long len)
1383{
1384 u64 maxsize = file_mmap_size_max(file, inode);
1385
1386 if (maxsize && len > maxsize)
1387 return false;
1388 maxsize -= len;
1389 if (pgoff > maxsize >> PAGE_SHIFT)
1390 return false;
1391 return true;
1392}
1393
1394/*
1395 * The caller must hold down_write(¤t->mm->mmap_sem).
1396 */
1397unsigned long do_mmap(struct file *file, unsigned long addr,
1398 unsigned long len, unsigned long prot,
1399 unsigned long flags, vm_flags_t vm_flags,
1400 unsigned long pgoff, unsigned long *populate,
1401 struct list_head *uf)
1402{
1403 struct mm_struct *mm = current->mm;
1404 int pkey = 0;
1405
1406 *populate = 0;
1407
1408 if (!len)
1409 return -EINVAL;
1410
1411 /*
1412 * Does the application expect PROT_READ to imply PROT_EXEC?
1413 *
1414 * (the exception is when the underlying filesystem is noexec
1415 * mounted, in which case we dont add PROT_EXEC.)
1416 */
1417 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1418 if (!(file && path_noexec(&file->f_path)))
1419 prot |= PROT_EXEC;
1420
1421 /* force arch specific MAP_FIXED handling in get_unmapped_area */
1422 if (flags & MAP_FIXED_NOREPLACE)
1423 flags |= MAP_FIXED;
1424
1425 if (!(flags & MAP_FIXED))
1426 addr = round_hint_to_min(addr);
1427
1428 /* Careful about overflows.. */
1429 len = PAGE_ALIGN(len);
1430 if (!len)
1431 return -ENOMEM;
1432
1433 /* offset overflow? */
1434 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1435 return -EOVERFLOW;
1436
1437 /* Too many mappings? */
1438 if (mm->map_count > sysctl_max_map_count)
1439 return -ENOMEM;
1440
1441 /* Obtain the address to map to. we verify (or select) it and ensure
1442 * that it represents a valid section of the address space.
1443 */
1444 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1445 if (offset_in_page(addr))
1446 return addr;
1447
1448 if (flags & MAP_FIXED_NOREPLACE) {
1449 struct vm_area_struct *vma = find_vma(mm, addr);
1450
1451 if (vma && vma->vm_start < addr + len)
1452 return -EEXIST;
1453 }
1454
1455 if (prot == PROT_EXEC) {
1456 pkey = execute_only_pkey(mm);
1457 if (pkey < 0)
1458 pkey = 0;
1459 }
1460
1461 /* Do simple checking here so the lower-level routines won't have
1462 * to. we assume access permissions have been handled by the open
1463 * of the memory object, so we don't do any here.
1464 */
1465 vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1466 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1467
1468 if (flags & MAP_LOCKED)
1469 if (!can_do_mlock())
1470 return -EPERM;
1471
1472 if (mlock_future_check(mm, vm_flags, len))
1473 return -EAGAIN;
1474
1475 if (file) {
1476 struct inode *inode = file_inode(file);
1477 unsigned long flags_mask;
1478
1479 if (!file_mmap_ok(file, inode, pgoff, len))
1480 return -EOVERFLOW;
1481
1482 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1483
1484 switch (flags & MAP_TYPE) {
1485 case MAP_SHARED:
1486 /*
1487 * Force use of MAP_SHARED_VALIDATE with non-legacy
1488 * flags. E.g. MAP_SYNC is dangerous to use with
1489 * MAP_SHARED as you don't know which consistency model
1490 * you will get. We silently ignore unsupported flags
1491 * with MAP_SHARED to preserve backward compatibility.
1492 */
1493 flags &= LEGACY_MAP_MASK;
1494 /* fall through */
1495 case MAP_SHARED_VALIDATE:
1496 if (flags & ~flags_mask)
1497 return -EOPNOTSUPP;
1498 if (prot & PROT_WRITE) {
1499 if (!(file->f_mode & FMODE_WRITE))
1500 return -EACCES;
1501 if (IS_SWAPFILE(file->f_mapping->host))
1502 return -ETXTBSY;
1503 }
1504
1505 /*
1506 * Make sure we don't allow writing to an append-only
1507 * file..
1508 */
1509 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1510 return -EACCES;
1511
1512 /*
1513 * Make sure there are no mandatory locks on the file.
1514 */
1515 if (locks_verify_locked(file))
1516 return -EAGAIN;
1517
1518 vm_flags |= VM_SHARED | VM_MAYSHARE;
1519 if (!(file->f_mode & FMODE_WRITE))
1520 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1521
1522 /* fall through */
1523 case MAP_PRIVATE:
1524 if (!(file->f_mode & FMODE_READ))
1525 return -EACCES;
1526 if (path_noexec(&file->f_path)) {
1527 if (vm_flags & VM_EXEC)
1528 return -EPERM;
1529 vm_flags &= ~VM_MAYEXEC;
1530 }
1531
1532 if (!file->f_op->mmap)
1533 return -ENODEV;
1534 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1535 return -EINVAL;
1536 break;
1537
1538 default:
1539 return -EINVAL;
1540 }
1541 } else {
1542 switch (flags & MAP_TYPE) {
1543 case MAP_SHARED:
1544 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1545 return -EINVAL;
1546 /*
1547 * Ignore pgoff.
1548 */
1549 pgoff = 0;
1550 vm_flags |= VM_SHARED | VM_MAYSHARE;
1551 break;
1552 case MAP_PRIVATE:
1553 /*
1554 * Set pgoff according to addr for anon_vma.
1555 */
1556 pgoff = addr >> PAGE_SHIFT;
1557 break;
1558 default:
1559 return -EINVAL;
1560 }
1561 }
1562
1563 /*
1564 * Set 'VM_NORESERVE' if we should not account for the
1565 * memory use of this mapping.
1566 */
1567 if (flags & MAP_NORESERVE) {
1568 /* We honor MAP_NORESERVE if allowed to overcommit */
1569 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1570 vm_flags |= VM_NORESERVE;
1571
1572 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1573 if (file && is_file_hugepages(file))
1574 vm_flags |= VM_NORESERVE;
1575 }
1576
1577 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1578 if (!IS_ERR_VALUE(addr) &&
1579 ((vm_flags & VM_LOCKED) ||
1580 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1581 *populate = len;
1582 return addr;
1583}
1584
1585unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1586 unsigned long prot, unsigned long flags,
1587 unsigned long fd, unsigned long pgoff)
1588{
1589 struct file *file = NULL;
1590 unsigned long retval;
1591
1592 addr = untagged_addr(addr);
1593
1594 if (!(flags & MAP_ANONYMOUS)) {
1595 audit_mmap_fd(fd, flags);
1596 file = fget(fd);
1597 if (!file)
1598 return -EBADF;
1599 if (is_file_hugepages(file))
1600 len = ALIGN(len, huge_page_size(hstate_file(file)));
1601 retval = -EINVAL;
1602 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1603 goto out_fput;
1604 } else if (flags & MAP_HUGETLB) {
1605 struct user_struct *user = NULL;
1606 struct hstate *hs;
1607
1608 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1609 if (!hs)
1610 return -EINVAL;
1611
1612 len = ALIGN(len, huge_page_size(hs));
1613 /*
1614 * VM_NORESERVE is used because the reservations will be
1615 * taken when vm_ops->mmap() is called
1616 * A dummy user value is used because we are not locking
1617 * memory so no accounting is necessary
1618 */
1619 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1620 VM_NORESERVE,
1621 &user, HUGETLB_ANONHUGE_INODE,
1622 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1623 if (IS_ERR(file))
1624 return PTR_ERR(file);
1625 }
1626
1627 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1628
1629 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1630out_fput:
1631 if (file)
1632 fput(file);
1633 return retval;
1634}
1635
1636SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1637 unsigned long, prot, unsigned long, flags,
1638 unsigned long, fd, unsigned long, pgoff)
1639{
1640 return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1641}
1642
1643#ifdef __ARCH_WANT_SYS_OLD_MMAP
1644struct mmap_arg_struct {
1645 unsigned long addr;
1646 unsigned long len;
1647 unsigned long prot;
1648 unsigned long flags;
1649 unsigned long fd;
1650 unsigned long offset;
1651};
1652
1653SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1654{
1655 struct mmap_arg_struct a;
1656
1657 if (copy_from_user(&a, arg, sizeof(a)))
1658 return -EFAULT;
1659 if (offset_in_page(a.offset))
1660 return -EINVAL;
1661
1662 return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1663 a.offset >> PAGE_SHIFT);
1664}
1665#endif /* __ARCH_WANT_SYS_OLD_MMAP */
1666
1667/*
1668 * Some shared mappings will want the pages marked read-only
1669 * to track write events. If so, we'll downgrade vm_page_prot
1670 * to the private version (using protection_map[] without the
1671 * VM_SHARED bit).
1672 */
1673int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1674{
1675 vm_flags_t vm_flags = vma->vm_flags;
1676 const struct vm_operations_struct *vm_ops = vma->vm_ops;
1677
1678 /* If it was private or non-writable, the write bit is already clear */
1679 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1680 return 0;
1681
1682 /* The backer wishes to know when pages are first written to? */
1683 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1684 return 1;
1685
1686 /* The open routine did something to the protections that pgprot_modify
1687 * won't preserve? */
1688 if (pgprot_val(vm_page_prot) !=
1689 pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1690 return 0;
1691
1692 /* Do we need to track softdirty? */
1693 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1694 return 1;
1695
1696 /* Specialty mapping? */
1697 if (vm_flags & VM_PFNMAP)
1698 return 0;
1699
1700 /* Can the mapping track the dirty pages? */
1701 return vma->vm_file && vma->vm_file->f_mapping &&
1702 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1703}
1704
1705/*
1706 * We account for memory if it's a private writeable mapping,
1707 * not hugepages and VM_NORESERVE wasn't set.
1708 */
1709static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1710{
1711 /*
1712 * hugetlb has its own accounting separate from the core VM
1713 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1714 */
1715 if (file && is_file_hugepages(file))
1716 return 0;
1717
1718 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1719}
1720
1721unsigned long mmap_region(struct file *file, unsigned long addr,
1722 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
1723 struct list_head *uf)
1724{
1725 struct mm_struct *mm = current->mm;
1726 struct vm_area_struct *vma, *prev;
1727 int error;
1728 struct rb_node **rb_link, *rb_parent;
1729 unsigned long charged = 0;
1730
1731 /* Check against address space limit. */
1732 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1733 unsigned long nr_pages;
1734
1735 /*
1736 * MAP_FIXED may remove pages of mappings that intersects with
1737 * requested mapping. Account for the pages it would unmap.
1738 */
1739 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1740
1741 if (!may_expand_vm(mm, vm_flags,
1742 (len >> PAGE_SHIFT) - nr_pages))
1743 return -ENOMEM;
1744 }
1745
1746 /* Clear old maps */
1747 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
1748 &rb_parent)) {
1749 if (do_munmap(mm, addr, len, uf))
1750 return -ENOMEM;
1751 }
1752
1753 /*
1754 * Private writable mapping: check memory availability
1755 */
1756 if (accountable_mapping(file, vm_flags)) {
1757 charged = len >> PAGE_SHIFT;
1758 if (security_vm_enough_memory_mm(mm, charged))
1759 return -ENOMEM;
1760 vm_flags |= VM_ACCOUNT;
1761 }
1762
1763 /*
1764 * Can we just expand an old mapping?
1765 */
1766 vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1767 NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1768 if (vma)
1769 goto out;
1770
1771 /*
1772 * Determine the object being mapped and call the appropriate
1773 * specific mapper. the address has already been validated, but
1774 * not unmapped, but the maps are removed from the list.
1775 */
1776 vma = vm_area_alloc(mm);
1777 if (!vma) {
1778 error = -ENOMEM;
1779 goto unacct_error;
1780 }
1781
1782 vma->vm_start = addr;
1783 vma->vm_end = addr + len;
1784 vma->vm_flags = vm_flags;
1785 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1786 vma->vm_pgoff = pgoff;
1787
1788 if (file) {
1789 if (vm_flags & VM_DENYWRITE) {
1790 error = deny_write_access(file);
1791 if (error)
1792 goto free_vma;
1793 }
1794 if (vm_flags & VM_SHARED) {
1795 error = mapping_map_writable(file->f_mapping);
1796 if (error)
1797 goto allow_write_and_free_vma;
1798 }
1799
1800 /* ->mmap() can change vma->vm_file, but must guarantee that
1801 * vma_link() below can deny write-access if VM_DENYWRITE is set
1802 * and map writably if VM_SHARED is set. This usually means the
1803 * new file must not have been exposed to user-space, yet.
1804 */
1805 vma->vm_file = get_file(file);
1806 error = call_mmap(file, vma);
1807 if (error)
1808 goto unmap_and_free_vma;
1809
1810 /* Can addr have changed??
1811 *
1812 * Answer: Yes, several device drivers can do it in their
1813 * f_op->mmap method. -DaveM
1814 * Bug: If addr is changed, prev, rb_link, rb_parent should
1815 * be updated for vma_link()
1816 */
1817 WARN_ON_ONCE(addr != vma->vm_start);
1818
1819 addr = vma->vm_start;
1820 vm_flags = vma->vm_flags;
1821 } else if (vm_flags & VM_SHARED) {
1822 error = shmem_zero_setup(vma);
1823 if (error)
1824 goto free_vma;
1825 } else {
1826 vma_set_anonymous(vma);
1827 }
1828
1829 vma_link(mm, vma, prev, rb_link, rb_parent);
1830 /* Once vma denies write, undo our temporary denial count */
1831 if (file) {
1832 if (vm_flags & VM_SHARED)
1833 mapping_unmap_writable(file->f_mapping);
1834 if (vm_flags & VM_DENYWRITE)
1835 allow_write_access(file);
1836 }
1837 file = vma->vm_file;
1838out:
1839 perf_event_mmap(vma);
1840
1841 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1842 if (vm_flags & VM_LOCKED) {
1843 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
1844 is_vm_hugetlb_page(vma) ||
1845 vma == get_gate_vma(current->mm))
1846 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1847 else
1848 mm->locked_vm += (len >> PAGE_SHIFT);
1849 }
1850
1851 if (file)
1852 uprobe_mmap(vma);
1853
1854 /*
1855 * New (or expanded) vma always get soft dirty status.
1856 * Otherwise user-space soft-dirty page tracker won't
1857 * be able to distinguish situation when vma area unmapped,
1858 * then new mapped in-place (which must be aimed as
1859 * a completely new data area).
1860 */
1861 vma->vm_flags |= VM_SOFTDIRTY;
1862
1863 vma_set_page_prot(vma);
1864
1865 return addr;
1866
1867unmap_and_free_vma:
1868 vma->vm_file = NULL;
1869 fput(file);
1870
1871 /* Undo any partial mapping done by a device driver. */
1872 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1873 charged = 0;
1874 if (vm_flags & VM_SHARED)
1875 mapping_unmap_writable(file->f_mapping);
1876allow_write_and_free_vma:
1877 if (vm_flags & VM_DENYWRITE)
1878 allow_write_access(file);
1879free_vma:
1880 vm_area_free(vma);
1881unacct_error:
1882 if (charged)
1883 vm_unacct_memory(charged);
1884 return error;
1885}
1886
1887unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1888{
1889 /*
1890 * We implement the search by looking for an rbtree node that
1891 * immediately follows a suitable gap. That is,
1892 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1893 * - gap_end = vma->vm_start >= info->low_limit + length;
1894 * - gap_end - gap_start >= length
1895 */
1896
1897 struct mm_struct *mm = current->mm;
1898 struct vm_area_struct *vma;
1899 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1900
1901 /* Adjust search length to account for worst case alignment overhead */
1902 length = info->length + info->align_mask;
1903 if (length < info->length)
1904 return -ENOMEM;
1905
1906 /* Adjust search limits by the desired length */
1907 if (info->high_limit < length)
1908 return -ENOMEM;
1909 high_limit = info->high_limit - length;
1910
1911 if (info->low_limit > high_limit)
1912 return -ENOMEM;
1913 low_limit = info->low_limit + length;
1914
1915 /* Check if rbtree root looks promising */
1916 if (RB_EMPTY_ROOT(&mm->mm_rb))
1917 goto check_highest;
1918 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1919 if (vma->rb_subtree_gap < length)
1920 goto check_highest;
1921
1922 while (true) {
1923 /* Visit left subtree if it looks promising */
1924 gap_end = vm_start_gap(vma);
1925 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1926 struct vm_area_struct *left =
1927 rb_entry(vma->vm_rb.rb_left,
1928 struct vm_area_struct, vm_rb);
1929 if (left->rb_subtree_gap >= length) {
1930 vma = left;
1931 continue;
1932 }
1933 }
1934
1935 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1936check_current:
1937 /* Check if current node has a suitable gap */
1938 if (gap_start > high_limit)
1939 return -ENOMEM;
1940 if (gap_end >= low_limit &&
1941 gap_end > gap_start && gap_end - gap_start >= length)
1942 goto found;
1943
1944 /* Visit right subtree if it looks promising */
1945 if (vma->vm_rb.rb_right) {
1946 struct vm_area_struct *right =
1947 rb_entry(vma->vm_rb.rb_right,
1948 struct vm_area_struct, vm_rb);
1949 if (right->rb_subtree_gap >= length) {
1950 vma = right;
1951 continue;
1952 }
1953 }
1954
1955 /* Go back up the rbtree to find next candidate node */
1956 while (true) {
1957 struct rb_node *prev = &vma->vm_rb;
1958 if (!rb_parent(prev))
1959 goto check_highest;
1960 vma = rb_entry(rb_parent(prev),
1961 struct vm_area_struct, vm_rb);
1962 if (prev == vma->vm_rb.rb_left) {
1963 gap_start = vm_end_gap(vma->vm_prev);
1964 gap_end = vm_start_gap(vma);
1965 goto check_current;
1966 }
1967 }
1968 }
1969
1970check_highest:
1971 /* Check highest gap, which does not precede any rbtree node */
1972 gap_start = mm->highest_vm_end;
1973 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */
1974 if (gap_start > high_limit)
1975 return -ENOMEM;
1976
1977found:
1978 /* We found a suitable gap. Clip it with the original low_limit. */
1979 if (gap_start < info->low_limit)
1980 gap_start = info->low_limit;
1981
1982 /* Adjust gap address to the desired alignment */
1983 gap_start += (info->align_offset - gap_start) & info->align_mask;
1984
1985 VM_BUG_ON(gap_start + info->length > info->high_limit);
1986 VM_BUG_ON(gap_start + info->length > gap_end);
1987 return gap_start;
1988}
1989
1990unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1991{
1992 struct mm_struct *mm = current->mm;
1993 struct vm_area_struct *vma;
1994 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1995
1996 /* Adjust search length to account for worst case alignment overhead */
1997 length = info->length + info->align_mask;
1998 if (length < info->length)
1999 return -ENOMEM;
2000
2001 /*
2002 * Adjust search limits by the desired length.
2003 * See implementation comment at top of unmapped_area().
2004 */
2005 gap_end = info->high_limit;
2006 if (gap_end < length)
2007 return -ENOMEM;
2008 high_limit = gap_end - length;
2009
2010 if (info->low_limit > high_limit)
2011 return -ENOMEM;
2012 low_limit = info->low_limit + length;
2013
2014 /* Check highest gap, which does not precede any rbtree node */
2015 gap_start = mm->highest_vm_end;
2016 if (gap_start <= high_limit)
2017 goto found_highest;
2018
2019 /* Check if rbtree root looks promising */
2020 if (RB_EMPTY_ROOT(&mm->mm_rb))
2021 return -ENOMEM;
2022 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
2023 if (vma->rb_subtree_gap < length)
2024 return -ENOMEM;
2025
2026 while (true) {
2027 /* Visit right subtree if it looks promising */
2028 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
2029 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
2030 struct vm_area_struct *right =
2031 rb_entry(vma->vm_rb.rb_right,
2032 struct vm_area_struct, vm_rb);
2033 if (right->rb_subtree_gap >= length) {
2034 vma = right;
2035 continue;
2036 }
2037 }
2038
2039check_current:
2040 /* Check if current node has a suitable gap */
2041 gap_end = vm_start_gap(vma);
2042 if (gap_end < low_limit)
2043 return -ENOMEM;
2044 if (gap_start <= high_limit &&
2045 gap_end > gap_start && gap_end - gap_start >= length)
2046 goto found;
2047
2048 /* Visit left subtree if it looks promising */
2049 if (vma->vm_rb.rb_left) {
2050 struct vm_area_struct *left =
2051 rb_entry(vma->vm_rb.rb_left,
2052 struct vm_area_struct, vm_rb);
2053 if (left->rb_subtree_gap >= length) {
2054 vma = left;
2055 continue;
2056 }
2057 }
2058
2059 /* Go back up the rbtree to find next candidate node */
2060 while (true) {
2061 struct rb_node *prev = &vma->vm_rb;
2062 if (!rb_parent(prev))
2063 return -ENOMEM;
2064 vma = rb_entry(rb_parent(prev),
2065 struct vm_area_struct, vm_rb);
2066 if (prev == vma->vm_rb.rb_right) {
2067 gap_start = vma->vm_prev ?
2068 vm_end_gap(vma->vm_prev) : 0;
2069 goto check_current;
2070 }
2071 }
2072 }
2073
2074found:
2075 /* We found a suitable gap. Clip it with the original high_limit. */
2076 if (gap_end > info->high_limit)
2077 gap_end = info->high_limit;
2078
2079found_highest:
2080 /* Compute highest gap address at the desired alignment */
2081 gap_end -= info->length;
2082 gap_end -= (gap_end - info->align_offset) & info->align_mask;
2083
2084 VM_BUG_ON(gap_end < info->low_limit);
2085 VM_BUG_ON(gap_end < gap_start);
2086 return gap_end;
2087}
2088
2089
2090#ifndef arch_get_mmap_end
2091#define arch_get_mmap_end(addr) (TASK_SIZE)
2092#endif
2093
2094#ifndef arch_get_mmap_base
2095#define arch_get_mmap_base(addr, base) (base)
2096#endif
2097
2098/* Get an address range which is currently unmapped.
2099 * For shmat() with addr=0.
2100 *
2101 * Ugly calling convention alert:
2102 * Return value with the low bits set means error value,
2103 * ie
2104 * if (ret & ~PAGE_MASK)
2105 * error = ret;
2106 *
2107 * This function "knows" that -ENOMEM has the bits set.
2108 */
2109#ifndef HAVE_ARCH_UNMAPPED_AREA
2110unsigned long
2111arch_get_unmapped_area(struct file *filp, unsigned long addr,
2112 unsigned long len, unsigned long pgoff, unsigned long flags)
2113{
2114 struct mm_struct *mm = current->mm;
2115 struct vm_area_struct *vma, *prev;
2116 struct vm_unmapped_area_info info;
2117 const unsigned long mmap_end = arch_get_mmap_end(addr);
2118
2119 if (len > mmap_end - mmap_min_addr)
2120 return -ENOMEM;
2121
2122 if (flags & MAP_FIXED)
2123 return addr;
2124
2125 if (addr) {
2126 addr = PAGE_ALIGN(addr);
2127 vma = find_vma_prev(mm, addr, &prev);
2128 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2129 (!vma || addr + len <= vm_start_gap(vma)) &&
2130 (!prev || addr >= vm_end_gap(prev)))
2131 return addr;
2132 }
2133
2134 info.flags = 0;
2135 info.length = len;
2136 info.low_limit = mm->mmap_base;
2137 info.high_limit = mmap_end;
2138 info.align_mask = 0;
2139 return vm_unmapped_area(&info);
2140}
2141#endif
2142
2143/*
2144 * This mmap-allocator allocates new areas top-down from below the
2145 * stack's low limit (the base):
2146 */
2147#ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2148unsigned long
2149arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
2150 unsigned long len, unsigned long pgoff,
2151 unsigned long flags)
2152{
2153 struct vm_area_struct *vma, *prev;
2154 struct mm_struct *mm = current->mm;
2155 struct vm_unmapped_area_info info;
2156 const unsigned long mmap_end = arch_get_mmap_end(addr);
2157
2158 /* requested length too big for entire address space */
2159 if (len > mmap_end - mmap_min_addr)
2160 return -ENOMEM;
2161
2162 if (flags & MAP_FIXED)
2163 return addr;
2164
2165 /* requesting a specific address */
2166 if (addr) {
2167 addr = PAGE_ALIGN(addr);
2168 vma = find_vma_prev(mm, addr, &prev);
2169 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2170 (!vma || addr + len <= vm_start_gap(vma)) &&
2171 (!prev || addr >= vm_end_gap(prev)))
2172 return addr;
2173 }
2174
2175 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2176 info.length = len;
2177 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2178 info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
2179 info.align_mask = 0;
2180 addr = vm_unmapped_area(&info);
2181
2182 /*
2183 * A failed mmap() very likely causes application failure,
2184 * so fall back to the bottom-up function here. This scenario
2185 * can happen with large stack limits and large mmap()
2186 * allocations.
