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/slab.h>
10#include <linux/backing-dev.h>
11#include <linux/mm.h>
12#include <linux/shm.h>
13#include <linux/mman.h>
14#include <linux/pagemap.h>
15#include <linux/swap.h>
16#include <linux/syscalls.h>
17#include <linux/capability.h>
18#include <linux/init.h>
19#include <linux/file.h>
20#include <linux/fs.h>
21#include <linux/personality.h>
22#include <linux/security.h>
23#include <linux/hugetlb.h>
24#include <linux/profile.h>
25#include <linux/export.h>
26#include <linux/mount.h>
27#include <linux/mempolicy.h>
28#include <linux/rmap.h>
29#include <linux/mmu_notifier.h>
30#include <linux/perf_event.h>
31#include <linux/audit.h>
32#include <linux/khugepaged.h>
33#include <linux/uprobes.h>
34
35#include <asm/uaccess.h>
36#include <asm/cacheflush.h>
37#include <asm/tlb.h>
38#include <asm/mmu_context.h>
39
40#include "internal.h"
41
42#ifndef arch_mmap_check
43#define arch_mmap_check(addr, len, flags) (0)
44#endif
45
46#ifndef arch_rebalance_pgtables
47#define arch_rebalance_pgtables(addr, len) (addr)
48#endif
49
50static void unmap_region(struct mm_struct *mm,
51 struct vm_area_struct *vma, struct vm_area_struct *prev,
52 unsigned long start, unsigned long end);
53
54/*
55 * WARNING: the debugging will use recursive algorithms so never enable this
56 * unless you know what you are doing.
57 */
58#undef DEBUG_MM_RB
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% */
90int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
91/*
92 * Make sure vm_committed_as in one cacheline and not cacheline shared with
93 * other variables. It can be updated by several CPUs frequently.
94 */
95struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
96
97/*
98 * Check that a process has enough memory to allocate a new virtual
99 * mapping. 0 means there is enough memory for the allocation to
100 * succeed and -ENOMEM implies there is not.
101 *
102 * We currently support three overcommit policies, which are set via the
103 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting
104 *
105 * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
106 * Additional code 2002 Jul 20 by Robert Love.
107 *
108 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
109 *
110 * Note this is a helper function intended to be used by LSMs which
111 * wish to use this logic.
112 */
113int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
114{
115 unsigned long free, allowed;
116
117 vm_acct_memory(pages);
118
119 /*
120 * Sometimes we want to use more memory than we have
121 */
122 if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
123 return 0;
124
125 if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
126 free = global_page_state(NR_FREE_PAGES);
127 free += global_page_state(NR_FILE_PAGES);
128
129 /*
130 * shmem pages shouldn't be counted as free in this
131 * case, they can't be purged, only swapped out, and
132 * that won't affect the overall amount of available
133 * memory in the system.
134 */
135 free -= global_page_state(NR_SHMEM);
136
137 free += nr_swap_pages;
138
139 /*
140 * Any slabs which are created with the
141 * SLAB_RECLAIM_ACCOUNT flag claim to have contents
142 * which are reclaimable, under pressure. The dentry
143 * cache and most inode caches should fall into this
144 */
145 free += global_page_state(NR_SLAB_RECLAIMABLE);
146
147 /*
148 * Leave reserved pages. The pages are not for anonymous pages.
149 */
150 if (free <= totalreserve_pages)
151 goto error;
152 else
153 free -= totalreserve_pages;
154
155 /*
156 * Leave the last 3% for root
157 */
158 if (!cap_sys_admin)
159 free -= free / 32;
160
161 if (free > pages)
162 return 0;
163
164 goto error;
165 }
166
167 allowed = (totalram_pages - hugetlb_total_pages())
168 * sysctl_overcommit_ratio / 100;
169 /*
170 * Leave the last 3% for root
171 */
172 if (!cap_sys_admin)
173 allowed -= allowed / 32;
174 allowed += total_swap_pages;
175
176 /* Don't let a single process grow too big:
177 leave 3% of the size of this process for other processes */
178 if (mm)
179 allowed -= mm->total_vm / 32;
180
181 if (percpu_counter_read_positive(&vm_committed_as) < allowed)
182 return 0;
183error:
184 vm_unacct_memory(pages);
185
186 return -ENOMEM;
187}
188
189/*
190 * Requires inode->i_mapping->i_mmap_mutex
191 */
192static void __remove_shared_vm_struct(struct vm_area_struct *vma,
193 struct file *file, struct address_space *mapping)
194{
195 if (vma->vm_flags & VM_DENYWRITE)
196 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
197 if (vma->vm_flags & VM_SHARED)
198 mapping->i_mmap_writable--;
199
200 flush_dcache_mmap_lock(mapping);
201 if (unlikely(vma->vm_flags & VM_NONLINEAR))
202 list_del_init(&vma->shared.vm_set.list);
203 else
204 vma_prio_tree_remove(vma, &mapping->i_mmap);
205 flush_dcache_mmap_unlock(mapping);
206}
207
208/*
209 * Unlink a file-based vm structure from its prio_tree, to hide
210 * vma from rmap and vmtruncate before freeing its page tables.
