Loading...
1/*
2 * mm/mmap.c
3 *
4 * Written by obz.
5 *
6 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
7 */
8
9#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
10
11#include <linux/kernel.h>
12#include <linux/slab.h>
13#include <linux/backing-dev.h>
14#include <linux/mm.h>
15#include <linux/vmacache.h>
16#include <linux/shm.h>
17#include <linux/mman.h>
18#include <linux/pagemap.h>
19#include <linux/swap.h>
20#include <linux/syscalls.h>
21#include <linux/capability.h>
22#include <linux/init.h>
23#include <linux/file.h>
24#include <linux/fs.h>
25#include <linux/personality.h>
26#include <linux/security.h>
27#include <linux/hugetlb.h>
28#include <linux/shmem_fs.h>
29#include <linux/profile.h>
30#include <linux/export.h>
31#include <linux/mount.h>
32#include <linux/mempolicy.h>
33#include <linux/rmap.h>
34#include <linux/mmu_notifier.h>
35#include <linux/mmdebug.h>
36#include <linux/perf_event.h>
37#include <linux/audit.h>
38#include <linux/khugepaged.h>
39#include <linux/uprobes.h>
40#include <linux/rbtree_augmented.h>
41#include <linux/notifier.h>
42#include <linux/memory.h>
43#include <linux/printk.h>
44#include <linux/userfaultfd_k.h>
45#include <linux/moduleparam.h>
46#include <linux/pkeys.h>
47#include <linux/oom.h>
48
49#include <linux/uaccess.h>
50#include <asm/cacheflush.h>
51#include <asm/tlb.h>
52#include <asm/mmu_context.h>
53
54#include "internal.h"
55
56#ifndef arch_mmap_check
57#define arch_mmap_check(addr, len, flags) (0)
58#endif
59
60#ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
61const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
62const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
63int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
64#endif
65#ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
66const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
67const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
68int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
69#endif
70
71static bool ignore_rlimit_data;
72core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
73
74static void unmap_region(struct mm_struct *mm,
75 struct vm_area_struct *vma, struct vm_area_struct *prev,
76 unsigned long start, unsigned long end);
77
78/* description of effects of mapping type and prot in current implementation.
79 * this is due to the limited x86 page protection hardware. The expected
80 * behavior is in parens:
81 *
82 * map_type prot
83 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
84 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
85 * w: (no) no w: (no) no w: (yes) yes w: (no) no
86 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
87 *
88 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
89 * w: (no) no w: (no) no w: (copy) copy w: (no) no
90 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
91 *
92 * On arm64, PROT_EXEC has the following behaviour for both MAP_SHARED and
93 * MAP_PRIVATE:
94 * r: (no) no
95 * w: (no) no
96 * x: (yes) yes
97 */
98pgprot_t protection_map[16] __ro_after_init = {
99 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
100 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
101};
102
103#ifndef CONFIG_ARCH_HAS_FILTER_PGPROT
104static inline pgprot_t arch_filter_pgprot(pgprot_t prot)
105{
106 return prot;
107}
108#endif
109
110pgprot_t vm_get_page_prot(unsigned long vm_flags)
111{
112 pgprot_t ret = __pgprot(pgprot_val(protection_map[vm_flags &
113 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
114 pgprot_val(arch_vm_get_page_prot(vm_flags)));
115
116 return arch_filter_pgprot(ret);
117}
118EXPORT_SYMBOL(vm_get_page_prot);
119
120static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
121{
122 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
123}
124
125/* Update vma->vm_page_prot to reflect vma->vm_flags. */
126void vma_set_page_prot(struct vm_area_struct *vma)
127{
128 unsigned long vm_flags = vma->vm_flags;
129 pgprot_t vm_page_prot;
130
131 vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
132 if (vma_wants_writenotify(vma, vm_page_prot)) {
133 vm_flags &= ~VM_SHARED;
134 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
135 }
136 /* remove_protection_ptes reads vma->vm_page_prot without mmap_sem */
137 WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
138}
139
140/*
141 * Requires inode->i_mapping->i_mmap_rwsem
142 */
143static void __remove_shared_vm_struct(struct vm_area_struct *vma,
144 struct file *file, struct address_space *mapping)
145{
146 if (vma->vm_flags & VM_DENYWRITE)
147 atomic_inc(&file_inode(file)->i_writecount);
148 if (vma->vm_flags & VM_SHARED)
149 mapping_unmap_writable(mapping);
150
151 flush_dcache_mmap_lock(mapping);
152 vma_interval_tree_remove(vma, &mapping->i_mmap);
153 flush_dcache_mmap_unlock(mapping);
154}
155
156/*
157 * Unlink a file-based vm structure from its interval tree, to hide
158 * vma from rmap and vmtruncate before freeing its page tables.
159 */
160void unlink_file_vma(struct vm_area_struct *vma)
161{
162 struct file *file = vma->vm_file;
163
164 if (file) {
165 struct address_space *mapping = file->f_mapping;
166 i_mmap_lock_write(mapping);
167 __remove_shared_vm_struct(vma, file, mapping);
168 i_mmap_unlock_write(mapping);
169 }
170}
171
172/*
173 * Close a vm structure and free it, returning the next.
174 */
175static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
176{
177 struct vm_area_struct *next = vma->vm_next;
178
179 might_sleep();
180 if (vma->vm_ops && vma->vm_ops->close)
181 vma->vm_ops->close(vma);
182 if (vma->vm_file)
183 fput(vma->vm_file);
184 mpol_put(vma_policy(vma));
185 kmem_cache_free(vm_area_cachep, vma);
186 return next;
187}
188
189static int do_brk(unsigned long addr, unsigned long len, struct list_head *uf);
190
191SYSCALL_DEFINE1(brk, unsigned long, brk)
192{
193 unsigned long retval;
194 unsigned long newbrk, oldbrk;
195 struct mm_struct *mm = current->mm;
196 struct vm_area_struct *next;
197 unsigned long min_brk;
198 bool populate;
199 LIST_HEAD(uf);
200
201 if (down_write_killable(&mm->mmap_sem))
202 return -EINTR;
203
204#ifdef CONFIG_COMPAT_BRK
205 /*
206 * CONFIG_COMPAT_BRK can still be overridden by setting
207 * randomize_va_space to 2, which will still cause mm->start_brk
208 * to be arbitrarily shifted
209 */
210 if (current->brk_randomized)
211 min_brk = mm->start_brk;
212 else
213 min_brk = mm->end_data;
214#else
215 min_brk = mm->start_brk;
216#endif
217 if (brk < min_brk)
218 goto out;
219
220 /*
221 * Check against rlimit here. If this check is done later after the test
222 * of oldbrk with newbrk then it can escape the test and let the data
223 * segment grow beyond its set limit the in case where the limit is
224 * not page aligned -Ram Gupta
225 */
226 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
227 mm->end_data, mm->start_data))
228 goto out;
229
230 newbrk = PAGE_ALIGN(brk);
231 oldbrk = PAGE_ALIGN(mm->brk);
232 if (oldbrk == newbrk)
233 goto set_brk;
234
235 /* Always allow shrinking brk. */
236 if (brk <= mm->brk) {
237 if (!do_munmap(mm, newbrk, oldbrk-newbrk, &uf))
238 goto set_brk;
239 goto out;
240 }
241
242 /* Check against existing mmap mappings. */
243 next = find_vma(mm, oldbrk);
244 if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
245 goto out;
246
247 /* Ok, looks good - let it rip. */
248 if (do_brk(oldbrk, newbrk-oldbrk, &uf) < 0)
249 goto out;
250
251set_brk:
252 mm->brk = brk;
253 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
254 up_write(&mm->mmap_sem);
255 userfaultfd_unmap_complete(mm, &uf);
256 if (populate)
257 mm_populate(oldbrk, newbrk - oldbrk);
258 return brk;
259
260out:
261 retval = mm->brk;
262 up_write(&mm->mmap_sem);
263 return retval;
264}
265
266static long vma_compute_subtree_gap(struct vm_area_struct *vma)
267{
268 unsigned long max, prev_end, subtree_gap;
269
270 /*
271 * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
272 * allow two stack_guard_gaps between them here, and when choosing
273 * an unmapped area; whereas when expanding we only require one.
274 * That's a little inconsistent, but keeps the code here simpler.
275 */
276 max = vm_start_gap(vma);
277 if (vma->vm_prev) {
278 prev_end = vm_end_gap(vma->vm_prev);
279 if (max > prev_end)
280 max -= prev_end;
281 else
282 max = 0;
283 }
284 if (vma->vm_rb.rb_left) {
285 subtree_gap = rb_entry(vma->vm_rb.rb_left,
286 struct vm_area_struct, vm_rb)->rb_subtree_gap;
287 if (subtree_gap > max)
288 max = subtree_gap;
289 }
290 if (vma->vm_rb.rb_right) {
291 subtree_gap = rb_entry(vma->vm_rb.rb_right,
292 struct vm_area_struct, vm_rb)->rb_subtree_gap;
293 if (subtree_gap > max)
294 max = subtree_gap;
295 }
296 return max;
297}
298
299#ifdef CONFIG_DEBUG_VM_RB
300static int browse_rb(struct mm_struct *mm)
301{
302 struct rb_root *root = &mm->mm_rb;
303 int i = 0, j, bug = 0;
304 struct rb_node *nd, *pn = NULL;
305 unsigned long prev = 0, pend = 0;
306
307 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
308 struct vm_area_struct *vma;
309 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
310 if (vma->vm_start < prev) {
311 pr_emerg("vm_start %lx < prev %lx\n",
312 vma->vm_start, prev);
313 bug = 1;
314 }
315 if (vma->vm_start < pend) {
316 pr_emerg("vm_start %lx < pend %lx\n",
317 vma->vm_start, pend);
318 bug = 1;
319 }
320 if (vma->vm_start > vma->vm_end) {
321 pr_emerg("vm_start %lx > vm_end %lx\n",
322 vma->vm_start, vma->vm_end);
323 bug = 1;
324 }
325 spin_lock(&mm->page_table_lock);
326 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
327 pr_emerg("free gap %lx, correct %lx\n",
328 vma->rb_subtree_gap,
329 vma_compute_subtree_gap(vma));
330 bug = 1;
331 }
332 spin_unlock(&mm->page_table_lock);
333 i++;
334 pn = nd;
335 prev = vma->vm_start;
336 pend = vma->vm_end;
337 }
338 j = 0;
339 for (nd = pn; nd; nd = rb_prev(nd))
340 j++;
341 if (i != j) {
342 pr_emerg("backwards %d, forwards %d\n", j, i);
343 bug = 1;
344 }
345 return bug ? -1 : i;
346}
347
348static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
349{
350 struct rb_node *nd;
351
352 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
353 struct vm_area_struct *vma;
354 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
355 VM_BUG_ON_VMA(vma != ignore &&
356 vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
357 vma);
358 }
359}
360
361static void validate_mm(struct mm_struct *mm)
362{
363 int bug = 0;
364 int i = 0;
365 unsigned long highest_address = 0;
366 struct vm_area_struct *vma = mm->mmap;
367
368 while (vma) {
369 struct anon_vma *anon_vma = vma->anon_vma;
370 struct anon_vma_chain *avc;
371
372 if (anon_vma) {
373 anon_vma_lock_read(anon_vma);
374 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
375 anon_vma_interval_tree_verify(avc);
376 anon_vma_unlock_read(anon_vma);
377 }
378
379 highest_address = vm_end_gap(vma);
380 vma = vma->vm_next;
381 i++;
382 }
383 if (i != mm->map_count) {
384 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
385 bug = 1;
386 }
387 if (highest_address != mm->highest_vm_end) {
388 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
389 mm->highest_vm_end, highest_address);
390 bug = 1;
391 }
392 i = browse_rb(mm);
393 if (i != mm->map_count) {
394 if (i != -1)
395 pr_emerg("map_count %d rb %d\n", mm->map_count, i);
396 bug = 1;
397 }
398 VM_BUG_ON_MM(bug, mm);
399}
400#else
401#define validate_mm_rb(root, ignore) do { } while (0)
402#define validate_mm(mm) do { } while (0)
403#endif
404
405RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
406 unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
407
408/*
409 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
410 * vma->vm_prev->vm_end values changed, without modifying the vma's position
411 * in the rbtree.
412 */
413static void vma_gap_update(struct vm_area_struct *vma)
414{
415 /*
416 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
417 * function that does exacltly what we want.
418 */
419 vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
420}
421
422static inline void vma_rb_insert(struct vm_area_struct *vma,
423 struct rb_root *root)
424{
425 /* All rb_subtree_gap values must be consistent prior to insertion */
426 validate_mm_rb(root, NULL);
427
428 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
429}
430
431static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
432{
433 /*
434 * Note rb_erase_augmented is a fairly large inline function,
435 * so make sure we instantiate it only once with our desired
436 * augmented rbtree callbacks.
437 */
438 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
439}
440
441static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
442 struct rb_root *root,
443 struct vm_area_struct *ignore)
444{
445 /*
446 * All rb_subtree_gap values must be consistent prior to erase,
447 * with the possible exception of the "next" vma being erased if
448 * next->vm_start was reduced.
449 */
450 validate_mm_rb(root, ignore);
451
452 __vma_rb_erase(vma, root);
453}
454
455static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
456 struct rb_root *root)
457{
458 /*
459 * All rb_subtree_gap values must be consistent prior to erase,
460 * with the possible exception of the vma being erased.
461 */
462 validate_mm_rb(root, vma);
463
464 __vma_rb_erase(vma, root);
465}
466
467/*
468 * vma has some anon_vma assigned, and is already inserted on that
469 * anon_vma's interval trees.
470 *
471 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
472 * vma must be removed from the anon_vma's interval trees using
473 * anon_vma_interval_tree_pre_update_vma().
474 *
475 * After the update, the vma will be reinserted using
476 * anon_vma_interval_tree_post_update_vma().
477 *
478 * The entire update must be protected by exclusive mmap_sem and by
479 * the root anon_vma's mutex.
480 */
481static inline void
482anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
483{
484 struct anon_vma_chain *avc;
485
486 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
487 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
488}
489
490static inline void
491anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
492{
493 struct anon_vma_chain *avc;
494
495 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
496 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
497}
498
499static int find_vma_links(struct mm_struct *mm, unsigned long addr,
500 unsigned long end, struct vm_area_struct **pprev,
501 struct rb_node ***rb_link, struct rb_node **rb_parent)
502{
503 struct rb_node **__rb_link, *__rb_parent, *rb_prev;
504
505 __rb_link = &mm->mm_rb.rb_node;
506 rb_prev = __rb_parent = NULL;
507
508 while (*__rb_link) {
509 struct vm_area_struct *vma_tmp;
510
511 __rb_parent = *__rb_link;
512 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
513
514 if (vma_tmp->vm_end > addr) {
515 /* Fail if an existing vma overlaps the area */
516 if (vma_tmp->vm_start < end)
517 return -ENOMEM;
518 __rb_link = &__rb_parent->rb_left;
519 } else {
520 rb_prev = __rb_parent;
521 __rb_link = &__rb_parent->rb_right;
522 }
523 }
524
525 *pprev = NULL;
526 if (rb_prev)
527 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
528 *rb_link = __rb_link;
529 *rb_parent = __rb_parent;
530 return 0;
531}
532
533static unsigned long count_vma_pages_range(struct mm_struct *mm,
534 unsigned long addr, unsigned long end)
535{
536 unsigned long nr_pages = 0;
537 struct vm_area_struct *vma;
538
539 /* Find first overlaping mapping */
540 vma = find_vma_intersection(mm, addr, end);
541 if (!vma)
542 return 0;
543
544 nr_pages = (min(end, vma->vm_end) -
545 max(addr, vma->vm_start)) >> PAGE_SHIFT;
546
547 /* Iterate over the rest of the overlaps */
548 for (vma = vma->vm_next; vma; vma = vma->vm_next) {
549 unsigned long overlap_len;
550
551 if (vma->vm_start > end)
552 break;
553
554 overlap_len = min(end, vma->vm_end) - vma->vm_start;
555 nr_pages += overlap_len >> PAGE_SHIFT;
556 }
557
558 return nr_pages;
559}
560
561void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
562 struct rb_node **rb_link, struct rb_node *rb_parent)
563{
564 /* Update tracking information for the gap following the new vma. */
565 if (vma->vm_next)
566 vma_gap_update(vma->vm_next);
567 else
568 mm->highest_vm_end = vm_end_gap(vma);
569
570 /*
571 * vma->vm_prev wasn't known when we followed the rbtree to find the
572 * correct insertion point for that vma. As a result, we could not
573 * update the vma vm_rb parents rb_subtree_gap values on the way down.
574 * So, we first insert the vma with a zero rb_subtree_gap value
575 * (to be consistent with what we did on the way down), and then
576 * immediately update the gap to the correct value. Finally we
577 * rebalance the rbtree after all augmented values have been set.
578 */
579 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
580 vma->rb_subtree_gap = 0;
581 vma_gap_update(vma);
582 vma_rb_insert(vma, &mm->mm_rb);
583}
584
585static void __vma_link_file(struct vm_area_struct *vma)
586{
587 struct file *file;
588
589 file = vma->vm_file;
590 if (file) {
591 struct address_space *mapping = file->f_mapping;
592
593 if (vma->vm_flags & VM_DENYWRITE)
594 atomic_dec(&file_inode(file)->i_writecount);
595 if (vma->vm_flags & VM_SHARED)
596 atomic_inc(&mapping->i_mmap_writable);
597
598 flush_dcache_mmap_lock(mapping);
599 vma_interval_tree_insert(vma, &mapping->i_mmap);
600 flush_dcache_mmap_unlock(mapping);
601 }
602}
603
604static void
605__vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
606 struct vm_area_struct *prev, struct rb_node **rb_link,
607 struct rb_node *rb_parent)
608{
609 __vma_link_list(mm, vma, prev, rb_parent);
610 __vma_link_rb(mm, vma, rb_link, rb_parent);
611}
612
613static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
614 struct vm_area_struct *prev, struct rb_node **rb_link,
615 struct rb_node *rb_parent)
616{
617 struct address_space *mapping = NULL;
618
619 if (vma->vm_file) {
620 mapping = vma->vm_file->f_mapping;
621 i_mmap_lock_write(mapping);
622 }
623
624 __vma_link(mm, vma, prev, rb_link, rb_parent);
625 __vma_link_file(vma);
626
627 if (mapping)
628 i_mmap_unlock_write(mapping);
629
630 mm->map_count++;
631 validate_mm(mm);
632}
633
634/*
635 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
636 * mm's list and rbtree. It has already been inserted into the interval tree.
637 */
638static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
639{
640 struct vm_area_struct *prev;
641 struct rb_node **rb_link, *rb_parent;
642
643 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
644 &prev, &rb_link, &rb_parent))
645 BUG();
646 __vma_link(mm, vma, prev, rb_link, rb_parent);
647 mm->map_count++;
648}
649
650static __always_inline void __vma_unlink_common(struct mm_struct *mm,
651 struct vm_area_struct *vma,
652 struct vm_area_struct *prev,
653 bool has_prev,
654 struct vm_area_struct *ignore)
655{
656 struct vm_area_struct *next;
657
658 vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
659 next = vma->vm_next;
660 if (has_prev)
661 prev->vm_next = next;
662 else {
663 prev = vma->vm_prev;
664 if (prev)
665 prev->vm_next = next;
666 else
667 mm->mmap = next;
668 }
669 if (next)
670 next->vm_prev = prev;
671
672 /* Kill the cache */
673 vmacache_invalidate(mm);
674}
675
676static inline void __vma_unlink_prev(struct mm_struct *mm,
677 struct vm_area_struct *vma,
678 struct vm_area_struct *prev)
679{
680 __vma_unlink_common(mm, vma, prev, true, vma);
681}
682
683/*
684 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
685 * is already present in an i_mmap tree without adjusting the tree.
686 * The following helper function should be used when such adjustments
687 * are necessary. The "insert" vma (if any) is to be inserted
688 * before we drop the necessary locks.
689 */
690int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
691 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
692 struct vm_area_struct *expand)
693{
694 struct mm_struct *mm = vma->vm_mm;
695 struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
696 struct address_space *mapping = NULL;
697 struct rb_root_cached *root = NULL;
698 struct anon_vma *anon_vma = NULL;
699 struct file *file = vma->vm_file;
700 bool start_changed = false, end_changed = false;
701 long adjust_next = 0;
702 int remove_next = 0;
703
704 if (next && !insert) {
705 struct vm_area_struct *exporter = NULL, *importer = NULL;
706
707 if (end >= next->vm_end) {
708 /*
709 * vma expands, overlapping all the next, and
710 * perhaps the one after too (mprotect case 6).
711 * The only other cases that gets here are
712 * case 1, case 7 and case 8.
713 */
714 if (next == expand) {
715 /*
716 * The only case where we don't expand "vma"
717 * and we expand "next" instead is case 8.
718 */
719 VM_WARN_ON(end != next->vm_end);
720 /*
721 * remove_next == 3 means we're
722 * removing "vma" and that to do so we
723 * swapped "vma" and "next".
724 */
725 remove_next = 3;
726 VM_WARN_ON(file != next->vm_file);
727 swap(vma, next);
728 } else {
729 VM_WARN_ON(expand != vma);
730 /*
731 * case 1, 6, 7, remove_next == 2 is case 6,
732 * remove_next == 1 is case 1 or 7.
733 */
734 remove_next = 1 + (end > next->vm_end);
735 VM_WARN_ON(remove_next == 2 &&
736 end != next->vm_next->vm_end);
737 VM_WARN_ON(remove_next == 1 &&
738 end != next->vm_end);
739 /* trim end to next, for case 6 first pass */
740 end = next->vm_end;
741 }
742
743 exporter = next;
744 importer = vma;
745
746 /*
747 * If next doesn't have anon_vma, import from vma after
748 * next, if the vma overlaps with it.
749 */
750 if (remove_next == 2 && !next->anon_vma)
751 exporter = next->vm_next;
752
753 } else if (end > next->vm_start) {
754 /*
755 * vma expands, overlapping part of the next:
756 * mprotect case 5 shifting the boundary up.
757 */
758 adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
759 exporter = next;
760 importer = vma;
761 VM_WARN_ON(expand != importer);
762 } else if (end < vma->vm_end) {
763 /*
764 * vma shrinks, and !insert tells it's not
765 * split_vma inserting another: so it must be
766 * mprotect case 4 shifting the boundary down.
767 */
768 adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT);
769 exporter = vma;
770 importer = next;
771 VM_WARN_ON(expand != importer);
772 }
773
774 /*
775 * Easily overlooked: when mprotect shifts the boundary,
776 * make sure the expanding vma has anon_vma set if the
777 * shrinking vma had, to cover any anon pages imported.
778 */
779 if (exporter && exporter->anon_vma && !importer->anon_vma) {
780 int error;
781
782 importer->anon_vma = exporter->anon_vma;
783 error = anon_vma_clone(importer, exporter);
784 if (error)
785 return error;
786 }
787 }
788again:
789 vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
790
791 if (file) {
792 mapping = file->f_mapping;
793 root = &mapping->i_mmap;
794 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
795
796 if (adjust_next)
797 uprobe_munmap(next, next->vm_start, next->vm_end);
798
799 i_mmap_lock_write(mapping);
800 if (insert) {
801 /*
802 * Put into interval tree now, so instantiated pages
803 * are visible to arm/parisc __flush_dcache_page
804 * throughout; but we cannot insert into address
805 * space until vma start or end is updated.
806 */
807 __vma_link_file(insert);
808 }
809 }
810
811 anon_vma = vma->anon_vma;
812 if (!anon_vma && adjust_next)
813 anon_vma = next->anon_vma;
814 if (anon_vma) {
815 VM_WARN_ON(adjust_next && next->anon_vma &&
816 anon_vma != next->anon_vma);
817 anon_vma_lock_write(anon_vma);
818 anon_vma_interval_tree_pre_update_vma(vma);
819 if (adjust_next)
820 anon_vma_interval_tree_pre_update_vma(next);
821 }
822
823 if (root) {
824 flush_dcache_mmap_lock(mapping);
825 vma_interval_tree_remove(vma, root);
826 if (adjust_next)
827 vma_interval_tree_remove(next, root);
828 }
829
830 if (start != vma->vm_start) {
831 vma->vm_start = start;
832 start_changed = true;
833 }
834 if (end != vma->vm_end) {
835 vma->vm_end = end;
836 end_changed = true;
837 }
838 vma->vm_pgoff = pgoff;
839 if (adjust_next) {
840 next->vm_start += adjust_next << PAGE_SHIFT;
841 next->vm_pgoff += adjust_next;
842 }
843
844 if (root) {
845 if (adjust_next)
846 vma_interval_tree_insert(next, root);
847 vma_interval_tree_insert(vma, root);
848 flush_dcache_mmap_unlock(mapping);
849 }
850
851 if (remove_next) {
852 /*
853 * vma_merge has merged next into vma, and needs
854 * us to remove next before dropping the locks.
855 */
856 if (remove_next != 3)
857 __vma_unlink_prev(mm, next, vma);
858 else
859 /*
860 * vma is not before next if they've been
861 * swapped.
862 *
863 * pre-swap() next->vm_start was reduced so
864 * tell validate_mm_rb to ignore pre-swap()
865 * "next" (which is stored in post-swap()
866 * "vma").
867 */
868 __vma_unlink_common(mm, next, NULL, false, vma);
869 if (file)
870 __remove_shared_vm_struct(next, file, mapping);
871 } else if (insert) {
872 /*
873 * split_vma has split insert from vma, and needs
874 * us to insert it before dropping the locks
875 * (it may either follow vma or precede it).
876 */
877 __insert_vm_struct(mm, insert);
878 } else {
879 if (start_changed)
880 vma_gap_update(vma);
881 if (end_changed) {
882 if (!next)
883 mm->highest_vm_end = vm_end_gap(vma);
884 else if (!adjust_next)
885 vma_gap_update(next);
886 }
887 }
888
889 if (anon_vma) {
890 anon_vma_interval_tree_post_update_vma(vma);
891 if (adjust_next)
892 anon_vma_interval_tree_post_update_vma(next);
893 anon_vma_unlock_write(anon_vma);
894 }
895 if (mapping)
896 i_mmap_unlock_write(mapping);
897
898 if (root) {
899 uprobe_mmap(vma);
900
901 if (adjust_next)
902 uprobe_mmap(next);
903 }
904
905 if (remove_next) {
906 if (file) {
907 uprobe_munmap(next, next->vm_start, next->vm_end);
908 fput(file);
909 }
910 if (next->anon_vma)
911 anon_vma_merge(vma, next);
912 mm->map_count--;
913 mpol_put(vma_policy(next));
914 kmem_cache_free(vm_area_cachep, next);
915 /*
916 * In mprotect's case 6 (see comments on vma_merge),
917 * we must remove another next too. It would clutter
918 * up the code too much to do both in one go.
919 */
920 if (remove_next != 3) {
921 /*
922 * If "next" was removed and vma->vm_end was
923 * expanded (up) over it, in turn
924 * "next->vm_prev->vm_end" changed and the
925 * "vma->vm_next" gap must be updated.
926 */
927 next = vma->vm_next;
928 } else {
929 /*
930 * For the scope of the comment "next" and
931 * "vma" considered pre-swap(): if "vma" was
932 * removed, next->vm_start was expanded (down)
933 * over it and the "next" gap must be updated.
934 * Because of the swap() the post-swap() "vma"
935 * actually points to pre-swap() "next"
936 * (post-swap() "next" as opposed is now a
937 * dangling pointer).
938 */
939 next = vma;
940 }
941 if (remove_next == 2) {
942 remove_next = 1;
943 end = next->vm_end;
944 goto again;
945 }
946 else if (next)
947 vma_gap_update(next);
948 else {
949 /*
950 * If remove_next == 2 we obviously can't
951 * reach this path.
952 *
953 * If remove_next == 3 we can't reach this
954 * path because pre-swap() next is always not
955 * NULL. pre-swap() "next" is not being
956 * removed and its next->vm_end is not altered
957 * (and furthermore "end" already matches
958 * next->vm_end in remove_next == 3).
959 *
960 * We reach this only in the remove_next == 1
961 * case if the "next" vma that was removed was
962 * the highest vma of the mm. However in such
963 * case next->vm_end == "end" and the extended
964 * "vma" has vma->vm_end == next->vm_end so
965 * mm->highest_vm_end doesn't need any update
966 * in remove_next == 1 case.
967 */
968 VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
969 }
970 }
971 if (insert && file)
972 uprobe_mmap(insert);
973
974 validate_mm(mm);
975
976 return 0;
977}
978
979/*
980 * If the vma has a ->close operation then the driver probably needs to release
981 * per-vma resources, so we don't attempt to merge those.
982 */
983static inline int is_mergeable_vma(struct vm_area_struct *vma,
984 struct file *file, unsigned long vm_flags,
985 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
986{
987 /*
988 * VM_SOFTDIRTY should not prevent from VMA merging, if we
989 * match the flags but dirty bit -- the caller should mark
990 * merged VMA as dirty. If dirty bit won't be excluded from
991 * comparison, we increase pressue on the memory system forcing
992 * the kernel to generate new VMAs when old one could be
993 * extended instead.
994 */
995 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
996 return 0;
997 if (vma->vm_file != file)
998 return 0;
999 if (vma->vm_ops && vma->vm_ops->close)
1000 return 0;
1001 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
1002 return 0;
1003 return 1;
1004}
1005
1006static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
1007 struct anon_vma *anon_vma2,
1008 struct vm_area_struct *vma)
1009{
1010 /*
1011 * The list_is_singular() test is to avoid merging VMA cloned from
1012 * parents. This can improve scalability caused by anon_vma lock.
1013 */
1014 if ((!anon_vma1 || !anon_vma2) && (!vma ||
1015 list_is_singular(&vma->anon_vma_chain)))
1016 return 1;
1017 return anon_vma1 == anon_vma2;
1018}
1019
1020/*
1021 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1022 * in front of (at a lower virtual address and file offset than) the vma.
1023 *
1024 * We cannot merge two vmas if they have differently assigned (non-NULL)
1025 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1026 *
1027 * We don't check here for the merged mmap wrapping around the end of pagecache
1028 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
1029 * wrap, nor mmaps which cover the final page at index -1UL.
1030 */
1031static int
1032can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1033 struct anon_vma *anon_vma, struct file *file,
1034 pgoff_t vm_pgoff,
1035 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1036{
1037 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1038 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1039 if (vma->vm_pgoff == vm_pgoff)
1040 return 1;
1041 }
1042 return 0;
1043}
1044
1045/*
1046 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1047 * beyond (at a higher virtual address and file offset than) the vma.
1048 *
1049 * We cannot merge two vmas if they have differently assigned (non-NULL)
1050 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1051 */
1052static int
1053can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1054 struct anon_vma *anon_vma, struct file *file,
1055 pgoff_t vm_pgoff,
1056 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1057{
1058 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1059 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1060 pgoff_t vm_pglen;
1061 vm_pglen = vma_pages(vma);
1062 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1063 return 1;
1064 }
1065 return 0;
1066}
1067
1068/*
1069 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1070 * whether that can be merged with its predecessor or its successor.