2187 */
2188 if (offset_in_page(addr)) {
2189 VM_BUG_ON(addr != -ENOMEM);
2190 info.flags = 0;
2191 info.low_limit = TASK_UNMAPPED_BASE;
2192 info.high_limit = mmap_end;
2193 addr = vm_unmapped_area(&info);
2194 }
2195
2196 return addr;
2197}
2198#endif
2199
2200unsigned long
2201get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2202 unsigned long pgoff, unsigned long flags)
2203{
2204 unsigned long (*get_area)(struct file *, unsigned long,
2205 unsigned long, unsigned long, unsigned long);
2206
2207 unsigned long error = arch_mmap_check(addr, len, flags);
2208 if (error)
2209 return error;
2210
2211 /* Careful about overflows.. */
2212 if (len > TASK_SIZE)
2213 return -ENOMEM;
2214
2215 get_area = current->mm->get_unmapped_area;
2216 if (file) {
2217 if (file->f_op->get_unmapped_area)
2218 get_area = file->f_op->get_unmapped_area;
2219 } else if (flags & MAP_SHARED) {
2220 /*
2221 * mmap_region() will call shmem_zero_setup() to create a file,
2222 * so use shmem's get_unmapped_area in case it can be huge.
2223 * do_mmap_pgoff() will clear pgoff, so match alignment.
2224 */
2225 pgoff = 0;
2226 get_area = shmem_get_unmapped_area;
2227 }
2228
2229 addr = get_area(file, addr, len, pgoff, flags);
2230 if (IS_ERR_VALUE(addr))
2231 return addr;
2232
2233 if (addr > TASK_SIZE - len)
2234 return -ENOMEM;
2235 if (offset_in_page(addr))
2236 return -EINVAL;
2237
2238 error = security_mmap_addr(addr);
2239 return error ? error : addr;
2240}
2241
2242EXPORT_SYMBOL(get_unmapped_area);
2243
2244/* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
2245struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2246{
2247 struct rb_node *rb_node;
2248 struct vm_area_struct *vma;
2249
2250 /* Check the cache first. */
2251 vma = vmacache_find(mm, addr);
2252 if (likely(vma))
2253 return vma;
2254
2255 rb_node = mm->mm_rb.rb_node;
2256
2257 while (rb_node) {
2258 struct vm_area_struct *tmp;
2259
2260 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2261
2262 if (tmp->vm_end > addr) {
2263 vma = tmp;
2264 if (tmp->vm_start <= addr)
2265 break;
2266 rb_node = rb_node->rb_left;
2267 } else
2268 rb_node = rb_node->rb_right;
2269 }
2270
2271 if (vma)
2272 vmacache_update(addr, vma);
2273 return vma;
2274}
2275
2276EXPORT_SYMBOL(find_vma);
2277
2278/*
2279 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2280 */
2281struct vm_area_struct *
2282find_vma_prev(struct mm_struct *mm, unsigned long addr,
2283 struct vm_area_struct **pprev)
2284{
2285 struct vm_area_struct *vma;
2286
2287 vma = find_vma(mm, addr);
2288 if (vma) {
2289 *pprev = vma->vm_prev;
2290 } else {
2291 struct rb_node *rb_node = rb_last(&mm->mm_rb);
2292
2293 *pprev = rb_node ? rb_entry(rb_node, struct vm_area_struct, vm_rb) : NULL;
2294 }
2295 return vma;
2296}
2297
2298/*
2299 * Verify that the stack growth is acceptable and
2300 * update accounting. This is shared with both the
2301 * grow-up and grow-down cases.
2302 */
2303static int acct_stack_growth(struct vm_area_struct *vma,
2304 unsigned long size, unsigned long grow)
2305{
2306 struct mm_struct *mm = vma->vm_mm;
2307 unsigned long new_start;
2308
2309 /* address space limit tests */
2310 if (!may_expand_vm(mm, vma->vm_flags, grow))
2311 return -ENOMEM;
2312
2313 /* Stack limit test */
2314 if (size > rlimit(RLIMIT_STACK))
2315 return -ENOMEM;
2316
2317 /* mlock limit tests */
2318 if (vma->vm_flags & VM_LOCKED) {
2319 unsigned long locked;
2320 unsigned long limit;
2321 locked = mm->locked_vm + grow;
2322 limit = rlimit(RLIMIT_MEMLOCK);
2323 limit >>= PAGE_SHIFT;
2324 if (locked > limit && !capable(CAP_IPC_LOCK))
2325 return -ENOMEM;
2326 }
2327
2328 /* Check to ensure the stack will not grow into a hugetlb-only region */
2329 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2330 vma->vm_end - size;
2331 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2332 return -EFAULT;
2333
2334 /*
2335 * Overcommit.. This must be the final test, as it will
2336 * update security statistics.
2337 */
2338 if (security_vm_enough_memory_mm(mm, grow))
2339 return -ENOMEM;
2340
2341 return 0;
2342}
2343
2344#if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2345/*
2346 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2347 * vma is the last one with address > vma->vm_end. Have to extend vma.
2348 */
2349int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2350{
2351 struct mm_struct *mm = vma->vm_mm;
2352 struct vm_area_struct *next;
2353 unsigned long gap_addr;
2354 int error = 0;
2355
2356 if (!(vma->vm_flags & VM_GROWSUP))
2357 return -EFAULT;
2358
2359 /* Guard against exceeding limits of the address space. */
2360 address &= PAGE_MASK;
2361 if (address >= (TASK_SIZE & PAGE_MASK))
2362 return -ENOMEM;
2363 address += PAGE_SIZE;
2364
2365 /* Enforce stack_guard_gap */
2366 gap_addr = address + stack_guard_gap;
2367
2368 /* Guard against overflow */
2369 if (gap_addr < address || gap_addr > TASK_SIZE)
2370 gap_addr = TASK_SIZE;
2371
2372 next = vma->vm_next;
2373 if (next && next->vm_start < gap_addr &&
2374 (next->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2375 if (!(next->vm_flags & VM_GROWSUP))
2376 return -ENOMEM;
2377 /* Check that both stack segments have the same anon_vma? */
2378 }
2379
2380 /* We must make sure the anon_vma is allocated. */
2381 if (unlikely(anon_vma_prepare(vma)))
2382 return -ENOMEM;
2383
2384 /*
2385 * vma->vm_start/vm_end cannot change under us because the caller
2386 * is required to hold the mmap_sem in read mode. We need the
2387 * anon_vma lock to serialize against concurrent expand_stacks.
2388 */
2389 anon_vma_lock_write(vma->anon_vma);
2390
2391 /* Somebody else might have raced and expanded it already */
2392 if (address > vma->vm_end) {
2393 unsigned long size, grow;
2394
2395 size = address - vma->vm_start;
2396 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2397
2398 error = -ENOMEM;
2399 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2400 error = acct_stack_growth(vma, size, grow);
2401 if (!error) {
2402 /*
2403 * vma_gap_update() doesn't support concurrent
2404 * updates, but we only hold a shared mmap_sem
2405 * lock here, so we need to protect against
2406 * concurrent vma expansions.
2407 * anon_vma_lock_write() doesn't help here, as
2408 * we don't guarantee that all growable vmas
2409 * in a mm share the same root anon vma.
2410 * So, we reuse mm->page_table_lock to guard
2411 * against concurrent vma expansions.
2412 */
2413 spin_lock(&mm->page_table_lock);
2414 if (vma->vm_flags & VM_LOCKED)
2415 mm->locked_vm += grow;
2416 vm_stat_account(mm, vma->vm_flags, grow);
2417 anon_vma_interval_tree_pre_update_vma(vma);
2418 vma->vm_end = address;
2419 anon_vma_interval_tree_post_update_vma(vma);
2420 if (vma->vm_next)
2421 vma_gap_update(vma->vm_next);
2422 else
2423 mm->highest_vm_end = vm_end_gap(vma);
2424 spin_unlock(&mm->page_table_lock);
2425
2426 perf_event_mmap(vma);
2427 }
2428 }
2429 }
2430 anon_vma_unlock_write(vma->anon_vma);
2431 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2432 validate_mm(mm);
2433 return error;
2434}
2435#endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2436
2437/*
2438 * vma is the first one with address < vma->vm_start. Have to extend vma.
2439 */
2440int expand_downwards(struct vm_area_struct *vma,
2441 unsigned long address)
2442{
2443 struct mm_struct *mm = vma->vm_mm;
2444 struct vm_area_struct *prev;
2445 int error = 0;
2446
2447 address &= PAGE_MASK;
2448 if (address < mmap_min_addr)
2449 return -EPERM;
2450
2451 /* Enforce stack_guard_gap */
2452 prev = vma->vm_prev;
2453 /* Check that both stack segments have the same anon_vma? */
2454 if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2455 (prev->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2456 if (address - prev->vm_end < stack_guard_gap)
2457 return -ENOMEM;
2458 }
2459
2460 /* We must make sure the anon_vma is allocated. */
2461 if (unlikely(anon_vma_prepare(vma)))
2462 return -ENOMEM;
2463
2464 /*
2465 * vma->vm_start/vm_end cannot change under us because the caller
2466 * is required to hold the mmap_sem in read mode. We need the
2467 * anon_vma lock to serialize against concurrent expand_stacks.
2468 */
2469 anon_vma_lock_write(vma->anon_vma);
2470
2471 /* Somebody else might have raced and expanded it already */
2472 if (address < vma->vm_start) {
2473 unsigned long size, grow;
2474
2475 size = vma->vm_end - address;
2476 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2477
2478 error = -ENOMEM;
2479 if (grow <= vma->vm_pgoff) {
2480 error = acct_stack_growth(vma, size, grow);
2481 if (!error) {
2482 /*
2483 * vma_gap_update() doesn't support concurrent
2484 * updates, but we only hold a shared mmap_sem
2485 * lock here, so we need to protect against
2486 * concurrent vma expansions.
2487 * anon_vma_lock_write() doesn't help here, as
2488 * we don't guarantee that all growable vmas
2489 * in a mm share the same root anon vma.
2490 * So, we reuse mm->page_table_lock to guard
2491 * against concurrent vma expansions.
2492 */
2493 spin_lock(&mm->page_table_lock);
2494 if (vma->vm_flags & VM_LOCKED)
2495 mm->locked_vm += grow;
2496 vm_stat_account(mm, vma->vm_flags, grow);
2497 anon_vma_interval_tree_pre_update_vma(vma);
2498 vma->vm_start = address;
2499 vma->vm_pgoff -= grow;
2500 anon_vma_interval_tree_post_update_vma(vma);
2501 vma_gap_update(vma);
2502 spin_unlock(&mm->page_table_lock);
2503
2504 perf_event_mmap(vma);
2505 }
2506 }
2507 }
2508 anon_vma_unlock_write(vma->anon_vma);
2509 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2510 validate_mm(mm);
2511 return error;
2512}
2513
2514/* enforced gap between the expanding stack and other mappings. */
2515unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2516
2517static int __init cmdline_parse_stack_guard_gap(char *p)
2518{
2519 unsigned long val;
2520 char *endptr;
2521
2522 val = simple_strtoul(p, &endptr, 10);
2523 if (!*endptr)
2524 stack_guard_gap = val << PAGE_SHIFT;
2525
2526 return 0;
2527}
2528__setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2529
2530#ifdef CONFIG_STACK_GROWSUP
2531int expand_stack(struct vm_area_struct *vma, unsigned long address)
2532{
2533 return expand_upwards(vma, address);
2534}
2535
2536struct vm_area_struct *
2537find_extend_vma(struct mm_struct *mm, unsigned long addr)
2538{
2539 struct vm_area_struct *vma, *prev;
2540
2541 addr &= PAGE_MASK;
2542 vma = find_vma_prev(mm, addr, &prev);
2543 if (vma && (vma->vm_start <= addr))
2544 return vma;
2545 /* don't alter vm_end if the coredump is running */
2546 if (!prev || !mmget_still_valid(mm) || expand_stack(prev, addr))
2547 return NULL;
2548 if (prev->vm_flags & VM_LOCKED)
2549 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2550 return prev;
2551}
2552#else
2553int expand_stack(struct vm_area_struct *vma, unsigned long address)
2554{
2555 return expand_downwards(vma, address);
2556}
2557
2558struct vm_area_struct *
2559find_extend_vma(struct mm_struct *mm, unsigned long addr)
2560{
2561 struct vm_area_struct *vma;
2562 unsigned long start;
2563
2564 addr &= PAGE_MASK;
2565 vma = find_vma(mm, addr);
2566 if (!vma)
2567 return NULL;
2568 if (vma->vm_start <= addr)
2569 return vma;
2570 if (!(vma->vm_flags & VM_GROWSDOWN))
2571 return NULL;
2572 /* don't alter vm_start if the coredump is running */
2573 if (!mmget_still_valid(mm))
2574 return NULL;
2575 start = vma->vm_start;
2576 if (expand_stack(vma, addr))
2577 return NULL;
2578 if (vma->vm_flags & VM_LOCKED)
2579 populate_vma_page_range(vma, addr, start, NULL);
2580 return vma;
2581}
2582#endif
2583
2584EXPORT_SYMBOL_GPL(find_extend_vma);
2585
2586/*
2587 * Ok - we have the memory areas we should free on the vma list,
2588 * so release them, and do the vma updates.
2589 *
2590 * Called with the mm semaphore held.
2591 */
2592static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2593{
2594 unsigned long nr_accounted = 0;
2595
2596 /* Update high watermark before we lower total_vm */
2597 update_hiwater_vm(mm);
2598 do {
2599 long nrpages = vma_pages(vma);
2600
2601 if (vma->vm_flags & VM_ACCOUNT)
2602 nr_accounted += nrpages;
2603 vm_stat_account(mm, vma->vm_flags, -nrpages);
2604 vma = remove_vma(vma);
2605 } while (vma);
2606 vm_unacct_memory(nr_accounted);
2607 validate_mm(mm);
2608}
2609
2610/*
2611 * Get rid of page table information in the indicated region.
2612 *
2613 * Called with the mm semaphore held.
2614 */
2615static void unmap_region(struct mm_struct *mm,
2616 struct vm_area_struct *vma, struct vm_area_struct *prev,
2617 unsigned long start, unsigned long end)
2618{
2619 struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap;
2620 struct mmu_gather tlb;
2621
2622 lru_add_drain();
2623 tlb_gather_mmu(&tlb, mm, start, end);
2624 update_hiwater_rss(mm);
2625 unmap_vmas(&tlb, vma, start, end);
2626 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2627 next ? next->vm_start : USER_PGTABLES_CEILING);
2628 tlb_finish_mmu(&tlb, start, end);
2629}
2630
2631/*
2632 * Create a list of vma's touched by the unmap, removing them from the mm's
2633 * vma list as we go..
2634 */
2635static void
2636detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2637 struct vm_area_struct *prev, unsigned long end)
2638{
2639 struct vm_area_struct **insertion_point;
2640 struct vm_area_struct *tail_vma = NULL;
2641
2642 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2643 vma->vm_prev = NULL;
2644 do {
2645 vma_rb_erase(vma, &mm->mm_rb);
2646 mm->map_count--;
2647 tail_vma = vma;
2648 vma = vma->vm_next;
2649 } while (vma && vma->vm_start < end);
2650 *insertion_point = vma;
2651 if (vma) {
2652 vma->vm_prev = prev;
2653 vma_gap_update(vma);
2654 } else
2655 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2656 tail_vma->vm_next = NULL;
2657
2658 /* Kill the cache */
2659 vmacache_invalidate(mm);
2660}
2661
2662/*
2663 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it
2664 * has already been checked or doesn't make sense to fail.
2665 */
2666int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2667 unsigned long addr, int new_below)
2668{
2669 struct vm_area_struct *new;
2670 int err;
2671
2672 if (vma->vm_ops && vma->vm_ops->split) {
2673 err = vma->vm_ops->split(vma, addr);
2674 if (err)
2675 return err;
2676 }
2677
2678 new = vm_area_dup(vma);
2679 if (!new)
2680 return -ENOMEM;
2681
2682 if (new_below)
2683 new->vm_end = addr;
2684 else {
2685 new->vm_start = addr;
2686 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2687 }
2688
2689 err = vma_dup_policy(vma, new);
2690 if (err)
2691 goto out_free_vma;
2692
2693 err = anon_vma_clone(new, vma);
2694 if (err)
2695 goto out_free_mpol;
2696
2697 if (new->vm_file)
2698 get_file(new->vm_file);
2699
2700 if (new->vm_ops && new->vm_ops->open)
2701 new->vm_ops->open(new);
2702
2703 if (new_below)
2704 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2705 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2706 else
2707 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2708
2709 /* Success. */
2710 if (!err)
2711 return 0;
2712
2713 /* Clean everything up if vma_adjust failed. */
2714 if (new->vm_ops && new->vm_ops->close)
2715 new->vm_ops->close(new);
2716 if (new->vm_file)
2717 fput(new->vm_file);
2718 unlink_anon_vmas(new);
2719 out_free_mpol:
2720 mpol_put(vma_policy(new));
2721 out_free_vma:
2722 vm_area_free(new);
2723 return err;
2724}
2725
2726/*
2727 * Split a vma into two pieces at address 'addr', a new vma is allocated
2728 * either for the first part or the tail.
2729 */
2730int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2731 unsigned long addr, int new_below)
2732{
2733 if (mm->map_count >= sysctl_max_map_count)
2734 return -ENOMEM;
2735
2736 return __split_vma(mm, vma, addr, new_below);
2737}
2738
2739/* Munmap is split into 2 main parts -- this part which finds
2740 * what needs doing, and the areas themselves, which do the
2741 * work. This now handles partial unmappings.
2742 * Jeremy Fitzhardinge <jeremy@goop.org>
2743 */
2744int __do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2745 struct list_head *uf, bool downgrade)
2746{
2747 unsigned long end;
2748 struct vm_area_struct *vma, *prev, *last;
2749
2750 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2751 return -EINVAL;
2752
2753 len = PAGE_ALIGN(len);
2754 end = start + len;
2755 if (len == 0)
2756 return -EINVAL;
2757
2758 /*
2759 * arch_unmap() might do unmaps itself. It must be called
2760 * and finish any rbtree manipulation before this code
2761 * runs and also starts to manipulate the rbtree.
2762 */
2763 arch_unmap(mm, start, end);
2764
2765 /* Find the first overlapping VMA */
2766 vma = find_vma(mm, start);
2767 if (!vma)
2768 return 0;
2769 prev = vma->vm_prev;
2770 /* we have start < vma->vm_end */
2771
2772 /* if it doesn't overlap, we have nothing.. */
2773 if (vma->vm_start >= end)
2774 return 0;
2775
2776 /*
2777 * If we need to split any vma, do it now to save pain later.
2778 *
2779 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2780 * unmapped vm_area_struct will remain in use: so lower split_vma
2781 * places tmp vma above, and higher split_vma places tmp vma below.
2782 */
2783 if (start > vma->vm_start) {
2784 int error;
2785
2786 /*
2787 * Make sure that map_count on return from munmap() will
2788 * not exceed its limit; but let map_count go just above
2789 * its limit temporarily, to help free resources as expected.
2790 */
2791 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2792 return -ENOMEM;
2793
2794 error = __split_vma(mm, vma, start, 0);
2795 if (error)
2796 return error;
2797 prev = vma;
2798 }
2799
2800 /* Does it split the last one? */
2801 last = find_vma(mm, end);
2802 if (last && end > last->vm_start) {
2803 int error = __split_vma(mm, last, end, 1);
2804 if (error)
2805 return error;
2806 }
2807 vma = prev ? prev->vm_next : mm->mmap;
2808
2809 if (unlikely(uf)) {
2810 /*
2811 * If userfaultfd_unmap_prep returns an error the vmas
2812 * will remain splitted, but userland will get a
2813 * highly unexpected error anyway. This is no
2814 * different than the case where the first of the two
2815 * __split_vma fails, but we don't undo the first
2816 * split, despite we could. This is unlikely enough
2817 * failure that it's not worth optimizing it for.
2818 */
2819 int error = userfaultfd_unmap_prep(vma, start, end, uf);
2820 if (error)
2821 return error;
2822 }
2823
2824 /*
2825 * unlock any mlock()ed ranges before detaching vmas
2826 */
2827 if (mm->locked_vm) {
2828 struct vm_area_struct *tmp = vma;
2829 while (tmp && tmp->vm_start < end) {
2830 if (tmp->vm_flags & VM_LOCKED) {
2831 mm->locked_vm -= vma_pages(tmp);
2832 munlock_vma_pages_all(tmp);
2833 }
2834
2835 tmp = tmp->vm_next;
2836 }
2837 }
2838
2839 /* Detach vmas from rbtree */
2840 detach_vmas_to_be_unmapped(mm, vma, prev, end);
2841
2842 if (downgrade)
2843 downgrade_write(&mm->mmap_sem);
2844
2845 unmap_region(mm, vma, prev, start, end);
2846
2847 /* Fix up all other VM information */
2848 remove_vma_list(mm, vma);
2849
2850 return downgrade ? 1 : 0;
2851}
2852
2853int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2854 struct list_head *uf)
2855{
2856 return __do_munmap(mm, start, len, uf, false);
2857}
2858
2859static int __vm_munmap(unsigned long start, size_t len, bool downgrade)
2860{
2861 int ret;
2862 struct mm_struct *mm = current->mm;
2863 LIST_HEAD(uf);
2864
2865 if (down_write_killable(&mm->mmap_sem))
2866 return -EINTR;
2867
2868 ret = __do_munmap(mm, start, len, &uf, downgrade);
2869 /*
2870 * Returning 1 indicates mmap_sem is downgraded.
2871 * But 1 is not legal return value of vm_munmap() and munmap(), reset
2872 * it to 0 before return.
2873 */
2874 if (ret == 1) {
2875 up_read(&mm->mmap_sem);
2876 ret = 0;
2877 } else
2878 up_write(&mm->mmap_sem);
2879
2880 userfaultfd_unmap_complete(mm, &uf);
2881 return ret;
2882}
2883
2884int vm_munmap(unsigned long start, size_t len)
2885{
2886 return __vm_munmap(start, len, false);
2887}
2888EXPORT_SYMBOL(vm_munmap);
2889
2890SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2891{
2892 addr = untagged_addr(addr);
2893 profile_munmap(addr);
2894 return __vm_munmap(addr, len, true);
2895}
2896
2897
2898/*
2899 * Emulation of deprecated remap_file_pages() syscall.
2900 */
2901SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2902 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2903{
2904
2905 struct mm_struct *mm = current->mm;
2906 struct vm_area_struct *vma;
2907 unsigned long populate = 0;
2908 unsigned long ret = -EINVAL;
2909 struct file *file;
2910
2911 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.rst.\n",
2912 current->comm, current->pid);
2913
2914 if (prot)
2915 return ret;
2916 start = start & PAGE_MASK;
2917 size = size & PAGE_MASK;
2918
2919 if (start + size <= start)
2920 return ret;
2921
2922 /* Does pgoff wrap? */
2923 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2924 return ret;
2925
2926 if (down_write_killable(&mm->mmap_sem))
2927 return -EINTR;
2928
2929 vma = find_vma(mm, start);
2930
2931 if (!vma || !(vma->vm_flags & VM_SHARED))
2932 goto out;
2933
2934 if (start < vma->vm_start)
2935 goto out;
2936
2937 if (start + size > vma->vm_end) {
2938 struct vm_area_struct *next;
2939
2940 for (next = vma->vm_next; next; next = next->vm_next) {
2941 /* hole between vmas ? */
2942 if (next->vm_start != next->vm_prev->vm_end)
2943 goto out;
2944
2945 if (next->vm_file != vma->vm_file)
2946 goto out;
2947
2948 if (next->vm_flags != vma->vm_flags)
2949 goto out;
2950
2951 if (start + size <= next->vm_end)
2952 break;
2953 }
2954
2955 if (!next)
2956 goto out;
2957 }
2958
2959 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2960 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2961 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2962
2963 flags &= MAP_NONBLOCK;
2964 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2965 if (vma->vm_flags & VM_LOCKED) {
2966 struct vm_area_struct *tmp;
2967 flags |= MAP_LOCKED;
2968
2969 /* drop PG_Mlocked flag for over-mapped range */
2970 for (tmp = vma; tmp->vm_start >= start + size;
2971 tmp = tmp->vm_next) {
2972 /*
2973 * Split pmd and munlock page on the border
2974 * of the range.