211 */
212void unlink_file_vma(struct vm_area_struct *vma)
213{
214 struct file *file = vma->vm_file;
215
216 if (file) {
217 struct address_space *mapping = file->f_mapping;
218 mutex_lock(&mapping->i_mmap_mutex);
219 __remove_shared_vm_struct(vma, file, mapping);
220 mutex_unlock(&mapping->i_mmap_mutex);
221 }
222}
223
224/*
225 * Close a vm structure and free it, returning the next.
226 */
227static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
228{
229 struct vm_area_struct *next = vma->vm_next;
230
231 might_sleep();
232 if (vma->vm_ops && vma->vm_ops->close)
233 vma->vm_ops->close(vma);
234 if (vma->vm_file) {
235 fput(vma->vm_file);
236 if (vma->vm_flags & VM_EXECUTABLE)
237 removed_exe_file_vma(vma->vm_mm);
238 }
239 mpol_put(vma_policy(vma));
240 kmem_cache_free(vm_area_cachep, vma);
241 return next;
242}
243
244static unsigned long do_brk(unsigned long addr, unsigned long len);
245
246SYSCALL_DEFINE1(brk, unsigned long, brk)
247{
248 unsigned long rlim, retval;
249 unsigned long newbrk, oldbrk;
250 struct mm_struct *mm = current->mm;
251 unsigned long min_brk;
252
253 down_write(&mm->mmap_sem);
254
255#ifdef CONFIG_COMPAT_BRK
256 /*
257 * CONFIG_COMPAT_BRK can still be overridden by setting
258 * randomize_va_space to 2, which will still cause mm->start_brk
259 * to be arbitrarily shifted
260 */
261 if (current->brk_randomized)
262 min_brk = mm->start_brk;
263 else
264 min_brk = mm->end_data;
265#else
266 min_brk = mm->start_brk;
267#endif
268 if (brk < min_brk)
269 goto out;
270
271 /*
272 * Check against rlimit here. If this check is done later after the test
273 * of oldbrk with newbrk then it can escape the test and let the data
274 * segment grow beyond its set limit the in case where the limit is
275 * not page aligned -Ram Gupta
276 */
277 rlim = rlimit(RLIMIT_DATA);
278 if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
279 (mm->end_data - mm->start_data) > rlim)
280 goto out;
281
282 newbrk = PAGE_ALIGN(brk);
283 oldbrk = PAGE_ALIGN(mm->brk);
284 if (oldbrk == newbrk)
285 goto set_brk;
286
287 /* Always allow shrinking brk. */
288 if (brk <= mm->brk) {
289 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
290 goto set_brk;
291 goto out;
292 }
293
294 /* Check against existing mmap mappings. */
295 if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
296 goto out;
297
298 /* Ok, looks good - let it rip. */
299 if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
300 goto out;
301set_brk:
302 mm->brk = brk;
303out:
304 retval = mm->brk;
305 up_write(&mm->mmap_sem);
306 return retval;
307}
308
309#ifdef DEBUG_MM_RB
310static int browse_rb(struct rb_root *root)
311{
312 int i = 0, j;
313 struct rb_node *nd, *pn = NULL;
314 unsigned long prev = 0, pend = 0;
315
316 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
317 struct vm_area_struct *vma;
318 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
319 if (vma->vm_start < prev)
320 printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
321 if (vma->vm_start < pend)
322 printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
323 if (vma->vm_start > vma->vm_end)
324 printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
325 i++;
326 pn = nd;
327 prev = vma->vm_start;
328 pend = vma->vm_end;
329 }
330 j = 0;
331 for (nd = pn; nd; nd = rb_prev(nd)) {
332 j++;
333 }
334 if (i != j)
335 printk("backwards %d, forwards %d\n", j, i), i = 0;
336 return i;
337}
338
339void validate_mm(struct mm_struct *mm)
340{
341 int bug = 0;
342 int i = 0;
343 struct vm_area_struct *tmp = mm->mmap;
344 while (tmp) {
345 tmp = tmp->vm_next;
346 i++;
347 }
348 if (i != mm->map_count)
349 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
350 i = browse_rb(&mm->mm_rb);
351 if (i != mm->map_count)
352 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
353 BUG_ON(bug);
354}
355#else
356#define validate_mm(mm) do { } while (0)
357#endif
358
359static struct vm_area_struct *
360find_vma_prepare(struct mm_struct *mm, unsigned long addr,
361 struct vm_area_struct **pprev, struct rb_node ***rb_link,
362 struct rb_node ** rb_parent)
363{
364 struct vm_area_struct * vma;
365 struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