1071 * Or both (it neatly fills a hole).
1072 *
1073 * In most cases - when called for mmap, brk or mremap - [addr,end) is
1074 * certain not to be mapped by the time vma_merge is called; but when
1075 * called for mprotect, it is certain to be already mapped (either at
1076 * an offset within prev, or at the start of next), and the flags of
1077 * this area are about to be changed to vm_flags - and the no-change
1078 * case has already been eliminated.
1079 *
1080 * The following mprotect cases have to be considered, where AAAA is
1081 * the area passed down from mprotect_fixup, never extending beyond one
1082 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1083 *
1084 * AAAA AAAA AAAA AAAA
1085 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
1086 * cannot merge might become might become might become
1087 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
1088 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
1089 * mremap move: PPPPXXXXXXXX 8
1090 * AAAA
1091 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
1092 * might become case 1 below case 2 below case 3 below
1093 *
1094 * It is important for case 8 that the the vma NNNN overlapping the
1095 * region AAAA is never going to extended over XXXX. Instead XXXX must
1096 * be extended in region AAAA and NNNN must be removed. This way in
1097 * all cases where vma_merge succeeds, the moment vma_adjust drops the
1098 * rmap_locks, the properties of the merged vma will be already
1099 * correct for the whole merged range. Some of those properties like
1100 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1101 * be correct for the whole merged range immediately after the
1102 * rmap_locks are released. Otherwise if XXXX would be removed and
1103 * NNNN would be extended over the XXXX range, remove_migration_ptes
1104 * or other rmap walkers (if working on addresses beyond the "end"
1105 * parameter) may establish ptes with the wrong permissions of NNNN
1106 * instead of the right permissions of XXXX.
1107 */
1108struct vm_area_struct *vma_merge(struct mm_struct *mm,
1109 struct vm_area_struct *prev, unsigned long addr,
1110 unsigned long end, unsigned long vm_flags,
1111 struct anon_vma *anon_vma, struct file *file,
1112 pgoff_t pgoff, struct mempolicy *policy,
1113 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1114{
1115 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1116 struct vm_area_struct *area, *next;
1117 int err;
1118
1119 /*
1120 * We later require that vma->vm_flags == vm_flags,
1121 * so this tests vma->vm_flags & VM_SPECIAL, too.
1122 */
1123 if (vm_flags & VM_SPECIAL)
1124 return NULL;
1125
1126 if (prev)
1127 next = prev->vm_next;
1128 else
1129 next = mm->mmap;
1130 area = next;
1131 if (area && area->vm_end == end) /* cases 6, 7, 8 */
1132 next = next->vm_next;
1133
1134 /* verify some invariant that must be enforced by the caller */
1135 VM_WARN_ON(prev && addr <= prev->vm_start);
1136 VM_WARN_ON(area && end > area->vm_end);
1137 VM_WARN_ON(addr >= end);
1138
1139 /*
1140 * Can it merge with the predecessor?
1141 */
1142 if (prev && prev->vm_end == addr &&
1143 mpol_equal(vma_policy(prev), policy) &&
1144 can_vma_merge_after(prev, vm_flags,
1145 anon_vma, file, pgoff,
1146 vm_userfaultfd_ctx)) {
1147 /*
1148 * OK, it can. Can we now merge in the successor as well?
1149 */
1150 if (next && end == next->vm_start &&
1151 mpol_equal(policy, vma_policy(next)) &&
1152 can_vma_merge_before(next, vm_flags,
1153 anon_vma, file,
1154 pgoff+pglen,
1155 vm_userfaultfd_ctx) &&
1156 is_mergeable_anon_vma(prev->anon_vma,
1157 next->anon_vma, NULL)) {
1158 /* cases 1, 6 */
1159 err = __vma_adjust(prev, prev->vm_start,
1160 next->vm_end, prev->vm_pgoff, NULL,
1161 prev);
1162 } else /* cases 2, 5, 7 */
1163 err = __vma_adjust(prev, prev->vm_start,
1164 end, prev->vm_pgoff, NULL, prev);
1165 if (err)
1166 return NULL;
1167 khugepaged_enter_vma_merge(prev, vm_flags);
1168 return prev;
1169 }
1170
1171 /*
1172 * Can this new request be merged in front of next?
1173 */
1174 if (next && end == next->vm_start &&
1175 mpol_equal(policy, vma_policy(next)) &&
1176 can_vma_merge_before(next, vm_flags,
1177 anon_vma, file, pgoff+pglen,
1178 vm_userfaultfd_ctx)) {
1179 if (prev && addr < prev->vm_end) /* case 4 */
1180 err = __vma_adjust(prev, prev->vm_start,
1181 addr, prev->vm_pgoff, NULL, next);
1182 else { /* cases 3, 8 */
1183 err = __vma_adjust(area, addr, next->vm_end,
1184 next->vm_pgoff - pglen, NULL, next);
1185 /*
1186 * In case 3 area is already equal to next and
1187 * this is a noop, but in case 8 "area" has
1188 * been removed and next was expanded over it.
1189 */
1190 area = next;
1191 }
1192 if (err)
1193 return NULL;
1194 khugepaged_enter_vma_merge(area, vm_flags);
1195 return area;
1196 }
1197
1198 return NULL;
1199}
1200
1201/*
1202 * Rough compatbility check to quickly see if it's even worth looking
1203 * at sharing an anon_vma.
1204 *
1205 * They need to have the same vm_file, and the flags can only differ
1206 * in things that mprotect may change.
1207 *
1208 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1209 * we can merge the two vma's. For example, we refuse to merge a vma if
1210 * there is a vm_ops->close() function, because that indicates that the
1211 * driver is doing some kind of reference counting. But that doesn't
1212 * really matter for the anon_vma sharing case.
1213 */
1214static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1215{
1216 return a->vm_end == b->vm_start &&
1217 mpol_equal(vma_policy(a), vma_policy(b)) &&
1218 a->vm_file == b->vm_file &&
1219 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) &&
1220 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1221}
1222
1223/*
1224 * Do some basic sanity checking to see if we can re-use the anon_vma
1225 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1226 * the same as 'old', the other will be the new one that is trying
1227 * to share the anon_vma.
1228 *
1229 * NOTE! This runs with mm_sem held for reading, so it is possible that
1230 * the anon_vma of 'old' is concurrently in the process of being set up
1231 * by another page fault trying to merge _that_. But that's ok: if it
1232 * is being set up, that automatically means that it will be a singleton
1233 * acceptable for merging, so we can do all of this optimistically. But
1234 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1235 *
1236 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1237 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1238 * is to return an anon_vma that is "complex" due to having gone through
1239 * a fork).
1240 *
1241 * We also make sure that the two vma's are compatible (adjacent,
1242 * and with the same memory policies). That's all stable, even with just
1243 * a read lock on the mm_sem.
1244 */
1245static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1246{
1247 if (anon_vma_compatible(a, b)) {
1248 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1249
1250 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1251 return anon_vma;
1252 }
1253 return NULL;
1254}
1255
1256/*
1257 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1258 * neighbouring vmas for a suitable anon_vma, before it goes off
1259 * to allocate a new anon_vma. It checks because a repetitive
1260 * sequence of mprotects and faults may otherwise lead to distinct
1261 * anon_vmas being allocated, preventing vma merge in subsequent
1262 * mprotect.
1263 */
1264struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1265{
1266 struct anon_vma *anon_vma;
1267 struct vm_area_struct *near;
1268
1269 near = vma->vm_next;
1270 if (!near)
1271 goto try_prev;
1272
1273 anon_vma = reusable_anon_vma(near, vma, near);
1274 if (anon_vma)
1275 return anon_vma;
1276try_prev:
1277 near = vma->vm_prev;
1278 if (!near)
1279 goto none;
1280
1281 anon_vma = reusable_anon_vma(near, near, vma);
1282 if (anon_vma)
1283 return anon_vma;
1284none:
1285 /*
1286 * There's no absolute need to look only at touching neighbours:
1287 * we could search further afield for "compatible" anon_vmas.
1288 * But it would probably just be a waste of time searching,
1289 * or lead to too many vmas hanging off the same anon_vma.
1290 * We're trying to allow mprotect remerging later on,
1291 * not trying to minimize memory used for anon_vmas.
1292 */
1293 return NULL;
1294}
1295
1296/*
1297 * If a hint addr is less than mmap_min_addr change hint to be as
1298 * low as possible but still greater than mmap_min_addr
1299 */
1300static inline unsigned long round_hint_to_min(unsigned long hint)
1301{
1302 hint &= PAGE_MASK;
1303 if (((void *)hint != NULL) &&
1304 (hint < mmap_min_addr))
1305 return PAGE_ALIGN(mmap_min_addr);
1306 return hint;
1307}
1308
1309static inline int mlock_future_check(struct mm_struct *mm,
1310 unsigned long flags,
1311 unsigned long len)
1312{
1313 unsigned long locked, lock_limit;
1314
1315 /* mlock MCL_FUTURE? */
1316 if (flags & VM_LOCKED) {
1317 locked = len >> PAGE_SHIFT;
1318 locked += mm->locked_vm;
1319 lock_limit = rlimit(RLIMIT_MEMLOCK);
1320 lock_limit >>= PAGE_SHIFT;
1321 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1322 return -EAGAIN;
1323 }
1324 return 0;
1325}
1326
1327static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1328{
1329 if (S_ISREG(inode->i_mode))
1330 return MAX_LFS_FILESIZE;
1331
1332 if (S_ISBLK(inode->i_mode))
1333 return MAX_LFS_FILESIZE;
1334
1335 /* Special "we do even unsigned file positions" case */
1336 if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1337 return 0;
1338
1339 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1340 return ULONG_MAX;
1341}
1342
1343static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1344 unsigned long pgoff, unsigned long len)
1345{
1346 u64 maxsize = file_mmap_size_max(file, inode);
1347
1348 if (maxsize && len > maxsize)
1349 return false;
1350 maxsize -= len;
1351 if (pgoff > maxsize >> PAGE_SHIFT)
1352 return false;
1353 return true;
1354}
1355
1356/*
1357 * The caller must hold down_write(¤t->mm->mmap_sem).
1358 */
1359unsigned long do_mmap(struct file *file, unsigned long addr,
1360 unsigned long len, unsigned long prot,
1361 unsigned long flags, vm_flags_t vm_flags,
1362 unsigned long pgoff, unsigned long *populate,
1363 struct list_head *uf)
1364{
1365 struct mm_struct *mm = current->mm;
1366 int pkey = 0;
1367
1368 *populate = 0;
1369
1370 if (!len)
1371 return -EINVAL;
1372
1373 /*
1374 * Does the application expect PROT_READ to imply PROT_EXEC?
1375 *
1376 * (the exception is when the underlying filesystem is noexec
1377 * mounted, in which case we dont add PROT_EXEC.)
1378 */
1379 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1380 if (!(file && path_noexec(&file->f_path)))
1381 prot |= PROT_EXEC;
1382
1383 /* force arch specific MAP_FIXED handling in get_unmapped_area */
1384 if (flags & MAP_FIXED_NOREPLACE)
1385 flags |= MAP_FIXED;
1386
1387 if (!(flags & MAP_FIXED))
1388 addr = round_hint_to_min(addr);
1389
1390 /* Careful about overflows.. */
1391 len = PAGE_ALIGN(len);
1392 if (!len)
1393 return -ENOMEM;
1394
1395 /* offset overflow? */
1396 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1397 return -EOVERFLOW;
1398
1399 /* Too many mappings? */
1400 if (mm->map_count > sysctl_max_map_count)
1401 return -ENOMEM;
1402
1403 /* Obtain the address to map to. we verify (or select) it and ensure
1404 * that it represents a valid section of the address space.
1405 */
1406 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1407 if (offset_in_page(addr))
1408 return addr;
1409
1410 if (flags & MAP_FIXED_NOREPLACE) {
1411 struct vm_area_struct *vma = find_vma(mm, addr);
1412
1413 if (vma && vma->vm_start <= addr)
1414 return -EEXIST;
1415 }
1416
1417 if (prot == PROT_EXEC) {
1418 pkey = execute_only_pkey(mm);
1419 if (pkey < 0)
1420 pkey = 0;
1421 }
1422
1423 /* Do simple checking here so the lower-level routines won't have
1424 * to. we assume access permissions have been handled by the open
1425 * of the memory object, so we don't do any here.
1426 */
1427 vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1428 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1429
1430 if (flags & MAP_LOCKED)
1431 if (!can_do_mlock())
1432 return -EPERM;
1433
1434 if (mlock_future_check(mm, vm_flags, len))
1435 return -EAGAIN;
1436
1437 if (file) {
1438 struct inode *inode = file_inode(file);
1439 unsigned long flags_mask;
1440
1441 if (!file_mmap_ok(file, inode, pgoff, len))
1442 return -EOVERFLOW;
1443
1444 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1445
1446 switch (flags & MAP_TYPE) {
1447 case MAP_SHARED:
1448 /*
1449 * Force use of MAP_SHARED_VALIDATE with non-legacy
1450 * flags. E.g. MAP_SYNC is dangerous to use with
1451 * MAP_SHARED as you don't know which consistency model
1452 * you will get. We silently ignore unsupported flags
1453 * with MAP_SHARED to preserve backward compatibility.
1454 */
1455 flags &= LEGACY_MAP_MASK;
1456 /* fall through */
1457 case MAP_SHARED_VALIDATE:
1458 if (flags & ~flags_mask)
1459 return -EOPNOTSUPP;
1460 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1461 return -EACCES;
1462
1463 /*
1464 * Make sure we don't allow writing to an append-only
1465 * file..
1466 */
1467 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1468 return -EACCES;
1469
1470 /*
1471 * Make sure there are no mandatory locks on the file.
1472 */
1473 if (locks_verify_locked(file))
1474 return -EAGAIN;
1475
1476 vm_flags |= VM_SHARED | VM_MAYSHARE;
1477 if (!(file->f_mode & FMODE_WRITE))
1478 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1479
1480 /* fall through */
1481 case MAP_PRIVATE:
1482 if (!(file->f_mode & FMODE_READ))
1483 return -EACCES;
1484 if (path_noexec(&file->f_path)) {
1485 if (vm_flags & VM_EXEC)
1486 return -EPERM;
1487 vm_flags &= ~VM_MAYEXEC;
1488 }
1489
1490 if (!file->f_op->mmap)
1491 return -ENODEV;
1492 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1493 return -EINVAL;
1494 break;
1495
1496 default:
1497 return -EINVAL;
1498 }
1499 } else {
1500 switch (flags & MAP_TYPE) {
1501 case MAP_SHARED:
1502 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1503 return -EINVAL;
1504 /*
1505 * Ignore pgoff.
1506 */
1507 pgoff = 0;
1508 vm_flags |= VM_SHARED | VM_MAYSHARE;
1509 break;
1510 case MAP_PRIVATE:
1511 /*
1512 * Set pgoff according to addr for anon_vma.
1513 */
1514 pgoff = addr >> PAGE_SHIFT;
1515 break;
1516 default:
1517 return -EINVAL;
1518 }
1519 }
1520
1521 /*
1522 * Set 'VM_NORESERVE' if we should not account for the
1523 * memory use of this mapping.
1524 */
1525 if (flags & MAP_NORESERVE) {
1526 /* We honor MAP_NORESERVE if allowed to overcommit */
1527 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1528 vm_flags |= VM_NORESERVE;
1529
1530 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1531 if (file && is_file_hugepages(file))
1532 vm_flags |= VM_NORESERVE;
1533 }
1534
1535 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1536 if (!IS_ERR_VALUE(addr) &&
1537 ((vm_flags & VM_LOCKED) ||
1538 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1539 *populate = len;
1540 return addr;
1541}
1542
1543unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1544 unsigned long prot, unsigned long flags,
1545 unsigned long fd, unsigned long pgoff)
1546{
1547 struct file *file = NULL;
1548 unsigned long retval;
1549
1550 if (!(flags & MAP_ANONYMOUS)) {
1551 audit_mmap_fd(fd, flags);
1552 file = fget(fd);
1553 if (!file)
1554 return -EBADF;
1555 if (is_file_hugepages(file))
1556 len = ALIGN(len, huge_page_size(hstate_file(file)));
1557 retval = -EINVAL;
1558 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1559 goto out_fput;
1560 } else if (flags & MAP_HUGETLB) {
1561 struct user_struct *user = NULL;
1562 struct hstate *hs;
1563
1564 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1565 if (!hs)
1566 return -EINVAL;
1567
1568 len = ALIGN(len, huge_page_size(hs));
1569 /*
1570 * VM_NORESERVE is used because the reservations will be
1571 * taken when vm_ops->mmap() is called
1572 * A dummy user value is used because we are not locking
1573 * memory so no accounting is necessary
1574 */
1575 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1576 VM_NORESERVE,
1577 &user, HUGETLB_ANONHUGE_INODE,
1578 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1579 if (IS_ERR(file))
1580 return PTR_ERR(file);
1581 }
1582
1583 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1584
1585 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1586out_fput:
1587 if (file)
1588 fput(file);
1589 return retval;
1590}
1591
1592SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1593 unsigned long, prot, unsigned long, flags,
1594 unsigned long, fd, unsigned long, pgoff)
1595{
1596 return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1597}
1598
1599#ifdef __ARCH_WANT_SYS_OLD_MMAP
1600struct mmap_arg_struct {
1601 unsigned long addr;
1602 unsigned long len;
1603 unsigned long prot;
1604 unsigned long flags;
1605 unsigned long fd;
1606 unsigned long offset;
1607};
1608
1609SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1610{
1611 struct mmap_arg_struct a;
1612
1613 if (copy_from_user(&a, arg, sizeof(a)))
1614 return -EFAULT;
1615 if (offset_in_page(a.offset))
1616 return -EINVAL;
1617
1618 return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1619 a.offset >> PAGE_SHIFT);
1620}
1621#endif /* __ARCH_WANT_SYS_OLD_MMAP */
1622
1623/*
1624 * Some shared mappigns will want the pages marked read-only
1625 * to track write events. If so, we'll downgrade vm_page_prot
1626 * to the private version (using protection_map[] without the
1627 * VM_SHARED bit).
1628 */
1629int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1630{
1631 vm_flags_t vm_flags = vma->vm_flags;
1632 const struct vm_operations_struct *vm_ops = vma->vm_ops;
1633
1634 /* If it was private or non-writable, the write bit is already clear */
1635 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1636 return 0;
1637
1638 /* The backer wishes to know when pages are first written to? */
1639 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1640 return 1;
1641
1642 /* The open routine did something to the protections that pgprot_modify
1643 * won't preserve? */
1644 if (pgprot_val(vm_page_prot) !=
1645 pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1646 return 0;
1647
1648 /* Do we need to track softdirty? */
1649 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1650 return 1;
1651
1652 /* Specialty mapping? */
1653 if (vm_flags & VM_PFNMAP)
1654 return 0;
1655
1656 /* Can the mapping track the dirty pages? */
1657 return vma->vm_file && vma->vm_file->f_mapping &&
1658 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1659}
1660
1661/*
1662 * We account for memory if it's a private writeable mapping,
1663 * not hugepages and VM_NORESERVE wasn't set.
1664 */
1665static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1666{
1667 /*
1668 * hugetlb has its own accounting separate from the core VM
1669 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1670 */
1671 if (file && is_file_hugepages(file))
1672 return 0;
1673
1674 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1675}
1676
1677unsigned long mmap_region(struct file *file, unsigned long addr,
1678 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
1679 struct list_head *uf)
1680{
1681 struct mm_struct *mm = current->mm;
1682 struct vm_area_struct *vma, *prev;
1683 int error;
1684 struct rb_node **rb_link, *rb_parent;
1685 unsigned long charged = 0;
1686
1687 /* Check against address space limit. */
1688 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1689 unsigned long nr_pages;
1690
1691 /*
1692 * MAP_FIXED may remove pages of mappings that intersects with
1693 * requested mapping. Account for the pages it would unmap.
1694 */
1695 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1696
1697 if (!may_expand_vm(mm, vm_flags,
1698 (len >> PAGE_SHIFT) - nr_pages))
1699 return -ENOMEM;
1700 }
1701
1702 /* Clear old maps */
1703 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
1704 &rb_parent)) {
1705 if (do_munmap(mm, addr, len, uf))
1706 return -ENOMEM;
1707 }
1708
1709 /*
1710 * Private writable mapping: check memory availability
1711 */
1712 if (accountable_mapping(file, vm_flags)) {
1713 charged = len >> PAGE_SHIFT;
1714 if (security_vm_enough_memory_mm(mm, charged))
1715 return -ENOMEM;
1716 vm_flags |= VM_ACCOUNT;
1717 }
1718
1719 /*
1720 * Can we just expand an old mapping?
1721 */
1722 vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1723 NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1724 if (vma)
1725 goto out;
1726
1727 /*
1728 * Determine the object being mapped and call the appropriate
1729 * specific mapper. the address has already been validated, but
1730 * not unmapped, but the maps are removed from the list.
1731 */
1732 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1733 if (!vma) {
1734 error = -ENOMEM;
1735 goto unacct_error;
1736 }
1737
1738 vma->vm_mm = mm;
1739 vma->vm_start = addr;
1740 vma->vm_end = addr + len;
1741 vma->vm_flags = vm_flags;
1742 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1743 vma->vm_pgoff = pgoff;
1744 INIT_LIST_HEAD(&vma->anon_vma_chain);
1745
1746 if (file) {
1747 if (vm_flags & VM_DENYWRITE) {
1748 error = deny_write_access(file);
1749 if (error)
1750 goto free_vma;
1751 }
1752 if (vm_flags & VM_SHARED) {
1753 error = mapping_map_writable(file->f_mapping);
1754 if (error)
1755 goto allow_write_and_free_vma;
1756 }
1757
1758 /* ->mmap() can change vma->vm_file, but must guarantee that
1759 * vma_link() below can deny write-access if VM_DENYWRITE is set
1760 * and map writably if VM_SHARED is set. This usually means the
1761 * new file must not have been exposed to user-space, yet.
1762 */
1763 vma->vm_file = get_file(file);
1764 error = call_mmap(file, vma);
1765 if (error)
1766 goto unmap_and_free_vma;
1767
1768 /* Can addr have changed??
1769 *
1770 * Answer: Yes, several device drivers can do it in their
1771 * f_op->mmap method. -DaveM
1772 * Bug: If addr is changed, prev, rb_link, rb_parent should
1773 * be updated for vma_link()
1774 */
1775 WARN_ON_ONCE(addr != vma->vm_start);
1776
1777 addr = vma->vm_start;
1778 vm_flags = vma->vm_flags;
1779 } else if (vm_flags & VM_SHARED) {
1780 error = shmem_zero_setup(vma);
1781 if (error)
1782 goto free_vma;
1783 }
1784
1785 vma_link(mm, vma, prev, rb_link, rb_parent);
1786 /* Once vma denies write, undo our temporary denial count */
1787 if (file) {
1788 if (vm_flags & VM_SHARED)
1789 mapping_unmap_writable(file->f_mapping);
1790 if (vm_flags & VM_DENYWRITE)
1791 allow_write_access(file);
1792 }
1793 file = vma->vm_file;
1794out:
1795 perf_event_mmap(vma);
1796
1797 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1798 if (vm_flags & VM_LOCKED) {
1799 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1800 vma == get_gate_vma(current->mm)))
1801 mm->locked_vm += (len >> PAGE_SHIFT);
1802 else
1803 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1804 }
1805
1806 if (file)
1807 uprobe_mmap(vma);
1808
1809 /*
1810 * New (or expanded) vma always get soft dirty status.
1811 * Otherwise user-space soft-dirty page tracker won't
1812 * be able to distinguish situation when vma area unmapped,
1813 * then new mapped in-place (which must be aimed as
1814 * a completely new data area).
1815 */
1816 vma->vm_flags |= VM_SOFTDIRTY;
1817
1818 vma_set_page_prot(vma);
1819
1820 return addr;
1821
1822unmap_and_free_vma:
1823 vma->vm_file = NULL;
1824 fput(file);
1825
1826 /* Undo any partial mapping done by a device driver. */
1827 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1828 charged = 0;
1829 if (vm_flags & VM_SHARED)
1830 mapping_unmap_writable(file->f_mapping);
1831allow_write_and_free_vma:
1832 if (vm_flags & VM_DENYWRITE)
1833 allow_write_access(file);
1834free_vma:
1835 kmem_cache_free(vm_area_cachep, vma);
1836unacct_error:
1837 if (charged)
1838 vm_unacct_memory(charged);
1839 return error;
1840}
1841
1842unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1843{
1844 /*
1845 * We implement the search by looking for an rbtree node that
1846 * immediately follows a suitable gap. That is,
1847 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1848 * - gap_end = vma->vm_start >= info->low_limit + length;
1849 * - gap_end - gap_start >= length
1850 */
1851
1852 struct mm_struct *mm = current->mm;
1853 struct vm_area_struct *vma;
1854 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1855
1856 /* Adjust search length to account for worst case alignment overhead */
1857 length = info->length + info->align_mask;
1858 if (length < info->length)
1859 return -ENOMEM;
1860
1861 /* Adjust search limits by the desired length */
1862 if (info->high_limit < length)
1863 return -ENOMEM;
1864 high_limit = info->high_limit - length;
1865
1866 if (info->low_limit > high_limit)
1867 return -ENOMEM;
1868 low_limit = info->low_limit + length;
1869
1870 /* Check if rbtree root looks promising */
1871 if (RB_EMPTY_ROOT(&mm->mm_rb))
1872 goto check_highest;
1873 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1874 if (vma->rb_subtree_gap < length)
1875 goto check_highest;
1876
1877 while (true) {
1878 /* Visit left subtree if it looks promising */
1879 gap_end = vm_start_gap(vma);
1880 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1881 struct vm_area_struct *left =
1882 rb_entry(vma->vm_rb.rb_left,
1883 struct vm_area_struct, vm_rb);
1884 if (left->rb_subtree_gap >= length) {
1885 vma = left;
1886 continue;
1887 }
1888 }
1889
1890 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1891check_current:
1892 /* Check if current node has a suitable gap */
1893 if (gap_start > high_limit)
1894 return -ENOMEM;
1895 if (gap_end >= low_limit &&
1896 gap_end > gap_start && gap_end - gap_start >= length)
1897 goto found;
1898
1899 /* Visit right subtree if it looks promising */
1900 if (vma->vm_rb.rb_right) {
1901 struct vm_area_struct *right =
1902 rb_entry(vma->vm_rb.rb_right,
1903 struct vm_area_struct, vm_rb);
1904 if (right->rb_subtree_gap >= length) {
1905 vma = right;
1906 continue;
1907 }
1908 }
1909
1910 /* Go back up the rbtree to find next candidate node */
1911 while (true) {
1912 struct rb_node *prev = &vma->vm_rb;
1913 if (!rb_parent(prev))
1914 goto check_highest;
1915 vma = rb_entry(rb_parent(prev),
1916 struct vm_area_struct, vm_rb);
1917 if (prev == vma->vm_rb.rb_left) {
1918 gap_start = vm_end_gap(vma->vm_prev);
1919 gap_end = vm_start_gap(vma);
1920 goto check_current;
1921 }
1922 }
1923 }
1924
1925check_highest:
1926 /* Check highest gap, which does not precede any rbtree node */
1927 gap_start = mm->highest_vm_end;
1928 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */
1929 if (gap_start > high_limit)
1930 return -ENOMEM;
1931
1932found:
1933 /* We found a suitable gap. Clip it with the original low_limit. */
1934 if (gap_start < info->low_limit)
1935 gap_start = info->low_limit;
1936
1937 /* Adjust gap address to the desired alignment */
1938 gap_start += (info->align_offset - gap_start) & info->align_mask;
1939
1940 VM_BUG_ON(gap_start + info->length > info->high_limit);
1941 VM_BUG_ON(gap_start + info->length > gap_end);
1942 return gap_start;
1943}
1944
1945unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1946{
1947 struct mm_struct *mm = current->mm;
1948 struct vm_area_struct *vma;
1949 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1950
1951 /* Adjust search length to account for worst case alignment overhead */
1952 length = info->length + info->align_mask;
1953 if (length < info->length)
1954 return -ENOMEM;
1955
1956 /*
1957 * Adjust search limits by the desired length.
1958 * See implementation comment at top of unmapped_area().
1959 */
1960 gap_end = info->high_limit;
1961 if (gap_end < length)
1962 return -ENOMEM;
1963 high_limit = gap_end - length;
1964
1965 if (info->low_limit > high_limit)
1966 return -ENOMEM;
1967 low_limit = info->low_limit + length;
1968
1969 /* Check highest gap, which does not precede any rbtree node */
1970 gap_start = mm->highest_vm_end;
1971 if (gap_start <= high_limit)
1972 goto found_highest;
1973
1974 /* Check if rbtree root looks promising */
1975 if (RB_EMPTY_ROOT(&mm->mm_rb))
1976 return -ENOMEM;
1977 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1978 if (vma->rb_subtree_gap < length)
1979 return -ENOMEM;
1980
1981 while (true) {
1982 /* Visit right subtree if it looks promising */
1983 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1984 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1985 struct vm_area_struct *right =
1986 rb_entry(vma->vm_rb.rb_right,
1987 struct vm_area_struct, vm_rb);
1988 if (right->rb_subtree_gap >= length) {
1989 vma = right;
1990 continue;
1991 }
1992 }
1993
1994check_current:
1995 /* Check if current node has a suitable gap */
1996 gap_end = vm_start_gap(vma);
1997 if (gap_end < low_limit)
1998 return -ENOMEM;
1999 if (gap_start <= high_limit &&
2000 gap_end > gap_start && gap_end - gap_start >= length)
2001 goto found;
2002
2003 /* Visit left subtree if it looks promising */
2004 if (vma->vm_rb.rb_left) {
2005 struct vm_area_struct *left =
2006 rb_entry(vma->vm_rb.rb_left,
2007 struct vm_area_struct, vm_rb);
2008 if (left->rb_subtree_gap >= length) {
2009 vma = left;
2010 continue;
2011 }
2012 }
2013
2014 /* Go back up the rbtree to find next candidate node */
2015 while (true) {
2016 struct rb_node *prev = &vma->vm_rb;
2017 if (!rb_parent(prev))
2018 return -ENOMEM;
2019 vma = rb_entry(rb_parent(prev),
2020 struct vm_area_struct, vm_rb);
2021 if (prev == vma->vm_rb.rb_right) {
2022 gap_start = vma->vm_prev ?