2975 */
2976 vma_adjust_trans_huge(tmp, start, start + size, 0);
2977
2978 munlock_vma_pages_range(tmp,
2979 max(tmp->vm_start, start),
2980 min(tmp->vm_end, start + size));
2981 }
2982 }
2983
2984 file = get_file(vma->vm_file);
2985 ret = do_mmap_pgoff(vma->vm_file, start, size,
2986 prot, flags, pgoff, &populate, NULL);
2987 fput(file);
2988out:
2989 up_write(&mm->mmap_sem);
2990 if (populate)
2991 mm_populate(ret, populate);
2992 if (!IS_ERR_VALUE(ret))
2993 ret = 0;
2994 return ret;
2995}
2996
2997/*
2998 * this is really a simplified "do_mmap". it only handles
2999 * anonymous maps. eventually we may be able to do some
3000 * brk-specific accounting here.
3001 */
3002static int do_brk_flags(unsigned long addr, unsigned long len, unsigned long flags, struct list_head *uf)
3003{
3004 struct mm_struct *mm = current->mm;
3005 struct vm_area_struct *vma, *prev;
3006 struct rb_node **rb_link, *rb_parent;
3007 pgoff_t pgoff = addr >> PAGE_SHIFT;
3008 int error;
3009
3010 /* Until we need other flags, refuse anything except VM_EXEC. */
3011 if ((flags & (~VM_EXEC)) != 0)
3012 return -EINVAL;
3013 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
3014
3015 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
3016 if (offset_in_page(error))
3017 return error;
3018
3019 error = mlock_future_check(mm, mm->def_flags, len);
3020 if (error)
3021 return error;
3022
3023 /*
3024 * Clear old maps. this also does some error checking for us
3025 */
3026 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
3027 &rb_parent)) {
3028 if (do_munmap(mm, addr, len, uf))
3029 return -ENOMEM;
3030 }
3031
3032 /* Check against address space limits *after* clearing old maps... */
3033 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
3034 return -ENOMEM;
3035
3036 if (mm->map_count > sysctl_max_map_count)
3037 return -ENOMEM;
3038
3039 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
3040 return -ENOMEM;
3041
3042 /* Can we just expand an old private anonymous mapping? */
3043 vma = vma_merge(mm, prev, addr, addr + len, flags,
3044 NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
3045 if (vma)
3046 goto out;
3047
3048 /*
3049 * create a vma struct for an anonymous mapping
3050 */
3051 vma = vm_area_alloc(mm);
3052 if (!vma) {
3053 vm_unacct_memory(len >> PAGE_SHIFT);
3054 return -ENOMEM;
3055 }
3056
3057 vma_set_anonymous(vma);
3058 vma->vm_start = addr;
3059 vma->vm_end = addr + len;
3060 vma->vm_pgoff = pgoff;
3061 vma->vm_flags = flags;
3062 vma->vm_page_prot = vm_get_page_prot(flags);
3063 vma_link(mm, vma, prev, rb_link, rb_parent);
3064out:
3065 perf_event_mmap(vma);
3066 mm->total_vm += len >> PAGE_SHIFT;
3067 mm->data_vm += len >> PAGE_SHIFT;
3068 if (flags & VM_LOCKED)
3069 mm->locked_vm += (len >> PAGE_SHIFT);
3070 vma->vm_flags |= VM_SOFTDIRTY;
3071 return 0;
3072}
3073
3074int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3075{
3076 struct mm_struct *mm = current->mm;
3077 unsigned long len;
3078 int ret;
3079 bool populate;
3080 LIST_HEAD(uf);
3081
3082 len = PAGE_ALIGN(request);
3083 if (len < request)
3084 return -ENOMEM;
3085 if (!len)
3086 return 0;
3087
3088 if (down_write_killable(&mm->mmap_sem))
3089 return -EINTR;
3090
3091 ret = do_brk_flags(addr, len, flags, &uf);
3092 populate = ((mm->def_flags & VM_LOCKED) != 0);
3093 up_write(&mm->mmap_sem);
3094 userfaultfd_unmap_complete(mm, &uf);
3095 if (populate && !ret)
3096 mm_populate(addr, len);
3097 return ret;
3098}
3099EXPORT_SYMBOL(vm_brk_flags);
3100
3101int vm_brk(unsigned long addr, unsigned long len)
3102{
3103 return vm_brk_flags(addr, len, 0);
3104}
3105EXPORT_SYMBOL(vm_brk);
3106
3107/* Release all mmaps. */
3108void exit_mmap(struct mm_struct *mm)
3109{
3110 struct mmu_gather tlb;
3111 struct vm_area_struct *vma;
3112 unsigned long nr_accounted = 0;
3113
3114 /* mm's last user has gone, and its about to be pulled down */
3115 mmu_notifier_release(mm);
3116
3117 if (unlikely(mm_is_oom_victim(mm))) {
3118 /*
3119 * Manually reap the mm to free as much memory as possible.
3120 * Then, as the oom reaper does, set MMF_OOM_SKIP to disregard
3121 * this mm from further consideration. Taking mm->mmap_sem for
3122 * write after setting MMF_OOM_SKIP will guarantee that the oom
3123 * reaper will not run on this mm again after mmap_sem is
3124 * dropped.
3125 *
3126 * Nothing can be holding mm->mmap_sem here and the above call
3127 * to mmu_notifier_release(mm) ensures mmu notifier callbacks in
3128 * __oom_reap_task_mm() will not block.
3129 *
3130 * This needs to be done before calling munlock_vma_pages_all(),
3131 * which clears VM_LOCKED, otherwise the oom reaper cannot
3132 * reliably test it.
3133 */
3134 (void)__oom_reap_task_mm(mm);
3135
3136 set_bit(MMF_OOM_SKIP, &mm->flags);
3137 down_write(&mm->mmap_sem);
3138 up_write(&mm->mmap_sem);
3139 }
3140
3141 if (mm->locked_vm) {
3142 vma = mm->mmap;
3143 while (vma) {
3144 if (vma->vm_flags & VM_LOCKED)
3145 munlock_vma_pages_all(vma);
3146 vma = vma->vm_next;
3147 }
3148 }
3149
3150 arch_exit_mmap(mm);
3151
3152 vma = mm->mmap;
3153 if (!vma) /* Can happen if dup_mmap() received an OOM */
3154 return;
3155
3156 lru_add_drain();
3157 flush_cache_mm(mm);
3158 tlb_gather_mmu(&tlb, mm, 0, -1);
3159 /* update_hiwater_rss(mm) here? but nobody should be looking */
3160 /* Use -1 here to ensure all VMAs in the mm are unmapped */
3161 unmap_vmas(&tlb, vma, 0, -1);
3162 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3163 tlb_finish_mmu(&tlb, 0, -1);
3164
3165 /*
3166 * Walk the list again, actually closing and freeing it,
3167 * with preemption enabled, without holding any MM locks.
3168 */
3169 while (vma) {
3170 if (vma->vm_flags & VM_ACCOUNT)
3171 nr_accounted += vma_pages(vma);
3172 vma = remove_vma(vma);
3173 }
3174 vm_unacct_memory(nr_accounted);
3175}
3176
3177/* Insert vm structure into process list sorted by address
3178 * and into the inode's i_mmap tree. If vm_file is non-NULL
3179 * then i_mmap_rwsem is taken here.
3180 */
3181int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3182{
3183 struct vm_area_struct *prev;
3184 struct rb_node **rb_link, *rb_parent;
3185
3186 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3187 &prev, &rb_link, &rb_parent))
3188 return -ENOMEM;
3189 if ((vma->vm_flags & VM_ACCOUNT) &&
3190 security_vm_enough_memory_mm(mm, vma_pages(vma)))
3191 return -ENOMEM;
3192
3193 /*
3194 * The vm_pgoff of a purely anonymous vma should be irrelevant
3195 * until its first write fault, when page's anon_vma and index
3196 * are set. But now set the vm_pgoff it will almost certainly
3197 * end up with (unless mremap moves it elsewhere before that
3198 * first wfault), so /proc/pid/maps tells a consistent story.
3199 *
3200 * By setting it to reflect the virtual start address of the
3201 * vma, merges and splits can happen in a seamless way, just
3202 * using the existing file pgoff checks and manipulations.
3203 * Similarly in do_mmap_pgoff and in do_brk.
3204 */
3205 if (vma_is_anonymous(vma)) {
3206 BUG_ON(vma->anon_vma);
3207 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3208 }
3209
3210 vma_link(mm, vma, prev, rb_link, rb_parent);
3211 return 0;
3212}
3213
3214/*
3215 * Copy the vma structure to a new location in the same mm,
3216 * prior to moving page table entries, to effect an mremap move.
3217 */
3218struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3219 unsigned long addr, unsigned long len, pgoff_t pgoff,
3220 bool *need_rmap_locks)
3221{
3222 struct vm_area_struct *vma = *vmap;
3223 unsigned long vma_start = vma->vm_start;
3224 struct mm_struct *mm = vma->vm_mm;
3225 struct vm_area_struct *new_vma, *prev;
3226 struct rb_node **rb_link, *rb_parent;
3227 bool faulted_in_anon_vma = true;
3228
3229 /*
3230 * If anonymous vma has not yet been faulted, update new pgoff
3231 * to match new location, to increase its chance of merging.
3232 */
3233 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3234 pgoff = addr >> PAGE_SHIFT;
3235 faulted_in_anon_vma = false;
3236 }
3237
3238 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3239 return NULL; /* should never get here */
3240 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3241 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3242 vma->vm_userfaultfd_ctx);
3243 if (new_vma) {
3244 /*
3245 * Source vma may have been merged into new_vma
3246 */
3247 if (unlikely(vma_start >= new_vma->vm_start &&
3248 vma_start < new_vma->vm_end)) {
3249 /*
3250 * The only way we can get a vma_merge with
3251 * self during an mremap is if the vma hasn't
3252 * been faulted in yet and we were allowed to
3253 * reset the dst vma->vm_pgoff to the
3254 * destination address of the mremap to allow
3255 * the merge to happen. mremap must change the
3256 * vm_pgoff linearity between src and dst vmas
3257 * (in turn preventing a vma_merge) to be
3258 * safe. It is only safe to keep the vm_pgoff
3259 * linear if there are no pages mapped yet.
3260 */
3261 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3262 *vmap = vma = new_vma;
3263 }
3264 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3265 } else {
3266 new_vma = vm_area_dup(vma);
3267 if (!new_vma)
3268 goto out;
3269 new_vma->vm_start = addr;
3270 new_vma->vm_end = addr + len;
3271 new_vma->vm_pgoff = pgoff;
3272 if (vma_dup_policy(vma, new_vma))
3273 goto out_free_vma;
3274 if (anon_vma_clone(new_vma, vma))
3275 goto out_free_mempol;
3276 if (new_vma->vm_file)
3277 get_file(new_vma->vm_file);
3278 if (new_vma->vm_ops && new_vma->vm_ops->open)
3279 new_vma->vm_ops->open(new_vma);
3280 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3281 *need_rmap_locks = false;
3282 }
3283 return new_vma;
3284
3285out_free_mempol:
3286 mpol_put(vma_policy(new_vma));
3287out_free_vma:
3288 vm_area_free(new_vma);
3289out:
3290 return NULL;
3291}
3292
3293/*
3294 * Return true if the calling process may expand its vm space by the passed
3295 * number of pages
3296 */
3297bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3298{
3299 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3300 return false;
3301
3302 if (is_data_mapping(flags) &&
3303 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3304 /* Workaround for Valgrind */
3305 if (rlimit(RLIMIT_DATA) == 0 &&
3306 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3307 return true;
3308
3309 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3310 current->comm, current->pid,
3311 (mm->data_vm + npages) << PAGE_SHIFT,
3312 rlimit(RLIMIT_DATA),
3313 ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3314
3315 if (!ignore_rlimit_data)
3316 return false;
3317 }
3318
3319 return true;
3320}
3321
3322void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3323{
3324 mm->total_vm += npages;
3325
3326 if (is_exec_mapping(flags))
3327 mm->exec_vm += npages;
3328 else if (is_stack_mapping(flags))
3329 mm->stack_vm += npages;
3330 else if (is_data_mapping(flags))
3331 mm->data_vm += npages;
3332}
3333
3334static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3335
3336/*
3337 * Having a close hook prevents vma merging regardless of flags.
3338 */
3339static void special_mapping_close(struct vm_area_struct *vma)
3340{
3341}
3342
3343static const char *special_mapping_name(struct vm_area_struct *vma)
3344{
3345 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3346}
3347
3348static int special_mapping_mremap(struct vm_area_struct *new_vma)
3349{
3350 struct vm_special_mapping *sm = new_vma->vm_private_data;
3351
3352 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3353 return -EFAULT;
3354
3355 if (sm->mremap)
3356 return sm->mremap(sm, new_vma);
3357
3358 return 0;
3359}
3360
3361static const struct vm_operations_struct special_mapping_vmops = {
3362 .close = special_mapping_close,
3363 .fault = special_mapping_fault,
3364 .mremap = special_mapping_mremap,
3365 .name = special_mapping_name,
3366};
3367
3368static const struct vm_operations_struct legacy_special_mapping_vmops = {
3369 .close = special_mapping_close,
3370 .fault = special_mapping_fault,
3371};
3372
3373static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3374{
3375 struct vm_area_struct *vma = vmf->vma;
3376 pgoff_t pgoff;
3377 struct page **pages;
3378
3379 if (vma->vm_ops == &legacy_special_mapping_vmops) {
3380 pages = vma->vm_private_data;
3381 } else {
3382 struct vm_special_mapping *sm = vma->vm_private_data;
3383
3384 if (sm->fault)
3385 return sm->fault(sm, vmf->vma, vmf);
3386
3387 pages = sm->pages;
3388 }
3389
3390 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3391 pgoff--;
3392
3393 if (*pages) {
3394 struct page *page = *pages;
3395 get_page(page);
3396 vmf->page = page;
3397 return 0;
3398 }
3399
3400 return VM_FAULT_SIGBUS;
3401}
3402
3403static struct vm_area_struct *__install_special_mapping(
3404 struct mm_struct *mm,
3405 unsigned long addr, unsigned long len,
3406 unsigned long vm_flags, void *priv,
3407 const struct vm_operations_struct *ops)
3408{
3409 int ret;
3410 struct vm_area_struct *vma;
3411
3412 vma = vm_area_alloc(mm);
3413 if (unlikely(vma == NULL))
3414 return ERR_PTR(-ENOMEM);
3415
3416 vma->vm_start = addr;
3417 vma->vm_end = addr + len;
3418
3419 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3420 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3421
3422 vma->vm_ops = ops;
3423 vma->vm_private_data = priv;
3424
3425 ret = insert_vm_struct(mm, vma);
3426 if (ret)
3427 goto out;
3428
3429 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3430
3431 perf_event_mmap(vma);
3432
3433 return vma;
3434
3435out:
3436 vm_area_free(vma);
3437 return ERR_PTR(ret);
3438}
3439
3440bool vma_is_special_mapping(const struct vm_area_struct *vma,
3441 const struct vm_special_mapping *sm)
3442{
3443 return vma->vm_private_data == sm &&
3444 (vma->vm_ops == &special_mapping_vmops ||
3445 vma->vm_ops == &legacy_special_mapping_vmops);
3446}
3447
3448/*
3449 * Called with mm->mmap_sem held for writing.
3450 * Insert a new vma covering the given region, with the given flags.
3451 * Its pages are supplied by the given array of struct page *.
3452 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3453 * The region past the last page supplied will always produce SIGBUS.
3454 * The array pointer and the pages it points to are assumed to stay alive
3455 * for as long as this mapping might exist.
3456 */
3457struct vm_area_struct *_install_special_mapping(
3458 struct mm_struct *mm,
3459 unsigned long addr, unsigned long len,
3460 unsigned long vm_flags, const struct vm_special_mapping *spec)
3461{
3462 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3463 &special_mapping_vmops);
3464}
3465
3466int install_special_mapping(struct mm_struct *mm,
3467 unsigned long addr, unsigned long len,
3468 unsigned long vm_flags, struct page **pages)
3469{
3470 struct vm_area_struct *vma = __install_special_mapping(
3471 mm, addr, len, vm_flags, (void *)pages,
3472 &legacy_special_mapping_vmops);
3473
3474 return PTR_ERR_OR_ZERO(vma);
3475}
3476
3477static DEFINE_MUTEX(mm_all_locks_mutex);
3478
3479static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3480{
3481 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3482 /*
3483 * The LSB of head.next can't change from under us
3484 * because we hold the mm_all_locks_mutex.
3485 */
3486 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
3487 /*
3488 * We can safely modify head.next after taking the
3489 * anon_vma->root->rwsem. If some other vma in this mm shares
3490 * the same anon_vma we won't take it again.
3491 *
3492 * No need of atomic instructions here, head.next
3493 * can't change from under us thanks to the
3494 * anon_vma->root->rwsem.
3495 */
3496 if (__test_and_set_bit(0, (unsigned long *)
3497 &anon_vma->root->rb_root.rb_root.rb_node))
3498 BUG();
3499 }
3500}
3501
3502static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3503{
3504 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3505 /*
3506 * AS_MM_ALL_LOCKS can't change from under us because
3507 * we hold the mm_all_locks_mutex.
3508 *
3509 * Operations on ->flags have to be atomic because
3510 * even if AS_MM_ALL_LOCKS is stable thanks to the
3511 * mm_all_locks_mutex, there may be other cpus
3512 * changing other bitflags in parallel to us.
3513 */
3514 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3515 BUG();
3516 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem);
3517 }
3518}
3519
3520/*
3521 * This operation locks against the VM for all pte/vma/mm related
3522 * operations that could ever happen on a certain mm. This includes
3523 * vmtruncate, try_to_unmap, and all page faults.
3524 *
3525 * The caller must take the mmap_sem in write mode before calling
3526 * mm_take_all_locks(). The caller isn't allowed to release the
3527 * mmap_sem until mm_drop_all_locks() returns.
3528 *
3529 * mmap_sem in write mode is required in order to block all operations
3530 * that could modify pagetables and free pages without need of
3531 * altering the vma layout. It's also needed in write mode to avoid new
3532 * anon_vmas to be associated with existing vmas.
3533 *
3534 * A single task can't take more than one mm_take_all_locks() in a row
3535 * or it would deadlock.
3536 *
3537 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3538 * mapping->flags avoid to take the same lock twice, if more than one
3539 * vma in this mm is backed by the same anon_vma or address_space.
3540 *
3541 * We take locks in following order, accordingly to comment at beginning
3542 * of mm/rmap.c:
3543 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3544 * hugetlb mapping);
3545 * - all i_mmap_rwsem locks;
3546 * - all anon_vma->rwseml
3547 *
3548 * We can take all locks within these types randomly because the VM code
3549 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3550 * mm_all_locks_mutex.
3551 *
3552 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3553 * that may have to take thousand of locks.
3554 *
3555 * mm_take_all_locks() can fail if it's interrupted by signals.
3556 */
3557int mm_take_all_locks(struct mm_struct *mm)
3558{
3559 struct vm_area_struct *vma;
3560 struct anon_vma_chain *avc;
3561
3562 BUG_ON(down_read_trylock(&mm->mmap_sem));
3563
3564 mutex_lock(&mm_all_locks_mutex);
3565
3566 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3567 if (signal_pending(current))
3568 goto out_unlock;
3569 if (vma->vm_file && vma->vm_file->f_mapping &&
3570 is_vm_hugetlb_page(vma))
3571 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3572 }
3573
3574 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3575 if (signal_pending(current))
3576 goto out_unlock;
3577 if (vma->vm_file && vma->vm_file->f_mapping &&
3578 !is_vm_hugetlb_page(vma))
3579 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3580 }
3581
3582 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3583 if (signal_pending(current))
3584 goto out_unlock;
3585 if (vma->anon_vma)
3586 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3587 vm_lock_anon_vma(mm, avc->anon_vma);
3588 }
3589
3590 return 0;
3591
3592out_unlock:
3593 mm_drop_all_locks(mm);
3594 return -EINTR;
3595}
3596
3597static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3598{
3599 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3600 /*
3601 * The LSB of head.next can't change to 0 from under
3602 * us because we hold the mm_all_locks_mutex.
3603 *
3604 * We must however clear the bitflag before unlocking
3605 * the vma so the users using the anon_vma->rb_root will
3606 * never see our bitflag.
3607 *
3608 * No need of atomic instructions here, head.next
3609 * can't change from under us until we release the
3610 * anon_vma->root->rwsem.
3611 */
3612 if (!__test_and_clear_bit(0, (unsigned long *)
3613 &anon_vma->root->rb_root.rb_root.rb_node))
3614 BUG();
3615 anon_vma_unlock_write(anon_vma);
3616 }
3617}
3618
3619static void vm_unlock_mapping(struct address_space *mapping)
3620{
3621 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3622 /*
3623 * AS_MM_ALL_LOCKS can't change to 0 from under us
3624 * because we hold the mm_all_locks_mutex.
3625 */
3626 i_mmap_unlock_write(mapping);
3627 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3628 &mapping->flags))
3629 BUG();
3630 }
3631}
3632
3633/*
3634 * The mmap_sem cannot be released by the caller until
3635 * mm_drop_all_locks() returns.
3636 */
3637void mm_drop_all_locks(struct mm_struct *mm)
3638{
3639 struct vm_area_struct *vma;
3640 struct anon_vma_chain *avc;
3641
3642 BUG_ON(down_read_trylock(&mm->mmap_sem));
3643 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3644
3645 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3646 if (vma->anon_vma)
3647 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3648 vm_unlock_anon_vma(avc->anon_vma);
3649 if (vma->vm_file && vma->vm_file->f_mapping)
3650 vm_unlock_mapping(vma->vm_file->f_mapping);
3651 }
3652
3653 mutex_unlock(&mm_all_locks_mutex);
3654}
3655
3656/*
3657 * initialise the percpu counter for VM
3658 */
3659void __init mmap_init(void)
3660{
3661 int ret;
3662
3663 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3664 VM_BUG_ON(ret);
3665}
3666
3667/*
3668 * Initialise sysctl_user_reserve_kbytes.
3669 *
3670 * This is intended to prevent a user from starting a single memory hogging
3671 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3672 * mode.
3673 *
3674 * The default value is min(3% of free memory, 128MB)
3675 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3676 */
3677static int init_user_reserve(void)
3678{
3679 unsigned long free_kbytes;
3680
3681 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3682
3683 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3684 return 0;
3685}
3686subsys_initcall(init_user_reserve);
3687
3688/*
3689 * Initialise sysctl_admin_reserve_kbytes.
3690 *
3691 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3692 * to log in and kill a memory hogging process.
3693 *
3694 * Systems with more than 256MB will reserve 8MB, enough to recover
3695 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3696 * only reserve 3% of free pages by default.
3697 */
3698static int init_admin_reserve(void)
3699{
3700 unsigned long free_kbytes;
3701
3702 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3703
3704 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3705 return 0;
3706}
3707subsys_initcall(init_admin_reserve);
3708
3709/*
3710 * Reinititalise user and admin reserves if memory is added or removed.
3711 *
3712 * The default user reserve max is 128MB, and the default max for the
3713 * admin reserve is 8MB. These are usually, but not always, enough to
3714 * enable recovery from a memory hogging process using login/sshd, a shell,
3715 * and tools like top. It may make sense to increase or even disable the
3716 * reserve depending on the existence of swap or variations in the recovery
3717 * tools. So, the admin may have changed them.
3718 *
3719 * If memory is added and the reserves have been eliminated or increased above
3720 * the default max, then we'll trust the admin.
3721 *
3722 * If memory is removed and there isn't enough free memory, then we
3723 * need to reset the reserves.
3724 *
3725 * Otherwise keep the reserve set by the admin.
3726 */
3727static int reserve_mem_notifier(struct notifier_block *nb,
3728 unsigned long action, void *data)
3729{
3730 unsigned long tmp, free_kbytes;
3731
3732 switch (action) {
3733 case MEM_ONLINE:
3734 /* Default max is 128MB. Leave alone if modified by operator. */
3735 tmp = sysctl_user_reserve_kbytes;
3736 if (0 < tmp && tmp < (1UL << 17))
3737 init_user_reserve();
3738
3739 /* Default max is 8MB. Leave alone if modified by operator. */
3740 tmp = sysctl_admin_reserve_kbytes;
3741 if (0 < tmp && tmp < (1UL << 13))
3742 init_admin_reserve();
3743
3744 break;
3745 case MEM_OFFLINE:
3746 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3747
3748 if (sysctl_user_reserve_kbytes > free_kbytes) {
3749 init_user_reserve();
3750 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3751 sysctl_user_reserve_kbytes);
3752 }
3753
3754 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3755 init_admin_reserve();
3756 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3757 sysctl_admin_reserve_kbytes);
3758 }
3759 break;
3760 default:
3761 break;
3762 }
3763 return NOTIFY_OK;
3764}
3765
3766static struct notifier_block reserve_mem_nb = {
3767 .notifier_call = reserve_mem_notifier,
3768};
3769
3770static int __meminit init_reserve_notifier(void)
3771{
3772 if (register_hotmemory_notifier(&reserve_mem_nb))
3773 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3774
3775 return 0;
3776}
3777subsys_initcall(init_reserve_notifier);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * mm/mmap.c
4 *
5 * Written by obz.
