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