2023 vm_end_gap(vma->vm_prev) : 0;
2024 goto check_current;
2025 }
2026 }
2027 }
2028
2029found:
2030 /* We found a suitable gap. Clip it with the original high_limit. */
2031 if (gap_end > info->high_limit)
2032 gap_end = info->high_limit;
2033
2034found_highest:
2035 /* Compute highest gap address at the desired alignment */
2036 gap_end -= info->length;
2037 gap_end -= (gap_end - info->align_offset) & info->align_mask;
2038
2039 VM_BUG_ON(gap_end < info->low_limit);
2040 VM_BUG_ON(gap_end < gap_start);
2041 return gap_end;
2042}
2043
2044/* Get an address range which is currently unmapped.
2045 * For shmat() with addr=0.
2046 *
2047 * Ugly calling convention alert:
2048 * Return value with the low bits set means error value,
2049 * ie
2050 * if (ret & ~PAGE_MASK)
2051 * error = ret;
2052 *
2053 * This function "knows" that -ENOMEM has the bits set.
2054 */
2055#ifndef HAVE_ARCH_UNMAPPED_AREA
2056unsigned long
2057arch_get_unmapped_area(struct file *filp, unsigned long addr,
2058 unsigned long len, unsigned long pgoff, unsigned long flags)
2059{
2060 struct mm_struct *mm = current->mm;
2061 struct vm_area_struct *vma, *prev;
2062 struct vm_unmapped_area_info info;
2063
2064 if (len > TASK_SIZE - mmap_min_addr)
2065 return -ENOMEM;
2066
2067 if (flags & MAP_FIXED)
2068 return addr;
2069
2070 if (addr) {
2071 addr = PAGE_ALIGN(addr);
2072 vma = find_vma_prev(mm, addr, &prev);
2073 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
2074 (!vma || addr + len <= vm_start_gap(vma)) &&
2075 (!prev || addr >= vm_end_gap(prev)))
2076 return addr;
2077 }
2078
2079 info.flags = 0;
2080 info.length = len;
2081 info.low_limit = mm->mmap_base;
2082 info.high_limit = TASK_SIZE;
2083 info.align_mask = 0;
2084 return vm_unmapped_area(&info);
2085}
2086#endif
2087
2088/*
2089 * This mmap-allocator allocates new areas top-down from below the
2090 * stack's low limit (the base):
2091 */
2092#ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2093unsigned long
2094arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
2095 const unsigned long len, const unsigned long pgoff,
2096 const unsigned long flags)
2097{
2098 struct vm_area_struct *vma, *prev;
2099 struct mm_struct *mm = current->mm;
2100 unsigned long addr = addr0;
2101 struct vm_unmapped_area_info info;
2102
2103 /* requested length too big for entire address space */
2104 if (len > TASK_SIZE - mmap_min_addr)
2105 return -ENOMEM;
2106
2107 if (flags & MAP_FIXED)
2108 return addr;
2109
2110 /* requesting a specific address */
2111 if (addr) {
2112 addr = PAGE_ALIGN(addr);
2113 vma = find_vma_prev(mm, addr, &prev);
2114 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
2115 (!vma || addr + len <= vm_start_gap(vma)) &&
2116 (!prev || addr >= vm_end_gap(prev)))
2117 return addr;
2118 }
2119
2120 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2121 info.length = len;
2122 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2123 info.high_limit = mm->mmap_base;
2124 info.align_mask = 0;
2125 addr = vm_unmapped_area(&info);
2126
2127 /*
2128 * A failed mmap() very likely causes application failure,
2129 * so fall back to the bottom-up function here. This scenario
2130 * can happen with large stack limits and large mmap()
2131 * allocations.
2132 */
2133 if (offset_in_page(addr)) {
2134 VM_BUG_ON(addr != -ENOMEM);
2135 info.flags = 0;
2136 info.low_limit = TASK_UNMAPPED_BASE;
2137 info.high_limit = TASK_SIZE;
2138 addr = vm_unmapped_area(&info);
2139 }
2140
2141 return addr;
2142}
2143#endif
2144
2145unsigned long
2146get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2147 unsigned long pgoff, unsigned long flags)
2148{
2149 unsigned long (*get_area)(struct file *, unsigned long,
2150 unsigned long, unsigned long, unsigned long);
2151
2152 unsigned long error = arch_mmap_check(addr, len, flags);
2153 if (error)
2154 return error;
2155
2156 /* Careful about overflows.. */
2157 if (len > TASK_SIZE)
2158 return -ENOMEM;
2159
2160 get_area = current->mm->get_unmapped_area;
2161 if (file) {
2162 if (file->f_op->get_unmapped_area)
2163 get_area = file->f_op->get_unmapped_area;
2164 } else if (flags & MAP_SHARED) {
2165 /*
2166 * mmap_region() will call shmem_zero_setup() to create a file,
2167 * so use shmem's get_unmapped_area in case it can be huge.
2168 * do_mmap_pgoff() will clear pgoff, so match alignment.
2169 */
2170 pgoff = 0;
2171 get_area = shmem_get_unmapped_area;
2172 }
2173
2174 addr = get_area(file, addr, len, pgoff, flags);
2175 if (IS_ERR_VALUE(addr))
2176 return addr;
2177
2178 if (addr > TASK_SIZE - len)
2179 return -ENOMEM;
2180 if (offset_in_page(addr))
2181 return -EINVAL;
2182
2183 error = security_mmap_addr(addr);
2184 return error ? error : addr;
2185}
2186
2187EXPORT_SYMBOL(get_unmapped_area);
2188
2189/* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
2190struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2191{
2192 struct rb_node *rb_node;
2193 struct vm_area_struct *vma;
2194
2195 /* Check the cache first. */
2196 vma = vmacache_find(mm, addr);
2197 if (likely(vma))
2198 return vma;
2199
2200 rb_node = mm->mm_rb.rb_node;
2201
2202 while (rb_node) {
2203 struct vm_area_struct *tmp;
2204
2205 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2206
2207 if (tmp->vm_end > addr) {
2208 vma = tmp;
2209 if (tmp->vm_start <= addr)
2210 break;
2211 rb_node = rb_node->rb_left;
2212 } else
2213 rb_node = rb_node->rb_right;
2214 }
2215
2216 if (vma)
2217 vmacache_update(addr, vma);
2218 return vma;
2219}
2220
2221EXPORT_SYMBOL(find_vma);
2222
2223/*
2224 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2225 */
2226struct vm_area_struct *
2227find_vma_prev(struct mm_struct *mm, unsigned long addr,
2228 struct vm_area_struct **pprev)
2229{
2230 struct vm_area_struct *vma;
2231
2232 vma = find_vma(mm, addr);
2233 if (vma) {
2234 *pprev = vma->vm_prev;
2235 } else {
2236 struct rb_node *rb_node = mm->mm_rb.rb_node;
2237 *pprev = NULL;
2238 while (rb_node) {
2239 *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2240 rb_node = rb_node->rb_right;
2241 }
2242 }
2243 return vma;
2244}
2245
2246/*
2247 * Verify that the stack growth is acceptable and
2248 * update accounting. This is shared with both the
2249 * grow-up and grow-down cases.
2250 */
2251static int acct_stack_growth(struct vm_area_struct *vma,
2252 unsigned long size, unsigned long grow)
2253{
2254 struct mm_struct *mm = vma->vm_mm;
2255 unsigned long new_start;
2256
2257 /* address space limit tests */
2258 if (!may_expand_vm(mm, vma->vm_flags, grow))
2259 return -ENOMEM;
2260
2261 /* Stack limit test */
2262 if (size > rlimit(RLIMIT_STACK))
2263 return -ENOMEM;
2264
2265 /* mlock limit tests */
2266 if (vma->vm_flags & VM_LOCKED) {
2267 unsigned long locked;
2268 unsigned long limit;
2269 locked = mm->locked_vm + grow;
2270 limit = rlimit(RLIMIT_MEMLOCK);
2271 limit >>= PAGE_SHIFT;
2272 if (locked > limit && !capable(CAP_IPC_LOCK))
2273 return -ENOMEM;
2274 }
2275
2276 /* Check to ensure the stack will not grow into a hugetlb-only region */
2277 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2278 vma->vm_end - size;
2279 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2280 return -EFAULT;
2281
2282 /*
2283 * Overcommit.. This must be the final test, as it will
2284 * update security statistics.
2285 */
2286 if (security_vm_enough_memory_mm(mm, grow))
2287 return -ENOMEM;
2288
2289 return 0;
2290}
2291
2292#if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2293/*
2294 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2295 * vma is the last one with address > vma->vm_end. Have to extend vma.
2296 */
2297int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2298{
2299 struct mm_struct *mm = vma->vm_mm;
2300 struct vm_area_struct *next;
2301 unsigned long gap_addr;
2302 int error = 0;
2303
2304 if (!(vma->vm_flags & VM_GROWSUP))
2305 return -EFAULT;
2306
2307 /* Guard against exceeding limits of the address space. */
2308 address &= PAGE_MASK;
2309 if (address >= (TASK_SIZE & PAGE_MASK))
2310 return -ENOMEM;
2311 address += PAGE_SIZE;
2312
2313 /* Enforce stack_guard_gap */
2314 gap_addr = address + stack_guard_gap;
2315
2316 /* Guard against overflow */
2317 if (gap_addr < address || gap_addr > TASK_SIZE)
2318 gap_addr = TASK_SIZE;
2319
2320 next = vma->vm_next;
2321 if (next && next->vm_start < gap_addr &&
2322 (next->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2323 if (!(next->vm_flags & VM_GROWSUP))
2324 return -ENOMEM;
2325 /* Check that both stack segments have the same anon_vma? */
2326 }
2327
2328 /* We must make sure the anon_vma is allocated. */
2329 if (unlikely(anon_vma_prepare(vma)))
2330 return -ENOMEM;
2331
2332 /*
2333 * vma->vm_start/vm_end cannot change under us because the caller
2334 * is required to hold the mmap_sem in read mode. We need the
2335 * anon_vma lock to serialize against concurrent expand_stacks.
2336 */
2337 anon_vma_lock_write(vma->anon_vma);
2338
2339 /* Somebody else might have raced and expanded it already */
2340 if (address > vma->vm_end) {
2341 unsigned long size, grow;
2342
2343 size = address - vma->vm_start;
2344 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2345
2346 error = -ENOMEM;
2347 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2348 error = acct_stack_growth(vma, size, grow);
2349 if (!error) {
2350 /*
2351 * vma_gap_update() doesn't support concurrent
2352 * updates, but we only hold a shared mmap_sem
2353 * lock here, so we need to protect against
2354 * concurrent vma expansions.
2355 * anon_vma_lock_write() doesn't help here, as
2356 * we don't guarantee that all growable vmas
2357 * in a mm share the same root anon vma.
2358 * So, we reuse mm->page_table_lock to guard
2359 * against concurrent vma expansions.
2360 */
2361 spin_lock(&mm->page_table_lock);
2362 if (vma->vm_flags & VM_LOCKED)
2363 mm->locked_vm += grow;
2364 vm_stat_account(mm, vma->vm_flags, grow);
2365 anon_vma_interval_tree_pre_update_vma(vma);
2366 vma->vm_end = address;
2367 anon_vma_interval_tree_post_update_vma(vma);
2368 if (vma->vm_next)
2369 vma_gap_update(vma->vm_next);
2370 else
2371 mm->highest_vm_end = vm_end_gap(vma);
2372 spin_unlock(&mm->page_table_lock);
2373
2374 perf_event_mmap(vma);
2375 }
2376 }
2377 }
2378 anon_vma_unlock_write(vma->anon_vma);
2379 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2380 validate_mm(mm);
2381 return error;
2382}
2383#endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2384
2385/*
2386 * vma is the first one with address < vma->vm_start. Have to extend vma.
2387 */
2388int expand_downwards(struct vm_area_struct *vma,
2389 unsigned long address)
2390{
2391 struct mm_struct *mm = vma->vm_mm;
2392 struct vm_area_struct *prev;
2393 int error;
2394
2395 address &= PAGE_MASK;
2396 error = security_mmap_addr(address);
2397 if (error)
2398 return error;
2399
2400 /* Enforce stack_guard_gap */
2401 prev = vma->vm_prev;
2402 /* Check that both stack segments have the same anon_vma? */
2403 if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2404 (prev->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2405 if (address - prev->vm_end < stack_guard_gap)
2406 return -ENOMEM;
2407 }
2408
2409 /* We must make sure the anon_vma is allocated. */
2410 if (unlikely(anon_vma_prepare(vma)))
2411 return -ENOMEM;
2412
2413 /*
2414 * vma->vm_start/vm_end cannot change under us because the caller
2415 * is required to hold the mmap_sem in read mode. We need the
2416 * anon_vma lock to serialize against concurrent expand_stacks.
2417 */
2418 anon_vma_lock_write(vma->anon_vma);
2419
2420 /* Somebody else might have raced and expanded it already */
2421 if (address < vma->vm_start) {
2422 unsigned long size, grow;
2423
2424 size = vma->vm_end - address;
2425 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2426
2427 error = -ENOMEM;
2428 if (grow <= vma->vm_pgoff) {
2429 error = acct_stack_growth(vma, size, grow);
2430 if (!error) {
2431 /*
2432 * vma_gap_update() doesn't support concurrent
2433 * updates, but we only hold a shared mmap_sem
2434 * lock here, so we need to protect against
2435 * concurrent vma expansions.
2436 * anon_vma_lock_write() doesn't help here, as
2437 * we don't guarantee that all growable vmas
2438 * in a mm share the same root anon vma.
2439 * So, we reuse mm->page_table_lock to guard
2440 * against concurrent vma expansions.
2441 */
2442 spin_lock(&mm->page_table_lock);
2443 if (vma->vm_flags & VM_LOCKED)
2444 mm->locked_vm += grow;
2445 vm_stat_account(mm, vma->vm_flags, grow);
2446 anon_vma_interval_tree_pre_update_vma(vma);
2447 vma->vm_start = address;
2448 vma->vm_pgoff -= grow;
2449 anon_vma_interval_tree_post_update_vma(vma);
2450 vma_gap_update(vma);
2451 spin_unlock(&mm->page_table_lock);
2452
2453 perf_event_mmap(vma);
2454 }
2455 }
2456 }
2457 anon_vma_unlock_write(vma->anon_vma);
2458 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2459 validate_mm(mm);
2460 return error;
2461}
2462
2463/* enforced gap between the expanding stack and other mappings. */
2464unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2465
2466static int __init cmdline_parse_stack_guard_gap(char *p)
2467{
2468 unsigned long val;
2469 char *endptr;
2470
2471 val = simple_strtoul(p, &endptr, 10);
2472 if (!*endptr)
2473 stack_guard_gap = val << PAGE_SHIFT;
2474
2475 return 0;
2476}
2477__setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2478
2479#ifdef CONFIG_STACK_GROWSUP
2480int expand_stack(struct vm_area_struct *vma, unsigned long address)
2481{
2482 return expand_upwards(vma, address);
2483}
2484
2485struct vm_area_struct *
2486find_extend_vma(struct mm_struct *mm, unsigned long addr)
2487{
2488 struct vm_area_struct *vma, *prev;
2489
2490 addr &= PAGE_MASK;
2491 vma = find_vma_prev(mm, addr, &prev);
2492 if (vma && (vma->vm_start <= addr))
2493 return vma;
2494 if (!prev || expand_stack(prev, addr))
2495 return NULL;
2496 if (prev->vm_flags & VM_LOCKED)
2497 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2498 return prev;
2499}
2500#else
2501int expand_stack(struct vm_area_struct *vma, unsigned long address)
2502{
2503 return expand_downwards(vma, address);
2504}
2505
2506struct vm_area_struct *
2507find_extend_vma(struct mm_struct *mm, unsigned long addr)
2508{
2509 struct vm_area_struct *vma;
2510 unsigned long start;
2511
2512 addr &= PAGE_MASK;
2513 vma = find_vma(mm, addr);
2514 if (!vma)
2515 return NULL;
2516 if (vma->vm_start <= addr)
2517 return vma;
2518 if (!(vma->vm_flags & VM_GROWSDOWN))
2519 return NULL;
2520 start = vma->vm_start;
2521 if (expand_stack(vma, addr))
2522 return NULL;
2523 if (vma->vm_flags & VM_LOCKED)
2524 populate_vma_page_range(vma, addr, start, NULL);
2525 return vma;
2526}
2527#endif
2528
2529EXPORT_SYMBOL_GPL(find_extend_vma);
2530
2531/*
2532 * Ok - we have the memory areas we should free on the vma list,
2533 * so release them, and do the vma updates.
2534 *
2535 * Called with the mm semaphore held.
2536 */
2537static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2538{
2539 unsigned long nr_accounted = 0;
2540
2541 /* Update high watermark before we lower total_vm */
2542 update_hiwater_vm(mm);
2543 do {
2544 long nrpages = vma_pages(vma);
2545
2546 if (vma->vm_flags & VM_ACCOUNT)
2547 nr_accounted += nrpages;
2548 vm_stat_account(mm, vma->vm_flags, -nrpages);
2549 vma = remove_vma(vma);
2550 } while (vma);
2551 vm_unacct_memory(nr_accounted);
2552 validate_mm(mm);
2553}
2554
2555/*
2556 * Get rid of page table information in the indicated region.
2557 *
2558 * Called with the mm semaphore held.
2559 */
2560static void unmap_region(struct mm_struct *mm,
2561 struct vm_area_struct *vma, struct vm_area_struct *prev,
2562 unsigned long start, unsigned long end)
2563{
2564 struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap;
2565 struct mmu_gather tlb;
2566
2567 lru_add_drain();
2568 tlb_gather_mmu(&tlb, mm, start, end);
2569 update_hiwater_rss(mm);
2570 unmap_vmas(&tlb, vma, start, end);
2571 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2572 next ? next->vm_start : USER_PGTABLES_CEILING);
2573 tlb_finish_mmu(&tlb, start, end);
2574}
2575
2576/*
2577 * Create a list of vma's touched by the unmap, removing them from the mm's
2578 * vma list as we go..
2579 */
2580static void
2581detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2582 struct vm_area_struct *prev, unsigned long end)
2583{
2584 struct vm_area_struct **insertion_point;
2585 struct vm_area_struct *tail_vma = NULL;
2586
2587 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2588 vma->vm_prev = NULL;
2589 do {
2590 vma_rb_erase(vma, &mm->mm_rb);
2591 mm->map_count--;
2592 tail_vma = vma;
2593 vma = vma->vm_next;
2594 } while (vma && vma->vm_start < end);
2595 *insertion_point = vma;
2596 if (vma) {
2597 vma->vm_prev = prev;
2598 vma_gap_update(vma);
2599 } else
2600 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2601 tail_vma->vm_next = NULL;
2602
2603 /* Kill the cache */
2604 vmacache_invalidate(mm);
2605}
2606
2607/*
2608 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it
2609 * has already been checked or doesn't make sense to fail.
2610 */
2611int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2612 unsigned long addr, int new_below)
2613{
2614 struct vm_area_struct *new;
2615 int err;
2616
2617 if (vma->vm_ops && vma->vm_ops->split) {
2618 err = vma->vm_ops->split(vma, addr);
2619 if (err)
2620 return err;
2621 }
2622
2623 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2624 if (!new)
2625 return -ENOMEM;
2626
2627 /* most fields are the same, copy all, and then fixup */
2628 *new = *vma;
2629
2630 INIT_LIST_HEAD(&new->anon_vma_chain);
2631
2632 if (new_below)
2633 new->vm_end = addr;
2634 else {
2635 new->vm_start = addr;
2636 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2637 }
2638
2639 err = vma_dup_policy(vma, new);
2640 if (err)
2641 goto out_free_vma;
2642
2643 err = anon_vma_clone(new, vma);
2644 if (err)
2645 goto out_free_mpol;
2646
2647 if (new->vm_file)
2648 get_file(new->vm_file);
2649
2650 if (new->vm_ops && new->vm_ops->open)
2651 new->vm_ops->open(new);
2652
2653 if (new_below)
2654 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2655 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2656 else
2657 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2658
2659 /* Success. */
2660 if (!err)
2661 return 0;
2662
2663 /* Clean everything up if vma_adjust failed. */
2664 if (new->vm_ops && new->vm_ops->close)
2665 new->vm_ops->close(new);
2666 if (new->vm_file)
2667 fput(new->vm_file);
2668 unlink_anon_vmas(new);
2669 out_free_mpol:
2670 mpol_put(vma_policy(new));
2671 out_free_vma:
2672 kmem_cache_free(vm_area_cachep, new);
2673 return err;
2674}
2675
2676/*
2677 * Split a vma into two pieces at address 'addr', a new vma is allocated
2678 * either for the first part or the tail.
2679 */
2680int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2681 unsigned long addr, int new_below)
2682{
2683 if (mm->map_count >= sysctl_max_map_count)
2684 return -ENOMEM;
2685
2686 return __split_vma(mm, vma, addr, new_below);
2687}
2688
2689/* Munmap is split into 2 main parts -- this part which finds
2690 * what needs doing, and the areas themselves, which do the
2691 * work. This now handles partial unmappings.
2692 * Jeremy Fitzhardinge <jeremy@goop.org>
2693 */
2694int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2695 struct list_head *uf)
2696{
2697 unsigned long end;
2698 struct vm_area_struct *vma, *prev, *last;
2699
2700 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2701 return -EINVAL;
2702
2703 len = PAGE_ALIGN(len);
2704 if (len == 0)
2705 return -EINVAL;
2706
2707 /* Find the first overlapping VMA */
2708 vma = find_vma(mm, start);
2709 if (!vma)
2710 return 0;
2711 prev = vma->vm_prev;
2712 /* we have start < vma->vm_end */
2713
2714 /* if it doesn't overlap, we have nothing.. */
2715 end = start + len;
2716 if (vma->vm_start >= end)
2717 return 0;
2718
2719 /*
2720 * If we need to split any vma, do it now to save pain later.
2721 *
2722 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2723 * unmapped vm_area_struct will remain in use: so lower split_vma
2724 * places tmp vma above, and higher split_vma places tmp vma below.
2725 */
2726 if (start > vma->vm_start) {
2727 int error;
2728
2729 /*
2730 * Make sure that map_count on return from munmap() will
2731 * not exceed its limit; but let map_count go just above
2732 * its limit temporarily, to help free resources as expected.
2733 */
2734 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2735 return -ENOMEM;
2736
2737 error = __split_vma(mm, vma, start, 0);
2738 if (error)
2739 return error;
2740 prev = vma;
2741 }
2742
2743 /* Does it split the last one? */
2744 last = find_vma(mm, end);
2745 if (last && end > last->vm_start) {
2746 int error = __split_vma(mm, last, end, 1);
2747 if (error)
2748 return error;
2749 }
2750 vma = prev ? prev->vm_next : mm->mmap;
2751
2752 if (unlikely(uf)) {
2753 /*
2754 * If userfaultfd_unmap_prep returns an error the vmas
2755 * will remain splitted, but userland will get a
2756 * highly unexpected error anyway. This is no
2757 * different than the case where the first of the two
2758 * __split_vma fails, but we don't undo the first
2759 * split, despite we could. This is unlikely enough
2760 * failure that it's not worth optimizing it for.
2761 */
2762 int error = userfaultfd_unmap_prep(vma, start, end, uf);
2763 if (error)
2764 return error;
2765 }
2766
2767 /*
2768 * unlock any mlock()ed ranges before detaching vmas
2769 */
2770 if (mm->locked_vm) {
2771 struct vm_area_struct *tmp = vma;
2772 while (tmp && tmp->vm_start < end) {
2773 if (tmp->vm_flags & VM_LOCKED) {
2774 mm->locked_vm -= vma_pages(tmp);
2775 munlock_vma_pages_all(tmp);
2776 }
2777 tmp = tmp->vm_next;
2778 }
2779 }
2780
2781 /*
2782 * Remove the vma's, and unmap the actual pages
2783 */
2784 detach_vmas_to_be_unmapped(mm, vma, prev, end);
2785 unmap_region(mm, vma, prev, start, end);
2786
2787 arch_unmap(mm, vma, start, end);
2788
2789 /* Fix up all other VM information */
2790 remove_vma_list(mm, vma);
2791
2792 return 0;
2793}
2794
2795int vm_munmap(unsigned long start, size_t len)
2796{
2797 int ret;
2798 struct mm_struct *mm = current->mm;
2799 LIST_HEAD(uf);
2800
2801 if (down_write_killable(&mm->mmap_sem))
2802 return -EINTR;
2803
2804 ret = do_munmap(mm, start, len, &uf);
2805 up_write(&mm->mmap_sem);
2806 userfaultfd_unmap_complete(mm, &uf);
2807 return ret;
2808}
2809EXPORT_SYMBOL(vm_munmap);
2810
2811SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2812{
2813 profile_munmap(addr);
2814 return vm_munmap(addr, len);
2815}
2816
2817
2818/*
2819 * Emulation of deprecated remap_file_pages() syscall.
2820 */
2821SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2822 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2823{
2824
2825 struct mm_struct *mm = current->mm;
2826 struct vm_area_struct *vma;
2827 unsigned long populate = 0;
2828 unsigned long ret = -EINVAL;
2829 struct file *file;
2830
2831 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.txt.\n",
2832 current->comm, current->pid);
2833
2834 if (prot)
2835 return ret;
2836 start = start & PAGE_MASK;
2837 size = size & PAGE_MASK;
2838
2839 if (start + size <= start)
2840 return ret;
2841
2842 /* Does pgoff wrap? */
2843 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2844 return ret;
2845
2846 if (down_write_killable(&mm->mmap_sem))
2847 return -EINTR;
2848
2849 vma = find_vma(mm, start);
2850
2851 if (!vma || !(vma->vm_flags & VM_SHARED))
2852 goto out;
2853
2854 if (start < vma->vm_start)
2855 goto out;
2856
2857 if (start + size > vma->vm_end) {
2858 struct vm_area_struct *next;
2859
2860 for (next = vma->vm_next; next; next = next->vm_next) {
2861 /* hole between vmas ? */
2862 if (next->vm_start != next->vm_prev->vm_end)
2863 goto out;
2864
2865 if (next->vm_file != vma->vm_file)
2866 goto out;
2867
2868 if (next->vm_flags != vma->vm_flags)
2869 goto out;
2870
2871 if (start + size <= next->vm_end)
2872 break;
2873 }
2874
2875 if (!next)
2876 goto out;
2877 }
2878
2879 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2880 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2881 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2882
2883 flags &= MAP_NONBLOCK;
2884 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2885 if (vma->vm_flags & VM_LOCKED) {
2886 struct vm_area_struct *tmp;
2887 flags |= MAP_LOCKED;
2888
2889 /* drop PG_Mlocked flag for over-mapped range */
2890 for (tmp = vma; tmp->vm_start >= start + size;
2891 tmp = tmp->vm_next) {
2892 /*
2893 * Split pmd and munlock page on the border
2894 * of the range.
2895 */
2896 vma_adjust_trans_huge(tmp, start, start + size, 0);
2897
2898 munlock_vma_pages_range(tmp,
2899 max(tmp->vm_start, start),
2900 min(tmp->vm_end, start + size));
2901 }
2902 }
2903
2904 file = get_file(vma->vm_file);
2905 ret = do_mmap_pgoff(vma->vm_file, start, size,
2906 prot, flags, pgoff, &populate, NULL);
2907 fput(file);
2908out:
2909 up_write(&mm->mmap_sem);
2910 if (populate)
2911 mm_populate(ret, populate);
2912 if (!IS_ERR_VALUE(ret))
2913 ret = 0;
2914 return ret;
2915}
2916
2917static inline void verify_mm_writelocked(struct mm_struct *mm)
2918{
2919#ifdef CONFIG_DEBUG_VM
2920 if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2921 WARN_ON(1);
2922 up_read(&mm->mmap_sem);
2923 }
2924#endif
2925}
2926
2927/*
2928 * this is really a simplified "do_mmap". it only handles
2929 * anonymous maps. eventually we may be able to do some
2930 * brk-specific accounting here.
2931 */
2932static int do_brk_flags(unsigned long addr, unsigned long request, unsigned long flags, struct list_head *uf)
2933{
2934 struct mm_struct *mm = current->mm;
2935 struct vm_area_struct *vma, *prev;
2936 unsigned long len;
2937 struct rb_node **rb_link, *rb_parent;
2938 pgoff_t pgoff = addr >> PAGE_SHIFT;
2939 int error;
2940
2941 len = PAGE_ALIGN(request);
2942 if (len < request)
2943 return -ENOMEM;
2944 if (!len)
2945 return 0;
2946
2947 /* Until we need other flags, refuse anything except VM_EXEC. */
2948 if ((flags & (~VM_EXEC)) != 0)
2949 return -EINVAL;
2950 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2951
2952 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2953 if (offset_in_page(error))
2954 return error;
2955
2956 error = mlock_future_check(mm, mm->def_flags, len);
2957 if (error)
2958 return error;
2959
2960 /*
2961 * mm->mmap_sem is required to protect against another thread
2962 * changing the mappings in case we sleep.
2963 */
2964 verify_mm_writelocked(mm);
2965
2966 /*
2967 * Clear old maps. this also does some error checking for us
2968 */
2969 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
2970 &rb_parent)) {
2971 if (do_munmap(mm, addr, len, uf))
2972 return -ENOMEM;
2973 }
2974
2975 /* Check against address space limits *after* clearing old maps... */
2976 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
2977 return -ENOMEM;
2978
2979 if (mm->map_count > sysctl_max_map_count)
2980 return -ENOMEM;
2981
2982 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2983 return -ENOMEM;
2984
2985 /* Can we just expand an old private anonymous mapping? */
2986 vma = vma_merge(mm, prev, addr, addr + len, flags,
2987 NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
2988 if (vma)
2989 goto out;
2990
2991 /*
2992 * create a vma struct for an anonymous mapping
2993 */
2994 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2995 if (!vma) {
2996 vm_unacct_memory(len >> PAGE_SHIFT);
2997 return -ENOMEM;
2998 }
2999
3000 INIT_LIST_HEAD(&vma->anon_vma_chain);
3001 vma->vm_mm = mm;
3002 vma->vm_start = addr;
3003 vma->vm_end = addr + len;
3004 vma->vm_pgoff = pgoff;
3005 vma->vm_flags = flags;
3006 vma->vm_page_prot = vm_get_page_prot(flags);
3007 vma_link(mm, vma, prev, rb_link, rb_parent);
3008out:
3009 perf_event_mmap(vma);
3010 mm->total_vm += len >> PAGE_SHIFT;
3011 mm->data_vm += len >> PAGE_SHIFT;
3012 if (flags & VM_LOCKED)
3013 mm->locked_vm += (len >> PAGE_SHIFT);
3014 vma->vm_flags |= VM_SOFTDIRTY;
3015 return 0;
3016}
3017
3018static int do_brk(unsigned long addr, unsigned long len, struct list_head *uf)
3019{
3020 return do_brk_flags(addr, len, 0, uf);
3021}
3022
3023int vm_brk_flags(unsigned long addr, unsigned long len, unsigned long flags)
3024{
3025 struct mm_struct *mm = current->mm;
3026 int ret;
3027 bool populate;
3028 LIST_HEAD(uf);
3029
3030 if (down_write_killable(&mm->mmap_sem))
3031 return -EINTR;
3032
3033 ret = do_brk_flags(addr, len, flags, &uf);
3034 populate = ((mm->def_flags & VM_LOCKED) != 0);
3035 up_write(&mm->mmap_sem);
3036 userfaultfd_unmap_complete(mm, &uf);
3037 if (populate && !ret)
3038 mm_populate(addr, len);
3039 return ret;
3040}
3041EXPORT_SYMBOL(vm_brk_flags);
3042
3043int vm_brk(unsigned long addr, unsigned long len)
3044{
3045 return vm_brk_flags(addr, len, 0);
3046}
3047EXPORT_SYMBOL(vm_brk);
3048
3049/* Release all mmaps. */
3050void exit_mmap(struct mm_struct *mm)
3051{
3052 struct mmu_gather tlb;
3053 struct vm_area_struct *vma;
3054 unsigned long nr_accounted = 0;
3055
3056 /* mm's last user has gone, and its about to be pulled down */
3057 mmu_notifier_release(mm);
3058
3059 if (unlikely(mm_is_oom_victim(mm))) {
3060 /*
3061 * Manually reap the mm to free as much memory as possible.