6 *
7 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
8 */
9
10#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11
12#include <linux/kernel.h>
13#include <linux/slab.h>
14#include <linux/backing-dev.h>
15#include <linux/mm.h>
16#include <linux/mm_inline.h>
17#include <linux/shm.h>
18#include <linux/mman.h>
19#include <linux/pagemap.h>
20#include <linux/swap.h>
21#include <linux/syscalls.h>
22#include <linux/capability.h>
23#include <linux/init.h>
24#include <linux/file.h>
25#include <linux/fs.h>
26#include <linux/personality.h>
27#include <linux/security.h>
28#include <linux/hugetlb.h>
29#include <linux/shmem_fs.h>
30#include <linux/profile.h>
31#include <linux/export.h>
32#include <linux/mount.h>
33#include <linux/mempolicy.h>
34#include <linux/rmap.h>
35#include <linux/mmu_notifier.h>
36#include <linux/mmdebug.h>
37#include <linux/perf_event.h>
38#include <linux/audit.h>
39#include <linux/khugepaged.h>
40#include <linux/uprobes.h>
41#include <linux/notifier.h>
42#include <linux/memory.h>
43#include <linux/printk.h>
44#include <linux/userfaultfd_k.h>
45#include <linux/moduleparam.h>
46#include <linux/pkeys.h>
47#include <linux/oom.h>
48#include <linux/sched/mm.h>
49
50#include <linux/uaccess.h>
51#include <asm/cacheflush.h>
52#include <asm/tlb.h>
53#include <asm/mmu_context.h>
54
55#define CREATE_TRACE_POINTS
56#include <trace/events/mmap.h>
57
58#include "internal.h"
59
60#ifndef arch_mmap_check
61#define arch_mmap_check(addr, len, flags) (0)
62#endif
63
64#ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
65const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
66const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
67int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
68#endif
69#ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
70const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
71const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
72int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
73#endif
74
75static bool ignore_rlimit_data;
76core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
77
78static void unmap_region(struct mm_struct *mm, struct maple_tree *mt,
79 struct vm_area_struct *vma, struct vm_area_struct *prev,
80 struct vm_area_struct *next, unsigned long start,
81 unsigned long end);
82
83static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
84{
85 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
86}
87
88/* Update vma->vm_page_prot to reflect vma->vm_flags. */
89void vma_set_page_prot(struct vm_area_struct *vma)
90{
91 unsigned long vm_flags = vma->vm_flags;
92 pgprot_t vm_page_prot;
93
94 vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
95 if (vma_wants_writenotify(vma, vm_page_prot)) {
96 vm_flags &= ~VM_SHARED;
97 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
98 }
99 /* remove_protection_ptes reads vma->vm_page_prot without mmap_lock */
100 WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
101}
102
103/*
104 * Requires inode->i_mapping->i_mmap_rwsem
105 */
106static void __remove_shared_vm_struct(struct vm_area_struct *vma,
107 struct file *file, struct address_space *mapping)
108{
109 if (vma->vm_flags & VM_SHARED)
110 mapping_unmap_writable(mapping);
111
112 flush_dcache_mmap_lock(mapping);
113 vma_interval_tree_remove(vma, &mapping->i_mmap);
114 flush_dcache_mmap_unlock(mapping);
115}
116
117/*
118 * Unlink a file-based vm structure from its interval tree, to hide
119 * vma from rmap and vmtruncate before freeing its page tables.
120 */
121void unlink_file_vma(struct vm_area_struct *vma)
122{
123 struct file *file = vma->vm_file;
124
125 if (file) {
126 struct address_space *mapping = file->f_mapping;
127 i_mmap_lock_write(mapping);
128 __remove_shared_vm_struct(vma, file, mapping);
129 i_mmap_unlock_write(mapping);
130 }
131}
132
133/*
134 * Close a vm structure and free it.
135 */
136static void remove_vma(struct vm_area_struct *vma)
137{
138 might_sleep();
139 if (vma->vm_ops && vma->vm_ops->close)
140 vma->vm_ops->close(vma);
141 if (vma->vm_file)
142 fput(vma->vm_file);
143 mpol_put(vma_policy(vma));
144 vm_area_free(vma);
145}
146
147/*
148 * check_brk_limits() - Use platform specific check of range & verify mlock
149 * limits.
150 * @addr: The address to check
151 * @len: The size of increase.
152 *
153 * Return: 0 on success.
154 */
155static int check_brk_limits(unsigned long addr, unsigned long len)
156{
157 unsigned long mapped_addr;
158
159 mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
160 if (IS_ERR_VALUE(mapped_addr))
161 return mapped_addr;
162
163 return mlock_future_check(current->mm, current->mm->def_flags, len);
164}
165static int do_brk_munmap(struct ma_state *mas, struct vm_area_struct *vma,
166 unsigned long newbrk, unsigned long oldbrk,
167 struct list_head *uf);
168static int do_brk_flags(struct ma_state *mas, struct vm_area_struct *brkvma,
169 unsigned long addr, unsigned long request, unsigned long flags);
170SYSCALL_DEFINE1(brk, unsigned long, brk)
171{
172 unsigned long newbrk, oldbrk, origbrk;
173 struct mm_struct *mm = current->mm;
174 struct vm_area_struct *brkvma, *next = NULL;
175 unsigned long min_brk;
176 bool populate;
177 bool downgraded = false;
178 LIST_HEAD(uf);
179 MA_STATE(mas, &mm->mm_mt, 0, 0);
180
181 if (mmap_write_lock_killable(mm))
182 return -EINTR;
183
184 origbrk = mm->brk;
185
186#ifdef CONFIG_COMPAT_BRK
187 /*
188 * CONFIG_COMPAT_BRK can still be overridden by setting
189 * randomize_va_space to 2, which will still cause mm->start_brk
190 * to be arbitrarily shifted
191 */
192 if (current->brk_randomized)
193 min_brk = mm->start_brk;
194 else
195 min_brk = mm->end_data;
196#else
197 min_brk = mm->start_brk;
198#endif
199 if (brk < min_brk)
200 goto out;
201
202 /*
203 * Check against rlimit here. If this check is done later after the test
204 * of oldbrk with newbrk then it can escape the test and let the data
205 * segment grow beyond its set limit the in case where the limit is
206 * not page aligned -Ram Gupta
207 */
208 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
209 mm->end_data, mm->start_data))
210 goto out;
211
212 newbrk = PAGE_ALIGN(brk);
213 oldbrk = PAGE_ALIGN(mm->brk);
214 if (oldbrk == newbrk) {
215 mm->brk = brk;
216 goto success;
217 }
218
219 /*
220 * Always allow shrinking brk.
221 * do_brk_munmap() may downgrade mmap_lock to read.
222 */
223 if (brk <= mm->brk) {
224 int ret;
225
226 /* Search one past newbrk */
227 mas_set(&mas, newbrk);
228 brkvma = mas_find(&mas, oldbrk);
229 if (!brkvma || brkvma->vm_start >= oldbrk)
230 goto out; /* mapping intersects with an existing non-brk vma. */
231 /*
232 * mm->brk must be protected by write mmap_lock.
233 * do_brk_munmap() may downgrade the lock, so update it
234 * before calling do_brk_munmap().
235 */
236 mm->brk = brk;
237 ret = do_brk_munmap(&mas, brkvma, newbrk, oldbrk, &uf);
238 if (ret == 1) {
239 downgraded = true;
240 goto success;
241 } else if (!ret)
242 goto success;
243
244 mm->brk = origbrk;
245 goto out;
246 }
247
248 if (check_brk_limits(oldbrk, newbrk - oldbrk))
249 goto out;
250
251 /*
252 * Only check if the next VMA is within the stack_guard_gap of the
253 * expansion area
254 */
255 mas_set(&mas, oldbrk);
256 next = mas_find(&mas, newbrk - 1 + PAGE_SIZE + stack_guard_gap);
257 if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
258 goto out;
259
260 brkvma = mas_prev(&mas, mm->start_brk);
261 /* Ok, looks good - let it rip. */
262 if (do_brk_flags(&mas, brkvma, oldbrk, newbrk - oldbrk, 0) < 0)
263 goto out;
264
265 mm->brk = brk;
266
267success:
268 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
269 if (downgraded)
270 mmap_read_unlock(mm);
271 else
272 mmap_write_unlock(mm);
273 userfaultfd_unmap_complete(mm, &uf);
274 if (populate)
275 mm_populate(oldbrk, newbrk - oldbrk);
276 return brk;
277
278out:
279 mmap_write_unlock(mm);
280 return origbrk;
281}
282
283#if defined(CONFIG_DEBUG_VM_MAPLE_TREE)
284extern void mt_validate(struct maple_tree *mt);
285extern void mt_dump(const struct maple_tree *mt);
286
287/* Validate the maple tree */
288static void validate_mm_mt(struct mm_struct *mm)
289{
290 struct maple_tree *mt = &mm->mm_mt;
291 struct vm_area_struct *vma_mt;
292
293 MA_STATE(mas, mt, 0, 0);
294
295 mt_validate(&mm->mm_mt);
296 mas_for_each(&mas, vma_mt, ULONG_MAX) {
297 if ((vma_mt->vm_start != mas.index) ||
298 (vma_mt->vm_end - 1 != mas.last)) {
299 pr_emerg("issue in %s\n", current->comm);
300 dump_stack();
301 dump_vma(vma_mt);
302 pr_emerg("mt piv: %p %lu - %lu\n", vma_mt,
303 mas.index, mas.last);
304 pr_emerg("mt vma: %p %lu - %lu\n", vma_mt,
305 vma_mt->vm_start, vma_mt->vm_end);
306
307 mt_dump(mas.tree);
308 if (vma_mt->vm_end != mas.last + 1) {
309 pr_err("vma: %p vma_mt %lu-%lu\tmt %lu-%lu\n",
310 mm, vma_mt->vm_start, vma_mt->vm_end,
311 mas.index, mas.last);
312 mt_dump(mas.tree);
313 }
314 VM_BUG_ON_MM(vma_mt->vm_end != mas.last + 1, mm);
315 if (vma_mt->vm_start != mas.index) {
316 pr_err("vma: %p vma_mt %p %lu - %lu doesn't match\n",
317 mm, vma_mt, vma_mt->vm_start, vma_mt->vm_end);
318 mt_dump(mas.tree);
319 }
320 VM_BUG_ON_MM(vma_mt->vm_start != mas.index, mm);
321 }
322 }
323}
324
325static void validate_mm(struct mm_struct *mm)
326{
327 int bug = 0;
328 int i = 0;
329 struct vm_area_struct *vma;
330 MA_STATE(mas, &mm->mm_mt, 0, 0);
331
332 validate_mm_mt(mm);
333
334 mas_for_each(&mas, vma, ULONG_MAX) {
335#ifdef CONFIG_DEBUG_VM_RB
336 struct anon_vma *anon_vma = vma->anon_vma;
337 struct anon_vma_chain *avc;
338
339 if (anon_vma) {
340 anon_vma_lock_read(anon_vma);
341 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
342 anon_vma_interval_tree_verify(avc);
343 anon_vma_unlock_read(anon_vma);
344 }
345#endif
346 i++;
347 }
348 if (i != mm->map_count) {
349 pr_emerg("map_count %d mas_for_each %d\n", mm->map_count, i);
350 bug = 1;
351 }
352 VM_BUG_ON_MM(bug, mm);
353}
354
355#else /* !CONFIG_DEBUG_VM_MAPLE_TREE */
356#define validate_mm_mt(root) do { } while (0)
357#define validate_mm(mm) do { } while (0)
358#endif /* CONFIG_DEBUG_VM_MAPLE_TREE */
359
360/*
361 * vma has some anon_vma assigned, and is already inserted on that
362 * anon_vma's interval trees.
363 *
364 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
365 * vma must be removed from the anon_vma's interval trees using
366 * anon_vma_interval_tree_pre_update_vma().
367 *
368 * After the update, the vma will be reinserted using
369 * anon_vma_interval_tree_post_update_vma().
370 *
371 * The entire update must be protected by exclusive mmap_lock and by
372 * the root anon_vma's mutex.
373 */
374static inline void
375anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
376{
377 struct anon_vma_chain *avc;
378
379 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
380 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
381}
382
383static inline void
384anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
385{
386 struct anon_vma_chain *avc;
387
388 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
389 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
390}
391
392static unsigned long count_vma_pages_range(struct mm_struct *mm,
393 unsigned long addr, unsigned long end)
394{
395 VMA_ITERATOR(vmi, mm, addr);
396 struct vm_area_struct *vma;
397 unsigned long nr_pages = 0;
398
399 for_each_vma_range(vmi, vma, end) {
400 unsigned long vm_start = max(addr, vma->vm_start);
401 unsigned long vm_end = min(end, vma->vm_end);
402
403 nr_pages += PHYS_PFN(vm_end - vm_start);
404 }
405
406 return nr_pages;
407}
408
409static void __vma_link_file(struct vm_area_struct *vma,
410 struct address_space *mapping)
411{
412 if (vma->vm_flags & VM_SHARED)
413 mapping_allow_writable(mapping);
414
415 flush_dcache_mmap_lock(mapping);
416 vma_interval_tree_insert(vma, &mapping->i_mmap);
417 flush_dcache_mmap_unlock(mapping);
418}
419
420/*
421 * vma_mas_store() - Store a VMA in the maple tree.
422 * @vma: The vm_area_struct
423 * @mas: The maple state
424 *
425 * Efficient way to store a VMA in the maple tree when the @mas has already
426 * walked to the correct location.
427 *
428 * Note: the end address is inclusive in the maple tree.
429 */
430void vma_mas_store(struct vm_area_struct *vma, struct ma_state *mas)
431{
432 trace_vma_store(mas->tree, vma);
433 mas_set_range(mas, vma->vm_start, vma->vm_end - 1);
434 mas_store_prealloc(mas, vma);
435}
436
437/*
438 * vma_mas_remove() - Remove a VMA from the maple tree.
439 * @vma: The vm_area_struct
440 * @mas: The maple state
441 *
442 * Efficient way to remove a VMA from the maple tree when the @mas has already
443 * been established and points to the correct location.
444 * Note: the end address is inclusive in the maple tree.
445 */
446void vma_mas_remove(struct vm_area_struct *vma, struct ma_state *mas)
447{
448 trace_vma_mas_szero(mas->tree, vma->vm_start, vma->vm_end - 1);
449 mas->index = vma->vm_start;
450 mas->last = vma->vm_end - 1;
451 mas_store_prealloc(mas, NULL);
452}
453
454/*
455 * vma_mas_szero() - Set a given range to zero. Used when modifying a
456 * vm_area_struct start or end.
457 *
458 * @mas: The maple tree ma_state
459 * @start: The start address to zero
460 * @end: The end address to zero.
461 */
462static inline void vma_mas_szero(struct ma_state *mas, unsigned long start,
463 unsigned long end)
464{
465 trace_vma_mas_szero(mas->tree, start, end - 1);
466 mas_set_range(mas, start, end - 1);
467 mas_store_prealloc(mas, NULL);
468}
469
470static int vma_link(struct mm_struct *mm, struct vm_area_struct *vma)
471{
472 MA_STATE(mas, &mm->mm_mt, 0, 0);
473 struct address_space *mapping = NULL;
474
475 if (mas_preallocate(&mas, vma, GFP_KERNEL))
476 return -ENOMEM;
477
478 if (vma->vm_file) {
479 mapping = vma->vm_file->f_mapping;
480 i_mmap_lock_write(mapping);
481 }
482
483 vma_mas_store(vma, &mas);
484
485 if (mapping) {
486 __vma_link_file(vma, mapping);
487 i_mmap_unlock_write(mapping);
488 }
489
490 mm->map_count++;
491 validate_mm(mm);
492 return 0;
493}
494
495/*
496 * vma_expand - Expand an existing VMA
497 *
498 * @mas: The maple state
499 * @vma: The vma to expand
500 * @start: The start of the vma
501 * @end: The exclusive end of the vma
502 * @pgoff: The page offset of vma
503 * @next: The current of next vma.
504 *
505 * Expand @vma to @start and @end. Can expand off the start and end. Will
506 * expand over @next if it's different from @vma and @end == @next->vm_end.
507 * Checking if the @vma can expand and merge with @next needs to be handled by
508 * the caller.
509 *
510 * Returns: 0 on success
511 */
512inline int vma_expand(struct ma_state *mas, struct vm_area_struct *vma,
513 unsigned long start, unsigned long end, pgoff_t pgoff,
514 struct vm_area_struct *next)
515{
516 struct mm_struct *mm = vma->vm_mm;
517 struct address_space *mapping = NULL;
518 struct rb_root_cached *root = NULL;
519 struct anon_vma *anon_vma = vma->anon_vma;
520 struct file *file = vma->vm_file;
521 bool remove_next = false;
522
523 if (next && (vma != next) && (end == next->vm_end)) {
524 remove_next = true;
525 if (next->anon_vma && !vma->anon_vma) {
526 int error;
527
528 anon_vma = next->anon_vma;
529 vma->anon_vma = anon_vma;
530 error = anon_vma_clone(vma, next);
531 if (error)
532 return error;
533 }
534 }
535
536 /* Not merging but overwriting any part of next is not handled. */
537 VM_BUG_ON(next && !remove_next && next != vma && end > next->vm_start);
538 /* Only handles expanding */
539 VM_BUG_ON(vma->vm_start < start || vma->vm_end > end);
540
541 if (mas_preallocate(mas, vma, GFP_KERNEL))
542 goto nomem;
543
544 vma_adjust_trans_huge(vma, start, end, 0);
545
546 if (file) {
547 mapping = file->f_mapping;
548 root = &mapping->i_mmap;
549 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
550 i_mmap_lock_write(mapping);
551 }
552
553 if (anon_vma) {
554 anon_vma_lock_write(anon_vma);
555 anon_vma_interval_tree_pre_update_vma(vma);
556 }
557
558 if (file) {
559 flush_dcache_mmap_lock(mapping);
560 vma_interval_tree_remove(vma, root);
561 }
562
563 vma->vm_start = start;
564 vma->vm_end = end;
565 vma->vm_pgoff = pgoff;
566 /* Note: mas must be pointing to the expanding VMA */
567 vma_mas_store(vma, mas);
568
569 if (file) {
570 vma_interval_tree_insert(vma, root);
571 flush_dcache_mmap_unlock(mapping);
572 }
573
574 /* Expanding over the next vma */
575 if (remove_next && file) {
576 __remove_shared_vm_struct(next, file, mapping);
577 }
578
579 if (anon_vma) {
580 anon_vma_interval_tree_post_update_vma(vma);
581 anon_vma_unlock_write(anon_vma);
582 }
583
584 if (file) {
585 i_mmap_unlock_write(mapping);
586 uprobe_mmap(vma);
587 }
588
589 if (remove_next) {
590 if (file) {
591 uprobe_munmap(next, next->vm_start, next->vm_end);
592 fput(file);
593 }
594 if (next->anon_vma)
595 anon_vma_merge(vma, next);
596 mm->map_count--;
597 mpol_put(vma_policy(next));
598 vm_area_free(next);
599 }
600
601 validate_mm(mm);
602 return 0;
603
604nomem:
605 return -ENOMEM;
606}
607
608/*
609 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
610 * is already present in an i_mmap tree without adjusting the tree.
611 * The following helper function should be used when such adjustments
612 * are necessary. The "insert" vma (if any) is to be inserted
613 * before we drop the necessary locks.
614 */
615int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
616 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
617 struct vm_area_struct *expand)
618{
619 struct mm_struct *mm = vma->vm_mm;
620 struct vm_area_struct *next_next = NULL; /* uninit var warning */
621 struct vm_area_struct *next = find_vma(mm, vma->vm_end);
622 struct vm_area_struct *orig_vma = vma;
623 struct address_space *mapping = NULL;
624 struct rb_root_cached *root = NULL;
625 struct anon_vma *anon_vma = NULL;
626 struct file *file = vma->vm_file;
627 bool vma_changed = false;
628 long adjust_next = 0;
629 int remove_next = 0;
630 MA_STATE(mas, &mm->mm_mt, 0, 0);
631 struct vm_area_struct *exporter = NULL, *importer = NULL;
632
633 if (next && !insert) {
634 if (end >= next->vm_end) {
635 /*
636 * vma expands, overlapping all the next, and
637 * perhaps the one after too (mprotect case 6).
638 * The only other cases that gets here are
639 * case 1, case 7 and case 8.
640 */
641 if (next == expand) {
642 /*
643 * The only case where we don't expand "vma"
644 * and we expand "next" instead is case 8.
645 */
646 VM_WARN_ON(end != next->vm_end);
647 /*
648 * remove_next == 3 means we're
649 * removing "vma" and that to do so we
650 * swapped "vma" and "next".
651 */
652 remove_next = 3;
653 VM_WARN_ON(file != next->vm_file);
654 swap(vma, next);
655 } else {
656 VM_WARN_ON(expand != vma);
657 /*
658 * case 1, 6, 7, remove_next == 2 is case 6,
659 * remove_next == 1 is case 1 or 7.
660 */
661 remove_next = 1 + (end > next->vm_end);
662 if (remove_next == 2)
663 next_next = find_vma(mm, next->vm_end);
664
665 VM_WARN_ON(remove_next == 2 &&
666 end != next_next->vm_end);
667 }
668
669 exporter = next;
670 importer = vma;
671
672 /*
673 * If next doesn't have anon_vma, import from vma after
674 * next, if the vma overlaps with it.
675 */
676 if (remove_next == 2 && !next->anon_vma)
677 exporter = next_next;
678
679 } else if (end > next->vm_start) {
680 /*
681 * vma expands, overlapping part of the next:
682 * mprotect case 5 shifting the boundary up.
683 */
684 adjust_next = (end - next->vm_start);
685 exporter = next;
686 importer = vma;
687 VM_WARN_ON(expand != importer);
688 } else if (end < vma->vm_end) {
689 /*
690 * vma shrinks, and !insert tells it's not
691 * split_vma inserting another: so it must be
692 * mprotect case 4 shifting the boundary down.
693 */
694 adjust_next = -(vma->vm_end - end);
695 exporter = vma;
696 importer = next;
697 VM_WARN_ON(expand != importer);
698 }
699
700 /*
701 * Easily overlooked: when mprotect shifts the boundary,
702 * make sure the expanding vma has anon_vma set if the
703 * shrinking vma had, to cover any anon pages imported.
704 */
705 if (exporter && exporter->anon_vma && !importer->anon_vma) {
706 int error;
707
708 importer->anon_vma = exporter->anon_vma;
709 error = anon_vma_clone(importer, exporter);
710 if (error)
711 return error;
712 }
713 }
714
715 if (mas_preallocate(&mas, vma, GFP_KERNEL))
716 return -ENOMEM;
717
718 vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
719 if (file) {
720 mapping = file->f_mapping;
721 root = &mapping->i_mmap;
722 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
723
724 if (adjust_next)
725 uprobe_munmap(next, next->vm_start, next->vm_end);
726
727 i_mmap_lock_write(mapping);
728 if (insert && insert->vm_file) {
729 /*
730 * Put into interval tree now, so instantiated pages
731 * are visible to arm/parisc __flush_dcache_page
732 * throughout; but we cannot insert into address
733 * space until vma start or end is updated.
