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