3062 * Then, as the oom reaper does, set MMF_OOM_SKIP to disregard
3063 * this mm from further consideration. Taking mm->mmap_sem for
3064 * write after setting MMF_OOM_SKIP will guarantee that the oom
3065 * reaper will not run on this mm again after mmap_sem is
3066 * dropped.
3067 *
3068 * Nothing can be holding mm->mmap_sem here and the above call
3069 * to mmu_notifier_release(mm) ensures mmu notifier callbacks in
3070 * __oom_reap_task_mm() will not block.
3071 *
3072 * This needs to be done before calling munlock_vma_pages_all(),
3073 * which clears VM_LOCKED, otherwise the oom reaper cannot
3074 * reliably test it.
3075 */
3076 mutex_lock(&oom_lock);
3077 __oom_reap_task_mm(mm);
3078 mutex_unlock(&oom_lock);
3079
3080 set_bit(MMF_OOM_SKIP, &mm->flags);
3081 down_write(&mm->mmap_sem);
3082 up_write(&mm->mmap_sem);
3083 }
3084
3085 if (mm->locked_vm) {
3086 vma = mm->mmap;
3087 while (vma) {
3088 if (vma->vm_flags & VM_LOCKED)
3089 munlock_vma_pages_all(vma);
3090 vma = vma->vm_next;
3091 }
3092 }
3093
3094 arch_exit_mmap(mm);
3095
3096 vma = mm->mmap;
3097 if (!vma) /* Can happen if dup_mmap() received an OOM */
3098 return;
3099
3100 lru_add_drain();
3101 flush_cache_mm(mm);
3102 tlb_gather_mmu(&tlb, mm, 0, -1);
3103 /* update_hiwater_rss(mm) here? but nobody should be looking */
3104 /* Use -1 here to ensure all VMAs in the mm are unmapped */
3105 unmap_vmas(&tlb, vma, 0, -1);
3106 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3107 tlb_finish_mmu(&tlb, 0, -1);
3108
3109 /*
3110 * Walk the list again, actually closing and freeing it,
3111 * with preemption enabled, without holding any MM locks.
3112 */
3113 while (vma) {
3114 if (vma->vm_flags & VM_ACCOUNT)
3115 nr_accounted += vma_pages(vma);
3116 vma = remove_vma(vma);
3117 }
3118 vm_unacct_memory(nr_accounted);
3119}
3120
3121/* Insert vm structure into process list sorted by address
3122 * and into the inode's i_mmap tree. If vm_file is non-NULL
3123 * then i_mmap_rwsem is taken here.
3124 */
3125int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3126{
3127 struct vm_area_struct *prev;
3128 struct rb_node **rb_link, *rb_parent;
3129
3130 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3131 &prev, &rb_link, &rb_parent))
3132 return -ENOMEM;
3133 if ((vma->vm_flags & VM_ACCOUNT) &&
3134 security_vm_enough_memory_mm(mm, vma_pages(vma)))
3135 return -ENOMEM;
3136
3137 /*
3138 * The vm_pgoff of a purely anonymous vma should be irrelevant
3139 * until its first write fault, when page's anon_vma and index
3140 * are set. But now set the vm_pgoff it will almost certainly
3141 * end up with (unless mremap moves it elsewhere before that
3142 * first wfault), so /proc/pid/maps tells a consistent story.
3143 *
3144 * By setting it to reflect the virtual start address of the
3145 * vma, merges and splits can happen in a seamless way, just
3146 * using the existing file pgoff checks and manipulations.
3147 * Similarly in do_mmap_pgoff and in do_brk.
3148 */
3149 if (vma_is_anonymous(vma)) {
3150 BUG_ON(vma->anon_vma);
3151 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3152 }
3153
3154 vma_link(mm, vma, prev, rb_link, rb_parent);
3155 return 0;
3156}
3157
3158/*
3159 * Copy the vma structure to a new location in the same mm,
3160 * prior to moving page table entries, to effect an mremap move.
3161 */
3162struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3163 unsigned long addr, unsigned long len, pgoff_t pgoff,
3164 bool *need_rmap_locks)
3165{
3166 struct vm_area_struct *vma = *vmap;
3167 unsigned long vma_start = vma->vm_start;
3168 struct mm_struct *mm = vma->vm_mm;
3169 struct vm_area_struct *new_vma, *prev;
3170 struct rb_node **rb_link, *rb_parent;
3171 bool faulted_in_anon_vma = true;
3172
3173 /*
3174 * If anonymous vma has not yet been faulted, update new pgoff
3175 * to match new location, to increase its chance of merging.
3176 */
3177 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3178 pgoff = addr >> PAGE_SHIFT;
3179 faulted_in_anon_vma = false;
3180 }
3181
3182 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3183 return NULL; /* should never get here */
3184 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3185 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3186 vma->vm_userfaultfd_ctx);
3187 if (new_vma) {
3188 /*
3189 * Source vma may have been merged into new_vma
3190 */
3191 if (unlikely(vma_start >= new_vma->vm_start &&
3192 vma_start < new_vma->vm_end)) {
3193 /*
3194 * The only way we can get a vma_merge with
3195 * self during an mremap is if the vma hasn't
3196 * been faulted in yet and we were allowed to
3197 * reset the dst vma->vm_pgoff to the
3198 * destination address of the mremap to allow
3199 * the merge to happen. mremap must change the
3200 * vm_pgoff linearity between src and dst vmas
3201 * (in turn preventing a vma_merge) to be
3202 * safe. It is only safe to keep the vm_pgoff
3203 * linear if there are no pages mapped yet.
3204 */
3205 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3206 *vmap = vma = new_vma;
3207 }
3208 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3209 } else {
3210 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
3211 if (!new_vma)
3212 goto out;
3213 *new_vma = *vma;
3214 new_vma->vm_start = addr;
3215 new_vma->vm_end = addr + len;
3216 new_vma->vm_pgoff = pgoff;
3217 if (vma_dup_policy(vma, new_vma))
3218 goto out_free_vma;
3219 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
3220 if (anon_vma_clone(new_vma, vma))
3221 goto out_free_mempol;
3222 if (new_vma->vm_file)
3223 get_file(new_vma->vm_file);
3224 if (new_vma->vm_ops && new_vma->vm_ops->open)
3225 new_vma->vm_ops->open(new_vma);
3226 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3227 *need_rmap_locks = false;
3228 }
3229 return new_vma;
3230
3231out_free_mempol:
3232 mpol_put(vma_policy(new_vma));
3233out_free_vma:
3234 kmem_cache_free(vm_area_cachep, new_vma);
3235out:
3236 return NULL;
3237}
3238
3239/*
3240 * Return true if the calling process may expand its vm space by the passed
3241 * number of pages
3242 */
3243bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3244{
3245 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3246 return false;
3247
3248 if (is_data_mapping(flags) &&
3249 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3250 /* Workaround for Valgrind */
3251 if (rlimit(RLIMIT_DATA) == 0 &&
3252 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3253 return true;
3254
3255 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3256 current->comm, current->pid,
3257 (mm->data_vm + npages) << PAGE_SHIFT,
3258 rlimit(RLIMIT_DATA),
3259 ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3260
3261 if (!ignore_rlimit_data)
3262 return false;
3263 }
3264
3265 return true;
3266}
3267
3268void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3269{
3270 mm->total_vm += npages;
3271
3272 if (is_exec_mapping(flags))
3273 mm->exec_vm += npages;
3274 else if (is_stack_mapping(flags))
3275 mm->stack_vm += npages;
3276 else if (is_data_mapping(flags))
3277 mm->data_vm += npages;
3278}
3279
3280static int special_mapping_fault(struct vm_fault *vmf);
3281
3282/*
3283 * Having a close hook prevents vma merging regardless of flags.
3284 */
3285static void special_mapping_close(struct vm_area_struct *vma)
3286{
3287}
3288
3289static const char *special_mapping_name(struct vm_area_struct *vma)
3290{
3291 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3292}
3293
3294static int special_mapping_mremap(struct vm_area_struct *new_vma)
3295{
3296 struct vm_special_mapping *sm = new_vma->vm_private_data;
3297
3298 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3299 return -EFAULT;
3300
3301 if (sm->mremap)
3302 return sm->mremap(sm, new_vma);
3303
3304 return 0;
3305}
3306
3307static const struct vm_operations_struct special_mapping_vmops = {
3308 .close = special_mapping_close,
3309 .fault = special_mapping_fault,
3310 .mremap = special_mapping_mremap,
3311 .name = special_mapping_name,
3312};
3313
3314static const struct vm_operations_struct legacy_special_mapping_vmops = {
3315 .close = special_mapping_close,
3316 .fault = special_mapping_fault,
3317};
3318
3319static int special_mapping_fault(struct vm_fault *vmf)
3320{
3321 struct vm_area_struct *vma = vmf->vma;
3322 pgoff_t pgoff;
3323 struct page **pages;
3324
3325 if (vma->vm_ops == &legacy_special_mapping_vmops) {
3326 pages = vma->vm_private_data;
3327 } else {
3328 struct vm_special_mapping *sm = vma->vm_private_data;
3329
3330 if (sm->fault)
3331 return sm->fault(sm, vmf->vma, vmf);
3332
3333 pages = sm->pages;
3334 }
3335
3336 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3337 pgoff--;
3338
3339 if (*pages) {
3340 struct page *page = *pages;
3341 get_page(page);
3342 vmf->page = page;
3343 return 0;
3344 }
3345
3346 return VM_FAULT_SIGBUS;
3347}
3348
3349static struct vm_area_struct *__install_special_mapping(
3350 struct mm_struct *mm,
3351 unsigned long addr, unsigned long len,
3352 unsigned long vm_flags, void *priv,
3353 const struct vm_operations_struct *ops)
3354{
3355 int ret;
3356 struct vm_area_struct *vma;
3357
3358 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
3359 if (unlikely(vma == NULL))
3360 return ERR_PTR(-ENOMEM);
3361
3362 INIT_LIST_HEAD(&vma->anon_vma_chain);
3363 vma->vm_mm = mm;
3364 vma->vm_start = addr;
3365 vma->vm_end = addr + len;
3366
3367 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3368 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3369
3370 vma->vm_ops = ops;
3371 vma->vm_private_data = priv;
3372
3373 ret = insert_vm_struct(mm, vma);
3374 if (ret)
3375 goto out;
3376
3377 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3378
3379 perf_event_mmap(vma);
3380
3381 return vma;
3382
3383out:
3384 kmem_cache_free(vm_area_cachep, vma);
3385 return ERR_PTR(ret);
3386}
3387
3388bool vma_is_special_mapping(const struct vm_area_struct *vma,
3389 const struct vm_special_mapping *sm)
3390{
3391 return vma->vm_private_data == sm &&
3392 (vma->vm_ops == &special_mapping_vmops ||
3393 vma->vm_ops == &legacy_special_mapping_vmops);
3394}
3395
3396/*
3397 * Called with mm->mmap_sem held for writing.
3398 * Insert a new vma covering the given region, with the given flags.
3399 * Its pages are supplied by the given array of struct page *.
3400 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3401 * The region past the last page supplied will always produce SIGBUS.
3402 * The array pointer and the pages it points to are assumed to stay alive
3403 * for as long as this mapping might exist.
3404 */
3405struct vm_area_struct *_install_special_mapping(
3406 struct mm_struct *mm,
3407 unsigned long addr, unsigned long len,
3408 unsigned long vm_flags, const struct vm_special_mapping *spec)
3409{
3410 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3411 &special_mapping_vmops);
3412}
3413
3414int install_special_mapping(struct mm_struct *mm,
3415 unsigned long addr, unsigned long len,
3416 unsigned long vm_flags, struct page **pages)
3417{
3418 struct vm_area_struct *vma = __install_special_mapping(
3419 mm, addr, len, vm_flags, (void *)pages,
3420 &legacy_special_mapping_vmops);
3421
3422 return PTR_ERR_OR_ZERO(vma);
3423}
3424
3425static DEFINE_MUTEX(mm_all_locks_mutex);
3426
3427static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3428{
3429 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3430 /*
3431 * The LSB of head.next can't change from under us
3432 * because we hold the mm_all_locks_mutex.
3433 */
3434 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
3435 /*
3436 * We can safely modify head.next after taking the
3437 * anon_vma->root->rwsem. If some other vma in this mm shares
3438 * the same anon_vma we won't take it again.
3439 *
3440 * No need of atomic instructions here, head.next
3441 * can't change from under us thanks to the
3442 * anon_vma->root->rwsem.
3443 */
3444 if (__test_and_set_bit(0, (unsigned long *)
3445 &anon_vma->root->rb_root.rb_root.rb_node))
3446 BUG();
3447 }
3448}
3449
3450static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3451{
3452 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3453 /*
3454 * AS_MM_ALL_LOCKS can't change from under us because
3455 * we hold the mm_all_locks_mutex.
3456 *
3457 * Operations on ->flags have to be atomic because
3458 * even if AS_MM_ALL_LOCKS is stable thanks to the
3459 * mm_all_locks_mutex, there may be other cpus
3460 * changing other bitflags in parallel to us.
3461 */
3462 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3463 BUG();
3464 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem);
3465 }
3466}
3467
3468/*
3469 * This operation locks against the VM for all pte/vma/mm related
3470 * operations that could ever happen on a certain mm. This includes
3471 * vmtruncate, try_to_unmap, and all page faults.
3472 *
3473 * The caller must take the mmap_sem in write mode before calling
3474 * mm_take_all_locks(). The caller isn't allowed to release the
3475 * mmap_sem until mm_drop_all_locks() returns.
3476 *
3477 * mmap_sem in write mode is required in order to block all operations
3478 * that could modify pagetables and free pages without need of
3479 * altering the vma layout. It's also needed in write mode to avoid new
3480 * anon_vmas to be associated with existing vmas.
3481 *
3482 * A single task can't take more than one mm_take_all_locks() in a row
3483 * or it would deadlock.
3484 *
3485 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3486 * mapping->flags avoid to take the same lock twice, if more than one
3487 * vma in this mm is backed by the same anon_vma or address_space.
3488 *
3489 * We take locks in following order, accordingly to comment at beginning
3490 * of mm/rmap.c:
3491 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3492 * hugetlb mapping);
3493 * - all i_mmap_rwsem locks;
3494 * - all anon_vma->rwseml
3495 *
3496 * We can take all locks within these types randomly because the VM code
3497 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3498 * mm_all_locks_mutex.
3499 *
3500 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3501 * that may have to take thousand of locks.
3502 *
3503 * mm_take_all_locks() can fail if it's interrupted by signals.
3504 */
3505int mm_take_all_locks(struct mm_struct *mm)
3506{
3507 struct vm_area_struct *vma;
3508 struct anon_vma_chain *avc;
3509
3510 BUG_ON(down_read_trylock(&mm->mmap_sem));
3511
3512 mutex_lock(&mm_all_locks_mutex);
3513
3514 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3515 if (signal_pending(current))
3516 goto out_unlock;
3517 if (vma->vm_file && vma->vm_file->f_mapping &&
3518 is_vm_hugetlb_page(vma))
3519 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3520 }
3521
3522 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3523 if (signal_pending(current))
3524 goto out_unlock;
3525 if (vma->vm_file && vma->vm_file->f_mapping &&
3526 !is_vm_hugetlb_page(vma))
3527 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3528 }
3529
3530 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3531 if (signal_pending(current))
3532 goto out_unlock;
3533 if (vma->anon_vma)
3534 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3535 vm_lock_anon_vma(mm, avc->anon_vma);
3536 }
3537
3538 return 0;
3539
3540out_unlock:
3541 mm_drop_all_locks(mm);
3542 return -EINTR;
3543}
3544
3545static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3546{
3547 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3548 /*
3549 * The LSB of head.next can't change to 0 from under
3550 * us because we hold the mm_all_locks_mutex.
3551 *
3552 * We must however clear the bitflag before unlocking
3553 * the vma so the users using the anon_vma->rb_root will
3554 * never see our bitflag.
3555 *
3556 * No need of atomic instructions here, head.next
3557 * can't change from under us until we release the
3558 * anon_vma->root->rwsem.
3559 */
3560 if (!__test_and_clear_bit(0, (unsigned long *)
3561 &anon_vma->root->rb_root.rb_root.rb_node))
3562 BUG();
3563 anon_vma_unlock_write(anon_vma);
3564 }
3565}
3566
3567static void vm_unlock_mapping(struct address_space *mapping)
3568{
3569 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3570 /*
3571 * AS_MM_ALL_LOCKS can't change to 0 from under us
3572 * because we hold the mm_all_locks_mutex.
3573 */
3574 i_mmap_unlock_write(mapping);
3575 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3576 &mapping->flags))
3577 BUG();
3578 }
3579}
3580
3581/*
3582 * The mmap_sem cannot be released by the caller until
3583 * mm_drop_all_locks() returns.
3584 */
3585void mm_drop_all_locks(struct mm_struct *mm)
3586{
3587 struct vm_area_struct *vma;
3588 struct anon_vma_chain *avc;
3589
3590 BUG_ON(down_read_trylock(&mm->mmap_sem));
3591 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3592
3593 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3594 if (vma->anon_vma)
3595 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3596 vm_unlock_anon_vma(avc->anon_vma);
3597 if (vma->vm_file && vma->vm_file->f_mapping)
3598 vm_unlock_mapping(vma->vm_file->f_mapping);
3599 }
3600
3601 mutex_unlock(&mm_all_locks_mutex);
3602}
3603
3604/*
3605 * initialise the percpu counter for VM
3606 */
3607void __init mmap_init(void)
3608{
3609 int ret;
3610
3611 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3612 VM_BUG_ON(ret);
3613}
3614
3615/*
3616 * Initialise sysctl_user_reserve_kbytes.
3617 *
3618 * This is intended to prevent a user from starting a single memory hogging
3619 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3620 * mode.
3621 *
3622 * The default value is min(3% of free memory, 128MB)
3623 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3624 */
3625static int init_user_reserve(void)
3626{
3627 unsigned long free_kbytes;
3628
3629 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3630
3631 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3632 return 0;
3633}
3634subsys_initcall(init_user_reserve);
3635
3636/*
3637 * Initialise sysctl_admin_reserve_kbytes.
3638 *
3639 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3640 * to log in and kill a memory hogging process.
3641 *
3642 * Systems with more than 256MB will reserve 8MB, enough to recover
3643 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3644 * only reserve 3% of free pages by default.
3645 */
3646static int init_admin_reserve(void)
3647{
3648 unsigned long free_kbytes;
3649
3650 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3651
3652 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3653 return 0;
3654}
3655subsys_initcall(init_admin_reserve);
3656
3657/*
3658 * Reinititalise user and admin reserves if memory is added or removed.
3659 *
3660 * The default user reserve max is 128MB, and the default max for the
3661 * admin reserve is 8MB. These are usually, but not always, enough to
3662 * enable recovery from a memory hogging process using login/sshd, a shell,
3663 * and tools like top. It may make sense to increase or even disable the
3664 * reserve depending on the existence of swap or variations in the recovery
3665 * tools. So, the admin may have changed them.
3666 *
3667 * If memory is added and the reserves have been eliminated or increased above
3668 * the default max, then we'll trust the admin.
3669 *
3670 * If memory is removed and there isn't enough free memory, then we
3671 * need to reset the reserves.
3672 *
3673 * Otherwise keep the reserve set by the admin.
3674 */
3675static int reserve_mem_notifier(struct notifier_block *nb,
3676 unsigned long action, void *data)
3677{
3678 unsigned long tmp, free_kbytes;
3679
3680 switch (action) {
3681 case MEM_ONLINE:
3682 /* Default max is 128MB. Leave alone if modified by operator. */
3683 tmp = sysctl_user_reserve_kbytes;
3684 if (0 < tmp && tmp < (1UL << 17))
3685 init_user_reserve();
3686
3687 /* Default max is 8MB. Leave alone if modified by operator. */
3688 tmp = sysctl_admin_reserve_kbytes;
3689 if (0 < tmp && tmp < (1UL << 13))
3690 init_admin_reserve();
3691
3692 break;
3693 case MEM_OFFLINE:
3694 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3695
3696 if (sysctl_user_reserve_kbytes > free_kbytes) {
3697 init_user_reserve();
3698 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3699 sysctl_user_reserve_kbytes);
3700 }
3701
3702 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3703 init_admin_reserve();
3704 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3705 sysctl_admin_reserve_kbytes);
3706 }
3707 break;
3708 default:
3709 break;
3710 }
3711 return NOTIFY_OK;
3712}
3713
3714static struct notifier_block reserve_mem_nb = {
3715 .notifier_call = reserve_mem_notifier,
3716};
3717
3718static int __meminit init_reserve_notifier(void)
3719{
3720 if (register_hotmemory_notifier(&reserve_mem_nb))
3721 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3722
3723 return 0;
3724}
3725subsys_initcall(init_reserve_notifier);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * mm/mmap.c
4 *
5 * Written by obz.
6 *
7 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
8 */
9
10#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11
12#include <linux/kernel.h>
13#include <linux/slab.h>
14#include <linux/backing-dev.h>
15#include <linux/mm.h>
16#include <linux/mm_inline.h>
17#include <linux/shm.h>
18#include <linux/mman.h>
19#include <linux/pagemap.h>
20#include <linux/swap.h>
21#include <linux/syscalls.h>
22#include <linux/capability.h>
23#include <linux/init.h>
24#include <linux/file.h>
25#include <linux/fs.h>
26#include <linux/personality.h>
27#include <linux/security.h>
28#include <linux/hugetlb.h>
29#include <linux/shmem_fs.h>
30#include <linux/profile.h>
31#include <linux/export.h>
32#include <linux/mount.h>
33#include <linux/mempolicy.h>
34#include <linux/rmap.h>
35#include <linux/mmu_notifier.h>
36#include <linux/mmdebug.h>
37#include <linux/perf_event.h>
38#include <linux/audit.h>
39#include <linux/khugepaged.h>
40#include <linux/uprobes.h>
41#include <linux/notifier.h>
42#include <linux/memory.h>
43#include <linux/printk.h>
44#include <linux/userfaultfd_k.h>
45#include <linux/moduleparam.h>
46#include <linux/pkeys.h>
47#include <linux/oom.h>
48#include <linux/sched/mm.h>
49#include <linux/ksm.h>
50
51#include <linux/uaccess.h>
52#include <asm/cacheflush.h>
53#include <asm/tlb.h>
54#include <asm/mmu_context.h>
55
56#define CREATE_TRACE_POINTS
57#include <trace/events/mmap.h>
58
59#include "internal.h"
60
61#ifndef arch_mmap_check
62#define arch_mmap_check(addr, len, flags) (0)
63#endif
64
65#ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
66const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
67int mmap_rnd_bits_max __ro_after_init = CONFIG_ARCH_MMAP_RND_BITS_MAX;
68int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
69#endif
70#ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
71const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
72const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
73int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
74#endif
75
76static bool ignore_rlimit_data;
77core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
78
79static void unmap_region(struct mm_struct *mm, struct ma_state *mas,
80 struct vm_area_struct *vma, struct vm_area_struct *prev,
81 struct vm_area_struct *next, unsigned long start,
82 unsigned long end, unsigned long tree_end, bool mm_wr_locked);
83
84static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
85{
86 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
87}
88
89/* Update vma->vm_page_prot to reflect vma->vm_flags. */
90void vma_set_page_prot(struct vm_area_struct *vma)
91{
92 unsigned long vm_flags = vma->vm_flags;
93 pgprot_t vm_page_prot;
94
95 vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
96 if (vma_wants_writenotify(vma, vm_page_prot)) {
97 vm_flags &= ~VM_SHARED;
98 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
99 }
100 /* remove_protection_ptes reads vma->vm_page_prot without mmap_lock */
101 WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
102}
103
104/*
105 * Requires inode->i_mapping->i_mmap_rwsem
106 */
107static void __remove_shared_vm_struct(struct vm_area_struct *vma,
108 struct address_space *mapping)
109{
110 if (vma_is_shared_maywrite(vma))
111 mapping_unmap_writable(mapping);
112
113 flush_dcache_mmap_lock(mapping);
114 vma_interval_tree_remove(vma, &mapping->i_mmap);
115 flush_dcache_mmap_unlock(mapping);
116}
117
118/*
119 * Unlink a file-based vm structure from its interval tree, to hide
120 * vma from rmap and vmtruncate before freeing its page tables.
121 */
122void unlink_file_vma(struct vm_area_struct *vma)
123{
124 struct file *file = vma->vm_file;
125
126 if (file) {
127 struct address_space *mapping = file->f_mapping;
128 i_mmap_lock_write(mapping);
129 __remove_shared_vm_struct(vma, mapping);
130 i_mmap_unlock_write(mapping);
131 }
132}
133
134/*
135 * Close a vm structure and free it.
136 */
137static void remove_vma(struct vm_area_struct *vma, bool unreachable)
138{
139 might_sleep();
140 if (vma->vm_ops && vma->vm_ops->close)
141 vma->vm_ops->close(vma);
142 if (vma->vm_file)
143 fput(vma->vm_file);
144 mpol_put(vma_policy(vma));
145 if (unreachable)
146 __vm_area_free(vma);
147 else
148 vm_area_free(vma);
149}
150
151static inline struct vm_area_struct *vma_prev_limit(struct vma_iterator *vmi,
152 unsigned long min)
153{
154 return mas_prev(&vmi->mas, min);
155}
156
157/*
158 * check_brk_limits() - Use platform specific check of range & verify mlock
159 * limits.
160 * @addr: The address to check
161 * @len: The size of increase.
162 *
163 * Return: 0 on success.
164 */
165static int check_brk_limits(unsigned long addr, unsigned long len)
166{
167 unsigned long mapped_addr;
168
169 mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
170 if (IS_ERR_VALUE(mapped_addr))
171 return mapped_addr;
172
173 return mlock_future_ok(current->mm, current->mm->def_flags, len)
174 ? 0 : -EAGAIN;
175}
176static int do_brk_flags(struct vma_iterator *vmi, struct vm_area_struct *brkvma,
177 unsigned long addr, unsigned long request, unsigned long flags);
178SYSCALL_DEFINE1(brk, unsigned long, brk)
179{
180 unsigned long newbrk, oldbrk, origbrk;
181 struct mm_struct *mm = current->mm;
182 struct vm_area_struct *brkvma, *next = NULL;
183 unsigned long min_brk;
184 bool populate = false;
185 LIST_HEAD(uf);
186 struct vma_iterator vmi;
187
188 if (mmap_write_lock_killable(mm))
189 return -EINTR;
190
191 origbrk = mm->brk;
192
193#ifdef CONFIG_COMPAT_BRK
194 /*
195 * CONFIG_COMPAT_BRK can still be overridden by setting
196 * randomize_va_space to 2, which will still cause mm->start_brk
197 * to be arbitrarily shifted
198 */
199 if (current->brk_randomized)
200 min_brk = mm->start_brk;
201 else
202 min_brk = mm->end_data;
203#else
204 min_brk = mm->start_brk;
205#endif
206 if (brk < min_brk)
207 goto out;
208
209 /*
210 * Check against rlimit here. If this check is done later after the test
211 * of oldbrk with newbrk then it can escape the test and let the data
212 * segment grow beyond its set limit the in case where the limit is
213 * not page aligned -Ram Gupta
214 */
215 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
216 mm->end_data, mm->start_data))
217 goto out;
218
219 newbrk = PAGE_ALIGN(brk);
220 oldbrk = PAGE_ALIGN(mm->brk);
221 if (oldbrk == newbrk) {
222 mm->brk = brk;
223 goto success;
224 }
225
226 /* Always allow shrinking brk. */
227 if (brk <= mm->brk) {
228 /* Search one past newbrk */
229 vma_iter_init(&vmi, mm, newbrk);
230 brkvma = vma_find(&vmi, oldbrk);
231 if (!brkvma || brkvma->vm_start >= oldbrk)
232 goto out; /* mapping intersects with an existing non-brk vma. */
233 /*
234 * mm->brk must be protected by write mmap_lock.
235 * do_vma_munmap() will drop the lock on success, so update it
236 * before calling do_vma_munmap().