366
367 __rb_link = &mm->mm_rb.rb_node;
368 rb_prev = __rb_parent = NULL;
369 vma = NULL;
370
371 while (*__rb_link) {
372 struct vm_area_struct *vma_tmp;
373
374 __rb_parent = *__rb_link;
375 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
376
377 if (vma_tmp->vm_end > addr) {
378 vma = vma_tmp;
379 if (vma_tmp->vm_start <= addr)
380 break;
381 __rb_link = &__rb_parent->rb_left;
382 } else {
383 rb_prev = __rb_parent;
384 __rb_link = &__rb_parent->rb_right;
385 }
386 }
387
388 *pprev = NULL;
389 if (rb_prev)
390 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
391 *rb_link = __rb_link;
392 *rb_parent = __rb_parent;
393 return vma;
394}
395
396void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
397 struct rb_node **rb_link, struct rb_node *rb_parent)
398{
399 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
400 rb_insert_color(&vma->vm_rb, &mm->mm_rb);
401}
402
403static void __vma_link_file(struct vm_area_struct *vma)
404{
405 struct file *file;
406
407 file = vma->vm_file;
408 if (file) {
409 struct address_space *mapping = file->f_mapping;
410
411 if (vma->vm_flags & VM_DENYWRITE)
412 atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
413 if (vma->vm_flags & VM_SHARED)
414 mapping->i_mmap_writable++;
415
416 flush_dcache_mmap_lock(mapping);
417 if (unlikely(vma->vm_flags & VM_NONLINEAR))
418 vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
419 else
420 vma_prio_tree_insert(vma, &mapping->i_mmap);
421 flush_dcache_mmap_unlock(mapping);
422 }
423}
424
425static void
426__vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
427 struct vm_area_struct *prev, struct rb_node **rb_link,
428 struct rb_node *rb_parent)
429{
430 __vma_link_list(mm, vma, prev, rb_parent);
431 __vma_link_rb(mm, vma, rb_link, rb_parent);
432}
433
434static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
435 struct vm_area_struct *prev, struct rb_node **rb_link,
436 struct rb_node *rb_parent)
437{
438 struct address_space *mapping = NULL;
439
440 if (vma->vm_file)
441 mapping = vma->vm_file->f_mapping;
442
443 if (mapping)
444 mutex_lock(&mapping->i_mmap_mutex);
445
446 __vma_link(mm, vma, prev, rb_link, rb_parent);
447 __vma_link_file(vma);
448
449 if (mapping)
450 mutex_unlock(&mapping->i_mmap_mutex);
451
452 mm->map_count++;
453 validate_mm(mm);
454}
455
456/*
457 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
458 * mm's list and rbtree. It has already been inserted into the prio_tree.
459 */
460static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
461{
462 struct vm_area_struct *__vma, *prev;
463 struct rb_node **rb_link, *rb_parent;
464
465 __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
466 BUG_ON(__vma && __vma->vm_start < vma->vm_end);
467 __vma_link(mm, vma, prev, rb_link, rb_parent);
468 mm->map_count++;
469}
470
471static inline void
472__vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
473 struct vm_area_struct *prev)
474{
475 struct vm_area_struct *next = vma->vm_next;
476
477 prev->vm_next = next;
478 if (next)
479 next->vm_prev = prev;
480 rb_erase(&vma->vm_rb, &mm->mm_rb);
481 if (mm->mmap_cache == vma)
482 mm->mmap_cache = prev;
483}
484
485/*
486 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
487 * is already present in an i_mmap tree without adjusting the tree.
488 * The following helper function should be used when such adjustments
489 * are necessary. The "insert" vma (if any) is to be inserted
490 * before we drop the necessary locks.
491 */
492int vma_adjust(struct vm_area_struct *vma, unsigned long start,
493 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
494{
495 struct mm_struct *mm = vma->vm_mm;
496 struct vm_area_struct *next = vma->vm_next;
497 struct vm_area_struct *importer = NULL;
498 struct address_space *mapping = NULL;
499 struct prio_tree_root *root = NULL;
500 struct anon_vma *anon_vma = NULL;
501 struct file *file = vma->vm_file;
502 long adjust_next = 0;
503 int remove_next = 0;
504
505 if (next && !insert) {
506 struct vm_area_struct *exporter = NULL;
507
508 if (end >= next->vm_end) {
509 /*
510 * vma expands, overlapping all the next, and
511 * perhaps the one after too (mprotect case 6).
