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