237 */
238 mm->brk = brk;
239 if (do_vma_munmap(&vmi, brkvma, newbrk, oldbrk, &uf, true))
240 goto out;
241
242 goto success_unlocked;
243 }
244
245 if (check_brk_limits(oldbrk, newbrk - oldbrk))
246 goto out;
247
248 /*
249 * Only check if the next VMA is within the stack_guard_gap of the
250 * expansion area
251 */
252 vma_iter_init(&vmi, mm, oldbrk);
253 next = vma_find(&vmi, newbrk + PAGE_SIZE + stack_guard_gap);
254 if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
255 goto out;
256
257 brkvma = vma_prev_limit(&vmi, mm->start_brk);
258 /* Ok, looks good - let it rip. */
259 if (do_brk_flags(&vmi, brkvma, oldbrk, newbrk - oldbrk, 0) < 0)
260 goto out;
261
262 mm->brk = brk;
263 if (mm->def_flags & VM_LOCKED)
264 populate = true;
265
266success:
267 mmap_write_unlock(mm);
268success_unlocked:
269 userfaultfd_unmap_complete(mm, &uf);
270 if (populate)
271 mm_populate(oldbrk, newbrk - oldbrk);
272 return brk;
273
274out:
275 mm->brk = origbrk;
276 mmap_write_unlock(mm);
277 return origbrk;
278}
279
280#if defined(CONFIG_DEBUG_VM_MAPLE_TREE)
281static void validate_mm(struct mm_struct *mm)
282{
283 int bug = 0;
284 int i = 0;
285 struct vm_area_struct *vma;
286 VMA_ITERATOR(vmi, mm, 0);
287
288 mt_validate(&mm->mm_mt);
289 for_each_vma(vmi, vma) {
290#ifdef CONFIG_DEBUG_VM_RB
291 struct anon_vma *anon_vma = vma->anon_vma;
292 struct anon_vma_chain *avc;
293#endif
294 unsigned long vmi_start, vmi_end;
295 bool warn = 0;
296
297 vmi_start = vma_iter_addr(&vmi);
298 vmi_end = vma_iter_end(&vmi);
299 if (VM_WARN_ON_ONCE_MM(vma->vm_end != vmi_end, mm))
300 warn = 1;
301
302 if (VM_WARN_ON_ONCE_MM(vma->vm_start != vmi_start, mm))
303 warn = 1;
304
305 if (warn) {
306 pr_emerg("issue in %s\n", current->comm);
307 dump_stack();
308 dump_vma(vma);
309 pr_emerg("tree range: %px start %lx end %lx\n", vma,
310 vmi_start, vmi_end - 1);
311 vma_iter_dump_tree(&vmi);
312 }
313
314#ifdef CONFIG_DEBUG_VM_RB
315 if (anon_vma) {
316 anon_vma_lock_read(anon_vma);
317 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
318 anon_vma_interval_tree_verify(avc);
319 anon_vma_unlock_read(anon_vma);
320 }
321#endif
322 i++;
323 }
324 if (i != mm->map_count) {
325 pr_emerg("map_count %d vma iterator %d\n", mm->map_count, i);
326 bug = 1;
327 }
328 VM_BUG_ON_MM(bug, mm);
329}
330
331#else /* !CONFIG_DEBUG_VM_MAPLE_TREE */
332#define validate_mm(mm) do { } while (0)
333#endif /* CONFIG_DEBUG_VM_MAPLE_TREE */
334
335/*
336 * vma has some anon_vma assigned, and is already inserted on that
337 * anon_vma's interval trees.
338 *
339 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
340 * vma must be removed from the anon_vma's interval trees using
341 * anon_vma_interval_tree_pre_update_vma().
342 *
343 * After the update, the vma will be reinserted using
344 * anon_vma_interval_tree_post_update_vma().
345 *
346 * The entire update must be protected by exclusive mmap_lock and by
347 * the root anon_vma's mutex.
348 */
349static inline void
350anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
351{
352 struct anon_vma_chain *avc;
353
354 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
355 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
356}
357
358static inline void
359anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
360{
361 struct anon_vma_chain *avc;
362
363 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
364 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
365}
366
367static unsigned long count_vma_pages_range(struct mm_struct *mm,
368 unsigned long addr, unsigned long end)
369{
370 VMA_ITERATOR(vmi, mm, addr);
371 struct vm_area_struct *vma;
372 unsigned long nr_pages = 0;
373
374 for_each_vma_range(vmi, vma, end) {
375 unsigned long vm_start = max(addr, vma->vm_start);
376 unsigned long vm_end = min(end, vma->vm_end);
377
378 nr_pages += PHYS_PFN(vm_end - vm_start);
379 }
380
381 return nr_pages;
382}
383
384static void __vma_link_file(struct vm_area_struct *vma,
385 struct address_space *mapping)
386{
387 if (vma_is_shared_maywrite(vma))
388 mapping_allow_writable(mapping);
389
390 flush_dcache_mmap_lock(mapping);
391 vma_interval_tree_insert(vma, &mapping->i_mmap);
392 flush_dcache_mmap_unlock(mapping);
393}
394
395static void vma_link_file(struct vm_area_struct *vma)
396{
397 struct file *file = vma->vm_file;
398 struct address_space *mapping;
399
400 if (file) {
401 mapping = file->f_mapping;
402 i_mmap_lock_write(mapping);
403 __vma_link_file(vma, mapping);
404 i_mmap_unlock_write(mapping);
405 }
406}
407
408static int vma_link(struct mm_struct *mm, struct vm_area_struct *vma)
409{
410 VMA_ITERATOR(vmi, mm, 0);
411
412 vma_iter_config(&vmi, vma->vm_start, vma->vm_end);
413 if (vma_iter_prealloc(&vmi, vma))
414 return -ENOMEM;
415
416 vma_start_write(vma);
417 vma_iter_store(&vmi, vma);
418 vma_link_file(vma);
419 mm->map_count++;
420 validate_mm(mm);
421 return 0;
422}
423
424/*
425 * init_multi_vma_prep() - Initializer for struct vma_prepare
426 * @vp: The vma_prepare struct
427 * @vma: The vma that will be altered once locked
428 * @next: The next vma if it is to be adjusted
429 * @remove: The first vma to be removed
430 * @remove2: The second vma to be removed
431 */
432static inline void init_multi_vma_prep(struct vma_prepare *vp,
433 struct vm_area_struct *vma, struct vm_area_struct *next,
434 struct vm_area_struct *remove, struct vm_area_struct *remove2)
435{
436 memset(vp, 0, sizeof(struct vma_prepare));
437 vp->vma = vma;
438 vp->anon_vma = vma->anon_vma;
439 vp->remove = remove;
440 vp->remove2 = remove2;
441 vp->adj_next = next;
442 if (!vp->anon_vma && next)
443 vp->anon_vma = next->anon_vma;
444
445 vp->file = vma->vm_file;
446 if (vp->file)
447 vp->mapping = vma->vm_file->f_mapping;
448
449}
450
451/*
452 * init_vma_prep() - Initializer wrapper for vma_prepare struct
453 * @vp: The vma_prepare struct
454 * @vma: The vma that will be altered once locked
455 */
456static inline void init_vma_prep(struct vma_prepare *vp,
457 struct vm_area_struct *vma)
458{
459 init_multi_vma_prep(vp, vma, NULL, NULL, NULL);
460}
461
462
463/*
464 * vma_prepare() - Helper function for handling locking VMAs prior to altering
465 * @vp: The initialized vma_prepare struct
466 */
467static inline void vma_prepare(struct vma_prepare *vp)
468{
469 if (vp->file) {
470 uprobe_munmap(vp->vma, vp->vma->vm_start, vp->vma->vm_end);
471
472 if (vp->adj_next)
473 uprobe_munmap(vp->adj_next, vp->adj_next->vm_start,
474 vp->adj_next->vm_end);
475
476 i_mmap_lock_write(vp->mapping);
477 if (vp->insert && vp->insert->vm_file) {
478 /*
479 * Put into interval tree now, so instantiated pages
480 * are visible to arm/parisc __flush_dcache_page
481 * throughout; but we cannot insert into address
482 * space until vma start or end is updated.
483 */
484 __vma_link_file(vp->insert,
485 vp->insert->vm_file->f_mapping);
486 }
487 }
488
489 if (vp->anon_vma) {
490 anon_vma_lock_write(vp->anon_vma);
491 anon_vma_interval_tree_pre_update_vma(vp->vma);
492 if (vp->adj_next)
493 anon_vma_interval_tree_pre_update_vma(vp->adj_next);
494 }
495
496 if (vp->file) {
497 flush_dcache_mmap_lock(vp->mapping);
498 vma_interval_tree_remove(vp->vma, &vp->mapping->i_mmap);
499 if (vp->adj_next)
500 vma_interval_tree_remove(vp->adj_next,
501 &vp->mapping->i_mmap);
502 }
503
504}
505
506/*
507 * vma_complete- Helper function for handling the unlocking after altering VMAs,
508 * or for inserting a VMA.
509 *
510 * @vp: The vma_prepare struct
511 * @vmi: The vma iterator
512 * @mm: The mm_struct
513 */
514static inline void vma_complete(struct vma_prepare *vp,
515 struct vma_iterator *vmi, struct mm_struct *mm)
516{
517 if (vp->file) {
518 if (vp->adj_next)
519 vma_interval_tree_insert(vp->adj_next,
520 &vp->mapping->i_mmap);
521 vma_interval_tree_insert(vp->vma, &vp->mapping->i_mmap);
522 flush_dcache_mmap_unlock(vp->mapping);
523 }
524
525 if (vp->remove && vp->file) {
526 __remove_shared_vm_struct(vp->remove, vp->mapping);
527 if (vp->remove2)
528 __remove_shared_vm_struct(vp->remove2, vp->mapping);
529 } else if (vp->insert) {
530 /*
531 * split_vma has split insert from vma, and needs
532 * us to insert it before dropping the locks
533 * (it may either follow vma or precede it).
534 */
535 vma_iter_store(vmi, vp->insert);
536 mm->map_count++;
537 }
538
539 if (vp->anon_vma) {
540 anon_vma_interval_tree_post_update_vma(vp->vma);
541 if (vp->adj_next)
542 anon_vma_interval_tree_post_update_vma(vp->adj_next);
543 anon_vma_unlock_write(vp->anon_vma);
544 }
545
546 if (vp->file) {
547 i_mmap_unlock_write(vp->mapping);
548 uprobe_mmap(vp->vma);
549
550 if (vp->adj_next)
551 uprobe_mmap(vp->adj_next);
552 }
553
554 if (vp->remove) {
555again:
556 vma_mark_detached(vp->remove, true);
557 if (vp->file) {
558 uprobe_munmap(vp->remove, vp->remove->vm_start,
559 vp->remove->vm_end);
560 fput(vp->file);
561 }
562 if (vp->remove->anon_vma)
563 anon_vma_merge(vp->vma, vp->remove);
564 mm->map_count--;
565 mpol_put(vma_policy(vp->remove));
566 if (!vp->remove2)
567 WARN_ON_ONCE(vp->vma->vm_end < vp->remove->vm_end);
568 vm_area_free(vp->remove);
569
570 /*
571 * In mprotect's case 6 (see comments on vma_merge),
572 * we are removing both mid and next vmas
573 */
574 if (vp->remove2) {
575 vp->remove = vp->remove2;
576 vp->remove2 = NULL;
577 goto again;
578 }
579 }
580 if (vp->insert && vp->file)
581 uprobe_mmap(vp->insert);
582 validate_mm(mm);
583}
584
585/*
586 * dup_anon_vma() - Helper function to duplicate anon_vma
587 * @dst: The destination VMA
588 * @src: The source VMA
589 * @dup: Pointer to the destination VMA when successful.
590 *
591 * Returns: 0 on success.
592 */
593static inline int dup_anon_vma(struct vm_area_struct *dst,
594 struct vm_area_struct *src, struct vm_area_struct **dup)
595{
596 /*
597 * Easily overlooked: when mprotect shifts the boundary, make sure the
598 * expanding vma has anon_vma set if the shrinking vma had, to cover any
599 * anon pages imported.
600 */
601 if (src->anon_vma && !dst->anon_vma) {
602 int ret;
603
604 vma_assert_write_locked(dst);
605 dst->anon_vma = src->anon_vma;
606 ret = anon_vma_clone(dst, src);
607 if (ret)
608 return ret;
609
610 *dup = dst;
611 }
612
613 return 0;
614}
615
616/*
617 * vma_expand - Expand an existing VMA
618 *
619 * @vmi: The vma iterator
620 * @vma: The vma to expand
621 * @start: The start of the vma
622 * @end: The exclusive end of the vma
623 * @pgoff: The page offset of vma
624 * @next: The current of next vma.
625 *
626 * Expand @vma to @start and @end. Can expand off the start and end. Will
627 * expand over @next if it's different from @vma and @end == @next->vm_end.
628 * Checking if the @vma can expand and merge with @next needs to be handled by
629 * the caller.
630 *
631 * Returns: 0 on success
632 */
633int vma_expand(struct vma_iterator *vmi, struct vm_area_struct *vma,
634 unsigned long start, unsigned long end, pgoff_t pgoff,
635 struct vm_area_struct *next)
636{
637 struct vm_area_struct *anon_dup = NULL;
638 bool remove_next = false;
639 struct vma_prepare vp;
640
641 vma_start_write(vma);
642 if (next && (vma != next) && (end == next->vm_end)) {
643 int ret;
644
645 remove_next = true;
646 vma_start_write(next);
647 ret = dup_anon_vma(vma, next, &anon_dup);
648 if (ret)
649 return ret;
650 }
651
652 init_multi_vma_prep(&vp, vma, NULL, remove_next ? next : NULL, NULL);
653 /* Not merging but overwriting any part of next is not handled. */
654 VM_WARN_ON(next && !vp.remove &&
655 next != vma && end > next->vm_start);
656 /* Only handles expanding */
657 VM_WARN_ON(vma->vm_start < start || vma->vm_end > end);
658
659 /* Note: vma iterator must be pointing to 'start' */
660 vma_iter_config(vmi, start, end);
661 if (vma_iter_prealloc(vmi, vma))
662 goto nomem;
663
664 vma_prepare(&vp);
665 vma_adjust_trans_huge(vma, start, end, 0);
666 vma_set_range(vma, start, end, pgoff);
667 vma_iter_store(vmi, vma);
668
669 vma_complete(&vp, vmi, vma->vm_mm);
670 return 0;
671
672nomem:
673 if (anon_dup)
674 unlink_anon_vmas(anon_dup);
675 return -ENOMEM;
676}
677
678/*
679 * vma_shrink() - Reduce an existing VMAs memory area
680 * @vmi: The vma iterator
681 * @vma: The VMA to modify
682 * @start: The new start
683 * @end: The new end
684 *
685 * Returns: 0 on success, -ENOMEM otherwise
686 */
687int vma_shrink(struct vma_iterator *vmi, struct vm_area_struct *vma,
688 unsigned long start, unsigned long end, pgoff_t pgoff)
689{
690 struct vma_prepare vp;
691
692 WARN_ON((vma->vm_start != start) && (vma->vm_end != end));
693
694 if (vma->vm_start < start)
695 vma_iter_config(vmi, vma->vm_start, start);
696 else
697 vma_iter_config(vmi, end, vma->vm_end);
698
699 if (vma_iter_prealloc(vmi, NULL))
700 return -ENOMEM;
701
702 vma_start_write(vma);
703
704 init_vma_prep(&vp, vma);
705 vma_prepare(&vp);
706 vma_adjust_trans_huge(vma, start, end, 0);
707
708 vma_iter_clear(vmi);
709 vma_set_range(vma, start, end, pgoff);
710 vma_complete(&vp, vmi, vma->vm_mm);
711 return 0;
712}
713
714/*
715 * If the vma has a ->close operation then the driver probably needs to release
716 * per-vma resources, so we don't attempt to merge those if the caller indicates
717 * the current vma may be removed as part of the merge.
718 */
719static inline bool is_mergeable_vma(struct vm_area_struct *vma,
720 struct file *file, unsigned long vm_flags,
721 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
722 struct anon_vma_name *anon_name, bool may_remove_vma)
723{
724 /*
725 * VM_SOFTDIRTY should not prevent from VMA merging, if we
726 * match the flags but dirty bit -- the caller should mark
727 * merged VMA as dirty. If dirty bit won't be excluded from
728 * comparison, we increase pressure on the memory system forcing
729 * the kernel to generate new VMAs when old one could be
730 * extended instead.
731 */
732 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
733 return false;
734 if (vma->vm_file != file)
735 return false;
736 if (may_remove_vma && vma->vm_ops && vma->vm_ops->close)
737 return false;
738 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
739 return false;
740 if (!anon_vma_name_eq(anon_vma_name(vma), anon_name))
741 return false;
742 return true;
743}
744
745static inline bool is_mergeable_anon_vma(struct anon_vma *anon_vma1,
746 struct anon_vma *anon_vma2, struct vm_area_struct *vma)
747{
748 /*
749 * The list_is_singular() test is to avoid merging VMA cloned from
750 * parents. This can improve scalability caused by anon_vma lock.
751 */
752 if ((!anon_vma1 || !anon_vma2) && (!vma ||
753 list_is_singular(&vma->anon_vma_chain)))
754 return true;
755 return anon_vma1 == anon_vma2;
756}
757
758/*
759 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
760 * in front of (at a lower virtual address and file offset than) the vma.
761 *
762 * We cannot merge two vmas if they have differently assigned (non-NULL)
763 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
764 *
765 * We don't check here for the merged mmap wrapping around the end of pagecache
766 * indices (16TB on ia32) because do_mmap() does not permit mmap's which
767 * wrap, nor mmaps which cover the final page at index -1UL.
768 *
769 * We assume the vma may be removed as part of the merge.
770 */
771static bool
772can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
773 struct anon_vma *anon_vma, struct file *file,
774 pgoff_t vm_pgoff, struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
775 struct anon_vma_name *anon_name)
776{
777 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name, true) &&
778 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
779 if (vma->vm_pgoff == vm_pgoff)
780 return true;
781 }
782 return false;
783}
784
785/*
786 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
787 * beyond (at a higher virtual address and file offset than) the vma.
788 *
789 * We cannot merge two vmas if they have differently assigned (non-NULL)
790 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
791 *
792 * We assume that vma is not removed as part of the merge.
793 */
794static bool
795can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
796 struct anon_vma *anon_vma, struct file *file,
797 pgoff_t vm_pgoff, struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
798 struct anon_vma_name *anon_name)
799{
800 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name, false) &&
801 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
802 pgoff_t vm_pglen;
803 vm_pglen = vma_pages(vma);
804 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
805 return true;
806 }
807 return false;
808}
809
810/*
811 * Given a mapping request (addr,end,vm_flags,file,pgoff,anon_name),
812 * figure out whether that can be merged with its predecessor or its
813 * successor. Or both (it neatly fills a hole).
814 *
815 * In most cases - when called for mmap, brk or mremap - [addr,end) is
816 * certain not to be mapped by the time vma_merge is called; but when
817 * called for mprotect, it is certain to be already mapped (either at
818 * an offset within prev, or at the start of next), and the flags of
819 * this area are about to be changed to vm_flags - and the no-change
820 * case has already been eliminated.
821 *
822 * The following mprotect cases have to be considered, where **** is
823 * the area passed down from mprotect_fixup, never extending beyond one
824 * vma, PPPP is the previous vma, CCCC is a concurrent vma that starts
825 * at the same address as **** and is of the same or larger span, and
826 * NNNN the next vma after ****:
827 *
828 * **** **** ****
829 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPCCCCCC
830 * cannot merge might become might become
831 * PPNNNNNNNNNN PPPPPPPPPPCC
832 * mmap, brk or case 4 below case 5 below
833 * mremap move:
834 * **** ****
835 * PPPP NNNN PPPPCCCCNNNN
836 * might become might become
837 * PPPPPPPPPPPP 1 or PPPPPPPPPPPP 6 or
838 * PPPPPPPPNNNN 2 or PPPPPPPPNNNN 7 or
839 * PPPPNNNNNNNN 3 PPPPNNNNNNNN 8
840 *
841 * It is important for case 8 that the vma CCCC overlapping the
842 * region **** is never going to extended over NNNN. Instead NNNN must
843 * be extended in region **** and CCCC must be removed. This way in
844 * all cases where vma_merge succeeds, the moment vma_merge drops the
845 * rmap_locks, the properties of the merged vma will be already
846 * correct for the whole merged range. Some of those properties like
847 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
848 * be correct for the whole merged range immediately after the
849 * rmap_locks are released. Otherwise if NNNN would be removed and
850 * CCCC would be extended over the NNNN range, remove_migration_ptes
851 * or other rmap walkers (if working on addresses beyond the "end"
852 * parameter) may establish ptes with the wrong permissions of CCCC
853 * instead of the right permissions of NNNN.
854 *
855 * In the code below:
856 * PPPP is represented by *prev
857 * CCCC is represented by *curr or not represented at all (NULL)
858 * NNNN is represented by *next or not represented at all (NULL)
859 * **** is not represented - it will be merged and the vma containing the
860 * area is returned, or the function will return NULL
861 */
862static struct vm_area_struct
863*vma_merge(struct vma_iterator *vmi, struct vm_area_struct *prev,
864 struct vm_area_struct *src, unsigned long addr, unsigned long end,
865 unsigned long vm_flags, pgoff_t pgoff, struct mempolicy *policy,
866 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
867 struct anon_vma_name *anon_name)
868{
869 struct mm_struct *mm = src->vm_mm;
870 struct anon_vma *anon_vma = src->anon_vma;
871 struct file *file = src->vm_file;
872 struct vm_area_struct *curr, *next, *res;
873 struct vm_area_struct *vma, *adjust, *remove, *remove2;
874 struct vm_area_struct *anon_dup = NULL;
875 struct vma_prepare vp;
876 pgoff_t vma_pgoff;
877 int err = 0;
878 bool merge_prev = false;
879 bool merge_next = false;
880 bool vma_expanded = false;
881 unsigned long vma_start = addr;
882 unsigned long vma_end = end;
883 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
884 long adj_start = 0;
885
886 /*
887 * We later require that vma->vm_flags == vm_flags,
888 * so this tests vma->vm_flags & VM_SPECIAL, too.
889 */
890 if (vm_flags & VM_SPECIAL)
891 return NULL;
892
893 /* Does the input range span an existing VMA? (cases 5 - 8) */
894 curr = find_vma_intersection(mm, prev ? prev->vm_end : 0, end);
895
896 if (!curr || /* cases 1 - 4 */
897 end == curr->vm_end) /* cases 6 - 8, adjacent VMA */
898 next = vma_lookup(mm, end);
899 else
900 next = NULL; /* case 5 */
901
902 if (prev) {
903 vma_start = prev->vm_start;
904 vma_pgoff = prev->vm_pgoff;
905
906 /* Can we merge the predecessor? */
907 if (addr == prev->vm_end && mpol_equal(vma_policy(prev), policy)
908 && can_vma_merge_after(prev, vm_flags, anon_vma, file,
909 pgoff, vm_userfaultfd_ctx, anon_name)) {
910 merge_prev = true;
911 vma_prev(vmi);
912 }
913 }
914
915 /* Can we merge the successor? */
916 if (next && mpol_equal(policy, vma_policy(next)) &&
917 can_vma_merge_before(next, vm_flags, anon_vma, file, pgoff+pglen,
918 vm_userfaultfd_ctx, anon_name)) {
919 merge_next = true;
920 }
921
922 /* Verify some invariant that must be enforced by the caller. */
923 VM_WARN_ON(prev && addr <= prev->vm_start);
924 VM_WARN_ON(curr && (addr != curr->vm_start || end > curr->vm_end));
925 VM_WARN_ON(addr >= end);
926
927 if (!merge_prev && !merge_next)
928 return NULL; /* Not mergeable. */
929
930 if (merge_prev)
931 vma_start_write(prev);
932
933 res = vma = prev;
934 remove = remove2 = adjust = NULL;
935
936 /* Can we merge both the predecessor and the successor? */
937 if (merge_prev && merge_next &&
938 is_mergeable_anon_vma(prev->anon_vma, next->anon_vma, NULL)) {
939 vma_start_write(next);
940 remove = next; /* case 1 */
941 vma_end = next->vm_end;
942 err = dup_anon_vma(prev, next, &anon_dup);
943 if (curr) { /* case 6 */
944 vma_start_write(curr);
945 remove = curr;
946 remove2 = next;
947 /*
948 * Note that the dup_anon_vma below cannot overwrite err
949 * since the first caller would do nothing unless next
950 * has an anon_vma.
951 */
952 if (!next->anon_vma)
953 err = dup_anon_vma(prev, curr, &anon_dup);
954 }
955 } else if (merge_prev) { /* case 2 */
956 if (curr) {
957 vma_start_write(curr);
958 if (end == curr->vm_end) { /* case 7 */
959 /*
960 * can_vma_merge_after() assumed we would not be
961 * removing prev vma, so it skipped the check
962 * for vm_ops->close, but we are removing curr
963 */
964 if (curr->vm_ops && curr->vm_ops->close)
965 err = -EINVAL;
966 remove = curr;
967 } else { /* case 5 */
968 adjust = curr;
969 adj_start = (end - curr->vm_start);
970 }
971 if (!err)
972 err = dup_anon_vma(prev, curr, &anon_dup);
973 }
974 } else { /* merge_next */
975 vma_start_write(next);
976 res = next;
977 if (prev && addr < prev->vm_end) { /* case 4 */
978 vma_start_write(prev);
979 vma_end = addr;
980 adjust = next;
981 adj_start = -(prev->vm_end - addr);
982 err = dup_anon_vma(next, prev, &anon_dup);
983 } else {
984 /*
985 * Note that cases 3 and 8 are the ONLY ones where prev
986 * is permitted to be (but is not necessarily) NULL.
987 */
988 vma = next; /* case 3 */
989 vma_start = addr;
990 vma_end = next->vm_end;
991 vma_pgoff = next->vm_pgoff - pglen;
992 if (curr) { /* case 8 */
993 vma_pgoff = curr->vm_pgoff;
994 vma_start_write(curr);
995 remove = curr;
996 err = dup_anon_vma(next, curr, &anon_dup);
997 }
998 }
999 }
1000
1001 /* Error in anon_vma clone. */
1002 if (err)
1003 goto anon_vma_fail;
1004
1005 if (vma_start < vma->vm_start || vma_end > vma->vm_end)
1006 vma_expanded = true;
1007
1008 if (vma_expanded) {
1009 vma_iter_config(vmi, vma_start, vma_end);
1010 } else {
1011 vma_iter_config(vmi, adjust->vm_start + adj_start,
1012 adjust->vm_end);
1013 }
1014
1015 if (vma_iter_prealloc(vmi, vma))
1016 goto prealloc_fail;
1017
1018 init_multi_vma_prep(&vp, vma, adjust, remove, remove2);
1019 VM_WARN_ON(vp.anon_vma && adjust && adjust->anon_vma &&
1020 vp.anon_vma != adjust->anon_vma);
1021
1022 vma_prepare(&vp);
1023 vma_adjust_trans_huge(vma, vma_start, vma_end, adj_start);
1024 vma_set_range(vma, vma_start, vma_end, vma_pgoff);
1025
1026 if (vma_expanded)
1027 vma_iter_store(vmi, vma);
1028
1029 if (adj_start) {
1030 adjust->vm_start += adj_start;
1031 adjust->vm_pgoff += adj_start >> PAGE_SHIFT;
1032 if (adj_start < 0) {
1033 WARN_ON(vma_expanded);
1034 vma_iter_store(vmi, next);
1035 }
1036 }
1037
1038 vma_complete(&vp, vmi, mm);
1039 khugepaged_enter_vma(res, vm_flags);
1040 return res;
1041
1042prealloc_fail:
1043 if (anon_dup)
1044 unlink_anon_vmas(anon_dup);
1045
1046anon_vma_fail:
1047 vma_iter_set(vmi, addr);
1048 vma_iter_load(vmi);
1049 return NULL;
1050}
1051
1052/*
1053 * Rough compatibility check to quickly see if it's even worth looking
1054 * at sharing an anon_vma.
1055 *
1056 * They need to have the same vm_file, and the flags can only differ
1057 * in things that mprotect may change.
1058 *
1059 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1060 * we can merge the two vma's. For example, we refuse to merge a vma if
1061 * there is a vm_ops->close() function, because that indicates that the
1062 * driver is doing some kind of reference counting. But that doesn't
1063 * really matter for the anon_vma sharing case.
1064 */
1065static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1066{
1067 return a->vm_end == b->vm_start &&
1068 mpol_equal(vma_policy(a), vma_policy(b)) &&
1069 a->vm_file == b->vm_file &&
1070 !((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) &&
1071 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1072}
1073
1074/*
1075 * Do some basic sanity checking to see if we can re-use the anon_vma
1076 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1077 * the same as 'old', the other will be the new one that is trying
1078 * to share the anon_vma.
1079 *
1080 * NOTE! This runs with mmap_lock held for reading, so it is possible that
1081 * the anon_vma of 'old' is concurrently in the process of being set up
1082 * by another page fault trying to merge _that_. But that's ok: if it
1083 * is being set up, that automatically means that it will be a singleton
1084 * acceptable for merging, so we can do all of this optimistically. But
1085 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1086 *
1087 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1088 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1089 * is to return an anon_vma that is "complex" due to having gone through
1090 * a fork).
1091 *
1092 * We also make sure that the two vma's are compatible (adjacent,
1093 * and with the same memory policies). That's all stable, even with just
1094 * a read lock on the mmap_lock.
1095 */
1096static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1097{
1098 if (anon_vma_compatible(a, b)) {
1099 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1100
1101 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1102 return anon_vma;
1103 }
1104 return NULL;
1105}
1106
1107/*
1108 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1109 * neighbouring vmas for a suitable anon_vma, before it goes off
1110 * to allocate a new anon_vma. It checks because a repetitive
1111 * sequence of mprotects and faults may otherwise lead to distinct
1112 * anon_vmas being allocated, preventing vma merge in subsequent
1113 * mprotect.
1114 */
1115struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1116{
1117 MA_STATE(mas, &vma->vm_mm->mm_mt, vma->vm_end, vma->vm_end);
1118 struct anon_vma *anon_vma = NULL;
1119 struct vm_area_struct *prev, *next;
1120
1121 /* Try next first. */
1122 next = mas_walk(&mas);
1123 if (next) {
1124 anon_vma = reusable_anon_vma(next, vma, next);
1125 if (anon_vma)
1126 return anon_vma;
1127 }
1128
1129 prev = mas_prev(&mas, 0);
1130 VM_BUG_ON_VMA(prev != vma, vma);
1131 prev = mas_prev(&mas, 0);
1132 /* Try prev next. */
1133 if (prev)
1134 anon_vma = reusable_anon_vma(prev, prev, vma);
1135
1136 /*
1137 * We might reach here with anon_vma == NULL if we can't find
1138 * any reusable anon_vma.
1139 * There's no absolute need to look only at touching neighbours:
1140 * we could search further afield for "compatible" anon_vmas.
1141 * But it would probably just be a waste of time searching,
1142 * or lead to too many vmas hanging off the same anon_vma.
1143 * We're trying to allow mprotect remerging later on,
1144 * not trying to minimize memory used for anon_vmas.
1145 */
1146 return anon_vma;
1147}
1148
1149/*
1150 * If a hint addr is less than mmap_min_addr change hint to be as
1151 * low as possible but still greater than mmap_min_addr
1152 */
1153static inline unsigned long round_hint_to_min(unsigned long hint)
1154{
1155 hint &= PAGE_MASK;
1156 if (((void *)hint != NULL) &&
1157 (hint < mmap_min_addr))
1158 return PAGE_ALIGN(mmap_min_addr);
1159 return hint;
1160}
1161
1162bool mlock_future_ok(struct mm_struct *mm, unsigned long flags,
1163 unsigned long bytes)
1164{
1165 unsigned long locked_pages, limit_pages;
1166
1167 if (!(flags & VM_LOCKED) || capable(CAP_IPC_LOCK))
1168 return true;
1169
1170 locked_pages = bytes >> PAGE_SHIFT;
1171 locked_pages += mm->locked_vm;
1172
1173 limit_pages = rlimit(RLIMIT_MEMLOCK);
1174 limit_pages >>= PAGE_SHIFT;
1175
1176 return locked_pages <= limit_pages;
1177}
1178
1179static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1180{
1181 if (S_ISREG(inode->i_mode))
1182 return MAX_LFS_FILESIZE;
1183
1184 if (S_ISBLK(inode->i_mode))
1185 return MAX_LFS_FILESIZE;
1186
1187 if (S_ISSOCK(inode->i_mode))
1188 return MAX_LFS_FILESIZE;
1189
1190 /* Special "we do even unsigned file positions" case */
1191 if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1192 return 0;
1193
1194 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1195 return ULONG_MAX;
1196}
1197
1198static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1199 unsigned long pgoff, unsigned long len)
1200{
1201 u64 maxsize = file_mmap_size_max(file, inode);
1202
1203 if (maxsize && len > maxsize)
1204 return false;
1205 maxsize -= len;
1206 if (pgoff > maxsize >> PAGE_SHIFT)
1207 return false;
1208 return true;
1209}
1210
1211/*
1212 * The caller must write-lock current->mm->mmap_lock.