734 */
735 __vma_link_file(insert, insert->vm_file->f_mapping);
736 }
737 }
738
739 anon_vma = vma->anon_vma;
740 if (!anon_vma && adjust_next)
741 anon_vma = next->anon_vma;
742 if (anon_vma) {
743 VM_WARN_ON(adjust_next && next->anon_vma &&
744 anon_vma != next->anon_vma);
745 anon_vma_lock_write(anon_vma);
746 anon_vma_interval_tree_pre_update_vma(vma);
747 if (adjust_next)
748 anon_vma_interval_tree_pre_update_vma(next);
749 }
750
751 if (file) {
752 flush_dcache_mmap_lock(mapping);
753 vma_interval_tree_remove(vma, root);
754 if (adjust_next)
755 vma_interval_tree_remove(next, root);
756 }
757
758 if (start != vma->vm_start) {
759 if ((vma->vm_start < start) &&
760 (!insert || (insert->vm_end != start))) {
761 vma_mas_szero(&mas, vma->vm_start, start);
762 VM_WARN_ON(insert && insert->vm_start > vma->vm_start);
763 } else {
764 vma_changed = true;
765 }
766 vma->vm_start = start;
767 }
768 if (end != vma->vm_end) {
769 if (vma->vm_end > end) {
770 if (!insert || (insert->vm_start != end)) {
771 vma_mas_szero(&mas, end, vma->vm_end);
772 mas_reset(&mas);
773 VM_WARN_ON(insert &&
774 insert->vm_end < vma->vm_end);
775 }
776 } else {
777 vma_changed = true;
778 }
779 vma->vm_end = end;
780 }
781
782 if (vma_changed)
783 vma_mas_store(vma, &mas);
784
785 vma->vm_pgoff = pgoff;
786 if (adjust_next) {
787 next->vm_start += adjust_next;
788 next->vm_pgoff += adjust_next >> PAGE_SHIFT;
789 vma_mas_store(next, &mas);
790 }
791
792 if (file) {
793 if (adjust_next)
794 vma_interval_tree_insert(next, root);
795 vma_interval_tree_insert(vma, root);
796 flush_dcache_mmap_unlock(mapping);
797 }
798
799 if (remove_next && file) {
800 __remove_shared_vm_struct(next, file, mapping);
801 if (remove_next == 2)
802 __remove_shared_vm_struct(next_next, file, mapping);
803 } else if (insert) {
804 /*
805 * split_vma has split insert from vma, and needs
806 * us to insert it before dropping the locks
807 * (it may either follow vma or precede it).
808 */
809 mas_reset(&mas);
810 vma_mas_store(insert, &mas);
811 mm->map_count++;
812 }
813
814 if (anon_vma) {
815 anon_vma_interval_tree_post_update_vma(vma);
816 if (adjust_next)
817 anon_vma_interval_tree_post_update_vma(next);
818 anon_vma_unlock_write(anon_vma);
819 }
820
821 if (file) {
822 i_mmap_unlock_write(mapping);
823 uprobe_mmap(vma);
824
825 if (adjust_next)
826 uprobe_mmap(next);
827 }
828
829 if (remove_next) {
830again:
831 if (file) {
832 uprobe_munmap(next, next->vm_start, next->vm_end);
833 fput(file);
834 }
835 if (next->anon_vma)
836 anon_vma_merge(vma, next);
837 mm->map_count--;
838 mpol_put(vma_policy(next));
839 if (remove_next != 2)
840 BUG_ON(vma->vm_end < next->vm_end);
841 vm_area_free(next);
842
843 /*
844 * In mprotect's case 6 (see comments on vma_merge),
845 * we must remove next_next too.
846 */
847 if (remove_next == 2) {
848 remove_next = 1;
849 next = next_next;
850 goto again;
851 }
852 }
853 if (insert && file)
854 uprobe_mmap(insert);
855
856 mas_destroy(&mas);
857 validate_mm(mm);
858
859 return 0;
860}
861
862/*
863 * If the vma has a ->close operation then the driver probably needs to release
864 * per-vma resources, so we don't attempt to merge those.
865 */
866static inline int is_mergeable_vma(struct vm_area_struct *vma,
867 struct file *file, unsigned long vm_flags,
868 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
869 struct anon_vma_name *anon_name)
870{
871 /*
872 * VM_SOFTDIRTY should not prevent from VMA merging, if we
873 * match the flags but dirty bit -- the caller should mark
874 * merged VMA as dirty. If dirty bit won't be excluded from
875 * comparison, we increase pressure on the memory system forcing
876 * the kernel to generate new VMAs when old one could be
877 * extended instead.
878 */
879 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
880 return 0;
881 if (vma->vm_file != file)
882 return 0;
883 if (vma->vm_ops && vma->vm_ops->close)
884 return 0;
885 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
886 return 0;
887 if (!anon_vma_name_eq(anon_vma_name(vma), anon_name))
888 return 0;
889 return 1;
890}
891
892static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
893 struct anon_vma *anon_vma2,
894 struct vm_area_struct *vma)
895{
896 /*
897 * The list_is_singular() test is to avoid merging VMA cloned from
898 * parents. This can improve scalability caused by anon_vma lock.
899 */
900 if ((!anon_vma1 || !anon_vma2) && (!vma ||
901 list_is_singular(&vma->anon_vma_chain)))
902 return 1;
903 return anon_vma1 == anon_vma2;
904}
905
906/*
907 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
908 * in front of (at a lower virtual address and file offset than) the vma.
909 *
910 * We cannot merge two vmas if they have differently assigned (non-NULL)
911 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
912 *
913 * We don't check here for the merged mmap wrapping around the end of pagecache
914 * indices (16TB on ia32) because do_mmap() does not permit mmap's which
915 * wrap, nor mmaps which cover the final page at index -1UL.
916 */
917static int
918can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
919 struct anon_vma *anon_vma, struct file *file,
920 pgoff_t vm_pgoff,
921 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
922 struct anon_vma_name *anon_name)
923{
924 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name) &&
925 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
926 if (vma->vm_pgoff == vm_pgoff)
927 return 1;
928 }
929 return 0;
930}
931
932/*
933 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
934 * beyond (at a higher virtual address and file offset than) the vma.
935 *
936 * We cannot merge two vmas if they have differently assigned (non-NULL)
937 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
938 */
939static int
940can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
941 struct anon_vma *anon_vma, struct file *file,
942 pgoff_t vm_pgoff,
943 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
944 struct anon_vma_name *anon_name)
945{
946 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name) &&
947 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
948 pgoff_t vm_pglen;
949 vm_pglen = vma_pages(vma);
950 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
951 return 1;
952 }
953 return 0;
954}
955
956/*
957 * Given a mapping request (addr,end,vm_flags,file,pgoff,anon_name),
958 * figure out whether that can be merged with its predecessor or its
959 * successor. Or both (it neatly fills a hole).
960 *
961 * In most cases - when called for mmap, brk or mremap - [addr,end) is
962 * certain not to be mapped by the time vma_merge is called; but when
963 * called for mprotect, it is certain to be already mapped (either at
964 * an offset within prev, or at the start of next), and the flags of
965 * this area are about to be changed to vm_flags - and the no-change
966 * case has already been eliminated.
967 *
968 * The following mprotect cases have to be considered, where AAAA is
969 * the area passed down from mprotect_fixup, never extending beyond one
970 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
971 *
972 * AAAA AAAA AAAA
973 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN
974 * cannot merge might become might become
975 * PPNNNNNNNNNN PPPPPPPPPPNN
976 * mmap, brk or case 4 below case 5 below
977 * mremap move:
978 * AAAA AAAA
979 * PPPP NNNN PPPPNNNNXXXX
980 * might become might become
981 * PPPPPPPPPPPP 1 or PPPPPPPPPPPP 6 or
982 * PPPPPPPPNNNN 2 or PPPPPPPPXXXX 7 or
983 * PPPPNNNNNNNN 3 PPPPXXXXXXXX 8
984 *
985 * It is important for case 8 that the vma NNNN overlapping the
986 * region AAAA is never going to extended over XXXX. Instead XXXX must
987 * be extended in region AAAA and NNNN must be removed. This way in
988 * all cases where vma_merge succeeds, the moment vma_adjust drops the
989 * rmap_locks, the properties of the merged vma will be already
990 * correct for the whole merged range. Some of those properties like
991 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
992 * be correct for the whole merged range immediately after the
993 * rmap_locks are released. Otherwise if XXXX would be removed and
994 * NNNN would be extended over the XXXX range, remove_migration_ptes
995 * or other rmap walkers (if working on addresses beyond the "end"
996 * parameter) may establish ptes with the wrong permissions of NNNN
997 * instead of the right permissions of XXXX.
998 */
999struct vm_area_struct *vma_merge(struct mm_struct *mm,
1000 struct vm_area_struct *prev, unsigned long addr,
1001 unsigned long end, unsigned long vm_flags,
1002 struct anon_vma *anon_vma, struct file *file,
1003 pgoff_t pgoff, struct mempolicy *policy,
1004 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
1005 struct anon_vma_name *anon_name)
1006{
1007 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1008 struct vm_area_struct *mid, *next, *res;
1009 int err = -1;
1010 bool merge_prev = false;
1011 bool merge_next = false;
1012
1013 /*
1014 * We later require that vma->vm_flags == vm_flags,
1015 * so this tests vma->vm_flags & VM_SPECIAL, too.
1016 */
1017 if (vm_flags & VM_SPECIAL)
1018 return NULL;
1019
1020 next = find_vma(mm, prev ? prev->vm_end : 0);
1021 mid = next;
1022 if (next && next->vm_end == end) /* cases 6, 7, 8 */
1023 next = find_vma(mm, next->vm_end);
1024
1025 /* verify some invariant that must be enforced by the caller */
1026 VM_WARN_ON(prev && addr <= prev->vm_start);
1027 VM_WARN_ON(mid && end > mid->vm_end);
1028 VM_WARN_ON(addr >= end);
1029
1030 /* Can we merge the predecessor? */
1031 if (prev && prev->vm_end == addr &&
1032 mpol_equal(vma_policy(prev), policy) &&
1033 can_vma_merge_after(prev, vm_flags,
1034 anon_vma, file, pgoff,
1035 vm_userfaultfd_ctx, anon_name)) {
1036 merge_prev = true;
1037 }
1038 /* Can we merge the successor? */
1039 if (next && end == next->vm_start &&
1040 mpol_equal(policy, vma_policy(next)) &&
1041 can_vma_merge_before(next, vm_flags,
1042 anon_vma, file, pgoff+pglen,
1043 vm_userfaultfd_ctx, anon_name)) {
1044 merge_next = true;
1045 }
1046 /* Can we merge both the predecessor and the successor? */
1047 if (merge_prev && merge_next &&
1048 is_mergeable_anon_vma(prev->anon_vma,
1049 next->anon_vma, NULL)) { /* cases 1, 6 */
1050 err = __vma_adjust(prev, prev->vm_start,
1051 next->vm_end, prev->vm_pgoff, NULL,
1052 prev);
1053 res = prev;
1054 } else if (merge_prev) { /* cases 2, 5, 7 */
1055 err = __vma_adjust(prev, prev->vm_start,
1056 end, prev->vm_pgoff, NULL, prev);
1057 res = prev;
1058 } else if (merge_next) {
1059 if (prev && addr < prev->vm_end) /* case 4 */
1060 err = __vma_adjust(prev, prev->vm_start,
1061 addr, prev->vm_pgoff, NULL, next);
1062 else /* cases 3, 8 */
1063 err = __vma_adjust(mid, addr, next->vm_end,
1064 next->vm_pgoff - pglen, NULL, next);
1065 res = next;
1066 }
1067
1068 /*
1069 * Cannot merge with predecessor or successor or error in __vma_adjust?
1070 */
1071 if (err)
1072 return NULL;
1073 khugepaged_enter_vma(res, vm_flags);
1074 return res;
1075}
1076
1077/*
1078 * Rough compatibility check to quickly see if it's even worth looking
1079 * at sharing an anon_vma.
1080 *
1081 * They need to have the same vm_file, and the flags can only differ
1082 * in things that mprotect may change.
1083 *
1084 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1085 * we can merge the two vma's. For example, we refuse to merge a vma if
1086 * there is a vm_ops->close() function, because that indicates that the
1087 * driver is doing some kind of reference counting. But that doesn't
1088 * really matter for the anon_vma sharing case.
1089 */
1090static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1091{
1092 return a->vm_end == b->vm_start &&
1093 mpol_equal(vma_policy(a), vma_policy(b)) &&
1094 a->vm_file == b->vm_file &&
1095 !((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) &&
1096 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1097}
1098
1099/*
1100 * Do some basic sanity checking to see if we can re-use the anon_vma
1101 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1102 * the same as 'old', the other will be the new one that is trying
1103 * to share the anon_vma.
1104 *
1105 * NOTE! This runs with mmap_lock held for reading, so it is possible that
1106 * the anon_vma of 'old' is concurrently in the process of being set up
1107 * by another page fault trying to merge _that_. But that's ok: if it
1108 * is being set up, that automatically means that it will be a singleton
1109 * acceptable for merging, so we can do all of this optimistically. But
1110 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1111 *
1112 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1113 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1114 * is to return an anon_vma that is "complex" due to having gone through
1115 * a fork).
1116 *
1117 * We also make sure that the two vma's are compatible (adjacent,
1118 * and with the same memory policies). That's all stable, even with just
1119 * a read lock on the mmap_lock.
1120 */
1121static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1122{
1123 if (anon_vma_compatible(a, b)) {
1124 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1125
1126 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1127 return anon_vma;
1128 }
1129 return NULL;
1130}
1131
1132/*
1133 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1134 * neighbouring vmas for a suitable anon_vma, before it goes off
1135 * to allocate a new anon_vma. It checks because a repetitive
1136 * sequence of mprotects and faults may otherwise lead to distinct
1137 * anon_vmas being allocated, preventing vma merge in subsequent
1138 * mprotect.
1139 */
1140struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1141{
1142 MA_STATE(mas, &vma->vm_mm->mm_mt, vma->vm_end, vma->vm_end);
1143 struct anon_vma *anon_vma = NULL;
1144 struct vm_area_struct *prev, *next;
1145
1146 /* Try next first. */
1147 next = mas_walk(&mas);
1148 if (next) {
1149 anon_vma = reusable_anon_vma(next, vma, next);
1150 if (anon_vma)
1151 return anon_vma;
1152 }
1153
1154 prev = mas_prev(&mas, 0);
1155 VM_BUG_ON_VMA(prev != vma, vma);
1156 prev = mas_prev(&mas, 0);
1157 /* Try prev next. */
1158 if (prev)
1159 anon_vma = reusable_anon_vma(prev, prev, vma);
1160
1161 /*
1162 * We might reach here with anon_vma == NULL if we can't find
1163 * any reusable anon_vma.
1164 * There's no absolute need to look only at touching neighbours:
1165 * we could search further afield for "compatible" anon_vmas.
1166 * But it would probably just be a waste of time searching,
1167 * or lead to too many vmas hanging off the same anon_vma.
1168 * We're trying to allow mprotect remerging later on,
1169 * not trying to minimize memory used for anon_vmas.
1170 */
1171 return anon_vma;
1172}
1173
1174/*
1175 * If a hint addr is less than mmap_min_addr change hint to be as
1176 * low as possible but still greater than mmap_min_addr
1177 */
1178static inline unsigned long round_hint_to_min(unsigned long hint)
1179{
1180 hint &= PAGE_MASK;
1181 if (((void *)hint != NULL) &&
1182 (hint < mmap_min_addr))
1183 return PAGE_ALIGN(mmap_min_addr);
1184 return hint;
1185}
1186
1187int mlock_future_check(struct mm_struct *mm, unsigned long flags,
1188 unsigned long len)
1189{
1190 unsigned long locked, lock_limit;
1191
1192 /* mlock MCL_FUTURE? */
1193 if (flags & VM_LOCKED) {
1194 locked = len >> PAGE_SHIFT;
1195 locked += mm->locked_vm;
1196 lock_limit = rlimit(RLIMIT_MEMLOCK);
1197 lock_limit >>= PAGE_SHIFT;
1198 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1199 return -EAGAIN;
1200 }
1201 return 0;
1202}
1203
1204static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1205{
1206 if (S_ISREG(inode->i_mode))
1207 return MAX_LFS_FILESIZE;
1208
1209 if (S_ISBLK(inode->i_mode))
1210 return MAX_LFS_FILESIZE;
1211
1212 if (S_ISSOCK(inode->i_mode))
1213 return MAX_LFS_FILESIZE;
1214
1215 /* Special "we do even unsigned file positions" case */
1216 if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1217 return 0;
1218
1219 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1220 return ULONG_MAX;
1221}
1222
1223static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1224 unsigned long pgoff, unsigned long len)
1225{
1226 u64 maxsize = file_mmap_size_max(file, inode);
1227
1228 if (maxsize && len > maxsize)
1229 return false;
1230 maxsize -= len;
1231 if (pgoff > maxsize >> PAGE_SHIFT)
1232 return false;
1233 return true;
1234}
1235
1236/*
1237 * The caller must write-lock current->mm->mmap_lock.
1238 */
1239unsigned long do_mmap(struct file *file, unsigned long addr,
1240 unsigned long len, unsigned long prot,
1241 unsigned long flags, unsigned long pgoff,
1242 unsigned long *populate, struct list_head *uf)
1243{
1244 struct mm_struct *mm = current->mm;
1245 vm_flags_t vm_flags;
1246 int pkey = 0;
1247
1248 validate_mm(mm);
1249 *populate = 0;
1250
1251 if (!len)
1252 return -EINVAL;
1253
1254 /*
1255 * Does the application expect PROT_READ to imply PROT_EXEC?
1256 *
1257 * (the exception is when the underlying filesystem is noexec
1258 * mounted, in which case we dont add PROT_EXEC.)
1259 */
1260 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1261 if (!(file && path_noexec(&file->f_path)))
1262 prot |= PROT_EXEC;
1263
1264 /* force arch specific MAP_FIXED handling in get_unmapped_area */
1265 if (flags & MAP_FIXED_NOREPLACE)
1266 flags |= MAP_FIXED;
1267
1268 if (!(flags & MAP_FIXED))
1269 addr = round_hint_to_min(addr);
1270
1271 /* Careful about overflows.. */
1272 len = PAGE_ALIGN(len);
1273 if (!len)
1274 return -ENOMEM;
1275
1276 /* offset overflow? */
1277 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1278 return -EOVERFLOW;
1279
1280 /* Too many mappings? */
1281 if (mm->map_count > sysctl_max_map_count)
1282 return -ENOMEM;
1283
1284 /* Obtain the address to map to. we verify (or select) it and ensure
1285 * that it represents a valid section of the address space.
1286 */
1287 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1288 if (IS_ERR_VALUE(addr))
1289 return addr;
1290
1291 if (flags & MAP_FIXED_NOREPLACE) {
1292 if (find_vma_intersection(mm, addr, addr + len))
1293 return -EEXIST;
1294 }
1295
1296 if (prot == PROT_EXEC) {
1297 pkey = execute_only_pkey(mm);
1298 if (pkey < 0)
1299 pkey = 0;
1300 }
1301
1302 /* Do simple checking here so the lower-level routines won't have
1303 * to. we assume access permissions have been handled by the open
1304 * of the memory object, so we don't do any here.
1305 */
1306 vm_flags = calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1307 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1308
1309 if (flags & MAP_LOCKED)
1310 if (!can_do_mlock())
1311 return -EPERM;
1312
1313 if (mlock_future_check(mm, vm_flags, len))
1314 return -EAGAIN;
1315
1316 if (file) {
1317 struct inode *inode = file_inode(file);
1318 unsigned long flags_mask;
1319
1320 if (!file_mmap_ok(file, inode, pgoff, len))
1321 return -EOVERFLOW;
1322
1323 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1324
1325 switch (flags & MAP_TYPE) {
1326 case MAP_SHARED:
1327 /*
1328 * Force use of MAP_SHARED_VALIDATE with non-legacy
1329 * flags. E.g. MAP_SYNC is dangerous to use with
1330 * MAP_SHARED as you don't know which consistency model
1331 * you will get. We silently ignore unsupported flags
1332 * with MAP_SHARED to preserve backward compatibility.
1333 */
1334 flags &= LEGACY_MAP_MASK;
1335 fallthrough;
1336 case MAP_SHARED_VALIDATE:
1337 if (flags & ~flags_mask)
1338 return -EOPNOTSUPP;
1339 if (prot & PROT_WRITE) {
1340 if (!(file->f_mode & FMODE_WRITE))
1341 return -EACCES;
1342 if (IS_SWAPFILE(file->f_mapping->host))
1343 return -ETXTBSY;
1344 }
1345
1346 /*
1347 * Make sure we don't allow writing to an append-only
1348 * file..
1349 */
1350 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1351 return -EACCES;
1352
1353 vm_flags |= VM_SHARED | VM_MAYSHARE;
1354 if (!(file->f_mode & FMODE_WRITE))
1355 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1356 fallthrough;
1357 case MAP_PRIVATE:
1358 if (!(file->f_mode & FMODE_READ))
1359 return -EACCES;
1360 if (path_noexec(&file->f_path)) {
1361 if (vm_flags & VM_EXEC)
1362 return -EPERM;
1363 vm_flags &= ~VM_MAYEXEC;
1364 }
1365
1366 if (!file->f_op->mmap)
1367 return -ENODEV;
1368 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1369 return -EINVAL;
1370 break;
1371
1372 default:
1373 return -EINVAL;
1374 }
1375 } else {
1376 switch (flags & MAP_TYPE) {
1377 case MAP_SHARED:
1378 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1379 return -EINVAL;
1380 /*
1381 * Ignore pgoff.
1382 */
1383 pgoff = 0;
1384 vm_flags |= VM_SHARED | VM_MAYSHARE;
1385 break;
1386 case MAP_PRIVATE:
1387 /*
1388 * Set pgoff according to addr for anon_vma.
1389 */
1390 pgoff = addr >> PAGE_SHIFT;
1391 break;
1392 default:
1393 return -EINVAL;
1394 }
1395 }
1396
1397 /*
1398 * Set 'VM_NORESERVE' if we should not account for the
1399 * memory use of this mapping.
1400 */
1401 if (flags & MAP_NORESERVE) {
1402 /* We honor MAP_NORESERVE if allowed to overcommit */
1403 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1404 vm_flags |= VM_NORESERVE;
1405
1406 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1407 if (file && is_file_hugepages(file))
1408 vm_flags |= VM_NORESERVE;
1409 }
1410
1411 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1412 if (!IS_ERR_VALUE(addr) &&
1413 ((vm_flags & VM_LOCKED) ||
1414 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1415 *populate = len;
1416 return addr;
1417}
1418
1419unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1420 unsigned long prot, unsigned long flags,
1421 unsigned long fd, unsigned long pgoff)
1422{
1423 struct file *file = NULL;
1424 unsigned long retval;
1425
1426 if (!(flags & MAP_ANONYMOUS)) {
1427 audit_mmap_fd(fd, flags);
1428 file = fget(fd);
1429 if (!file)
1430 return -EBADF;
1431 if (is_file_hugepages(file)) {
1432 len = ALIGN(len, huge_page_size(hstate_file(file)));
1433 } else if (unlikely(flags & MAP_HUGETLB)) {
1434 retval = -EINVAL;
1435 goto out_fput;
1436 }
1437 } else if (flags & MAP_HUGETLB) {
1438 struct hstate *hs;
1439
1440 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1441 if (!hs)
1442 return -EINVAL;
1443
1444 len = ALIGN(len, huge_page_size(hs));
1445 /*
1446 * VM_NORESERVE is used because the reservations will be
1447 * taken when vm_ops->mmap() is called
1448 */
1449 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1450 VM_NORESERVE,
1451 HUGETLB_ANONHUGE_INODE,
1452 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1453 if (IS_ERR(file))
1454 return PTR_ERR(file);
1455 }
1456
1457 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1458out_fput:
1459 if (file)
1460 fput(file);
1461 return retval;
1462}
1463
1464SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1465 unsigned long, prot, unsigned long, flags,
1466 unsigned long, fd, unsigned long, pgoff)
1467{
1468 return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1469}
1470
1471#ifdef __ARCH_WANT_SYS_OLD_MMAP
1472struct mmap_arg_struct {
1473 unsigned long addr;
1474 unsigned long len;
1475 unsigned long prot;
1476 unsigned long flags;
1477 unsigned long fd;
1478 unsigned long offset;
1479};
1480
1481SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1482{
1483 struct mmap_arg_struct a;
1484
1485 if (copy_from_user(&a, arg, sizeof(a)))
1486 return -EFAULT;
1487 if (offset_in_page(a.offset))
1488 return -EINVAL;
1489
1490 return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1491 a.offset >> PAGE_SHIFT);
1492}
1493#endif /* __ARCH_WANT_SYS_OLD_MMAP */
1494
1495/*
1496 * Some shared mappings will want the pages marked read-only
1497 * to track write events. If so, we'll downgrade vm_page_prot
1498 * to the private version (using protection_map[] without the
1499 * VM_SHARED bit).
1500 */
1501int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1502{
1503 vm_flags_t vm_flags = vma->vm_flags;
1504 const struct vm_operations_struct *vm_ops = vma->vm_ops;
1505
1506 /* If it was private or non-writable, the write bit is already clear */
1507 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1508 return 0;
1509
1510 /* The backer wishes to know when pages are first written to? */
1511 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1512 return 1;
1513
1514 /* The open routine did something to the protections that pgprot_modify
1515 * won't preserve? */
1516 if (pgprot_val(vm_page_prot) !=
1517 pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1518 return 0;
1519
1520 /*
1521 * Do we need to track softdirty? hugetlb does not support softdirty
1522 * tracking yet.