1213 */
1214unsigned long do_mmap(struct file *file, unsigned long addr,
1215 unsigned long len, unsigned long prot,
1216 unsigned long flags, vm_flags_t vm_flags,
1217 unsigned long pgoff, unsigned long *populate,
1218 struct list_head *uf)
1219{
1220 struct mm_struct *mm = current->mm;
1221 int pkey = 0;
1222
1223 *populate = 0;
1224
1225 if (!len)
1226 return -EINVAL;
1227
1228 /*
1229 * Does the application expect PROT_READ to imply PROT_EXEC?
1230 *
1231 * (the exception is when the underlying filesystem is noexec
1232 * mounted, in which case we don't add PROT_EXEC.)
1233 */
1234 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1235 if (!(file && path_noexec(&file->f_path)))
1236 prot |= PROT_EXEC;
1237
1238 /* force arch specific MAP_FIXED handling in get_unmapped_area */
1239 if (flags & MAP_FIXED_NOREPLACE)
1240 flags |= MAP_FIXED;
1241
1242 if (!(flags & MAP_FIXED))
1243 addr = round_hint_to_min(addr);
1244
1245 /* Careful about overflows.. */
1246 len = PAGE_ALIGN(len);
1247 if (!len)
1248 return -ENOMEM;
1249
1250 /* offset overflow? */
1251 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1252 return -EOVERFLOW;
1253
1254 /* Too many mappings? */
1255 if (mm->map_count > sysctl_max_map_count)
1256 return -ENOMEM;
1257
1258 /* Obtain the address to map to. we verify (or select) it and ensure
1259 * that it represents a valid section of the address space.
1260 */
1261 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1262 if (IS_ERR_VALUE(addr))
1263 return addr;
1264
1265 if (flags & MAP_FIXED_NOREPLACE) {
1266 if (find_vma_intersection(mm, addr, addr + len))
1267 return -EEXIST;
1268 }
1269
1270 if (prot == PROT_EXEC) {
1271 pkey = execute_only_pkey(mm);
1272 if (pkey < 0)
1273 pkey = 0;
1274 }
1275
1276 /* Do simple checking here so the lower-level routines won't have
1277 * to. we assume access permissions have been handled by the open
1278 * of the memory object, so we don't do any here.
1279 */
1280 vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1281 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1282
1283 if (flags & MAP_LOCKED)
1284 if (!can_do_mlock())
1285 return -EPERM;
1286
1287 if (!mlock_future_ok(mm, vm_flags, len))
1288 return -EAGAIN;
1289
1290 if (file) {
1291 struct inode *inode = file_inode(file);
1292 unsigned long flags_mask;
1293
1294 if (!file_mmap_ok(file, inode, pgoff, len))
1295 return -EOVERFLOW;
1296
1297 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1298
1299 switch (flags & MAP_TYPE) {
1300 case MAP_SHARED:
1301 /*
1302 * Force use of MAP_SHARED_VALIDATE with non-legacy
1303 * flags. E.g. MAP_SYNC is dangerous to use with
1304 * MAP_SHARED as you don't know which consistency model
1305 * you will get. We silently ignore unsupported flags
1306 * with MAP_SHARED to preserve backward compatibility.
1307 */
1308 flags &= LEGACY_MAP_MASK;
1309 fallthrough;
1310 case MAP_SHARED_VALIDATE:
1311 if (flags & ~flags_mask)
1312 return -EOPNOTSUPP;
1313 if (prot & PROT_WRITE) {
1314 if (!(file->f_mode & FMODE_WRITE))
1315 return -EACCES;
1316 if (IS_SWAPFILE(file->f_mapping->host))
1317 return -ETXTBSY;
1318 }
1319
1320 /*
1321 * Make sure we don't allow writing to an append-only
1322 * file..
1323 */
1324 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1325 return -EACCES;
1326
1327 vm_flags |= VM_SHARED | VM_MAYSHARE;
1328 if (!(file->f_mode & FMODE_WRITE))
1329 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1330 fallthrough;
1331 case MAP_PRIVATE:
1332 if (!(file->f_mode & FMODE_READ))
1333 return -EACCES;
1334 if (path_noexec(&file->f_path)) {
1335 if (vm_flags & VM_EXEC)
1336 return -EPERM;
1337 vm_flags &= ~VM_MAYEXEC;
1338 }
1339
1340 if (!file->f_op->mmap)
1341 return -ENODEV;
1342 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1343 return -EINVAL;
1344 break;
1345
1346 default:
1347 return -EINVAL;
1348 }
1349 } else {
1350 switch (flags & MAP_TYPE) {
1351 case MAP_SHARED:
1352 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1353 return -EINVAL;
1354 /*
1355 * Ignore pgoff.
1356 */
1357 pgoff = 0;
1358 vm_flags |= VM_SHARED | VM_MAYSHARE;
1359 break;
1360 case MAP_PRIVATE:
1361 /*
1362 * Set pgoff according to addr for anon_vma.
1363 */
1364 pgoff = addr >> PAGE_SHIFT;
1365 break;
1366 default:
1367 return -EINVAL;
1368 }
1369 }
1370
1371 /*
1372 * Set 'VM_NORESERVE' if we should not account for the
1373 * memory use of this mapping.
1374 */
1375 if (flags & MAP_NORESERVE) {
1376 /* We honor MAP_NORESERVE if allowed to overcommit */
1377 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1378 vm_flags |= VM_NORESERVE;
1379
1380 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1381 if (file && is_file_hugepages(file))
1382 vm_flags |= VM_NORESERVE;
1383 }
1384
1385 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1386 if (!IS_ERR_VALUE(addr) &&
1387 ((vm_flags & VM_LOCKED) ||
1388 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1389 *populate = len;
1390 return addr;
1391}
1392
1393unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1394 unsigned long prot, unsigned long flags,
1395 unsigned long fd, unsigned long pgoff)
1396{
1397 struct file *file = NULL;
1398 unsigned long retval;
1399
1400 if (!(flags & MAP_ANONYMOUS)) {
1401 audit_mmap_fd(fd, flags);
1402 file = fget(fd);
1403 if (!file)
1404 return -EBADF;
1405 if (is_file_hugepages(file)) {
1406 len = ALIGN(len, huge_page_size(hstate_file(file)));
1407 } else if (unlikely(flags & MAP_HUGETLB)) {
1408 retval = -EINVAL;
1409 goto out_fput;
1410 }
1411 } else if (flags & MAP_HUGETLB) {
1412 struct hstate *hs;
1413
1414 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1415 if (!hs)
1416 return -EINVAL;
1417
1418 len = ALIGN(len, huge_page_size(hs));
1419 /*
1420 * VM_NORESERVE is used because the reservations will be
1421 * taken when vm_ops->mmap() is called
1422 */
1423 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1424 VM_NORESERVE,
1425 HUGETLB_ANONHUGE_INODE,
1426 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1427 if (IS_ERR(file))
1428 return PTR_ERR(file);
1429 }
1430
1431 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1432out_fput:
1433 if (file)
1434 fput(file);
1435 return retval;
1436}
1437
1438SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1439 unsigned long, prot, unsigned long, flags,
1440 unsigned long, fd, unsigned long, pgoff)
1441{
1442 return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1443}
1444
1445#ifdef __ARCH_WANT_SYS_OLD_MMAP
1446struct mmap_arg_struct {
1447 unsigned long addr;
1448 unsigned long len;
1449 unsigned long prot;
1450 unsigned long flags;
1451 unsigned long fd;
1452 unsigned long offset;
1453};
1454
1455SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1456{
1457 struct mmap_arg_struct a;
1458
1459 if (copy_from_user(&a, arg, sizeof(a)))
1460 return -EFAULT;
1461 if (offset_in_page(a.offset))
1462 return -EINVAL;
1463
1464 return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1465 a.offset >> PAGE_SHIFT);
1466}
1467#endif /* __ARCH_WANT_SYS_OLD_MMAP */
1468
1469static bool vm_ops_needs_writenotify(const struct vm_operations_struct *vm_ops)
1470{
1471 return vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite);
1472}
1473
1474static bool vma_is_shared_writable(struct vm_area_struct *vma)
1475{
1476 return (vma->vm_flags & (VM_WRITE | VM_SHARED)) ==
1477 (VM_WRITE | VM_SHARED);
1478}
1479
1480static bool vma_fs_can_writeback(struct vm_area_struct *vma)
1481{
1482 /* No managed pages to writeback. */
1483 if (vma->vm_flags & VM_PFNMAP)
1484 return false;
1485
1486 return vma->vm_file && vma->vm_file->f_mapping &&
1487 mapping_can_writeback(vma->vm_file->f_mapping);
1488}
1489
1490/*
1491 * Does this VMA require the underlying folios to have their dirty state
1492 * tracked?
1493 */
1494bool vma_needs_dirty_tracking(struct vm_area_struct *vma)
1495{
1496 /* Only shared, writable VMAs require dirty tracking. */
1497 if (!vma_is_shared_writable(vma))
1498 return false;
1499
1500 /* Does the filesystem need to be notified? */
1501 if (vm_ops_needs_writenotify(vma->vm_ops))
1502 return true;
1503
1504 /*
1505 * Even if the filesystem doesn't indicate a need for writenotify, if it
1506 * can writeback, dirty tracking is still required.
1507 */
1508 return vma_fs_can_writeback(vma);
1509}
1510
1511/*
1512 * Some shared mappings will want the pages marked read-only
1513 * to track write events. If so, we'll downgrade vm_page_prot
1514 * to the private version (using protection_map[] without the
1515 * VM_SHARED bit).
1516 */
1517int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1518{
1519 /* If it was private or non-writable, the write bit is already clear */
1520 if (!vma_is_shared_writable(vma))
1521 return 0;
1522
1523 /* The backer wishes to know when pages are first written to? */
1524 if (vm_ops_needs_writenotify(vma->vm_ops))
1525 return 1;
1526
1527 /* The open routine did something to the protections that pgprot_modify
1528 * won't preserve? */
1529 if (pgprot_val(vm_page_prot) !=
1530 pgprot_val(vm_pgprot_modify(vm_page_prot, vma->vm_flags)))
1531 return 0;
1532
1533 /*
1534 * Do we need to track softdirty? hugetlb does not support softdirty
1535 * tracking yet.
1536 */
1537 if (vma_soft_dirty_enabled(vma) && !is_vm_hugetlb_page(vma))
1538 return 1;
1539
1540 /* Do we need write faults for uffd-wp tracking? */
1541 if (userfaultfd_wp(vma))
1542 return 1;
1543
1544 /* Can the mapping track the dirty pages? */
1545 return vma_fs_can_writeback(vma);
1546}
1547
1548/*
1549 * We account for memory if it's a private writeable mapping,
1550 * not hugepages and VM_NORESERVE wasn't set.
1551 */
1552static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1553{
1554 /*
1555 * hugetlb has its own accounting separate from the core VM
1556 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1557 */
1558 if (file && is_file_hugepages(file))
1559 return 0;
1560
1561 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1562}
1563
1564/**
1565 * unmapped_area() - Find an area between the low_limit and the high_limit with
1566 * the correct alignment and offset, all from @info. Note: current->mm is used
1567 * for the search.
1568 *
1569 * @info: The unmapped area information including the range [low_limit -
1570 * high_limit), the alignment offset and mask.
1571 *
1572 * Return: A memory address or -ENOMEM.
1573 */
1574static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1575{
1576 unsigned long length, gap;
1577 unsigned long low_limit, high_limit;
1578 struct vm_area_struct *tmp;
1579
1580 MA_STATE(mas, ¤t->mm->mm_mt, 0, 0);
1581
1582 /* Adjust search length to account for worst case alignment overhead */
1583 length = info->length + info->align_mask;
1584 if (length < info->length)
1585 return -ENOMEM;
1586
1587 low_limit = info->low_limit;
1588 if (low_limit < mmap_min_addr)
1589 low_limit = mmap_min_addr;
1590 high_limit = info->high_limit;
1591retry:
1592 if (mas_empty_area(&mas, low_limit, high_limit - 1, length))
1593 return -ENOMEM;
1594
1595 gap = mas.index;
1596 gap += (info->align_offset - gap) & info->align_mask;
1597 tmp = mas_next(&mas, ULONG_MAX);
1598 if (tmp && (tmp->vm_flags & VM_STARTGAP_FLAGS)) { /* Avoid prev check if possible */
1599 if (vm_start_gap(tmp) < gap + length - 1) {
1600 low_limit = tmp->vm_end;
1601 mas_reset(&mas);
1602 goto retry;
1603 }
1604 } else {
1605 tmp = mas_prev(&mas, 0);
1606 if (tmp && vm_end_gap(tmp) > gap) {
1607 low_limit = vm_end_gap(tmp);
1608 mas_reset(&mas);
1609 goto retry;
1610 }
1611 }
1612
1613 return gap;
1614}
1615
1616/**
1617 * unmapped_area_topdown() - Find an area between the low_limit and the
1618 * high_limit with the correct alignment and offset at the highest available
1619 * address, all from @info. Note: current->mm is used for the search.
1620 *
1621 * @info: The unmapped area information including the range [low_limit -
1622 * high_limit), the alignment offset and mask.
1623 *
1624 * Return: A memory address or -ENOMEM.
1625 */
1626static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1627{
1628 unsigned long length, gap, gap_end;
1629 unsigned long low_limit, high_limit;
1630 struct vm_area_struct *tmp;
1631
1632 MA_STATE(mas, ¤t->mm->mm_mt, 0, 0);
1633 /* Adjust search length to account for worst case alignment overhead */
1634 length = info->length + info->align_mask;
1635 if (length < info->length)
1636 return -ENOMEM;
1637
1638 low_limit = info->low_limit;
1639 if (low_limit < mmap_min_addr)
1640 low_limit = mmap_min_addr;
1641 high_limit = info->high_limit;
1642retry:
1643 if (mas_empty_area_rev(&mas, low_limit, high_limit - 1, length))
1644 return -ENOMEM;
1645
1646 gap = mas.last + 1 - info->length;
1647 gap -= (gap - info->align_offset) & info->align_mask;
1648 gap_end = mas.last;
1649 tmp = mas_next(&mas, ULONG_MAX);
1650 if (tmp && (tmp->vm_flags & VM_STARTGAP_FLAGS)) { /* Avoid prev check if possible */
1651 if (vm_start_gap(tmp) <= gap_end) {
1652 high_limit = vm_start_gap(tmp);
1653 mas_reset(&mas);
1654 goto retry;
1655 }
1656 } else {
1657 tmp = mas_prev(&mas, 0);
1658 if (tmp && vm_end_gap(tmp) > gap) {
1659 high_limit = tmp->vm_start;
1660 mas_reset(&mas);
1661 goto retry;
1662 }
1663 }
1664
1665 return gap;
1666}
1667
1668/*
1669 * Search for an unmapped address range.
1670 *
1671 * We are looking for a range that:
1672 * - does not intersect with any VMA;
1673 * - is contained within the [low_limit, high_limit) interval;
1674 * - is at least the desired size.
1675 * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
1676 */
1677unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
1678{
1679 unsigned long addr;
1680
1681 if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
1682 addr = unmapped_area_topdown(info);
1683 else
1684 addr = unmapped_area(info);
1685
1686 trace_vm_unmapped_area(addr, info);
1687 return addr;
1688}
1689
1690/* Get an address range which is currently unmapped.
1691 * For shmat() with addr=0.
1692 *
1693 * Ugly calling convention alert:
1694 * Return value with the low bits set means error value,
1695 * ie
1696 * if (ret & ~PAGE_MASK)
1697 * error = ret;
1698 *
1699 * This function "knows" that -ENOMEM has the bits set.
1700 */
1701unsigned long
1702generic_get_unmapped_area(struct file *filp, unsigned long addr,
1703 unsigned long len, unsigned long pgoff,
1704 unsigned long flags)
1705{
1706 struct mm_struct *mm = current->mm;
1707 struct vm_area_struct *vma, *prev;
1708 struct vm_unmapped_area_info info;
1709 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
1710
1711 if (len > mmap_end - mmap_min_addr)
1712 return -ENOMEM;
1713
1714 if (flags & MAP_FIXED)
1715 return addr;
1716
1717 if (addr) {
1718 addr = PAGE_ALIGN(addr);
1719 vma = find_vma_prev(mm, addr, &prev);
1720 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
1721 (!vma || addr + len <= vm_start_gap(vma)) &&
1722 (!prev || addr >= vm_end_gap(prev)))
1723 return addr;
1724 }
1725
1726 info.flags = 0;
1727 info.length = len;
1728 info.low_limit = mm->mmap_base;
1729 info.high_limit = mmap_end;
1730 info.align_mask = 0;
1731 info.align_offset = 0;
1732 return vm_unmapped_area(&info);
1733}
1734
1735#ifndef HAVE_ARCH_UNMAPPED_AREA
1736unsigned long
1737arch_get_unmapped_area(struct file *filp, unsigned long addr,
1738 unsigned long len, unsigned long pgoff,
1739 unsigned long flags)
1740{
1741 return generic_get_unmapped_area(filp, addr, len, pgoff, flags);
1742}
1743#endif
1744
1745/*
1746 * This mmap-allocator allocates new areas top-down from below the
1747 * stack's low limit (the base):
1748 */
1749unsigned long
1750generic_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
1751 unsigned long len, unsigned long pgoff,
1752 unsigned long flags)
1753{
1754 struct vm_area_struct *vma, *prev;
1755 struct mm_struct *mm = current->mm;
1756 struct vm_unmapped_area_info info;
1757 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
1758
1759 /* requested length too big for entire address space */
1760 if (len > mmap_end - mmap_min_addr)
1761 return -ENOMEM;
1762
1763 if (flags & MAP_FIXED)
1764 return addr;
1765
1766 /* requesting a specific address */
1767 if (addr) {
1768 addr = PAGE_ALIGN(addr);
1769 vma = find_vma_prev(mm, addr, &prev);
1770 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
1771 (!vma || addr + len <= vm_start_gap(vma)) &&
1772 (!prev || addr >= vm_end_gap(prev)))
1773 return addr;
1774 }
1775
1776 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
1777 info.length = len;
1778 info.low_limit = PAGE_SIZE;
1779 info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
1780 info.align_mask = 0;
1781 info.align_offset = 0;
1782 addr = vm_unmapped_area(&info);
1783
1784 /*
1785 * A failed mmap() very likely causes application failure,
1786 * so fall back to the bottom-up function here. This scenario
1787 * can happen with large stack limits and large mmap()
1788 * allocations.
1789 */
1790 if (offset_in_page(addr)) {
1791 VM_BUG_ON(addr != -ENOMEM);
1792 info.flags = 0;
1793 info.low_limit = TASK_UNMAPPED_BASE;
1794 info.high_limit = mmap_end;
1795 addr = vm_unmapped_area(&info);
1796 }
1797
1798 return addr;
1799}
1800
1801#ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1802unsigned long
1803arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
1804 unsigned long len, unsigned long pgoff,
1805 unsigned long flags)
1806{
1807 return generic_get_unmapped_area_topdown(filp, addr, len, pgoff, flags);
1808}
1809#endif
1810
1811unsigned long
1812get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1813 unsigned long pgoff, unsigned long flags)
1814{
1815 unsigned long (*get_area)(struct file *, unsigned long,
1816 unsigned long, unsigned long, unsigned long);
1817
1818 unsigned long error = arch_mmap_check(addr, len, flags);
1819 if (error)
1820 return error;
1821
1822 /* Careful about overflows.. */
1823 if (len > TASK_SIZE)
1824 return -ENOMEM;
1825
1826 get_area = current->mm->get_unmapped_area;
1827 if (file) {
1828 if (file->f_op->get_unmapped_area)
1829 get_area = file->f_op->get_unmapped_area;
1830 } else if (flags & MAP_SHARED) {
1831 /*
1832 * mmap_region() will call shmem_zero_setup() to create a file,
1833 * so use shmem's get_unmapped_area in case it can be huge.
1834 */
1835 get_area = shmem_get_unmapped_area;
1836 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
1837 /* Ensures that larger anonymous mappings are THP aligned. */
1838 get_area = thp_get_unmapped_area;
1839 }
1840
1841 /* Always treat pgoff as zero for anonymous memory. */
1842 if (!file)
1843 pgoff = 0;
1844
1845 addr = get_area(file, addr, len, pgoff, flags);
1846 if (IS_ERR_VALUE(addr))
1847 return addr;
1848
1849 if (addr > TASK_SIZE - len)
1850 return -ENOMEM;
1851 if (offset_in_page(addr))
1852 return -EINVAL;
1853
1854 error = security_mmap_addr(addr);
1855 return error ? error : addr;
1856}
1857
1858EXPORT_SYMBOL(get_unmapped_area);
1859
1860/**
1861 * find_vma_intersection() - Look up the first VMA which intersects the interval
1862 * @mm: The process address space.
1863 * @start_addr: The inclusive start user address.
1864 * @end_addr: The exclusive end user address.
1865 *
1866 * Returns: The first VMA within the provided range, %NULL otherwise. Assumes
1867 * start_addr < end_addr.
1868 */
1869struct vm_area_struct *find_vma_intersection(struct mm_struct *mm,
1870 unsigned long start_addr,
1871 unsigned long end_addr)
1872{
1873 unsigned long index = start_addr;
1874
1875 mmap_assert_locked(mm);
1876 return mt_find(&mm->mm_mt, &index, end_addr - 1);
1877}
1878EXPORT_SYMBOL(find_vma_intersection);
1879
1880/**
1881 * find_vma() - Find the VMA for a given address, or the next VMA.
1882 * @mm: The mm_struct to check
1883 * @addr: The address
1884 *
1885 * Returns: The VMA associated with addr, or the next VMA.
1886 * May return %NULL in the case of no VMA at addr or above.
1887 */
1888struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1889{
1890 unsigned long index = addr;
1891
1892 mmap_assert_locked(mm);
1893 return mt_find(&mm->mm_mt, &index, ULONG_MAX);
1894}
1895EXPORT_SYMBOL(find_vma);
1896
1897/**
1898 * find_vma_prev() - Find the VMA for a given address, or the next vma and
1899 * set %pprev to the previous VMA, if any.
1900 * @mm: The mm_struct to check
1901 * @addr: The address
1902 * @pprev: The pointer to set to the previous VMA
1903 *
1904 * Note that RCU lock is missing here since the external mmap_lock() is used
1905 * instead.
1906 *
1907 * Returns: The VMA associated with @addr, or the next vma.
1908 * May return %NULL in the case of no vma at addr or above.
1909 */
1910struct vm_area_struct *
1911find_vma_prev(struct mm_struct *mm, unsigned long addr,
1912 struct vm_area_struct **pprev)
1913{
1914 struct vm_area_struct *vma;
1915 MA_STATE(mas, &mm->mm_mt, addr, addr);
1916
1917 vma = mas_walk(&mas);
1918 *pprev = mas_prev(&mas, 0);
1919 if (!vma)
1920 vma = mas_next(&mas, ULONG_MAX);
1921 return vma;
1922}
1923
1924/*
1925 * Verify that the stack growth is acceptable and
1926 * update accounting. This is shared with both the
1927 * grow-up and grow-down cases.
1928 */
1929static int acct_stack_growth(struct vm_area_struct *vma,
1930 unsigned long size, unsigned long grow)
1931{
1932 struct mm_struct *mm = vma->vm_mm;
1933 unsigned long new_start;
1934
1935 /* address space limit tests */
1936 if (!may_expand_vm(mm, vma->vm_flags, grow))
1937 return -ENOMEM;
1938
1939 /* Stack limit test */
1940 if (size > rlimit(RLIMIT_STACK))
1941 return -ENOMEM;
1942
1943 /* mlock limit tests */
1944 if (!mlock_future_ok(mm, vma->vm_flags, grow << PAGE_SHIFT))
1945 return -ENOMEM;
1946
1947 /* Check to ensure the stack will not grow into a hugetlb-only region */
1948 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1949 vma->vm_end - size;
1950 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1951 return -EFAULT;
1952
1953 /*
1954 * Overcommit.. This must be the final test, as it will
1955 * update security statistics.
1956 */
1957 if (security_vm_enough_memory_mm(mm, grow))
1958 return -ENOMEM;
1959
1960 return 0;
1961}
1962
1963#if defined(CONFIG_STACK_GROWSUP)
1964/*
1965 * PA-RISC uses this for its stack.
1966 * vma is the last one with address > vma->vm_end. Have to extend vma.
1967 */
1968static int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1969{
1970 struct mm_struct *mm = vma->vm_mm;
1971 struct vm_area_struct *next;
1972 unsigned long gap_addr;
1973 int error = 0;
1974 MA_STATE(mas, &mm->mm_mt, vma->vm_start, address);
1975
1976 if (!(vma->vm_flags & VM_GROWSUP))
1977 return -EFAULT;
1978
1979 /* Guard against exceeding limits of the address space. */
1980 address &= PAGE_MASK;
1981 if (address >= (TASK_SIZE & PAGE_MASK))
1982 return -ENOMEM;
1983 address += PAGE_SIZE;
1984
1985 /* Enforce stack_guard_gap */
1986 gap_addr = address + stack_guard_gap;
1987
1988 /* Guard against overflow */
1989 if (gap_addr < address || gap_addr > TASK_SIZE)
1990 gap_addr = TASK_SIZE;
1991
1992 next = find_vma_intersection(mm, vma->vm_end, gap_addr);
1993 if (next && vma_is_accessible(next)) {
1994 if (!(next->vm_flags & VM_GROWSUP))
1995 return -ENOMEM;
1996 /* Check that both stack segments have the same anon_vma? */
1997 }
1998
1999 if (next)
2000 mas_prev_range(&mas, address);
2001
2002 __mas_set_range(&mas, vma->vm_start, address - 1);
2003 if (mas_preallocate(&mas, vma, GFP_KERNEL))
2004 return -ENOMEM;
2005
2006 /* We must make sure the anon_vma is allocated. */
2007 if (unlikely(anon_vma_prepare(vma))) {
2008 mas_destroy(&mas);
2009 return -ENOMEM;
2010 }
2011
2012 /* Lock the VMA before expanding to prevent concurrent page faults */
2013 vma_start_write(vma);
2014 /*
2015 * vma->vm_start/vm_end cannot change under us because the caller
2016 * is required to hold the mmap_lock in read mode. We need the
2017 * anon_vma lock to serialize against concurrent expand_stacks.
2018 */
2019 anon_vma_lock_write(vma->anon_vma);
2020
2021 /* Somebody else might have raced and expanded it already */
2022 if (address > vma->vm_end) {
2023 unsigned long size, grow;
2024
2025 size = address - vma->vm_start;
2026 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2027
2028 error = -ENOMEM;
2029 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2030 error = acct_stack_growth(vma, size, grow);
2031 if (!error) {
2032 /*
2033 * We only hold a shared mmap_lock lock here, so
2034 * we need to protect against concurrent vma
2035 * expansions. anon_vma_lock_write() doesn't
2036 * help here, as we don't guarantee that all
2037 * growable vmas in a mm share the same root
2038 * anon vma. So, we reuse mm->page_table_lock
2039 * to guard against concurrent vma expansions.
2040 */
2041 spin_lock(&mm->page_table_lock);
2042 if (vma->vm_flags & VM_LOCKED)
2043 mm->locked_vm += grow;
2044 vm_stat_account(mm, vma->vm_flags, grow);
2045 anon_vma_interval_tree_pre_update_vma(vma);
2046 vma->vm_end = address;
2047 /* Overwrite old entry in mtree. */
2048 mas_store_prealloc(&mas, vma);
2049 anon_vma_interval_tree_post_update_vma(vma);
2050 spin_unlock(&mm->page_table_lock);
2051
2052 perf_event_mmap(vma);
2053 }
2054 }
2055 }
2056 anon_vma_unlock_write(vma->anon_vma);
2057 mas_destroy(&mas);
2058 validate_mm(mm);
2059 return error;
2060}
2061#endif /* CONFIG_STACK_GROWSUP */
2062
2063/*
2064 * vma is the first one with address < vma->vm_start. Have to extend vma.
2065 * mmap_lock held for writing.
2066 */
2067int expand_downwards(struct vm_area_struct *vma, unsigned long address)
2068{
2069 struct mm_struct *mm = vma->vm_mm;
2070 MA_STATE(mas, &mm->mm_mt, vma->vm_start, vma->vm_start);
2071 struct vm_area_struct *prev;
2072 int error = 0;
2073
2074 if (!(vma->vm_flags & VM_GROWSDOWN))
2075 return -EFAULT;
2076
2077 address &= PAGE_MASK;
2078 if (address < mmap_min_addr || address < FIRST_USER_ADDRESS)
2079 return -EPERM;
2080
2081 /* Enforce stack_guard_gap */
2082 prev = mas_prev(&mas, 0);
2083 /* Check that both stack segments have the same anon_vma? */
2084 if (prev) {
2085 if (!(prev->vm_flags & VM_GROWSDOWN) &&
2086 vma_is_accessible(prev) &&
2087 (address - prev->vm_end < stack_guard_gap))
2088 return -ENOMEM;
2089 }
2090
2091 if (prev)
2092 mas_next_range(&mas, vma->vm_start);
2093
2094 __mas_set_range(&mas, address, vma->vm_end - 1);
2095 if (mas_preallocate(&mas, vma, GFP_KERNEL))
2096 return -ENOMEM;
2097
2098 /* We must make sure the anon_vma is allocated. */
2099 if (unlikely(anon_vma_prepare(vma))) {
2100 mas_destroy(&mas);
2101 return -ENOMEM;
2102 }
2103
2104 /* Lock the VMA before expanding to prevent concurrent page faults */
2105 vma_start_write(vma);
2106 /*
2107 * vma->vm_start/vm_end cannot change under us because the caller
2108 * is required to hold the mmap_lock in read mode. We need the
2109 * anon_vma lock to serialize against concurrent expand_stacks.
2110 */
2111 anon_vma_lock_write(vma->anon_vma);
2112
2113 /* Somebody else might have raced and expanded it already */
2114 if (address < vma->vm_start) {
2115 unsigned long size, grow;
2116
2117 size = vma->vm_end - address;
2118 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2119
2120 error = -ENOMEM;
2121 if (grow <= vma->vm_pgoff) {
2122 error = acct_stack_growth(vma, size, grow);
2123 if (!error) {
2124 /*
2125 * We only hold a shared mmap_lock lock here, so
2126 * we need to protect against concurrent vma
2127 * expansions. anon_vma_lock_write() doesn't
2128 * help here, as we don't guarantee that all
2129 * growable vmas in a mm share the same root
2130 * anon vma. So, we reuse mm->page_table_lock
2131 * to guard against concurrent vma expansions.
2132 */
2133 spin_lock(&mm->page_table_lock);
2134 if (vma->vm_flags & VM_LOCKED)
2135 mm->locked_vm += grow;
2136 vm_stat_account(mm, vma->vm_flags, grow);
2137 anon_vma_interval_tree_pre_update_vma(vma);
2138 vma->vm_start = address;
2139 vma->vm_pgoff -= grow;
2140 /* Overwrite old entry in mtree. */
2141 mas_store_prealloc(&mas, vma);
2142 anon_vma_interval_tree_post_update_vma(vma);
2143 spin_unlock(&mm->page_table_lock);
2144
2145 perf_event_mmap(vma);
2146 }
2147 }
2148 }
2149 anon_vma_unlock_write(vma->anon_vma);
2150 mas_destroy(&mas);
2151 validate_mm(mm);
2152 return error;
2153}
2154
2155/* enforced gap between the expanding stack and other mappings. */
2156unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2157
2158static int __init cmdline_parse_stack_guard_gap(char *p)
2159{
2160 unsigned long val;
2161 char *endptr;
2162
2163 val = simple_strtoul(p, &endptr, 10);
2164 if (!*endptr)
2165 stack_guard_gap = val << PAGE_SHIFT;
2166
2167 return 1;
2168}
2169__setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2170
2171#ifdef CONFIG_STACK_GROWSUP
2172int expand_stack_locked(struct vm_area_struct *vma, unsigned long address)
2173{
2174 return expand_upwards(vma, address);
2175}
2176
2177struct vm_area_struct *find_extend_vma_locked(struct mm_struct *mm, unsigned long addr)
2178{
2179 struct vm_area_struct *vma, *prev;
2180
2181 addr &= PAGE_MASK;
2182 vma = find_vma_prev(mm, addr, &prev);
2183 if (vma && (vma->vm_start <= addr))
2184 return vma;
2185 if (!prev)
2186 return NULL;
2187 if (expand_stack_locked(prev, addr))
2188 return NULL;
2189 if (prev->vm_flags & VM_LOCKED)
2190 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2191 return prev;
2192}
2193#else
2194int expand_stack_locked(struct vm_area_struct *vma, unsigned long address)
2195{
2196 return expand_downwards(vma, address);
2197}
2198
2199struct vm_area_struct *find_extend_vma_locked(struct mm_struct *mm, unsigned long addr)
2200{
2201 struct vm_area_struct *vma;
2202 unsigned long start;
2203
2204 addr &= PAGE_MASK;
2205 vma = find_vma(mm, addr);
2206 if (!vma)
2207 return NULL;
2208 if (vma->vm_start <= addr)
2209 return vma;
2210 start = vma->vm_start;
2211 if (expand_stack_locked(vma, addr))
2212 return NULL;
2213 if (vma->vm_flags & VM_LOCKED)
2214 populate_vma_page_range(vma, addr, start, NULL);
2215 return vma;
2216}
2217#endif
2218
2219#if defined(CONFIG_STACK_GROWSUP)
2220
2221#define vma_expand_up(vma,addr) expand_upwards(vma, addr)
2222#define vma_expand_down(vma, addr) (-EFAULT)
2223
2224#else
2225
2226#define vma_expand_up(vma,addr) (-EFAULT)
2227#define vma_expand_down(vma, addr) expand_downwards(vma, addr)
2228
2229#endif
2230
2231/*
2232 * expand_stack(): legacy interface for page faulting. Don't use unless
2233 * you have to.
2234 *
2235 * This is called with the mm locked for reading, drops the lock, takes
2236 * the lock for writing, tries to look up a vma again, expands it if
2237 * necessary, and downgrades the lock to reading again.
2238 *
2239 * If no vma is found or it can't be expanded, it returns NULL and has
2240 * dropped the lock.
2241 */
2242struct vm_area_struct *expand_stack(struct mm_struct *mm, unsigned long addr)
2243{
2244 struct vm_area_struct *vma, *prev;
2245
2246 mmap_read_unlock(mm);
2247 if (mmap_write_lock_killable(mm))
2248 return NULL;
2249
2250 vma = find_vma_prev(mm, addr, &prev);
2251 if (vma && vma->vm_start <= addr)
2252 goto success;
2253
2254 if (prev && !vma_expand_up(prev, addr)) {
2255 vma = prev;
2256 goto success;
2257 }
2258
2259 if (vma && !vma_expand_down(vma, addr))
2260 goto success;
2261
2262 mmap_write_unlock(mm);
2263 return NULL;
2264
2265success:
2266 mmap_write_downgrade(mm);
2267 return vma;
2268}
2269
2270/*
2271 * Ok - we have the memory areas we should free on a maple tree so release them,
2272 * and do the vma updates.
2273 *
2274 * Called with the mm semaphore held.
2275 */
2276static inline void remove_mt(struct mm_struct *mm, struct ma_state *mas)
2277{
2278 unsigned long nr_accounted = 0;
2279 struct vm_area_struct *vma;
2280
2281 /* Update high watermark before we lower total_vm */
2282 update_hiwater_vm(mm);
2283 mas_for_each(mas, vma, ULONG_MAX) {
2284 long nrpages = vma_pages(vma);
2285
2286 if (vma->vm_flags & VM_ACCOUNT)
2287 nr_accounted += nrpages;
2288 vm_stat_account(mm, vma->vm_flags, -nrpages);
2289 remove_vma(vma, false);
2290 }
2291 vm_unacct_memory(nr_accounted);
2292}
2293
2294/*
2295 * Get rid of page table information in the indicated region.
2296 *
2297 * Called with the mm semaphore held.
2298 */
2299static void unmap_region(struct mm_struct *mm, struct ma_state *mas,
2300 struct vm_area_struct *vma, struct vm_area_struct *prev,
2301 struct vm_area_struct *next, unsigned long start,
2302 unsigned long end, unsigned long tree_end, bool mm_wr_locked)
2303{
2304 struct mmu_gather tlb;
2305 unsigned long mt_start = mas->index;
2306
2307 lru_add_drain();
2308 tlb_gather_mmu(&tlb, mm);
2309 update_hiwater_rss(mm);
2310 unmap_vmas(&tlb, mas, vma, start, end, tree_end, mm_wr_locked);
2311 mas_set(mas, mt_start);
2312 free_pgtables(&tlb, mas, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2313 next ? next->vm_start : USER_PGTABLES_CEILING,
2314 mm_wr_locked);
2315 tlb_finish_mmu(&tlb);
2316}
2317
2318/*
2319 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it
2320 * has already been checked or doesn't make sense to fail.
2321 * VMA Iterator will point to the end VMA.
2322 */
2323static int __split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma,
2324 unsigned long addr, int new_below)
2325{
2326 struct vma_prepare vp;
2327 struct vm_area_struct *new;
2328 int err;
2329
2330 WARN_ON(vma->vm_start >= addr);
2331 WARN_ON(vma->vm_end <= addr);
2332
2333 if (vma->vm_ops && vma->vm_ops->may_split) {
2334 err = vma->vm_ops->may_split(vma, addr);
2335 if (err)
2336 return err;
2337 }
2338
2339 new = vm_area_dup(vma);
2340 if (!new)
2341 return -ENOMEM;
2342
2343 if (new_below) {
2344 new->vm_end = addr;
2345 } else {
2346 new->vm_start = addr;
2347 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2348 }
2349
2350 err = -ENOMEM;
2351 vma_iter_config(vmi, new->vm_start, new->vm_end);
2352 if (vma_iter_prealloc(vmi, new))
2353 goto out_free_vma;
2354
2355 err = vma_dup_policy(vma, new);
2356 if (err)
2357 goto out_free_vmi;
2358
2359 err = anon_vma_clone(new, vma);
2360 if (err)
2361 goto out_free_mpol;
2362
2363 if (new->vm_file)
2364 get_file(new->vm_file);
2365
2366 if (new->vm_ops && new->vm_ops->open)
2367 new->vm_ops->open(new);
2368
2369 vma_start_write(vma);
2370 vma_start_write(new);
2371
2372 init_vma_prep(&vp, vma);
2373 vp.insert = new;
2374 vma_prepare(&vp);
2375 vma_adjust_trans_huge(vma, vma->vm_start, addr, 0);
2376
2377 if (new_below) {
2378 vma->vm_start = addr;
2379 vma->vm_pgoff += (addr - new->vm_start) >> PAGE_SHIFT;
2380 } else {
2381 vma->vm_end = addr;
2382 }
2383
2384 /* vma_complete stores the new vma */
2385 vma_complete(&vp, vmi, vma->vm_mm);
2386
2387 /* Success. */
2388 if (new_below)
2389 vma_next(vmi);
2390 return 0;
2391
2392out_free_mpol:
2393 mpol_put(vma_policy(new));
2394out_free_vmi:
2395 vma_iter_free(vmi);
2396out_free_vma:
2397 vm_area_free(new);
2398 return err;
2399}
2400
2401/*
2402 * Split a vma into two pieces at address 'addr', a new vma is allocated
2403 * either for the first part or the tail.
2404 */
2405static int split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma,
2406 unsigned long addr, int new_below)
2407{
2408 if (vma->vm_mm->map_count >= sysctl_max_map_count)
2409 return -ENOMEM;
2410
2411 return __split_vma(vmi, vma, addr, new_below);
2412}
2413
2414/*
2415 * We are about to modify one or multiple of a VMA's flags, policy, userfaultfd
2416 * context and anonymous VMA name within the range [start, end).
2417 *
2418 * As a result, we might be able to merge the newly modified VMA range with an
2419 * adjacent VMA with identical properties.
2420 *
2421 * If no merge is possible and the range does not span the entirety of the VMA,
2422 * we then need to split the VMA to accommodate the change.
2423 *
2424 * The function returns either the merged VMA, the original VMA if a split was
2425 * required instead, or an error if the split failed.
2426 */
2427struct vm_area_struct *vma_modify(struct vma_iterator *vmi,
2428 struct vm_area_struct *prev,
2429 struct vm_area_struct *vma,
2430 unsigned long start, unsigned long end,
2431 unsigned long vm_flags,
2432 struct mempolicy *policy,
2433 struct vm_userfaultfd_ctx uffd_ctx,
2434 struct anon_vma_name *anon_name)
2435{
2436 pgoff_t pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
2437 struct vm_area_struct *merged;
2438
2439 merged = vma_merge(vmi, prev, vma, start, end, vm_flags,
2440 pgoff, policy, uffd_ctx, anon_name);
2441 if (merged)
2442 return merged;
2443
2444 if (vma->vm_start < start) {
2445 int err = split_vma(vmi, vma, start, 1);
2446
2447 if (err)
2448 return ERR_PTR(err);
2449 }
2450
2451 if (vma->vm_end > end) {
2452 int err = split_vma(vmi, vma, end, 0);
2453
2454 if (err)
2455 return ERR_PTR(err);
2456 }
2457
2458 return vma;
2459}
2460
2461/*
2462 * Attempt to merge a newly mapped VMA with those adjacent to it. The caller
2463 * must ensure that [start, end) does not overlap any existing VMA.
2464 */
2465static struct vm_area_struct
2466*vma_merge_new_vma(struct vma_iterator *vmi, struct vm_area_struct *prev,
2467 struct vm_area_struct *vma, unsigned long start,
2468 unsigned long end, pgoff_t pgoff)
2469{
2470 return vma_merge(vmi, prev, vma, start, end, vma->vm_flags, pgoff,
2471 vma_policy(vma), vma->vm_userfaultfd_ctx, anon_vma_name(vma));
2472}
2473
2474/*
2475 * Expand vma by delta bytes, potentially merging with an immediately adjacent
2476 * VMA with identical properties.
2477 */
2478struct vm_area_struct *vma_merge_extend(struct vma_iterator *vmi,
2479 struct vm_area_struct *vma,
2480 unsigned long delta)
2481{
2482 pgoff_t pgoff = vma->vm_pgoff + vma_pages(vma);
2483
2484 /* vma is specified as prev, so case 1 or 2 will apply. */
2485 return vma_merge(vmi, vma, vma, vma->vm_end, vma->vm_end + delta,
2486 vma->vm_flags, pgoff, vma_policy(vma),
2487 vma->vm_userfaultfd_ctx, anon_vma_name(vma));
2488}
2489
2490/*
2491 * do_vmi_align_munmap() - munmap the aligned region from @start to @end.
2492 * @vmi: The vma iterator
2493 * @vma: The starting vm_area_struct
2494 * @mm: The mm_struct
2495 * @start: The aligned start address to munmap.
2496 * @end: The aligned end address to munmap.
2497 * @uf: The userfaultfd list_head
2498 * @unlock: Set to true to drop the mmap_lock. unlocking only happens on
2499 * success.
2500 *
2501 * Return: 0 on success and drops the lock if so directed, error and leaves the
2502 * lock held otherwise.
2503 */
2504static int
2505do_vmi_align_munmap(struct vma_iterator *vmi, struct vm_area_struct *vma,
2506 struct mm_struct *mm, unsigned long start,
2507 unsigned long end, struct list_head *uf, bool unlock)
2508{
2509 struct vm_area_struct *prev, *next = NULL;
2510 struct maple_tree mt_detach;
2511 int count = 0;
2512 int error = -ENOMEM;
2513 unsigned long locked_vm = 0;
2514 MA_STATE(mas_detach, &mt_detach, 0, 0);
2515 mt_init_flags(&mt_detach, vmi->mas.tree->ma_flags & MT_FLAGS_LOCK_MASK);
2516 mt_on_stack(mt_detach);
2517
2518 /*
2519 * If we need to split any vma, do it now to save pain later.
2520 *
2521 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2522 * unmapped vm_area_struct will remain in use: so lower split_vma
2523 * places tmp vma above, and higher split_vma places tmp vma below.
2524 */
2525
2526 /* Does it split the first one? */
2527 if (start > vma->vm_start) {
2528
2529 /*
2530 * Make sure that map_count on return from munmap() will
2531 * not exceed its limit; but let map_count go just above
2532 * its limit temporarily, to help free resources as expected.
2533 */
2534 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2535 goto map_count_exceeded;
2536
2537 error = __split_vma(vmi, vma, start, 1);
2538 if (error)
2539 goto start_split_failed;
2540 }
2541
2542 /*
2543 * Detach a range of VMAs from the mm. Using next as a temp variable as
2544 * it is always overwritten.
2545 */
2546 next = vma;
2547 do {
2548 /* Does it split the end? */
2549 if (next->vm_end > end) {
2550 error = __split_vma(vmi, next, end, 0);
2551 if (error)
2552 goto end_split_failed;
2553 }
2554 vma_start_write(next);
2555 mas_set(&mas_detach, count);
2556 error = mas_store_gfp(&mas_detach, next, GFP_KERNEL);
2557 if (error)
2558 goto munmap_gather_failed;
2559 vma_mark_detached(next, true);
2560 if (next->vm_flags & VM_LOCKED)
2561 locked_vm += vma_pages(next);
2562
2563 count++;
2564 if (unlikely(uf)) {
2565 /*
2566 * If userfaultfd_unmap_prep returns an error the vmas
2567 * will remain split, but userland will get a
2568 * highly unexpected error anyway. This is no
2569 * different than the case where the first of the two
2570 * __split_vma fails, but we don't undo the first
2571 * split, despite we could. This is unlikely enough
2572 * failure that it's not worth optimizing it for.
2573 */
2574 error = userfaultfd_unmap_prep(next, start, end, uf);
2575
2576 if (error)
2577 goto userfaultfd_error;
2578 }
2579#ifdef CONFIG_DEBUG_VM_MAPLE_TREE
2580 BUG_ON(next->vm_start < start);
2581 BUG_ON(next->vm_start > end);
2582#endif
2583 } for_each_vma_range(*vmi, next, end);
2584
2585#if defined(CONFIG_DEBUG_VM_MAPLE_TREE)
2586 /* Make sure no VMAs are about to be lost. */
2587 {
2588 MA_STATE(test, &mt_detach, 0, 0);
2589 struct vm_area_struct *vma_mas, *vma_test;
2590 int test_count = 0;
2591
2592 vma_iter_set(vmi, start);
2593 rcu_read_lock();
2594 vma_test = mas_find(&test, count - 1);
2595 for_each_vma_range(*vmi, vma_mas, end) {
2596 BUG_ON(vma_mas != vma_test);
2597 test_count++;
2598 vma_test = mas_next(&test, count - 1);
2599 }
2600 rcu_read_unlock();
2601 BUG_ON(count != test_count);
2602 }
2603#endif
2604
2605 while (vma_iter_addr(vmi) > start)
2606 vma_iter_prev_range(vmi);
2607
2608 error = vma_iter_clear_gfp(vmi, start, end, GFP_KERNEL);
2609 if (error)
2610 goto clear_tree_failed;
2611
2612 /* Point of no return */
2613 mm->locked_vm -= locked_vm;
2614 mm->map_count -= count;
2615 if (unlock)
2616 mmap_write_downgrade(mm);
2617
2618 prev = vma_iter_prev_range(vmi);
2619 next = vma_next(vmi);
2620 if (next)
2621 vma_iter_prev_range(vmi);
2622
2623 /*
2624 * We can free page tables without write-locking mmap_lock because VMAs
2625 * were isolated before we downgraded mmap_lock.
2626 */
2627 mas_set(&mas_detach, 1);
2628 unmap_region(mm, &mas_detach, vma, prev, next, start, end, count,
2629 !unlock);
2630 /* Statistics and freeing VMAs */
2631 mas_set(&mas_detach, 0);
2632 remove_mt(mm, &mas_detach);
2633 validate_mm(mm);
2634 if (unlock)
2635 mmap_read_unlock(mm);
2636
2637 __mt_destroy(&mt_detach);
2638 return 0;
2639
2640clear_tree_failed:
2641userfaultfd_error:
2642munmap_gather_failed:
2643end_split_failed:
2644 mas_set(&mas_detach, 0);
2645 mas_for_each(&mas_detach, next, end)
2646 vma_mark_detached(next, false);
2647
2648 __mt_destroy(&mt_detach);
2649start_split_failed:
2650map_count_exceeded:
2651 validate_mm(mm);
2652 return error;
2653}
2654
2655/*
2656 * do_vmi_munmap() - munmap a given range.
2657 * @vmi: The vma iterator
2658 * @mm: The mm_struct
2659 * @start: The start address to munmap
2660 * @len: The length of the range to munmap
2661 * @uf: The userfaultfd list_head
2662 * @unlock: set to true if the user wants to drop the mmap_lock on success
2663 *
2664 * This function takes a @mas that is either pointing to the previous VMA or set
2665 * to MA_START and sets it up to remove the mapping(s). The @len will be
2666 * aligned and any arch_unmap work will be preformed.
2667 *
2668 * Return: 0 on success and drops the lock if so directed, error and leaves the
2669 * lock held otherwise.
2670 */
2671int do_vmi_munmap(struct vma_iterator *vmi, struct mm_struct *mm,
2672 unsigned long start, size_t len, struct list_head *uf,
2673 bool unlock)
2674{
2675 unsigned long end;
2676 struct vm_area_struct *vma;
2677
2678 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2679 return -EINVAL;
2680
2681 end = start + PAGE_ALIGN(len);
2682 if (end == start)
2683 return -EINVAL;
2684
2685 /* arch_unmap() might do unmaps itself. */
2686 arch_unmap(mm, start, end);
2687
2688 /* Find the first overlapping VMA */
2689 vma = vma_find(vmi, end);
2690 if (!vma) {
2691 if (unlock)
2692 mmap_write_unlock(mm);
2693 return 0;
2694 }
2695
2696 return do_vmi_align_munmap(vmi, vma, mm, start, end, uf, unlock);
2697}
2698
2699/* do_munmap() - Wrapper function for non-maple tree aware do_munmap() calls.
2700 * @mm: The mm_struct
2701 * @start: The start address to munmap
2702 * @len: The length to be munmapped.
2703 * @uf: The userfaultfd list_head
2704 *
2705 * Return: 0 on success, error otherwise.
2706 */
2707int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2708 struct list_head *uf)
2709{
2710 VMA_ITERATOR(vmi, mm, start);
2711
2712 return do_vmi_munmap(&vmi, mm, start, len, uf, false);
2713}
2714
2715unsigned long mmap_region(struct file *file, unsigned long addr,
2716 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
2717 struct list_head *uf)
2718{
2719 struct mm_struct *mm = current->mm;
2720 struct vm_area_struct *vma = NULL;
2721 struct vm_area_struct *next, *prev, *merge;
2722 pgoff_t pglen = len >> PAGE_SHIFT;
2723 unsigned long charged = 0;
2724 unsigned long end = addr + len;
2725 unsigned long merge_start = addr, merge_end = end;
2726 bool writable_file_mapping = false;
2727 pgoff_t vm_pgoff;
2728 int error;
2729 VMA_ITERATOR(vmi, mm, addr);
2730
2731 /* Check against address space limit. */
2732 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
2733 unsigned long nr_pages;
2734
2735 /*
2736 * MAP_FIXED may remove pages of mappings that intersects with
2737 * requested mapping. Account for the pages it would unmap.
2738 */
2739 nr_pages = count_vma_pages_range(mm, addr, end);
2740
2741 if (!may_expand_vm(mm, vm_flags,
2742 (len >> PAGE_SHIFT) - nr_pages))
2743 return -ENOMEM;
2744 }
2745
2746 /* Unmap any existing mapping in the area */
2747 if (do_vmi_munmap(&vmi, mm, addr, len, uf, false))
2748 return -ENOMEM;
2749
2750 /*
2751 * Private writable mapping: check memory availability
2752 */
2753 if (accountable_mapping(file, vm_flags)) {
2754 charged = len >> PAGE_SHIFT;
2755 if (security_vm_enough_memory_mm(mm, charged))
2756 return -ENOMEM;
2757 vm_flags |= VM_ACCOUNT;
2758 }
2759
2760 next = vma_next(&vmi);
2761 prev = vma_prev(&vmi);
2762 if (vm_flags & VM_SPECIAL) {
2763 if (prev)
2764 vma_iter_next_range(&vmi);
2765 goto cannot_expand;
2766 }
2767
2768 /* Attempt to expand an old mapping */
2769 /* Check next */
2770 if (next && next->vm_start == end && !vma_policy(next) &&
2771 can_vma_merge_before(next, vm_flags, NULL, file, pgoff+pglen,
2772 NULL_VM_UFFD_CTX, NULL)) {
2773 merge_end = next->vm_end;
2774 vma = next;
2775 vm_pgoff = next->vm_pgoff - pglen;
2776 }
2777
2778 /* Check prev */
2779 if (prev && prev->vm_end == addr && !vma_policy(prev) &&
2780 (vma ? can_vma_merge_after(prev, vm_flags, vma->anon_vma, file,
2781 pgoff, vma->vm_userfaultfd_ctx, NULL) :
2782 can_vma_merge_after(prev, vm_flags, NULL, file, pgoff,
2783 NULL_VM_UFFD_CTX, NULL))) {
2784 merge_start = prev->vm_start;
2785 vma = prev;
2786 vm_pgoff = prev->vm_pgoff;
2787 } else if (prev) {
2788 vma_iter_next_range(&vmi);
2789 }
2790
2791 /* Actually expand, if possible */
2792 if (vma &&
2793 !vma_expand(&vmi, vma, merge_start, merge_end, vm_pgoff, next)) {
2794 khugepaged_enter_vma(vma, vm_flags);
2795 goto expanded;
2796 }
2797
2798 if (vma == prev)
2799 vma_iter_set(&vmi, addr);
2800cannot_expand:
2801
2802 /*
2803 * Determine the object being mapped and call the appropriate
2804 * specific mapper. the address has already been validated, but
2805 * not unmapped, but the maps are removed from the list.
2806 */
2807 vma = vm_area_alloc(mm);
2808 if (!vma) {
2809 error = -ENOMEM;
2810 goto unacct_error;
2811 }
2812
2813 vma_iter_config(&vmi, addr, end);
2814 vma_set_range(vma, addr, end, pgoff);
2815 vm_flags_init(vma, vm_flags);
2816 vma->vm_page_prot = vm_get_page_prot(vm_flags);
2817
2818 if (file) {
2819 vma->vm_file = get_file(file);
2820 error = call_mmap(file, vma);
2821 if (error)
2822 goto unmap_and_free_vma;
2823
2824 if (vma_is_shared_maywrite(vma)) {
2825 error = mapping_map_writable(file->f_mapping);
2826 if (error)
2827 goto close_and_free_vma;
2828
2829 writable_file_mapping = true;
2830 }
2831
2832 /*
2833 * Expansion is handled above, merging is handled below.
2834 * Drivers should not alter the address of the VMA.
2835 */
2836 error = -EINVAL;
2837 if (WARN_ON((addr != vma->vm_start)))
2838 goto close_and_free_vma;
2839
2840 vma_iter_config(&vmi, addr, end);
2841 /*
2842 * If vm_flags changed after call_mmap(), we should try merge
2843 * vma again as we may succeed this time.
2844 */
2845 if (unlikely(vm_flags != vma->vm_flags && prev)) {
2846 merge = vma_merge_new_vma(&vmi, prev, vma,
2847 vma->vm_start, vma->vm_end,
2848 vma->vm_pgoff);
2849 if (merge) {
2850 /*
2851 * ->mmap() can change vma->vm_file and fput
2852 * the original file. So fput the vma->vm_file
2853 * here or we would add an extra fput for file
2854 * and cause general protection fault
2855 * ultimately.
2856 */
2857 fput(vma->vm_file);
2858 vm_area_free(vma);
2859 vma = merge;
2860 /* Update vm_flags to pick up the change. */
2861 vm_flags = vma->vm_flags;
2862 goto unmap_writable;
2863 }
2864 }
2865
2866 vm_flags = vma->vm_flags;
2867 } else if (vm_flags & VM_SHARED) {
2868 error = shmem_zero_setup(vma);
2869 if (error)
2870 goto free_vma;
2871 } else {
2872 vma_set_anonymous(vma);
2873 }
2874
2875 if (map_deny_write_exec(vma, vma->vm_flags)) {
2876 error = -EACCES;
2877 goto close_and_free_vma;
2878 }
2879
2880 /* Allow architectures to sanity-check the vm_flags */
2881 error = -EINVAL;
2882 if (!arch_validate_flags(vma->vm_flags))
2883 goto close_and_free_vma;
2884
2885 error = -ENOMEM;
2886 if (vma_iter_prealloc(&vmi, vma))
2887 goto close_and_free_vma;
2888
2889 /* Lock the VMA since it is modified after insertion into VMA tree */
2890 vma_start_write(vma);
2891 vma_iter_store(&vmi, vma);
2892 mm->map_count++;
2893 vma_link_file(vma);
2894
2895 /*
2896 * vma_merge() calls khugepaged_enter_vma() either, the below
2897 * call covers the non-merge case.
2898 */
2899 khugepaged_enter_vma(vma, vma->vm_flags);
2900
2901 /* Once vma denies write, undo our temporary denial count */
2902unmap_writable:
2903 if (writable_file_mapping)
2904 mapping_unmap_writable(file->f_mapping);
2905 file = vma->vm_file;
2906 ksm_add_vma(vma);
2907expanded:
2908 perf_event_mmap(vma);
2909
2910 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
2911 if (vm_flags & VM_LOCKED) {
2912 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
2913 is_vm_hugetlb_page(vma) ||
2914 vma == get_gate_vma(current->mm))
2915 vm_flags_clear(vma, VM_LOCKED_MASK);
2916 else
2917 mm->locked_vm += (len >> PAGE_SHIFT);
2918 }
2919
2920 if (file)
2921 uprobe_mmap(vma);
2922
2923 /*
2924 * New (or expanded) vma always get soft dirty status.
2925 * Otherwise user-space soft-dirty page tracker won't
2926 * be able to distinguish situation when vma area unmapped,
2927 * then new mapped in-place (which must be aimed as
2928 * a completely new data area).
2929 */
2930 vm_flags_set(vma, VM_SOFTDIRTY);
2931
2932 vma_set_page_prot(vma);
2933
2934 validate_mm(mm);
2935 return addr;
2936
2937close_and_free_vma:
2938 if (file && vma->vm_ops && vma->vm_ops->close)
2939 vma->vm_ops->close(vma);
2940
2941 if (file || vma->vm_file) {
2942unmap_and_free_vma:
2943 fput(vma->vm_file);
2944 vma->vm_file = NULL;
2945
2946 vma_iter_set(&vmi, vma->vm_end);
2947 /* Undo any partial mapping done by a device driver. */
2948 unmap_region(mm, &vmi.mas, vma, prev, next, vma->vm_start,
2949 vma->vm_end, vma->vm_end, true);
2950 }
2951 if (writable_file_mapping)
2952 mapping_unmap_writable(file->f_mapping);
2953free_vma:
2954 vm_area_free(vma);
2955unacct_error:
2956 if (charged)
2957 vm_unacct_memory(charged);
2958 validate_mm(mm);
2959 return error;
2960}
2961
2962static int __vm_munmap(unsigned long start, size_t len, bool unlock)
2963{
2964 int ret;
2965 struct mm_struct *mm = current->mm;
2966 LIST_HEAD(uf);
2967 VMA_ITERATOR(vmi, mm, start);
2968
2969 if (mmap_write_lock_killable(mm))
2970 return -EINTR;
2971
2972 ret = do_vmi_munmap(&vmi, mm, start, len, &uf, unlock);
2973 if (ret || !unlock)
2974 mmap_write_unlock(mm);
2975
2976 userfaultfd_unmap_complete(mm, &uf);
2977 return ret;
2978}
2979
2980int vm_munmap(unsigned long start, size_t len)
2981{
2982 return __vm_munmap(start, len, false);
2983}
2984EXPORT_SYMBOL(vm_munmap);
2985
2986SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2987{
2988 addr = untagged_addr(addr);
2989 return __vm_munmap(addr, len, true);
2990}
2991
2992
2993/*
2994 * Emulation of deprecated remap_file_pages() syscall.