1523 */
1524 if (vma_soft_dirty_enabled(vma) && !is_vm_hugetlb_page(vma))
1525 return 1;
1526
1527 /* Do we need write faults for uffd-wp tracking? */
1528 if (userfaultfd_wp(vma))
1529 return 1;
1530
1531 /* Specialty mapping? */
1532 if (vm_flags & VM_PFNMAP)
1533 return 0;
1534
1535 /* Can the mapping track the dirty pages? */
1536 return vma->vm_file && vma->vm_file->f_mapping &&
1537 mapping_can_writeback(vma->vm_file->f_mapping);
1538}
1539
1540/*
1541 * We account for memory if it's a private writeable mapping,
1542 * not hugepages and VM_NORESERVE wasn't set.
1543 */
1544static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1545{
1546 /*
1547 * hugetlb has its own accounting separate from the core VM
1548 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1549 */
1550 if (file && is_file_hugepages(file))
1551 return 0;
1552
1553 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1554}
1555
1556/**
1557 * unmapped_area() - Find an area between the low_limit and the high_limit with
1558 * the correct alignment and offset, all from @info. Note: current->mm is used
1559 * for the search.
1560 *
1561 * @info: The unmapped area information including the range (low_limit -
1562 * hight_limit), the alignment offset and mask.
1563 *
1564 * Return: A memory address or -ENOMEM.
1565 */
1566static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1567{
1568 unsigned long length, gap;
1569
1570 MA_STATE(mas, ¤t->mm->mm_mt, 0, 0);
1571
1572 /* Adjust search length to account for worst case alignment overhead */
1573 length = info->length + info->align_mask;
1574 if (length < info->length)
1575 return -ENOMEM;
1576
1577 if (mas_empty_area(&mas, info->low_limit, info->high_limit - 1,
1578 length))
1579 return -ENOMEM;
1580
1581 gap = mas.index;
1582 gap += (info->align_offset - gap) & info->align_mask;
1583 return gap;
1584}
1585
1586/**
1587 * unmapped_area_topdown() - Find an area between the low_limit and the
1588 * high_limit with * the correct alignment and offset at the highest available
1589 * address, all from @info. Note: current->mm is used for the search.
1590 *
1591 * @info: The unmapped area information including the range (low_limit -
1592 * hight_limit), the alignment offset and mask.
1593 *
1594 * Return: A memory address or -ENOMEM.
1595 */
1596static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1597{
1598 unsigned long length, gap;
1599
1600 MA_STATE(mas, ¤t->mm->mm_mt, 0, 0);
1601 /* Adjust search length to account for worst case alignment overhead */
1602 length = info->length + info->align_mask;
1603 if (length < info->length)
1604 return -ENOMEM;
1605
1606 if (mas_empty_area_rev(&mas, info->low_limit, info->high_limit - 1,
1607 length))
1608 return -ENOMEM;
1609
1610 gap = mas.last + 1 - info->length;
1611 gap -= (gap - info->align_offset) & info->align_mask;
1612 return gap;
1613}
1614
1615/*
1616 * Search for an unmapped address range.
1617 *
1618 * We are looking for a range that:
1619 * - does not intersect with any VMA;
1620 * - is contained within the [low_limit, high_limit) interval;
1621 * - is at least the desired size.
1622 * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
1623 */
1624unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
1625{
1626 unsigned long addr;
1627
1628 if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
1629 addr = unmapped_area_topdown(info);
1630 else
1631 addr = unmapped_area(info);
1632
1633 trace_vm_unmapped_area(addr, info);
1634 return addr;
1635}
1636
1637/* Get an address range which is currently unmapped.
1638 * For shmat() with addr=0.
1639 *
1640 * Ugly calling convention alert:
1641 * Return value with the low bits set means error value,
1642 * ie
1643 * if (ret & ~PAGE_MASK)
1644 * error = ret;
1645 *
1646 * This function "knows" that -ENOMEM has the bits set.
1647 */
1648unsigned long
1649generic_get_unmapped_area(struct file *filp, unsigned long addr,
1650 unsigned long len, unsigned long pgoff,
1651 unsigned long flags)
1652{
1653 struct mm_struct *mm = current->mm;
1654 struct vm_area_struct *vma, *prev;
1655 struct vm_unmapped_area_info info;
1656 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
1657
1658 if (len > mmap_end - mmap_min_addr)
1659 return -ENOMEM;
1660
1661 if (flags & MAP_FIXED)
1662 return addr;
1663
1664 if (addr) {
1665 addr = PAGE_ALIGN(addr);
1666 vma = find_vma_prev(mm, addr, &prev);
1667 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
1668 (!vma || addr + len <= vm_start_gap(vma)) &&
1669 (!prev || addr >= vm_end_gap(prev)))
1670 return addr;
1671 }
1672
1673 info.flags = 0;
1674 info.length = len;
1675 info.low_limit = mm->mmap_base;
1676 info.high_limit = mmap_end;
1677 info.align_mask = 0;
1678 info.align_offset = 0;
1679 return vm_unmapped_area(&info);
1680}
1681
1682#ifndef HAVE_ARCH_UNMAPPED_AREA
1683unsigned long
1684arch_get_unmapped_area(struct file *filp, unsigned long addr,
1685 unsigned long len, unsigned long pgoff,
1686 unsigned long flags)
1687{
1688 return generic_get_unmapped_area(filp, addr, len, pgoff, flags);
1689}
1690#endif
1691
1692/*
1693 * This mmap-allocator allocates new areas top-down from below the
1694 * stack's low limit (the base):
1695 */
1696unsigned long
1697generic_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
1698 unsigned long len, unsigned long pgoff,
1699 unsigned long flags)
1700{
1701 struct vm_area_struct *vma, *prev;
1702 struct mm_struct *mm = current->mm;
1703 struct vm_unmapped_area_info info;
1704 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
1705
1706 /* requested length too big for entire address space */
1707 if (len > mmap_end - mmap_min_addr)
1708 return -ENOMEM;
1709
1710 if (flags & MAP_FIXED)
1711 return addr;
1712
1713 /* requesting a specific address */
1714 if (addr) {
1715 addr = PAGE_ALIGN(addr);
1716 vma = find_vma_prev(mm, addr, &prev);
1717 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
1718 (!vma || addr + len <= vm_start_gap(vma)) &&
1719 (!prev || addr >= vm_end_gap(prev)))
1720 return addr;
1721 }
1722
1723 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
1724 info.length = len;
1725 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
1726 info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
1727 info.align_mask = 0;
1728 info.align_offset = 0;
1729 addr = vm_unmapped_area(&info);
1730
1731 /*
1732 * A failed mmap() very likely causes application failure,
1733 * so fall back to the bottom-up function here. This scenario
1734 * can happen with large stack limits and large mmap()
1735 * allocations.
1736 */
1737 if (offset_in_page(addr)) {
1738 VM_BUG_ON(addr != -ENOMEM);
1739 info.flags = 0;
1740 info.low_limit = TASK_UNMAPPED_BASE;
1741 info.high_limit = mmap_end;
1742 addr = vm_unmapped_area(&info);
1743 }
1744
1745 return addr;
1746}
1747
1748#ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1749unsigned long
1750arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
1751 unsigned long len, unsigned long pgoff,
1752 unsigned long flags)
1753{
1754 return generic_get_unmapped_area_topdown(filp, addr, len, pgoff, flags);
1755}
1756#endif
1757
1758unsigned long
1759get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1760 unsigned long pgoff, unsigned long flags)
1761{
1762 unsigned long (*get_area)(struct file *, unsigned long,
1763 unsigned long, unsigned long, unsigned long);
1764
1765 unsigned long error = arch_mmap_check(addr, len, flags);
1766 if (error)
1767 return error;
1768
1769 /* Careful about overflows.. */
1770 if (len > TASK_SIZE)
1771 return -ENOMEM;
1772
1773 get_area = current->mm->get_unmapped_area;
1774 if (file) {
1775 if (file->f_op->get_unmapped_area)
1776 get_area = file->f_op->get_unmapped_area;
1777 } else if (flags & MAP_SHARED) {
1778 /*
1779 * mmap_region() will call shmem_zero_setup() to create a file,
1780 * so use shmem's get_unmapped_area in case it can be huge.
1781 * do_mmap() will clear pgoff, so match alignment.
1782 */
1783 pgoff = 0;
1784 get_area = shmem_get_unmapped_area;
1785 }
1786
1787 addr = get_area(file, addr, len, pgoff, flags);
1788 if (IS_ERR_VALUE(addr))
1789 return addr;
1790
1791 if (addr > TASK_SIZE - len)
1792 return -ENOMEM;
1793 if (offset_in_page(addr))
1794 return -EINVAL;
1795
1796 error = security_mmap_addr(addr);
1797 return error ? error : addr;
1798}
1799
1800EXPORT_SYMBOL(get_unmapped_area);
1801
1802/**
1803 * find_vma_intersection() - Look up the first VMA which intersects the interval
1804 * @mm: The process address space.
1805 * @start_addr: The inclusive start user address.
1806 * @end_addr: The exclusive end user address.
1807 *
1808 * Returns: The first VMA within the provided range, %NULL otherwise. Assumes
1809 * start_addr < end_addr.
1810 */
1811struct vm_area_struct *find_vma_intersection(struct mm_struct *mm,
1812 unsigned long start_addr,
1813 unsigned long end_addr)
1814{
1815 unsigned long index = start_addr;
1816
1817 mmap_assert_locked(mm);
1818 return mt_find(&mm->mm_mt, &index, end_addr - 1);
1819}
1820EXPORT_SYMBOL(find_vma_intersection);
1821
1822/**
1823 * find_vma() - Find the VMA for a given address, or the next VMA.
1824 * @mm: The mm_struct to check
1825 * @addr: The address
1826 *
1827 * Returns: The VMA associated with addr, or the next VMA.
1828 * May return %NULL in the case of no VMA at addr or above.
1829 */
1830struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1831{
1832 unsigned long index = addr;
1833
1834 mmap_assert_locked(mm);
1835 return mt_find(&mm->mm_mt, &index, ULONG_MAX);
1836}
1837EXPORT_SYMBOL(find_vma);
1838
1839/**
1840 * find_vma_prev() - Find the VMA for a given address, or the next vma and
1841 * set %pprev to the previous VMA, if any.
1842 * @mm: The mm_struct to check
1843 * @addr: The address
1844 * @pprev: The pointer to set to the previous VMA
1845 *
1846 * Note that RCU lock is missing here since the external mmap_lock() is used
1847 * instead.
1848 *
1849 * Returns: The VMA associated with @addr, or the next vma.
1850 * May return %NULL in the case of no vma at addr or above.
1851 */
1852struct vm_area_struct *
1853find_vma_prev(struct mm_struct *mm, unsigned long addr,
1854 struct vm_area_struct **pprev)
1855{
1856 struct vm_area_struct *vma;
1857 MA_STATE(mas, &mm->mm_mt, addr, addr);
1858
1859 vma = mas_walk(&mas);
1860 *pprev = mas_prev(&mas, 0);
1861 if (!vma)
1862 vma = mas_next(&mas, ULONG_MAX);
1863 return vma;
1864}
1865
1866/*
1867 * Verify that the stack growth is acceptable and
1868 * update accounting. This is shared with both the
1869 * grow-up and grow-down cases.
1870 */
1871static int acct_stack_growth(struct vm_area_struct *vma,
1872 unsigned long size, unsigned long grow)
1873{
1874 struct mm_struct *mm = vma->vm_mm;
1875 unsigned long new_start;
1876
1877 /* address space limit tests */
1878 if (!may_expand_vm(mm, vma->vm_flags, grow))
1879 return -ENOMEM;
1880
1881 /* Stack limit test */
1882 if (size > rlimit(RLIMIT_STACK))
1883 return -ENOMEM;
1884
1885 /* mlock limit tests */
1886 if (mlock_future_check(mm, vma->vm_flags, grow << PAGE_SHIFT))
1887 return -ENOMEM;
1888
1889 /* Check to ensure the stack will not grow into a hugetlb-only region */
1890 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1891 vma->vm_end - size;
1892 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1893 return -EFAULT;
1894
1895 /*
1896 * Overcommit.. This must be the final test, as it will
1897 * update security statistics.
1898 */
1899 if (security_vm_enough_memory_mm(mm, grow))
1900 return -ENOMEM;
1901
1902 return 0;
1903}
1904
1905#if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1906/*
1907 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1908 * vma is the last one with address > vma->vm_end. Have to extend vma.
1909 */
1910int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1911{
1912 struct mm_struct *mm = vma->vm_mm;
1913 struct vm_area_struct *next;
1914 unsigned long gap_addr;
1915 int error = 0;
1916 MA_STATE(mas, &mm->mm_mt, 0, 0);
1917
1918 if (!(vma->vm_flags & VM_GROWSUP))
1919 return -EFAULT;
1920
1921 /* Guard against exceeding limits of the address space. */
1922 address &= PAGE_MASK;
1923 if (address >= (TASK_SIZE & PAGE_MASK))
1924 return -ENOMEM;
1925 address += PAGE_SIZE;
1926
1927 /* Enforce stack_guard_gap */
1928 gap_addr = address + stack_guard_gap;
1929
1930 /* Guard against overflow */
1931 if (gap_addr < address || gap_addr > TASK_SIZE)
1932 gap_addr = TASK_SIZE;
1933
1934 next = find_vma_intersection(mm, vma->vm_end, gap_addr);
1935 if (next && vma_is_accessible(next)) {
1936 if (!(next->vm_flags & VM_GROWSUP))
1937 return -ENOMEM;
1938 /* Check that both stack segments have the same anon_vma? */
1939 }
1940
1941 if (mas_preallocate(&mas, vma, GFP_KERNEL))
1942 return -ENOMEM;
1943
1944 /* We must make sure the anon_vma is allocated. */
1945 if (unlikely(anon_vma_prepare(vma))) {
1946 mas_destroy(&mas);
1947 return -ENOMEM;
1948 }
1949
1950 /*
1951 * vma->vm_start/vm_end cannot change under us because the caller
1952 * is required to hold the mmap_lock in read mode. We need the
1953 * anon_vma lock to serialize against concurrent expand_stacks.
1954 */
1955 anon_vma_lock_write(vma->anon_vma);
1956
1957 /* Somebody else might have raced and expanded it already */
1958 if (address > vma->vm_end) {
1959 unsigned long size, grow;
1960
1961 size = address - vma->vm_start;
1962 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1963
1964 error = -ENOMEM;
1965 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
1966 error = acct_stack_growth(vma, size, grow);
1967 if (!error) {
1968 /*
1969 * We only hold a shared mmap_lock lock here, so
1970 * we need to protect against concurrent vma
1971 * expansions. anon_vma_lock_write() doesn't
1972 * help here, as we don't guarantee that all
1973 * growable vmas in a mm share the same root
1974 * anon vma. So, we reuse mm->page_table_lock
1975 * to guard against concurrent vma expansions.
1976 */
1977 spin_lock(&mm->page_table_lock);
1978 if (vma->vm_flags & VM_LOCKED)
1979 mm->locked_vm += grow;
1980 vm_stat_account(mm, vma->vm_flags, grow);
1981 anon_vma_interval_tree_pre_update_vma(vma);
1982 vma->vm_end = address;
1983 /* Overwrite old entry in mtree. */
1984 vma_mas_store(vma, &mas);
1985 anon_vma_interval_tree_post_update_vma(vma);
1986 spin_unlock(&mm->page_table_lock);
1987
1988 perf_event_mmap(vma);
1989 }
1990 }
1991 }
1992 anon_vma_unlock_write(vma->anon_vma);
1993 khugepaged_enter_vma(vma, vma->vm_flags);
1994 mas_destroy(&mas);
1995 return error;
1996}
1997#endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1998
1999/*
2000 * vma is the first one with address < vma->vm_start. Have to extend vma.
2001 */
2002int expand_downwards(struct vm_area_struct *vma, unsigned long address)
2003{
2004 struct mm_struct *mm = vma->vm_mm;
2005 MA_STATE(mas, &mm->mm_mt, vma->vm_start, vma->vm_start);
2006 struct vm_area_struct *prev;
2007 int error = 0;
2008
2009 address &= PAGE_MASK;
2010 if (address < mmap_min_addr)
2011 return -EPERM;
2012
2013 /* Enforce stack_guard_gap */
2014 prev = mas_prev(&mas, 0);
2015 /* Check that both stack segments have the same anon_vma? */
2016 if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2017 vma_is_accessible(prev)) {
2018 if (address - prev->vm_end < stack_guard_gap)
2019 return -ENOMEM;
2020 }
2021
2022 if (mas_preallocate(&mas, vma, GFP_KERNEL))
2023 return -ENOMEM;
2024
2025 /* We must make sure the anon_vma is allocated. */
2026 if (unlikely(anon_vma_prepare(vma))) {
2027 mas_destroy(&mas);
2028 return -ENOMEM;
2029 }
2030
2031 /*
2032 * vma->vm_start/vm_end cannot change under us because the caller
2033 * is required to hold the mmap_lock in read mode. We need the
2034 * anon_vma lock to serialize against concurrent expand_stacks.
2035 */
2036 anon_vma_lock_write(vma->anon_vma);
2037
2038 /* Somebody else might have raced and expanded it already */
2039 if (address < vma->vm_start) {
2040 unsigned long size, grow;
2041
2042 size = vma->vm_end - address;
2043 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2044
2045 error = -ENOMEM;
2046 if (grow <= vma->vm_pgoff) {
2047 error = acct_stack_growth(vma, size, grow);
2048 if (!error) {
2049 /*
2050 * We only hold a shared mmap_lock lock here, so
2051 * we need to protect against concurrent vma
2052 * expansions. anon_vma_lock_write() doesn't
2053 * help here, as we don't guarantee that all
2054 * growable vmas in a mm share the same root
2055 * anon vma. So, we reuse mm->page_table_lock
2056 * to guard against concurrent vma expansions.
2057 */
2058 spin_lock(&mm->page_table_lock);
2059 if (vma->vm_flags & VM_LOCKED)
2060 mm->locked_vm += grow;
2061 vm_stat_account(mm, vma->vm_flags, grow);
2062 anon_vma_interval_tree_pre_update_vma(vma);
2063 vma->vm_start = address;
2064 vma->vm_pgoff -= grow;
2065 /* Overwrite old entry in mtree. */
2066 vma_mas_store(vma, &mas);
2067 anon_vma_interval_tree_post_update_vma(vma);
2068 spin_unlock(&mm->page_table_lock);
2069
2070 perf_event_mmap(vma);
2071 }
2072 }
2073 }
2074 anon_vma_unlock_write(vma->anon_vma);
2075 khugepaged_enter_vma(vma, vma->vm_flags);
2076 mas_destroy(&mas);
2077 return error;
2078}
2079
2080/* enforced gap between the expanding stack and other mappings. */
2081unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2082
2083static int __init cmdline_parse_stack_guard_gap(char *p)
2084{
2085 unsigned long val;
2086 char *endptr;
2087
2088 val = simple_strtoul(p, &endptr, 10);
2089 if (!*endptr)
2090 stack_guard_gap = val << PAGE_SHIFT;
2091
2092 return 1;
2093}
2094__setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2095
2096#ifdef CONFIG_STACK_GROWSUP
2097int expand_stack(struct vm_area_struct *vma, unsigned long address)
2098{
2099 return expand_upwards(vma, address);
2100}
2101
2102struct vm_area_struct *
2103find_extend_vma(struct mm_struct *mm, unsigned long addr)
2104{
2105 struct vm_area_struct *vma, *prev;
2106
2107 addr &= PAGE_MASK;
2108 vma = find_vma_prev(mm, addr, &prev);
2109 if (vma && (vma->vm_start <= addr))
2110 return vma;
2111 if (!prev || expand_stack(prev, addr))
2112 return NULL;
2113 if (prev->vm_flags & VM_LOCKED)
2114 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2115 return prev;
2116}
2117#else
2118int expand_stack(struct vm_area_struct *vma, unsigned long address)
2119{
2120 return expand_downwards(vma, address);
2121}
2122
2123struct vm_area_struct *
2124find_extend_vma(struct mm_struct *mm, unsigned long addr)
2125{
2126 struct vm_area_struct *vma;
2127 unsigned long start;
2128
2129 addr &= PAGE_MASK;
2130 vma = find_vma(mm, addr);
2131 if (!vma)
2132 return NULL;
2133 if (vma->vm_start <= addr)
2134 return vma;
2135 if (!(vma->vm_flags & VM_GROWSDOWN))
2136 return NULL;
2137 start = vma->vm_start;
2138 if (expand_stack(vma, addr))
2139 return NULL;
2140 if (vma->vm_flags & VM_LOCKED)
2141 populate_vma_page_range(vma, addr, start, NULL);
2142 return vma;
2143}
2144#endif
2145
2146EXPORT_SYMBOL_GPL(find_extend_vma);
2147
2148/*
2149 * Ok - we have the memory areas we should free on a maple tree so release them,
2150 * and do the vma updates.
2151 *
2152 * Called with the mm semaphore held.
2153 */
2154static inline void remove_mt(struct mm_struct *mm, struct ma_state *mas)
2155{
2156 unsigned long nr_accounted = 0;
2157 struct vm_area_struct *vma;
2158
2159 /* Update high watermark before we lower total_vm */
2160 update_hiwater_vm(mm);
2161 mas_for_each(mas, vma, ULONG_MAX) {
2162 long nrpages = vma_pages(vma);
2163
2164 if (vma->vm_flags & VM_ACCOUNT)
2165 nr_accounted += nrpages;
2166 vm_stat_account(mm, vma->vm_flags, -nrpages);
2167 remove_vma(vma);
2168 }
2169 vm_unacct_memory(nr_accounted);
2170 validate_mm(mm);
2171}
2172
2173/*
2174 * Get rid of page table information in the indicated region.
2175 *
2176 * Called with the mm semaphore held.
2177 */
2178static void unmap_region(struct mm_struct *mm, struct maple_tree *mt,
2179 struct vm_area_struct *vma, struct vm_area_struct *prev,
2180 struct vm_area_struct *next,
2181 unsigned long start, unsigned long end)
2182{
2183 struct mmu_gather tlb;
2184
2185 lru_add_drain();
2186 tlb_gather_mmu(&tlb, mm);
2187 update_hiwater_rss(mm);
2188 unmap_vmas(&tlb, mt, vma, start, end);
2189 free_pgtables(&tlb, mt, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2190 next ? next->vm_start : USER_PGTABLES_CEILING);
2191 tlb_finish_mmu(&tlb);
2192}
2193
2194/*
2195 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it
2196 * has already been checked or doesn't make sense to fail.
2197 */
2198int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2199 unsigned long addr, int new_below)
2200{
2201 struct vm_area_struct *new;
2202 int err;
2203 validate_mm_mt(mm);
2204
2205 if (vma->vm_ops && vma->vm_ops->may_split) {
2206 err = vma->vm_ops->may_split(vma, addr);
2207 if (err)
2208 return err;
2209 }
2210
2211 new = vm_area_dup(vma);
2212 if (!new)
2213 return -ENOMEM;
2214
2215 if (new_below)
2216 new->vm_end = addr;
2217 else {
2218 new->vm_start = addr;
2219 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2220 }
2221
2222 err = vma_dup_policy(vma, new);
2223 if (err)
2224 goto out_free_vma;
2225
2226 err = anon_vma_clone(new, vma);
2227 if (err)
2228 goto out_free_mpol;
2229
2230 if (new->vm_file)
2231 get_file(new->vm_file);
2232
2233 if (new->vm_ops && new->vm_ops->open)
2234 new->vm_ops->open(new);
2235
2236 if (new_below)
2237 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2238 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2239 else
2240 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2241
2242 /* Success. */
2243 if (!err)
2244 return 0;
2245
2246 /* Avoid vm accounting in close() operation */
2247 new->vm_start = new->vm_end;
2248 new->vm_pgoff = 0;
2249 /* Clean everything up if vma_adjust failed. */
2250 if (new->vm_ops && new->vm_ops->close)
2251 new->vm_ops->close(new);
2252 if (new->vm_file)
2253 fput(new->vm_file);
2254 unlink_anon_vmas(new);
2255 out_free_mpol:
2256 mpol_put(vma_policy(new));
2257 out_free_vma:
2258 vm_area_free(new);
2259 validate_mm_mt(mm);
2260 return err;
2261}
2262
2263/*
2264 * Split a vma into two pieces at address 'addr', a new vma is allocated
2265 * either for the first part or the tail.
2266 */
2267int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2268 unsigned long addr, int new_below)
2269{
2270 if (mm->map_count >= sysctl_max_map_count)
2271 return -ENOMEM;
2272
2273 return __split_vma(mm, vma, addr, new_below);
2274}
2275
2276static inline int munmap_sidetree(struct vm_area_struct *vma,
2277 struct ma_state *mas_detach)
2278{
2279 mas_set_range(mas_detach, vma->vm_start, vma->vm_end - 1);
2280 if (mas_store_gfp(mas_detach, vma, GFP_KERNEL))
2281 return -ENOMEM;
2282
2283 if (vma->vm_flags & VM_LOCKED)
2284 vma->vm_mm->locked_vm -= vma_pages(vma);
2285
2286 return 0;
2287}
2288
2289/*
2290 * do_mas_align_munmap() - munmap the aligned region from @start to @end.