2995 */
2996SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2997 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2998{
2999
3000 struct mm_struct *mm = current->mm;
3001 struct vm_area_struct *vma;
3002 unsigned long populate = 0;
3003 unsigned long ret = -EINVAL;
3004 struct file *file;
3005
3006 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/mm/remap_file_pages.rst.\n",
3007 current->comm, current->pid);
3008
3009 if (prot)
3010 return ret;
3011 start = start & PAGE_MASK;
3012 size = size & PAGE_MASK;
3013
3014 if (start + size <= start)
3015 return ret;
3016
3017 /* Does pgoff wrap? */
3018 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
3019 return ret;
3020
3021 if (mmap_write_lock_killable(mm))
3022 return -EINTR;
3023
3024 vma = vma_lookup(mm, start);
3025
3026 if (!vma || !(vma->vm_flags & VM_SHARED))
3027 goto out;
3028
3029 if (start + size > vma->vm_end) {
3030 VMA_ITERATOR(vmi, mm, vma->vm_end);
3031 struct vm_area_struct *next, *prev = vma;
3032
3033 for_each_vma_range(vmi, next, start + size) {
3034 /* hole between vmas ? */
3035 if (next->vm_start != prev->vm_end)
3036 goto out;
3037
3038 if (next->vm_file != vma->vm_file)
3039 goto out;
3040
3041 if (next->vm_flags != vma->vm_flags)
3042 goto out;
3043
3044 if (start + size <= next->vm_end)
3045 break;
3046
3047 prev = next;
3048 }
3049
3050 if (!next)
3051 goto out;
3052 }
3053
3054 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
3055 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
3056 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
3057
3058 flags &= MAP_NONBLOCK;
3059 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
3060 if (vma->vm_flags & VM_LOCKED)
3061 flags |= MAP_LOCKED;
3062
3063 file = get_file(vma->vm_file);
3064 ret = do_mmap(vma->vm_file, start, size,
3065 prot, flags, 0, pgoff, &populate, NULL);
3066 fput(file);
3067out:
3068 mmap_write_unlock(mm);
3069 if (populate)
3070 mm_populate(ret, populate);
3071 if (!IS_ERR_VALUE(ret))
3072 ret = 0;
3073 return ret;
3074}
3075
3076/*
3077 * do_vma_munmap() - Unmap a full or partial vma.
3078 * @vmi: The vma iterator pointing at the vma
3079 * @vma: The first vma to be munmapped
3080 * @start: the start of the address to unmap
3081 * @end: The end of the address to unmap
3082 * @uf: The userfaultfd list_head
3083 * @unlock: Drop the lock on success
3084 *
3085 * unmaps a VMA mapping when the vma iterator is already in position.
3086 * Does not handle alignment.
3087 *
3088 * Return: 0 on success drops the lock of so directed, error on failure and will
3089 * still hold the lock.
3090 */
3091int do_vma_munmap(struct vma_iterator *vmi, struct vm_area_struct *vma,
3092 unsigned long start, unsigned long end, struct list_head *uf,
3093 bool unlock)
3094{
3095 struct mm_struct *mm = vma->vm_mm;
3096
3097 arch_unmap(mm, start, end);
3098 return do_vmi_align_munmap(vmi, vma, mm, start, end, uf, unlock);
3099}
3100
3101/*
3102 * do_brk_flags() - Increase the brk vma if the flags match.
3103 * @vmi: The vma iterator
3104 * @addr: The start address
3105 * @len: The length of the increase
3106 * @vma: The vma,
3107 * @flags: The VMA Flags
3108 *
3109 * Extend the brk VMA from addr to addr + len. If the VMA is NULL or the flags
3110 * do not match then create a new anonymous VMA. Eventually we may be able to
3111 * do some brk-specific accounting here.
3112 */
3113static int do_brk_flags(struct vma_iterator *vmi, struct vm_area_struct *vma,
3114 unsigned long addr, unsigned long len, unsigned long flags)
3115{
3116 struct mm_struct *mm = current->mm;
3117 struct vma_prepare vp;
3118
3119 /*
3120 * Check against address space limits by the changed size
3121 * Note: This happens *after* clearing old mappings in some code paths.
3122 */
3123 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
3124 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
3125 return -ENOMEM;
3126
3127 if (mm->map_count > sysctl_max_map_count)
3128 return -ENOMEM;
3129
3130 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
3131 return -ENOMEM;
3132
3133 /*
3134 * Expand the existing vma if possible; Note that singular lists do not
3135 * occur after forking, so the expand will only happen on new VMAs.
3136 */
3137 if (vma && vma->vm_end == addr && !vma_policy(vma) &&
3138 can_vma_merge_after(vma, flags, NULL, NULL,
3139 addr >> PAGE_SHIFT, NULL_VM_UFFD_CTX, NULL)) {
3140 vma_iter_config(vmi, vma->vm_start, addr + len);
3141 if (vma_iter_prealloc(vmi, vma))
3142 goto unacct_fail;
3143
3144 vma_start_write(vma);
3145
3146 init_vma_prep(&vp, vma);
3147 vma_prepare(&vp);
3148 vma_adjust_trans_huge(vma, vma->vm_start, addr + len, 0);
3149 vma->vm_end = addr + len;
3150 vm_flags_set(vma, VM_SOFTDIRTY);
3151 vma_iter_store(vmi, vma);
3152
3153 vma_complete(&vp, vmi, mm);
3154 khugepaged_enter_vma(vma, flags);
3155 goto out;
3156 }
3157
3158 if (vma)
3159 vma_iter_next_range(vmi);
3160 /* create a vma struct for an anonymous mapping */
3161 vma = vm_area_alloc(mm);
3162 if (!vma)
3163 goto unacct_fail;
3164
3165 vma_set_anonymous(vma);
3166 vma_set_range(vma, addr, addr + len, addr >> PAGE_SHIFT);
3167 vm_flags_init(vma, flags);
3168 vma->vm_page_prot = vm_get_page_prot(flags);
3169 vma_start_write(vma);
3170 if (vma_iter_store_gfp(vmi, vma, GFP_KERNEL))
3171 goto mas_store_fail;
3172
3173 mm->map_count++;
3174 validate_mm(mm);
3175 ksm_add_vma(vma);
3176out:
3177 perf_event_mmap(vma);
3178 mm->total_vm += len >> PAGE_SHIFT;
3179 mm->data_vm += len >> PAGE_SHIFT;
3180 if (flags & VM_LOCKED)
3181 mm->locked_vm += (len >> PAGE_SHIFT);
3182 vm_flags_set(vma, VM_SOFTDIRTY);
3183 return 0;
3184
3185mas_store_fail:
3186 vm_area_free(vma);
3187unacct_fail:
3188 vm_unacct_memory(len >> PAGE_SHIFT);
3189 return -ENOMEM;
3190}
3191
3192int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3193{
3194 struct mm_struct *mm = current->mm;
3195 struct vm_area_struct *vma = NULL;
3196 unsigned long len;
3197 int ret;
3198 bool populate;
3199 LIST_HEAD(uf);
3200 VMA_ITERATOR(vmi, mm, addr);
3201
3202 len = PAGE_ALIGN(request);
3203 if (len < request)
3204 return -ENOMEM;
3205 if (!len)
3206 return 0;
3207
3208 /* Until we need other flags, refuse anything except VM_EXEC. */
3209 if ((flags & (~VM_EXEC)) != 0)
3210 return -EINVAL;
3211
3212 if (mmap_write_lock_killable(mm))
3213 return -EINTR;
3214
3215 ret = check_brk_limits(addr, len);
3216 if (ret)
3217 goto limits_failed;
3218
3219 ret = do_vmi_munmap(&vmi, mm, addr, len, &uf, 0);
3220 if (ret)
3221 goto munmap_failed;
3222
3223 vma = vma_prev(&vmi);
3224 ret = do_brk_flags(&vmi, vma, addr, len, flags);
3225 populate = ((mm->def_flags & VM_LOCKED) != 0);
3226 mmap_write_unlock(mm);
3227 userfaultfd_unmap_complete(mm, &uf);
3228 if (populate && !ret)
3229 mm_populate(addr, len);
3230 return ret;
3231
3232munmap_failed:
3233limits_failed:
3234 mmap_write_unlock(mm);
3235 return ret;
3236}
3237EXPORT_SYMBOL(vm_brk_flags);
3238
3239/* Release all mmaps. */
3240void exit_mmap(struct mm_struct *mm)
3241{
3242 struct mmu_gather tlb;
3243 struct vm_area_struct *vma;
3244 unsigned long nr_accounted = 0;
3245 MA_STATE(mas, &mm->mm_mt, 0, 0);
3246 int count = 0;
3247
3248 /* mm's last user has gone, and its about to be pulled down */
3249 mmu_notifier_release(mm);
3250
3251 mmap_read_lock(mm);
3252 arch_exit_mmap(mm);
3253
3254 vma = mas_find(&mas, ULONG_MAX);
3255 if (!vma || unlikely(xa_is_zero(vma))) {
3256 /* Can happen if dup_mmap() received an OOM */
3257 mmap_read_unlock(mm);
3258 mmap_write_lock(mm);
3259 goto destroy;
3260 }
3261
3262 lru_add_drain();
3263 flush_cache_mm(mm);
3264 tlb_gather_mmu_fullmm(&tlb, mm);
3265 /* update_hiwater_rss(mm) here? but nobody should be looking */
3266 /* Use ULONG_MAX here to ensure all VMAs in the mm are unmapped */
3267 unmap_vmas(&tlb, &mas, vma, 0, ULONG_MAX, ULONG_MAX, false);
3268 mmap_read_unlock(mm);
3269
3270 /*
3271 * Set MMF_OOM_SKIP to hide this task from the oom killer/reaper
3272 * because the memory has been already freed.
3273 */
3274 set_bit(MMF_OOM_SKIP, &mm->flags);
3275 mmap_write_lock(mm);
3276 mt_clear_in_rcu(&mm->mm_mt);
3277 mas_set(&mas, vma->vm_end);
3278 free_pgtables(&tlb, &mas, vma, FIRST_USER_ADDRESS,
3279 USER_PGTABLES_CEILING, true);
3280 tlb_finish_mmu(&tlb);
3281
3282 /*
3283 * Walk the list again, actually closing and freeing it, with preemption
3284 * enabled, without holding any MM locks besides the unreachable
3285 * mmap_write_lock.
3286 */
3287 mas_set(&mas, vma->vm_end);
3288 do {
3289 if (vma->vm_flags & VM_ACCOUNT)
3290 nr_accounted += vma_pages(vma);
3291 remove_vma(vma, true);
3292 count++;
3293 cond_resched();
3294 vma = mas_find(&mas, ULONG_MAX);
3295 } while (vma && likely(!xa_is_zero(vma)));
3296
3297 BUG_ON(count != mm->map_count);
3298
3299 trace_exit_mmap(mm);
3300destroy:
3301 __mt_destroy(&mm->mm_mt);
3302 mmap_write_unlock(mm);
3303 vm_unacct_memory(nr_accounted);
3304}
3305
3306/* Insert vm structure into process list sorted by address
3307 * and into the inode's i_mmap tree. If vm_file is non-NULL
3308 * then i_mmap_rwsem is taken here.
3309 */
3310int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3311{
3312 unsigned long charged = vma_pages(vma);
3313
3314
3315 if (find_vma_intersection(mm, vma->vm_start, vma->vm_end))
3316 return -ENOMEM;
3317
3318 if ((vma->vm_flags & VM_ACCOUNT) &&
3319 security_vm_enough_memory_mm(mm, charged))
3320 return -ENOMEM;
3321
3322 /*
3323 * The vm_pgoff of a purely anonymous vma should be irrelevant
3324 * until its first write fault, when page's anon_vma and index
3325 * are set. But now set the vm_pgoff it will almost certainly
3326 * end up with (unless mremap moves it elsewhere before that
3327 * first wfault), so /proc/pid/maps tells a consistent story.
3328 *
3329 * By setting it to reflect the virtual start address of the
3330 * vma, merges and splits can happen in a seamless way, just
3331 * using the existing file pgoff checks and manipulations.
3332 * Similarly in do_mmap and in do_brk_flags.
3333 */
3334 if (vma_is_anonymous(vma)) {
3335 BUG_ON(vma->anon_vma);
3336 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3337 }
3338
3339 if (vma_link(mm, vma)) {
3340 if (vma->vm_flags & VM_ACCOUNT)
3341 vm_unacct_memory(charged);
3342 return -ENOMEM;
3343 }
3344
3345 return 0;
3346}
3347
3348/*
3349 * Copy the vma structure to a new location in the same mm,
3350 * prior to moving page table entries, to effect an mremap move.
3351 */
3352struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3353 unsigned long addr, unsigned long len, pgoff_t pgoff,
3354 bool *need_rmap_locks)
3355{
3356 struct vm_area_struct *vma = *vmap;
3357 unsigned long vma_start = vma->vm_start;
3358 struct mm_struct *mm = vma->vm_mm;
3359 struct vm_area_struct *new_vma, *prev;
3360 bool faulted_in_anon_vma = true;
3361 VMA_ITERATOR(vmi, mm, addr);
3362
3363 /*
3364 * If anonymous vma has not yet been faulted, update new pgoff
3365 * to match new location, to increase its chance of merging.
3366 */
3367 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3368 pgoff = addr >> PAGE_SHIFT;
3369 faulted_in_anon_vma = false;
3370 }
3371
3372 new_vma = find_vma_prev(mm, addr, &prev);
3373 if (new_vma && new_vma->vm_start < addr + len)
3374 return NULL; /* should never get here */
3375
3376 new_vma = vma_merge_new_vma(&vmi, prev, vma, addr, addr + len, pgoff);
3377 if (new_vma) {
3378 /*
3379 * Source vma may have been merged into new_vma
3380 */
3381 if (unlikely(vma_start >= new_vma->vm_start &&
3382 vma_start < new_vma->vm_end)) {
3383 /*
3384 * The only way we can get a vma_merge with
3385 * self during an mremap is if the vma hasn't
3386 * been faulted in yet and we were allowed to
3387 * reset the dst vma->vm_pgoff to the
3388 * destination address of the mremap to allow
3389 * the merge to happen. mremap must change the
3390 * vm_pgoff linearity between src and dst vmas
3391 * (in turn preventing a vma_merge) to be
3392 * safe. It is only safe to keep the vm_pgoff
3393 * linear if there are no pages mapped yet.
3394 */
3395 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3396 *vmap = vma = new_vma;
3397 }
3398 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3399 } else {
3400 new_vma = vm_area_dup(vma);
3401 if (!new_vma)
3402 goto out;
3403 vma_set_range(new_vma, addr, addr + len, pgoff);
3404 if (vma_dup_policy(vma, new_vma))
3405 goto out_free_vma;
3406 if (anon_vma_clone(new_vma, vma))
3407 goto out_free_mempol;
3408 if (new_vma->vm_file)
3409 get_file(new_vma->vm_file);
3410 if (new_vma->vm_ops && new_vma->vm_ops->open)
3411 new_vma->vm_ops->open(new_vma);
3412 if (vma_link(mm, new_vma))
3413 goto out_vma_link;
3414 *need_rmap_locks = false;
3415 }
3416 return new_vma;
3417
3418out_vma_link:
3419 if (new_vma->vm_ops && new_vma->vm_ops->close)
3420 new_vma->vm_ops->close(new_vma);
3421
3422 if (new_vma->vm_file)
3423 fput(new_vma->vm_file);
3424
3425 unlink_anon_vmas(new_vma);
3426out_free_mempol:
3427 mpol_put(vma_policy(new_vma));
3428out_free_vma:
3429 vm_area_free(new_vma);
3430out:
3431 return NULL;
3432}
3433
3434/*
3435 * Return true if the calling process may expand its vm space by the passed
3436 * number of pages
3437 */
3438bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3439{
3440 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3441 return false;
3442
3443 if (is_data_mapping(flags) &&
3444 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3445 /* Workaround for Valgrind */
3446 if (rlimit(RLIMIT_DATA) == 0 &&
3447 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3448 return true;
3449
3450 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3451 current->comm, current->pid,
3452 (mm->data_vm + npages) << PAGE_SHIFT,
3453 rlimit(RLIMIT_DATA),
3454 ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3455
3456 if (!ignore_rlimit_data)
3457 return false;
3458 }
3459
3460 return true;
3461}
3462
3463void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3464{
3465 WRITE_ONCE(mm->total_vm, READ_ONCE(mm->total_vm)+npages);
3466
3467 if (is_exec_mapping(flags))
3468 mm->exec_vm += npages;
3469 else if (is_stack_mapping(flags))
3470 mm->stack_vm += npages;
3471 else if (is_data_mapping(flags))
3472 mm->data_vm += npages;
3473}
3474
3475static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3476
3477/*
3478 * Having a close hook prevents vma merging regardless of flags.
3479 */
3480static void special_mapping_close(struct vm_area_struct *vma)
3481{
3482}
3483
3484static const char *special_mapping_name(struct vm_area_struct *vma)
3485{
3486 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3487}
3488
3489static int special_mapping_mremap(struct vm_area_struct *new_vma)
3490{
3491 struct vm_special_mapping *sm = new_vma->vm_private_data;
3492
3493 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3494 return -EFAULT;
3495
3496 if (sm->mremap)
3497 return sm->mremap(sm, new_vma);
3498
3499 return 0;
3500}
3501
3502static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr)
3503{
3504 /*
3505 * Forbid splitting special mappings - kernel has expectations over
3506 * the number of pages in mapping. Together with VM_DONTEXPAND
3507 * the size of vma should stay the same over the special mapping's
3508 * lifetime.
3509 */
3510 return -EINVAL;
3511}
3512
3513static const struct vm_operations_struct special_mapping_vmops = {
3514 .close = special_mapping_close,
3515 .fault = special_mapping_fault,
3516 .mremap = special_mapping_mremap,
3517 .name = special_mapping_name,
3518 /* vDSO code relies that VVAR can't be accessed remotely */
3519 .access = NULL,
3520 .may_split = special_mapping_split,
3521};
3522
3523static const struct vm_operations_struct legacy_special_mapping_vmops = {
3524 .close = special_mapping_close,
3525 .fault = special_mapping_fault,
3526};
3527
3528static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3529{
3530 struct vm_area_struct *vma = vmf->vma;
3531 pgoff_t pgoff;
3532 struct page **pages;
3533
3534 if (vma->vm_ops == &legacy_special_mapping_vmops) {
3535 pages = vma->vm_private_data;
3536 } else {
3537 struct vm_special_mapping *sm = vma->vm_private_data;
3538
3539 if (sm->fault)
3540 return sm->fault(sm, vmf->vma, vmf);
3541
3542 pages = sm->pages;
3543 }
3544
3545 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3546 pgoff--;
3547
3548 if (*pages) {
3549 struct page *page = *pages;
3550 get_page(page);
3551 vmf->page = page;
3552 return 0;
3553 }
3554
3555 return VM_FAULT_SIGBUS;
3556}
3557
3558static struct vm_area_struct *__install_special_mapping(
3559 struct mm_struct *mm,
3560 unsigned long addr, unsigned long len,
3561 unsigned long vm_flags, void *priv,
3562 const struct vm_operations_struct *ops)
3563{
3564 int ret;
3565 struct vm_area_struct *vma;
3566
3567 vma = vm_area_alloc(mm);
3568 if (unlikely(vma == NULL))
3569 return ERR_PTR(-ENOMEM);
3570
3571 vma_set_range(vma, addr, addr + len, 0);
3572 vm_flags_init(vma, (vm_flags | mm->def_flags |
3573 VM_DONTEXPAND | VM_SOFTDIRTY) & ~VM_LOCKED_MASK);
3574 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3575
3576 vma->vm_ops = ops;
3577 vma->vm_private_data = priv;
3578
3579 ret = insert_vm_struct(mm, vma);
3580 if (ret)
3581 goto out;
3582
3583 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3584
3585 perf_event_mmap(vma);
3586
3587 return vma;
3588
3589out:
3590 vm_area_free(vma);
3591 return ERR_PTR(ret);
3592}
3593
3594bool vma_is_special_mapping(const struct vm_area_struct *vma,
3595 const struct vm_special_mapping *sm)
3596{
3597 return vma->vm_private_data == sm &&
3598 (vma->vm_ops == &special_mapping_vmops ||
3599 vma->vm_ops == &legacy_special_mapping_vmops);
3600}
3601
3602/*
3603 * Called with mm->mmap_lock held for writing.
3604 * Insert a new vma covering the given region, with the given flags.
3605 * Its pages are supplied by the given array of struct page *.
3606 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3607 * The region past the last page supplied will always produce SIGBUS.
3608 * The array pointer and the pages it points to are assumed to stay alive
3609 * for as long as this mapping might exist.
3610 */
3611struct vm_area_struct *_install_special_mapping(
3612 struct mm_struct *mm,
3613 unsigned long addr, unsigned long len,
3614 unsigned long vm_flags, const struct vm_special_mapping *spec)
3615{
3616 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3617 &special_mapping_vmops);
3618}
3619
3620int install_special_mapping(struct mm_struct *mm,
3621 unsigned long addr, unsigned long len,
3622 unsigned long vm_flags, struct page **pages)
3623{
3624 struct vm_area_struct *vma = __install_special_mapping(
3625 mm, addr, len, vm_flags, (void *)pages,
3626 &legacy_special_mapping_vmops);
3627
3628 return PTR_ERR_OR_ZERO(vma);
3629}
3630
3631static DEFINE_MUTEX(mm_all_locks_mutex);
3632
3633static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3634{
3635 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3636 /*
3637 * The LSB of head.next can't change from under us
3638 * because we hold the mm_all_locks_mutex.
3639 */
3640 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
3641 /*
3642 * We can safely modify head.next after taking the
3643 * anon_vma->root->rwsem. If some other vma in this mm shares
3644 * the same anon_vma we won't take it again.
3645 *
3646 * No need of atomic instructions here, head.next
3647 * can't change from under us thanks to the
3648 * anon_vma->root->rwsem.
3649 */
3650 if (__test_and_set_bit(0, (unsigned long *)
3651 &anon_vma->root->rb_root.rb_root.rb_node))
3652 BUG();
3653 }
3654}
3655
3656static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3657{
3658 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3659 /*
3660 * AS_MM_ALL_LOCKS can't change from under us because
3661 * we hold the mm_all_locks_mutex.
3662 *
3663 * Operations on ->flags have to be atomic because
3664 * even if AS_MM_ALL_LOCKS is stable thanks to the
3665 * mm_all_locks_mutex, there may be other cpus
3666 * changing other bitflags in parallel to us.
3667 */
3668 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3669 BUG();
3670 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
3671 }
3672}
3673
3674/*
3675 * This operation locks against the VM for all pte/vma/mm related
3676 * operations that could ever happen on a certain mm. This includes
3677 * vmtruncate, try_to_unmap, and all page faults.
3678 *
3679 * The caller must take the mmap_lock in write mode before calling
3680 * mm_take_all_locks(). The caller isn't allowed to release the
3681 * mmap_lock until mm_drop_all_locks() returns.
3682 *
3683 * mmap_lock in write mode is required in order to block all operations
3684 * that could modify pagetables and free pages without need of
3685 * altering the vma layout. It's also needed in write mode to avoid new
3686 * anon_vmas to be associated with existing vmas.
3687 *
3688 * A single task can't take more than one mm_take_all_locks() in a row
3689 * or it would deadlock.
3690 *
3691 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3692 * mapping->flags avoid to take the same lock twice, if more than one
3693 * vma in this mm is backed by the same anon_vma or address_space.
3694 *
3695 * We take locks in following order, accordingly to comment at beginning
3696 * of mm/rmap.c:
3697 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3698 * hugetlb mapping);
3699 * - all vmas marked locked
3700 * - all i_mmap_rwsem locks;
3701 * - all anon_vma->rwseml
3702 *
3703 * We can take all locks within these types randomly because the VM code
3704 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3705 * mm_all_locks_mutex.
3706 *
3707 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3708 * that may have to take thousand of locks.
3709 *
3710 * mm_take_all_locks() can fail if it's interrupted by signals.
3711 */
3712int mm_take_all_locks(struct mm_struct *mm)
3713{
3714 struct vm_area_struct *vma;
3715 struct anon_vma_chain *avc;
3716 MA_STATE(mas, &mm->mm_mt, 0, 0);
3717
3718 mmap_assert_write_locked(mm);
3719
3720 mutex_lock(&mm_all_locks_mutex);
3721
3722 /*
3723 * vma_start_write() does not have a complement in mm_drop_all_locks()
3724 * because vma_start_write() is always asymmetrical; it marks a VMA as
3725 * being written to until mmap_write_unlock() or mmap_write_downgrade()
3726 * is reached.
3727 */
3728 mas_for_each(&mas, vma, ULONG_MAX) {
3729 if (signal_pending(current))
3730 goto out_unlock;
3731 vma_start_write(vma);
3732 }
3733
3734 mas_set(&mas, 0);
3735 mas_for_each(&mas, vma, ULONG_MAX) {
3736 if (signal_pending(current))
3737 goto out_unlock;
3738 if (vma->vm_file && vma->vm_file->f_mapping &&
3739 is_vm_hugetlb_page(vma))
3740 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3741 }
3742
3743 mas_set(&mas, 0);
3744 mas_for_each(&mas, vma, ULONG_MAX) {
3745 if (signal_pending(current))
3746 goto out_unlock;
3747 if (vma->vm_file && vma->vm_file->f_mapping &&
3748 !is_vm_hugetlb_page(vma))
3749 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3750 }
3751
3752 mas_set(&mas, 0);
3753 mas_for_each(&mas, vma, ULONG_MAX) {
3754 if (signal_pending(current))
3755 goto out_unlock;
3756 if (vma->anon_vma)
3757 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3758 vm_lock_anon_vma(mm, avc->anon_vma);
3759 }
3760
3761 return 0;
3762
3763out_unlock:
3764 mm_drop_all_locks(mm);
3765 return -EINTR;
3766}
3767
3768static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3769{
3770 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3771 /*
3772 * The LSB of head.next can't change to 0 from under
3773 * us because we hold the mm_all_locks_mutex.
3774 *
3775 * We must however clear the bitflag before unlocking
3776 * the vma so the users using the anon_vma->rb_root will
3777 * never see our bitflag.
3778 *
3779 * No need of atomic instructions here, head.next
3780 * can't change from under us until we release the
3781 * anon_vma->root->rwsem.
3782 */
3783 if (!__test_and_clear_bit(0, (unsigned long *)
3784 &anon_vma->root->rb_root.rb_root.rb_node))
3785 BUG();
3786 anon_vma_unlock_write(anon_vma);
3787 }
3788}
3789
3790static void vm_unlock_mapping(struct address_space *mapping)
3791{
3792 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3793 /*
3794 * AS_MM_ALL_LOCKS can't change to 0 from under us
3795 * because we hold the mm_all_locks_mutex.
3796 */
3797 i_mmap_unlock_write(mapping);
3798 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3799 &mapping->flags))
3800 BUG();
3801 }
3802}
3803
3804/*
3805 * The mmap_lock cannot be released by the caller until
3806 * mm_drop_all_locks() returns.
3807 */
3808void mm_drop_all_locks(struct mm_struct *mm)
3809{
3810 struct vm_area_struct *vma;
3811 struct anon_vma_chain *avc;
3812 MA_STATE(mas, &mm->mm_mt, 0, 0);
3813
3814 mmap_assert_write_locked(mm);
3815 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3816
3817 mas_for_each(&mas, vma, ULONG_MAX) {
3818 if (vma->anon_vma)
3819 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3820 vm_unlock_anon_vma(avc->anon_vma);
3821 if (vma->vm_file && vma->vm_file->f_mapping)
3822 vm_unlock_mapping(vma->vm_file->f_mapping);
3823 }
3824
3825 mutex_unlock(&mm_all_locks_mutex);
3826}
3827
3828/*
3829 * initialise the percpu counter for VM
3830 */
3831void __init mmap_init(void)
3832{
3833 int ret;
3834
3835 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3836 VM_BUG_ON(ret);
3837}
3838
3839/*
3840 * Initialise sysctl_user_reserve_kbytes.
3841 *
3842 * This is intended to prevent a user from starting a single memory hogging
3843 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3844 * mode.
3845 *
3846 * The default value is min(3% of free memory, 128MB)
3847 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3848 */
3849static int init_user_reserve(void)
3850{
3851 unsigned long free_kbytes;
3852
3853 free_kbytes = K(global_zone_page_state(NR_FREE_PAGES));
3854
3855 sysctl_user_reserve_kbytes = min(free_kbytes / 32, SZ_128K);
3856 return 0;
3857}
3858subsys_initcall(init_user_reserve);
3859
3860/*
3861 * Initialise sysctl_admin_reserve_kbytes.
3862 *
3863 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3864 * to log in and kill a memory hogging process.
3865 *
3866 * Systems with more than 256MB will reserve 8MB, enough to recover
3867 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3868 * only reserve 3% of free pages by default.
3869 */
3870static int init_admin_reserve(void)
3871{
3872 unsigned long free_kbytes;
3873
3874 free_kbytes = K(global_zone_page_state(NR_FREE_PAGES));
3875
3876 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, SZ_8K);
3877 return 0;
3878}
3879subsys_initcall(init_admin_reserve);
3880
3881/*
3882 * Reinititalise user and admin reserves if memory is added or removed.
3883 *
3884 * The default user reserve max is 128MB, and the default max for the
3885 * admin reserve is 8MB. These are usually, but not always, enough to
3886 * enable recovery from a memory hogging process using login/sshd, a shell,
3887 * and tools like top. It may make sense to increase or even disable the
3888 * reserve depending on the existence of swap or variations in the recovery
3889 * tools. So, the admin may have changed them.
3890 *
3891 * If memory is added and the reserves have been eliminated or increased above
3892 * the default max, then we'll trust the admin.
3893 *
3894 * If memory is removed and there isn't enough free memory, then we
3895 * need to reset the reserves.
3896 *
3897 * Otherwise keep the reserve set by the admin.
3898 */
3899static int reserve_mem_notifier(struct notifier_block *nb,
3900 unsigned long action, void *data)
3901{
3902 unsigned long tmp, free_kbytes;
3903
3904 switch (action) {
3905 case MEM_ONLINE:
3906 /* Default max is 128MB. Leave alone if modified by operator. */
3907 tmp = sysctl_user_reserve_kbytes;
3908 if (tmp > 0 && tmp < SZ_128K)
3909 init_user_reserve();
3910
3911 /* Default max is 8MB. Leave alone if modified by operator. */
3912 tmp = sysctl_admin_reserve_kbytes;
3913 if (tmp > 0 && tmp < SZ_8K)
3914 init_admin_reserve();
3915
3916 break;
3917 case MEM_OFFLINE:
3918 free_kbytes = K(global_zone_page_state(NR_FREE_PAGES));
3919
3920 if (sysctl_user_reserve_kbytes > free_kbytes) {
3921 init_user_reserve();
3922 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3923 sysctl_user_reserve_kbytes);
3924 }
3925
3926 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3927 init_admin_reserve();
3928 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3929 sysctl_admin_reserve_kbytes);
3930 }
3931 break;
3932 default:
3933 break;
3934 }
3935 return NOTIFY_OK;
3936}
3937
3938static int __meminit init_reserve_notifier(void)
3939{
3940 if (hotplug_memory_notifier(reserve_mem_notifier, DEFAULT_CALLBACK_PRI))
3941 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3942
3943 return 0;
3944}
3945subsys_initcall(init_reserve_notifier);