2291 * @mas: The maple_state, ideally set up to alter the correct tree location.
2292 * @vma: The starting vm_area_struct
2293 * @mm: The mm_struct
2294 * @start: The aligned start address to munmap.
2295 * @end: The aligned end address to munmap.
2296 * @uf: The userfaultfd list_head
2297 * @downgrade: Set to true to attempt a write downgrade of the mmap_lock
2298 *
2299 * If @downgrade is true, check return code for potential release of the lock.
2300 */
2301static int
2302do_mas_align_munmap(struct ma_state *mas, struct vm_area_struct *vma,
2303 struct mm_struct *mm, unsigned long start,
2304 unsigned long end, struct list_head *uf, bool downgrade)
2305{
2306 struct vm_area_struct *prev, *next = NULL;
2307 struct maple_tree mt_detach;
2308 int count = 0;
2309 int error = -ENOMEM;
2310 MA_STATE(mas_detach, &mt_detach, 0, 0);
2311 mt_init_flags(&mt_detach, MT_FLAGS_LOCK_EXTERN);
2312 mt_set_external_lock(&mt_detach, &mm->mmap_lock);
2313
2314 if (mas_preallocate(mas, vma, GFP_KERNEL))
2315 return -ENOMEM;
2316
2317 mas->last = end - 1;
2318 /*
2319 * If we need to split any vma, do it now to save pain later.
2320 *
2321 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2322 * unmapped vm_area_struct will remain in use: so lower split_vma
2323 * places tmp vma above, and higher split_vma places tmp vma below.
2324 */
2325
2326 /* Does it split the first one? */
2327 if (start > vma->vm_start) {
2328
2329 /*
2330 * Make sure that map_count on return from munmap() will
2331 * not exceed its limit; but let map_count go just above
2332 * its limit temporarily, to help free resources as expected.
2333 */
2334 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2335 goto map_count_exceeded;
2336
2337 /*
2338 * mas_pause() is not needed since mas->index needs to be set
2339 * differently than vma->vm_end anyways.
2340 */
2341 error = __split_vma(mm, vma, start, 0);
2342 if (error)
2343 goto start_split_failed;
2344
2345 mas_set(mas, start);
2346 vma = mas_walk(mas);
2347 }
2348
2349 prev = mas_prev(mas, 0);
2350 if (unlikely((!prev)))
2351 mas_set(mas, start);
2352
2353 /*
2354 * Detach a range of VMAs from the mm. Using next as a temp variable as
2355 * it is always overwritten.
2356 */
2357 mas_for_each(mas, next, end - 1) {
2358 /* Does it split the end? */
2359 if (next->vm_end > end) {
2360 struct vm_area_struct *split;
2361
2362 error = __split_vma(mm, next, end, 1);
2363 if (error)
2364 goto end_split_failed;
2365
2366 mas_set(mas, end);
2367 split = mas_prev(mas, 0);
2368 error = munmap_sidetree(split, &mas_detach);
2369 if (error)
2370 goto munmap_sidetree_failed;
2371
2372 count++;
2373 if (vma == next)
2374 vma = split;
2375 break;
2376 }
2377 error = munmap_sidetree(next, &mas_detach);
2378 if (error)
2379 goto munmap_sidetree_failed;
2380
2381 count++;
2382#ifdef CONFIG_DEBUG_VM_MAPLE_TREE
2383 BUG_ON(next->vm_start < start);
2384 BUG_ON(next->vm_start > end);
2385#endif
2386 }
2387
2388 if (!next)
2389 next = mas_next(mas, ULONG_MAX);
2390
2391 if (unlikely(uf)) {
2392 /*
2393 * If userfaultfd_unmap_prep returns an error the vmas
2394 * will remain split, but userland will get a
2395 * highly unexpected error anyway. This is no
2396 * different than the case where the first of the two
2397 * __split_vma fails, but we don't undo the first
2398 * split, despite we could. This is unlikely enough
2399 * failure that it's not worth optimizing it for.
2400 */
2401 error = userfaultfd_unmap_prep(mm, start, end, uf);
2402
2403 if (error)
2404 goto userfaultfd_error;
2405 }
2406
2407 /* Point of no return */
2408 mas_set_range(mas, start, end - 1);
2409#if defined(CONFIG_DEBUG_VM_MAPLE_TREE)
2410 /* Make sure no VMAs are about to be lost. */
2411 {
2412 MA_STATE(test, &mt_detach, start, end - 1);
2413 struct vm_area_struct *vma_mas, *vma_test;
2414 int test_count = 0;
2415
2416 rcu_read_lock();
2417 vma_test = mas_find(&test, end - 1);
2418 mas_for_each(mas, vma_mas, end - 1) {
2419 BUG_ON(vma_mas != vma_test);
2420 test_count++;
2421 vma_test = mas_next(&test, end - 1);
2422 }
2423 rcu_read_unlock();
2424 BUG_ON(count != test_count);
2425 mas_set_range(mas, start, end - 1);
2426 }
2427#endif
2428 mas_store_prealloc(mas, NULL);
2429 mm->map_count -= count;
2430 /*
2431 * Do not downgrade mmap_lock if we are next to VM_GROWSDOWN or
2432 * VM_GROWSUP VMA. Such VMAs can change their size under
2433 * down_read(mmap_lock) and collide with the VMA we are about to unmap.
2434 */
2435 if (downgrade) {
2436 if (next && (next->vm_flags & VM_GROWSDOWN))
2437 downgrade = false;
2438 else if (prev && (prev->vm_flags & VM_GROWSUP))
2439 downgrade = false;
2440 else
2441 mmap_write_downgrade(mm);
2442 }
2443
2444 unmap_region(mm, &mt_detach, vma, prev, next, start, end);
2445 /* Statistics and freeing VMAs */
2446 mas_set(&mas_detach, start);
2447 remove_mt(mm, &mas_detach);
2448 __mt_destroy(&mt_detach);
2449
2450
2451 validate_mm(mm);
2452 return downgrade ? 1 : 0;
2453
2454userfaultfd_error:
2455munmap_sidetree_failed:
2456end_split_failed:
2457 __mt_destroy(&mt_detach);
2458start_split_failed:
2459map_count_exceeded:
2460 mas_destroy(mas);
2461 return error;
2462}
2463
2464/*
2465 * do_mas_munmap() - munmap a given range.
2466 * @mas: The maple state
2467 * @mm: The mm_struct
2468 * @start: The start address to munmap
2469 * @len: The length of the range to munmap
2470 * @uf: The userfaultfd list_head
2471 * @downgrade: set to true if the user wants to attempt to write_downgrade the
2472 * mmap_lock
2473 *
2474 * This function takes a @mas that is either pointing to the previous VMA or set
2475 * to MA_START and sets it up to remove the mapping(s). The @len will be
2476 * aligned and any arch_unmap work will be preformed.
2477 *
2478 * Returns: -EINVAL on failure, 1 on success and unlock, 0 otherwise.
2479 */
2480int do_mas_munmap(struct ma_state *mas, struct mm_struct *mm,
2481 unsigned long start, size_t len, struct list_head *uf,
2482 bool downgrade)
2483{
2484 unsigned long end;
2485 struct vm_area_struct *vma;
2486
2487 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2488 return -EINVAL;
2489
2490 end = start + PAGE_ALIGN(len);
2491 if (end == start)
2492 return -EINVAL;
2493
2494 /* arch_unmap() might do unmaps itself. */
2495 arch_unmap(mm, start, end);
2496
2497 /* Find the first overlapping VMA */
2498 vma = mas_find(mas, end - 1);
2499 if (!vma)
2500 return 0;
2501
2502 return do_mas_align_munmap(mas, vma, mm, start, end, uf, downgrade);
2503}
2504
2505/* do_munmap() - Wrapper function for non-maple tree aware do_munmap() calls.
2506 * @mm: The mm_struct
2507 * @start: The start address to munmap
2508 * @len: The length to be munmapped.
2509 * @uf: The userfaultfd list_head
2510 */
2511int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2512 struct list_head *uf)
2513{
2514 MA_STATE(mas, &mm->mm_mt, start, start);
2515
2516 return do_mas_munmap(&mas, mm, start, len, uf, false);
2517}
2518
2519unsigned long mmap_region(struct file *file, unsigned long addr,
2520 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
2521 struct list_head *uf)
2522{
2523 struct mm_struct *mm = current->mm;
2524 struct vm_area_struct *vma = NULL;
2525 struct vm_area_struct *next, *prev, *merge;
2526 pgoff_t pglen = len >> PAGE_SHIFT;
2527 unsigned long charged = 0;
2528 unsigned long end = addr + len;
2529 unsigned long merge_start = addr, merge_end = end;
2530 pgoff_t vm_pgoff;
2531 int error;
2532 MA_STATE(mas, &mm->mm_mt, addr, end - 1);
2533
2534 /* Check against address space limit. */
2535 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
2536 unsigned long nr_pages;
2537
2538 /*
2539 * MAP_FIXED may remove pages of mappings that intersects with
2540 * requested mapping. Account for the pages it would unmap.
2541 */
2542 nr_pages = count_vma_pages_range(mm, addr, end);
2543
2544 if (!may_expand_vm(mm, vm_flags,
2545 (len >> PAGE_SHIFT) - nr_pages))
2546 return -ENOMEM;
2547 }
2548
2549 /* Unmap any existing mapping in the area */
2550 if (do_mas_munmap(&mas, mm, addr, len, uf, false))
2551 return -ENOMEM;
2552
2553 /*
2554 * Private writable mapping: check memory availability
2555 */
2556 if (accountable_mapping(file, vm_flags)) {
2557 charged = len >> PAGE_SHIFT;
2558 if (security_vm_enough_memory_mm(mm, charged))
2559 return -ENOMEM;
2560 vm_flags |= VM_ACCOUNT;
2561 }
2562
2563 next = mas_next(&mas, ULONG_MAX);
2564 prev = mas_prev(&mas, 0);
2565 if (vm_flags & VM_SPECIAL)
2566 goto cannot_expand;
2567
2568 /* Attempt to expand an old mapping */
2569 /* Check next */
2570 if (next && next->vm_start == end && !vma_policy(next) &&
2571 can_vma_merge_before(next, vm_flags, NULL, file, pgoff+pglen,
2572 NULL_VM_UFFD_CTX, NULL)) {
2573 merge_end = next->vm_end;
2574 vma = next;
2575 vm_pgoff = next->vm_pgoff - pglen;
2576 }
2577
2578 /* Check prev */
2579 if (prev && prev->vm_end == addr && !vma_policy(prev) &&
2580 (vma ? can_vma_merge_after(prev, vm_flags, vma->anon_vma, file,
2581 pgoff, vma->vm_userfaultfd_ctx, NULL) :
2582 can_vma_merge_after(prev, vm_flags, NULL, file, pgoff,
2583 NULL_VM_UFFD_CTX, NULL))) {
2584 merge_start = prev->vm_start;
2585 vma = prev;
2586 vm_pgoff = prev->vm_pgoff;
2587 }
2588
2589
2590 /* Actually expand, if possible */
2591 if (vma &&
2592 !vma_expand(&mas, vma, merge_start, merge_end, vm_pgoff, next)) {
2593 khugepaged_enter_vma(vma, vm_flags);
2594 goto expanded;
2595 }
2596
2597 mas.index = addr;
2598 mas.last = end - 1;
2599cannot_expand:
2600 /*
2601 * Determine the object being mapped and call the appropriate
2602 * specific mapper. the address has already been validated, but
2603 * not unmapped, but the maps are removed from the list.
2604 */
2605 vma = vm_area_alloc(mm);
2606 if (!vma) {
2607 error = -ENOMEM;
2608 goto unacct_error;
2609 }
2610
2611 vma->vm_start = addr;
2612 vma->vm_end = end;
2613 vma->vm_flags = vm_flags;
2614 vma->vm_page_prot = vm_get_page_prot(vm_flags);
2615 vma->vm_pgoff = pgoff;
2616
2617 if (file) {
2618 if (vm_flags & VM_SHARED) {
2619 error = mapping_map_writable(file->f_mapping);
2620 if (error)
2621 goto free_vma;
2622 }
2623
2624 vma->vm_file = get_file(file);
2625 error = call_mmap(file, vma);
2626 if (error)
2627 goto unmap_and_free_vma;
2628
2629 /*
2630 * Expansion is handled above, merging is handled below.
2631 * Drivers should not alter the address of the VMA.
2632 */
2633 if (WARN_ON((addr != vma->vm_start))) {
2634 error = -EINVAL;
2635 goto close_and_free_vma;
2636 }
2637 mas_reset(&mas);
2638
2639 /*
2640 * If vm_flags changed after call_mmap(), we should try merge
2641 * vma again as we may succeed this time.
2642 */
2643 if (unlikely(vm_flags != vma->vm_flags && prev)) {
2644 merge = vma_merge(mm, prev, vma->vm_start, vma->vm_end, vma->vm_flags,
2645 NULL, vma->vm_file, vma->vm_pgoff, NULL, NULL_VM_UFFD_CTX, NULL);
2646 if (merge) {
2647 /*
2648 * ->mmap() can change vma->vm_file and fput
2649 * the original file. So fput the vma->vm_file
2650 * here or we would add an extra fput for file
2651 * and cause general protection fault
2652 * ultimately.
2653 */
2654 fput(vma->vm_file);
2655 vm_area_free(vma);
2656 vma = merge;
2657 /* Update vm_flags to pick up the change. */
2658 vm_flags = vma->vm_flags;
2659 goto unmap_writable;
2660 }
2661 }
2662
2663 vm_flags = vma->vm_flags;
2664 } else if (vm_flags & VM_SHARED) {
2665 error = shmem_zero_setup(vma);
2666 if (error)
2667 goto free_vma;
2668 } else {
2669 vma_set_anonymous(vma);
2670 }
2671
2672 /* Allow architectures to sanity-check the vm_flags */
2673 if (!arch_validate_flags(vma->vm_flags)) {
2674 error = -EINVAL;
2675 if (file)
2676 goto close_and_free_vma;
2677 else if (vma->vm_file)
2678 goto unmap_and_free_vma;
2679 else
2680 goto free_vma;
2681 }
2682
2683 if (mas_preallocate(&mas, vma, GFP_KERNEL)) {
2684 error = -ENOMEM;
2685 if (file)
2686 goto close_and_free_vma;
2687 else if (vma->vm_file)
2688 goto unmap_and_free_vma;
2689 else
2690 goto free_vma;
2691 }
2692
2693 if (vma->vm_file)
2694 i_mmap_lock_write(vma->vm_file->f_mapping);
2695
2696 vma_mas_store(vma, &mas);
2697 mm->map_count++;
2698 if (vma->vm_file) {
2699 if (vma->vm_flags & VM_SHARED)
2700 mapping_allow_writable(vma->vm_file->f_mapping);
2701
2702 flush_dcache_mmap_lock(vma->vm_file->f_mapping);
2703 vma_interval_tree_insert(vma, &vma->vm_file->f_mapping->i_mmap);
2704 flush_dcache_mmap_unlock(vma->vm_file->f_mapping);
2705 i_mmap_unlock_write(vma->vm_file->f_mapping);
2706 }
2707
2708 /*
2709 * vma_merge() calls khugepaged_enter_vma() either, the below
2710 * call covers the non-merge case.
2711 */
2712 khugepaged_enter_vma(vma, vma->vm_flags);
2713
2714 /* Once vma denies write, undo our temporary denial count */
2715unmap_writable:
2716 if (file && vm_flags & VM_SHARED)
2717 mapping_unmap_writable(file->f_mapping);
2718 file = vma->vm_file;
2719expanded:
2720 perf_event_mmap(vma);
2721
2722 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
2723 if (vm_flags & VM_LOCKED) {
2724 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
2725 is_vm_hugetlb_page(vma) ||
2726 vma == get_gate_vma(current->mm))
2727 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
2728 else
2729 mm->locked_vm += (len >> PAGE_SHIFT);
2730 }
2731
2732 if (file)
2733 uprobe_mmap(vma);
2734
2735 /*
2736 * New (or expanded) vma always get soft dirty status.
2737 * Otherwise user-space soft-dirty page tracker won't
2738 * be able to distinguish situation when vma area unmapped,
2739 * then new mapped in-place (which must be aimed as
2740 * a completely new data area).
2741 */
2742 vma->vm_flags |= VM_SOFTDIRTY;
2743
2744 vma_set_page_prot(vma);
2745
2746 validate_mm(mm);
2747 return addr;
2748
2749close_and_free_vma:
2750 if (vma->vm_ops && vma->vm_ops->close)
2751 vma->vm_ops->close(vma);
2752unmap_and_free_vma:
2753 fput(vma->vm_file);
2754 vma->vm_file = NULL;
2755
2756 /* Undo any partial mapping done by a device driver. */
2757 unmap_region(mm, mas.tree, vma, prev, next, vma->vm_start, vma->vm_end);
2758 if (file && (vm_flags & VM_SHARED))
2759 mapping_unmap_writable(file->f_mapping);
2760free_vma:
2761 vm_area_free(vma);
2762unacct_error:
2763 if (charged)
2764 vm_unacct_memory(charged);
2765 validate_mm(mm);
2766 return error;
2767}
2768
2769static int __vm_munmap(unsigned long start, size_t len, bool downgrade)
2770{
2771 int ret;
2772 struct mm_struct *mm = current->mm;
2773 LIST_HEAD(uf);
2774 MA_STATE(mas, &mm->mm_mt, start, start);
2775
2776 if (mmap_write_lock_killable(mm))
2777 return -EINTR;
2778
2779 ret = do_mas_munmap(&mas, mm, start, len, &uf, downgrade);
2780 /*
2781 * Returning 1 indicates mmap_lock is downgraded.
2782 * But 1 is not legal return value of vm_munmap() and munmap(), reset
2783 * it to 0 before return.
2784 */
2785 if (ret == 1) {
2786 mmap_read_unlock(mm);
2787 ret = 0;
2788 } else
2789 mmap_write_unlock(mm);
2790
2791 userfaultfd_unmap_complete(mm, &uf);
2792 return ret;
2793}
2794
2795int vm_munmap(unsigned long start, size_t len)
2796{
2797 return __vm_munmap(start, len, false);
2798}
2799EXPORT_SYMBOL(vm_munmap);
2800
2801SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2802{
2803 addr = untagged_addr(addr);
2804 return __vm_munmap(addr, len, true);
2805}
2806
2807
2808/*
2809 * Emulation of deprecated remap_file_pages() syscall.
2810 */
2811SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2812 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2813{
2814
2815 struct mm_struct *mm = current->mm;
2816 struct vm_area_struct *vma;
2817 unsigned long populate = 0;
2818 unsigned long ret = -EINVAL;
2819 struct file *file;
2820
2821 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/mm/remap_file_pages.rst.\n",
2822 current->comm, current->pid);
2823
2824 if (prot)
2825 return ret;
2826 start = start & PAGE_MASK;
2827 size = size & PAGE_MASK;
2828
2829 if (start + size <= start)
2830 return ret;
2831
2832 /* Does pgoff wrap? */
2833 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2834 return ret;
2835
2836 if (mmap_write_lock_killable(mm))
2837 return -EINTR;
2838
2839 vma = vma_lookup(mm, start);
2840
2841 if (!vma || !(vma->vm_flags & VM_SHARED))
2842 goto out;
2843
2844 if (start + size > vma->vm_end) {
2845 VMA_ITERATOR(vmi, mm, vma->vm_end);
2846 struct vm_area_struct *next, *prev = vma;
2847
2848 for_each_vma_range(vmi, next, start + size) {
2849 /* hole between vmas ? */
2850 if (next->vm_start != prev->vm_end)
2851 goto out;
2852
2853 if (next->vm_file != vma->vm_file)
2854 goto out;
2855
2856 if (next->vm_flags != vma->vm_flags)
2857 goto out;
2858
2859 if (start + size <= next->vm_end)
2860 break;
2861
2862 prev = next;
2863 }
2864
2865 if (!next)
2866 goto out;
2867 }
2868
2869 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2870 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2871 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2872
2873 flags &= MAP_NONBLOCK;
2874 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2875 if (vma->vm_flags & VM_LOCKED)
2876 flags |= MAP_LOCKED;
2877
2878 file = get_file(vma->vm_file);
2879 ret = do_mmap(vma->vm_file, start, size,
2880 prot, flags, pgoff, &populate, NULL);
2881 fput(file);
2882out:
2883 mmap_write_unlock(mm);
2884 if (populate)
2885 mm_populate(ret, populate);
2886 if (!IS_ERR_VALUE(ret))
2887 ret = 0;
2888 return ret;
2889}
2890
2891/*
2892 * brk_munmap() - Unmap a parital vma.
2893 * @mas: The maple tree state.
2894 * @vma: The vma to be modified
2895 * @newbrk: the start of the address to unmap
2896 * @oldbrk: The end of the address to unmap
2897 * @uf: The userfaultfd list_head
2898 *
2899 * Returns: 1 on success.
2900 * unmaps a partial VMA mapping. Does not handle alignment, downgrades lock if
2901 * possible.
2902 */
2903static int do_brk_munmap(struct ma_state *mas, struct vm_area_struct *vma,
2904 unsigned long newbrk, unsigned long oldbrk,
2905 struct list_head *uf)
2906{
2907 struct mm_struct *mm = vma->vm_mm;
2908 int ret;
2909
2910 arch_unmap(mm, newbrk, oldbrk);
2911 ret = do_mas_align_munmap(mas, vma, mm, newbrk, oldbrk, uf, true);
2912 validate_mm_mt(mm);
2913 return ret;
2914}
2915
2916/*
2917 * do_brk_flags() - Increase the brk vma if the flags match.
2918 * @mas: The maple tree state.
2919 * @addr: The start address
2920 * @len: The length of the increase
2921 * @vma: The vma,
2922 * @flags: The VMA Flags
2923 *
2924 * Extend the brk VMA from addr to addr + len. If the VMA is NULL or the flags
2925 * do not match then create a new anonymous VMA. Eventually we may be able to
2926 * do some brk-specific accounting here.
2927 */
2928static int do_brk_flags(struct ma_state *mas, struct vm_area_struct *vma,
2929 unsigned long addr, unsigned long len, unsigned long flags)
2930{
2931 struct mm_struct *mm = current->mm;
2932
2933 validate_mm_mt(mm);
2934 /*
2935 * Check against address space limits by the changed size
2936 * Note: This happens *after* clearing old mappings in some code paths.
2937 */
2938 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2939 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
2940 return -ENOMEM;
2941
2942 if (mm->map_count > sysctl_max_map_count)
2943 return -ENOMEM;
2944
2945 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2946 return -ENOMEM;
2947
2948 /*
2949 * Expand the existing vma if possible; Note that singular lists do not
2950 * occur after forking, so the expand will only happen on new VMAs.
2951 */
2952 if (vma && vma->vm_end == addr && !vma_policy(vma) &&
2953 can_vma_merge_after(vma, flags, NULL, NULL,
2954 addr >> PAGE_SHIFT, NULL_VM_UFFD_CTX, NULL)) {
2955 mas_set_range(mas, vma->vm_start, addr + len - 1);
2956 if (mas_preallocate(mas, vma, GFP_KERNEL))
2957 goto unacct_fail;
2958
2959 vma_adjust_trans_huge(vma, vma->vm_start, addr + len, 0);
2960 if (vma->anon_vma) {
2961 anon_vma_lock_write(vma->anon_vma);
2962 anon_vma_interval_tree_pre_update_vma(vma);
2963 }
2964 vma->vm_end = addr + len;
2965 vma->vm_flags |= VM_SOFTDIRTY;
2966 mas_store_prealloc(mas, vma);
2967
2968 if (vma->anon_vma) {
2969 anon_vma_interval_tree_post_update_vma(vma);
2970 anon_vma_unlock_write(vma->anon_vma);
2971 }
2972 khugepaged_enter_vma(vma, flags);
2973 goto out;
2974 }
2975
2976 /* create a vma struct for an anonymous mapping */
2977 vma = vm_area_alloc(mm);
2978 if (!vma)
2979 goto unacct_fail;
2980
2981 vma_set_anonymous(vma);
2982 vma->vm_start = addr;
2983 vma->vm_end = addr + len;
2984 vma->vm_pgoff = addr >> PAGE_SHIFT;
2985 vma->vm_flags = flags;
2986 vma->vm_page_prot = vm_get_page_prot(flags);
2987 mas_set_range(mas, vma->vm_start, addr + len - 1);
2988 if (mas_store_gfp(mas, vma, GFP_KERNEL))
2989 goto mas_store_fail;
2990
2991 mm->map_count++;
2992out:
2993 perf_event_mmap(vma);
2994 mm->total_vm += len >> PAGE_SHIFT;
2995 mm->data_vm += len >> PAGE_SHIFT;
2996 if (flags & VM_LOCKED)
2997 mm->locked_vm += (len >> PAGE_SHIFT);
2998 vma->vm_flags |= VM_SOFTDIRTY;
2999 validate_mm(mm);
3000 return 0;
3001
3002mas_store_fail:
3003 vm_area_free(vma);
3004unacct_fail:
3005 vm_unacct_memory(len >> PAGE_SHIFT);
3006 return -ENOMEM;
3007}
3008
3009int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3010{
3011 struct mm_struct *mm = current->mm;
3012 struct vm_area_struct *vma = NULL;
3013 unsigned long len;
3014 int ret;
3015 bool populate;
3016 LIST_HEAD(uf);
3017 MA_STATE(mas, &mm->mm_mt, addr, addr);
3018
3019 len = PAGE_ALIGN(request);
3020 if (len < request)
3021 return -ENOMEM;
3022 if (!len)
3023 return 0;
3024
3025 if (mmap_write_lock_killable(mm))
3026 return -EINTR;
3027
3028 /* Until we need other flags, refuse anything except VM_EXEC. */
3029 if ((flags & (~VM_EXEC)) != 0)
3030 return -EINVAL;
3031
3032 ret = check_brk_limits(addr, len);
3033 if (ret)
3034 goto limits_failed;
3035
3036 ret = do_mas_munmap(&mas, mm, addr, len, &uf, 0);
3037 if (ret)
3038 goto munmap_failed;
3039
3040 vma = mas_prev(&mas, 0);
3041 ret = do_brk_flags(&mas, vma, addr, len, flags);
3042 populate = ((mm->def_flags & VM_LOCKED) != 0);
3043 mmap_write_unlock(mm);
3044 userfaultfd_unmap_complete(mm, &uf);
3045 if (populate && !ret)
3046 mm_populate(addr, len);
3047 return ret;
3048
3049munmap_failed:
3050limits_failed:
3051 mmap_write_unlock(mm);
3052 return ret;
3053}
3054EXPORT_SYMBOL(vm_brk_flags);
3055
3056int vm_brk(unsigned long addr, unsigned long len)
3057{
3058 return vm_brk_flags(addr, len, 0);
3059}
3060EXPORT_SYMBOL(vm_brk);
3061
3062/* Release all mmaps. */
3063void exit_mmap(struct mm_struct *mm)
3064{
3065 struct mmu_gather tlb;
3066 struct vm_area_struct *vma;
3067 unsigned long nr_accounted = 0;
3068 MA_STATE(mas, &mm->mm_mt, 0, 0);
3069 int count = 0;
3070
3071 /* mm's last user has gone, and its about to be pulled down */
3072 mmu_notifier_release(mm);
3073
3074 mmap_read_lock(mm);
3075 arch_exit_mmap(mm);
3076
3077 vma = mas_find(&mas, ULONG_MAX);
3078 if (!vma) {
3079 /* Can happen if dup_mmap() received an OOM */
3080 mmap_read_unlock(mm);
3081 return;
3082 }
3083
3084 lru_add_drain();
3085 flush_cache_mm(mm);
3086 tlb_gather_mmu_fullmm(&tlb, mm);
3087 /* update_hiwater_rss(mm) here? but nobody should be looking */
3088 /* Use ULONG_MAX here to ensure all VMAs in the mm are unmapped */
3089 unmap_vmas(&tlb, &mm->mm_mt, vma, 0, ULONG_MAX);
3090 mmap_read_unlock(mm);
3091
3092 /*
3093 * Set MMF_OOM_SKIP to hide this task from the oom killer/reaper
3094 * because the memory has been already freed.
3095 */
3096 set_bit(MMF_OOM_SKIP, &mm->flags);
3097 mmap_write_lock(mm);
3098 free_pgtables(&tlb, &mm->mm_mt, vma, FIRST_USER_ADDRESS,
3099 USER_PGTABLES_CEILING);
3100 tlb_finish_mmu(&tlb);
3101
3102 /*
3103 * Walk the list again, actually closing and freeing it, with preemption
3104 * enabled, without holding any MM locks besides the unreachable
3105 * mmap_write_lock.
3106 */
3107 do {
3108 if (vma->vm_flags & VM_ACCOUNT)
3109 nr_accounted += vma_pages(vma);
3110 remove_vma(vma);
3111 count++;
3112 cond_resched();
3113 } while ((vma = mas_find(&mas, ULONG_MAX)) != NULL);
3114
3115 BUG_ON(count != mm->map_count);
3116
3117 trace_exit_mmap(mm);
3118 __mt_destroy(&mm->mm_mt);
3119 mmap_write_unlock(mm);
3120 vm_unacct_memory(nr_accounted);
3121}
3122
3123/* Insert vm structure into process list sorted by address
3124 * and into the inode's i_mmap tree. If vm_file is non-NULL
3125 * then i_mmap_rwsem is taken here.
3126 */
3127int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3128{
3129 unsigned long charged = vma_pages(vma);
3130
3131
3132 if (find_vma_intersection(mm, vma->vm_start, vma->vm_end))
3133 return -ENOMEM;
3134
3135 if ((vma->vm_flags & VM_ACCOUNT) &&
3136 security_vm_enough_memory_mm(mm, charged))
3137 return -ENOMEM;
3138
3139 /*
3140 * The vm_pgoff of a purely anonymous vma should be irrelevant
3141 * until its first write fault, when page's anon_vma and index
3142 * are set. But now set the vm_pgoff it will almost certainly
3143 * end up with (unless mremap moves it elsewhere before that
3144 * first wfault), so /proc/pid/maps tells a consistent story.
3145 *
3146 * By setting it to reflect the virtual start address of the
3147 * vma, merges and splits can happen in a seamless way, just
3148 * using the existing file pgoff checks and manipulations.
3149 * Similarly in do_mmap and in do_brk_flags.
3150 */
3151 if (vma_is_anonymous(vma)) {
3152 BUG_ON(vma->anon_vma);
3153 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3154 }
3155
3156 if (vma_link(mm, vma)) {
3157 vm_unacct_memory(charged);
3158 return -ENOMEM;
3159 }
3160
3161 return 0;
3162}
3163
3164/*
3165 * Copy the vma structure to a new location in the same mm,
3166 * prior to moving page table entries, to effect an mremap move.
3167 */
3168struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3169 unsigned long addr, unsigned long len, pgoff_t pgoff,
3170 bool *need_rmap_locks)
3171{
3172 struct vm_area_struct *vma = *vmap;
3173 unsigned long vma_start = vma->vm_start;
3174 struct mm_struct *mm = vma->vm_mm;
3175 struct vm_area_struct *new_vma, *prev;
3176 bool faulted_in_anon_vma = true;
3177
3178 validate_mm_mt(mm);
3179 /*
3180 * If anonymous vma has not yet been faulted, update new pgoff
3181 * to match new location, to increase its chance of merging.
3182 */
3183 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3184 pgoff = addr >> PAGE_SHIFT;
3185 faulted_in_anon_vma = false;
3186 }
3187
3188 new_vma = find_vma_prev(mm, addr, &prev);
3189 if (new_vma && new_vma->vm_start < addr + len)
3190 return NULL; /* should never get here */
3191
3192 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3193 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3194 vma->vm_userfaultfd_ctx, anon_vma_name(vma));
3195 if (new_vma) {
3196 /*
3197 * Source vma may have been merged into new_vma
3198 */
3199 if (unlikely(vma_start >= new_vma->vm_start &&
3200 vma_start < new_vma->vm_end)) {
3201 /*
3202 * The only way we can get a vma_merge with
3203 * self during an mremap is if the vma hasn't
3204 * been faulted in yet and we were allowed to
3205 * reset the dst vma->vm_pgoff to the
3206 * destination address of the mremap to allow
3207 * the merge to happen. mremap must change the
3208 * vm_pgoff linearity between src and dst vmas
3209 * (in turn preventing a vma_merge) to be
3210 * safe. It is only safe to keep the vm_pgoff
3211 * linear if there are no pages mapped yet.
3212 */
3213 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3214 *vmap = vma = new_vma;
3215 }
3216 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3217 } else {
3218 new_vma = vm_area_dup(vma);
3219 if (!new_vma)
3220 goto out;
3221 new_vma->vm_start = addr;
3222 new_vma->vm_end = addr + len;
3223 new_vma->vm_pgoff = pgoff;
3224 if (vma_dup_policy(vma, new_vma))
3225 goto out_free_vma;
3226 if (anon_vma_clone(new_vma, vma))
3227 goto out_free_mempol;
3228 if (new_vma->vm_file)
3229 get_file(new_vma->vm_file);
3230 if (new_vma->vm_ops && new_vma->vm_ops->open)
3231 new_vma->vm_ops->open(new_vma);
3232 if (vma_link(mm, new_vma))
3233 goto out_vma_link;
3234 *need_rmap_locks = false;
3235 }
3236 validate_mm_mt(mm);
3237 return new_vma;
3238
3239out_vma_link:
3240 if (new_vma->vm_ops && new_vma->vm_ops->close)
3241 new_vma->vm_ops->close(new_vma);
3242
3243 if (new_vma->vm_file)
3244 fput(new_vma->vm_file);
3245
3246 unlink_anon_vmas(new_vma);
3247out_free_mempol:
3248 mpol_put(vma_policy(new_vma));
3249out_free_vma:
3250 vm_area_free(new_vma);
3251out:
3252 validate_mm_mt(mm);
3253 return NULL;
3254}
3255
3256/*
3257 * Return true if the calling process may expand its vm space by the passed
3258 * number of pages
3259 */
3260bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3261{
3262 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3263 return false;
3264
3265 if (is_data_mapping(flags) &&
3266 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3267 /* Workaround for Valgrind */
3268 if (rlimit(RLIMIT_DATA) == 0 &&
3269 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3270 return true;
3271
3272 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3273 current->comm, current->pid,
3274 (mm->data_vm + npages) << PAGE_SHIFT,
3275 rlimit(RLIMIT_DATA),
3276 ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3277
3278 if (!ignore_rlimit_data)
3279 return false;
3280 }
3281
3282 return true;
3283}
3284
3285void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3286{
3287 WRITE_ONCE(mm->total_vm, READ_ONCE(mm->total_vm)+npages);
3288
3289 if (is_exec_mapping(flags))
3290 mm->exec_vm += npages;
3291 else if (is_stack_mapping(flags))
3292 mm->stack_vm += npages;
3293 else if (is_data_mapping(flags))
3294 mm->data_vm += npages;
3295}
3296
3297static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3298
3299/*
3300 * Having a close hook prevents vma merging regardless of flags.
3301 */
3302static void special_mapping_close(struct vm_area_struct *vma)
3303{
3304}
3305
3306static const char *special_mapping_name(struct vm_area_struct *vma)
3307{
3308 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3309}
3310
3311static int special_mapping_mremap(struct vm_area_struct *new_vma)
3312{
3313 struct vm_special_mapping *sm = new_vma->vm_private_data;
3314
3315 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3316 return -EFAULT;
3317
3318 if (sm->mremap)
3319 return sm->mremap(sm, new_vma);
3320
3321 return 0;
3322}
3323
3324static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr)
3325{
3326 /*
3327 * Forbid splitting special mappings - kernel has expectations over
3328 * the number of pages in mapping. Together with VM_DONTEXPAND
3329 * the size of vma should stay the same over the special mapping's
3330 * lifetime.
3331 */
3332 return -EINVAL;
3333}
3334
3335static const struct vm_operations_struct special_mapping_vmops = {
3336 .close = special_mapping_close,
3337 .fault = special_mapping_fault,
3338 .mremap = special_mapping_mremap,
3339 .name = special_mapping_name,
3340 /* vDSO code relies that VVAR can't be accessed remotely */
3341 .access = NULL,
3342 .may_split = special_mapping_split,
3343};
3344
3345static const struct vm_operations_struct legacy_special_mapping_vmops = {
3346 .close = special_mapping_close,
3347 .fault = special_mapping_fault,
3348};
3349
3350static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3351{
3352 struct vm_area_struct *vma = vmf->vma;
3353 pgoff_t pgoff;
3354 struct page **pages;
3355
3356 if (vma->vm_ops == &legacy_special_mapping_vmops) {
3357 pages = vma->vm_private_data;
3358 } else {
3359 struct vm_special_mapping *sm = vma->vm_private_data;
3360
3361 if (sm->fault)
3362 return sm->fault(sm, vmf->vma, vmf);
3363
3364 pages = sm->pages;
3365 }
3366
3367 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3368 pgoff--;
3369
3370 if (*pages) {
3371 struct page *page = *pages;
3372 get_page(page);
3373 vmf->page = page;
3374 return 0;
3375 }
3376
3377 return VM_FAULT_SIGBUS;
3378}
3379
3380static struct vm_area_struct *__install_special_mapping(
3381 struct mm_struct *mm,
3382 unsigned long addr, unsigned long len,
3383 unsigned long vm_flags, void *priv,
3384 const struct vm_operations_struct *ops)
3385{
3386 int ret;
3387 struct vm_area_struct *vma;
3388
3389 validate_mm_mt(mm);
3390 vma = vm_area_alloc(mm);
3391 if (unlikely(vma == NULL))
3392 return ERR_PTR(-ENOMEM);
3393
3394 vma->vm_start = addr;
3395 vma->vm_end = addr + len;
3396
3397 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3398 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
3399 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3400
3401 vma->vm_ops = ops;
3402 vma->vm_private_data = priv;
3403
3404 ret = insert_vm_struct(mm, vma);
3405 if (ret)
3406 goto out;
3407
3408 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3409
3410 perf_event_mmap(vma);
3411
3412 validate_mm_mt(mm);
3413 return vma;
3414
3415out:
3416 vm_area_free(vma);
3417 validate_mm_mt(mm);
3418 return ERR_PTR(ret);
3419}
3420
3421bool vma_is_special_mapping(const struct vm_area_struct *vma,
3422 const struct vm_special_mapping *sm)
3423{
3424 return vma->vm_private_data == sm &&
3425 (vma->vm_ops == &special_mapping_vmops ||
3426 vma->vm_ops == &legacy_special_mapping_vmops);
3427}
3428
3429/*
3430 * Called with mm->mmap_lock held for writing.
3431 * Insert a new vma covering the given region, with the given flags.
3432 * Its pages are supplied by the given array of struct page *.
3433 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3434 * The region past the last page supplied will always produce SIGBUS.
3435 * The array pointer and the pages it points to are assumed to stay alive
3436 * for as long as this mapping might exist.
3437 */
3438struct vm_area_struct *_install_special_mapping(
3439 struct mm_struct *mm,
3440 unsigned long addr, unsigned long len,
3441 unsigned long vm_flags, const struct vm_special_mapping *spec)
3442{
3443 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3444 &special_mapping_vmops);
3445}
3446
3447int install_special_mapping(struct mm_struct *mm,
3448 unsigned long addr, unsigned long len,
3449 unsigned long vm_flags, struct page **pages)
3450{
3451 struct vm_area_struct *vma = __install_special_mapping(
3452 mm, addr, len, vm_flags, (void *)pages,
3453 &legacy_special_mapping_vmops);
3454
3455 return PTR_ERR_OR_ZERO(vma);
3456}
3457
3458static DEFINE_MUTEX(mm_all_locks_mutex);
3459
3460static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3461{
3462 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3463 /*
3464 * The LSB of head.next can't change from under us
3465 * because we hold the mm_all_locks_mutex.
3466 */
3467 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
3468 /*
3469 * We can safely modify head.next after taking the
3470 * anon_vma->root->rwsem. If some other vma in this mm shares
3471 * the same anon_vma we won't take it again.
3472 *
3473 * No need of atomic instructions here, head.next
3474 * can't change from under us thanks to the
3475 * anon_vma->root->rwsem.
3476 */
3477 if (__test_and_set_bit(0, (unsigned long *)
3478 &anon_vma->root->rb_root.rb_root.rb_node))
3479 BUG();
3480 }
3481}
3482
3483static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3484{
3485 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3486 /*
3487 * AS_MM_ALL_LOCKS can't change from under us because
3488 * we hold the mm_all_locks_mutex.
3489 *
3490 * Operations on ->flags have to be atomic because
3491 * even if AS_MM_ALL_LOCKS is stable thanks to the
3492 * mm_all_locks_mutex, there may be other cpus
3493 * changing other bitflags in parallel to us.
3494 */
3495 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3496 BUG();
3497 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
3498 }
3499}
3500
3501/*
3502 * This operation locks against the VM for all pte/vma/mm related
3503 * operations that could ever happen on a certain mm. This includes
3504 * vmtruncate, try_to_unmap, and all page faults.
3505 *
3506 * The caller must take the mmap_lock in write mode before calling
3507 * mm_take_all_locks(). The caller isn't allowed to release the
3508 * mmap_lock until mm_drop_all_locks() returns.
3509 *
3510 * mmap_lock in write mode is required in order to block all operations
3511 * that could modify pagetables and free pages without need of
3512 * altering the vma layout. It's also needed in write mode to avoid new
3513 * anon_vmas to be associated with existing vmas.
3514 *
3515 * A single task can't take more than one mm_take_all_locks() in a row
3516 * or it would deadlock.
3517 *
3518 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3519 * mapping->flags avoid to take the same lock twice, if more than one
3520 * vma in this mm is backed by the same anon_vma or address_space.
3521 *
3522 * We take locks in following order, accordingly to comment at beginning
3523 * of mm/rmap.c:
3524 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3525 * hugetlb mapping);
3526 * - all i_mmap_rwsem locks;
3527 * - all anon_vma->rwseml
3528 *
3529 * We can take all locks within these types randomly because the VM code
3530 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3531 * mm_all_locks_mutex.
3532 *
3533 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3534 * that may have to take thousand of locks.
3535 *
3536 * mm_take_all_locks() can fail if it's interrupted by signals.
3537 */
3538int mm_take_all_locks(struct mm_struct *mm)
3539{
3540 struct vm_area_struct *vma;
3541 struct anon_vma_chain *avc;
3542 MA_STATE(mas, &mm->mm_mt, 0, 0);
3543
3544 mmap_assert_write_locked(mm);
3545
3546 mutex_lock(&mm_all_locks_mutex);
3547
3548 mas_for_each(&mas, vma, ULONG_MAX) {
3549 if (signal_pending(current))
3550 goto out_unlock;
3551 if (vma->vm_file && vma->vm_file->f_mapping &&
3552 is_vm_hugetlb_page(vma))
3553 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3554 }
3555
3556 mas_set(&mas, 0);
3557 mas_for_each(&mas, vma, ULONG_MAX) {
3558 if (signal_pending(current))
3559 goto out_unlock;
3560 if (vma->vm_file && vma->vm_file->f_mapping &&
3561 !is_vm_hugetlb_page(vma))
3562 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3563 }
3564
3565 mas_set(&mas, 0);
3566 mas_for_each(&mas, vma, ULONG_MAX) {
3567 if (signal_pending(current))
3568 goto out_unlock;
3569 if (vma->anon_vma)
3570 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3571 vm_lock_anon_vma(mm, avc->anon_vma);
3572 }
3573
3574 return 0;
3575
3576out_unlock:
3577 mm_drop_all_locks(mm);
3578 return -EINTR;
3579}
3580
3581static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3582{
3583 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3584 /*
3585 * The LSB of head.next can't change to 0 from under
3586 * us because we hold the mm_all_locks_mutex.
3587 *
3588 * We must however clear the bitflag before unlocking
3589 * the vma so the users using the anon_vma->rb_root will
3590 * never see our bitflag.
3591 *
3592 * No need of atomic instructions here, head.next
3593 * can't change from under us until we release the
3594 * anon_vma->root->rwsem.
3595 */
3596 if (!__test_and_clear_bit(0, (unsigned long *)
3597 &anon_vma->root->rb_root.rb_root.rb_node))
3598 BUG();
3599 anon_vma_unlock_write(anon_vma);
3600 }
3601}
3602
3603static void vm_unlock_mapping(struct address_space *mapping)
3604{
3605 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3606 /*
3607 * AS_MM_ALL_LOCKS can't change to 0 from under us
3608 * because we hold the mm_all_locks_mutex.
3609 */
3610 i_mmap_unlock_write(mapping);
3611 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3612 &mapping->flags))
3613 BUG();
3614 }
3615}
3616
3617/*
3618 * The mmap_lock cannot be released by the caller until
3619 * mm_drop_all_locks() returns.
3620 */
3621void mm_drop_all_locks(struct mm_struct *mm)
3622{
3623 struct vm_area_struct *vma;
3624 struct anon_vma_chain *avc;
3625 MA_STATE(mas, &mm->mm_mt, 0, 0);
3626
3627 mmap_assert_write_locked(mm);
3628 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3629
3630 mas_for_each(&mas, vma, ULONG_MAX) {
3631 if (vma->anon_vma)
3632 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3633 vm_unlock_anon_vma(avc->anon_vma);
3634 if (vma->vm_file && vma->vm_file->f_mapping)
3635 vm_unlock_mapping(vma->vm_file->f_mapping);
3636 }
3637
3638 mutex_unlock(&mm_all_locks_mutex);
3639}
3640
3641/*
3642 * initialise the percpu counter for VM
3643 */
3644void __init mmap_init(void)
3645{
3646 int ret;
3647
3648 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3649 VM_BUG_ON(ret);
3650}
3651
3652/*
3653 * Initialise sysctl_user_reserve_kbytes.
3654 *
3655 * This is intended to prevent a user from starting a single memory hogging
3656 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3657 * mode.
3658 *
3659 * The default value is min(3% of free memory, 128MB)
3660 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3661 */
3662static int init_user_reserve(void)
3663{
3664 unsigned long free_kbytes;
3665
3666 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3667
3668 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3669 return 0;
3670}
3671subsys_initcall(init_user_reserve);
3672
3673/*
3674 * Initialise sysctl_admin_reserve_kbytes.
3675 *
3676 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3677 * to log in and kill a memory hogging process.
3678 *
3679 * Systems with more than 256MB will reserve 8MB, enough to recover
3680 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3681 * only reserve 3% of free pages by default.
3682 */
3683static int init_admin_reserve(void)
3684{
3685 unsigned long free_kbytes;
3686
3687 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3688
3689 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3690 return 0;
3691}
3692subsys_initcall(init_admin_reserve);
3693
3694/*
3695 * Reinititalise user and admin reserves if memory is added or removed.
3696 *
3697 * The default user reserve max is 128MB, and the default max for the
3698 * admin reserve is 8MB. These are usually, but not always, enough to
3699 * enable recovery from a memory hogging process using login/sshd, a shell,
3700 * and tools like top. It may make sense to increase or even disable the
3701 * reserve depending on the existence of swap or variations in the recovery
3702 * tools. So, the admin may have changed them.
3703 *
3704 * If memory is added and the reserves have been eliminated or increased above
3705 * the default max, then we'll trust the admin.
3706 *
3707 * If memory is removed and there isn't enough free memory, then we
3708 * need to reset the reserves.
3709 *
3710 * Otherwise keep the reserve set by the admin.
3711 */
3712static int reserve_mem_notifier(struct notifier_block *nb,
3713 unsigned long action, void *data)
3714{
3715 unsigned long tmp, free_kbytes;
3716
3717 switch (action) {
3718 case MEM_ONLINE:
3719 /* Default max is 128MB. Leave alone if modified by operator. */
3720 tmp = sysctl_user_reserve_kbytes;
3721 if (0 < tmp && tmp < (1UL << 17))
3722 init_user_reserve();
3723
3724 /* Default max is 8MB. Leave alone if modified by operator. */
3725 tmp = sysctl_admin_reserve_kbytes;
3726 if (0 < tmp && tmp < (1UL << 13))
3727 init_admin_reserve();
3728
3729 break;
3730 case MEM_OFFLINE:
3731 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3732
3733 if (sysctl_user_reserve_kbytes > free_kbytes) {
3734 init_user_reserve();
3735 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3736 sysctl_user_reserve_kbytes);
3737 }
3738
3739 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3740 init_admin_reserve();
3741 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3742 sysctl_admin_reserve_kbytes);
3743 }
3744 break;
3745 default:
3746 break;
3747 }
3748 return NOTIFY_OK;
3749}
3750
3751static int __meminit init_reserve_notifier(void)
3752{
3753 if (hotplug_memory_notifier(reserve_mem_notifier, DEFAULT_CALLBACK_PRI))
3754 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3755
3756 return 0;
3757}
3758subsys_initcall(init_reserve_notifier);