512 */
513again: remove_next = 1 + (end > next->vm_end);
514 end = next->vm_end;
515 exporter = next;
516 importer = vma;
517 } else if (end > next->vm_start) {
518 /*
519 * vma expands, overlapping part of the next:
520 * mprotect case 5 shifting the boundary up.
521 */
522 adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
523 exporter = next;
524 importer = vma;
525 } else if (end < vma->vm_end) {
526 /*
527 * vma shrinks, and !insert tells it's not
528 * split_vma inserting another: so it must be
529 * mprotect case 4 shifting the boundary down.
530 */
531 adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
532 exporter = vma;
533 importer = next;
534 }
535
536 /*
537 * Easily overlooked: when mprotect shifts the boundary,
538 * make sure the expanding vma has anon_vma set if the
539 * shrinking vma had, to cover any anon pages imported.
540 */
541 if (exporter && exporter->anon_vma && !importer->anon_vma) {
542 if (anon_vma_clone(importer, exporter))
543 return -ENOMEM;
544 importer->anon_vma = exporter->anon_vma;
545 }
546 }
547
548 if (file) {
549 mapping = file->f_mapping;
550 if (!(vma->vm_flags & VM_NONLINEAR)) {
551 root = &mapping->i_mmap;
552 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
553
554 if (adjust_next)
555 uprobe_munmap(next, next->vm_start,
556 next->vm_end);
557 }
558
559 mutex_lock(&mapping->i_mmap_mutex);
560 if (insert) {
561 /*
562 * Put into prio_tree now, so instantiated pages
563 * are visible to arm/parisc __flush_dcache_page
564 * throughout; but we cannot insert into address
565 * space until vma start or end is updated.
566 */
567 __vma_link_file(insert);
568 }
569 }
570
571 vma_adjust_trans_huge(vma, start, end, adjust_next);
572
573 /*
574 * When changing only vma->vm_end, we don't really need anon_vma
575 * lock. This is a fairly rare case by itself, but the anon_vma
576 * lock may be shared between many sibling processes. Skipping
577 * the lock for brk adjustments makes a difference sometimes.
578 */
579 if (vma->anon_vma && (importer || start != vma->vm_start)) {
580 anon_vma = vma->anon_vma;
581 anon_vma_lock(anon_vma);
582 }
583
584 if (root) {
585 flush_dcache_mmap_lock(mapping);
586 vma_prio_tree_remove(vma, root);
587 if (adjust_next)
588 vma_prio_tree_remove(next, root);
589 }
590
591 vma->vm_start = start;
592 vma->vm_end = end;
593 vma->vm_pgoff = pgoff;
594 if (adjust_next) {
595 next->vm_start += adjust_next << PAGE_SHIFT;
596 next->vm_pgoff += adjust_next;
597 }
598
599 if (root) {
600 if (adjust_next)
601 vma_prio_tree_insert(next, root);
602 vma_prio_tree_insert(vma, root);
603 flush_dcache_mmap_unlock(mapping);
604 }
605
606 if (remove_next) {
607 /*
608 * vma_merge has merged next into vma, and needs
609 * us to remove next before dropping the locks.
610 */
611 __vma_unlink(mm, next, vma);
612 if (file)
613 __remove_shared_vm_struct(next, file, mapping);
614 } else if (insert) {
615 /*
616 * split_vma has split insert from vma, and needs
617 * us to insert it before dropping the locks
618 * (it may either follow vma or precede it).
619 */
620 __insert_vm_struct(mm, insert);
621 }
622
623 if (anon_vma)
624 anon_vma_unlock(anon_vma);
625 if (mapping)
626 mutex_unlock(&mapping->i_mmap_mutex);
627
628 if (root) {
629 uprobe_mmap(vma);
630
631 if (adjust_next)
632 uprobe_mmap(next);
633 }
634
635 if (remove_next) {
636 if (file) {
637 uprobe_munmap(next, next->vm_start, next->vm_end);
638 fput(file);
639 if (next->vm_flags & VM_EXECUTABLE)
640 removed_exe_file_vma(mm);
641 }
642 if (next->anon_vma)
643 anon_vma_merge(vma, next);
644 mm->map_count--;
645 mpol_put(vma_policy(next));
646 kmem_cache_free(vm_area_cachep, next);
647 /*
648 * In mprotect's case 6 (see comments on vma_merge),
649 * we must remove another next too. It would clutter
650 * up the code too much to do both in one go.
651 */
652 if (remove_next == 2) {
653 next = vma->vm_next;
654 goto again;
655 }
656 }
657 if (insert && file)
658 uprobe_mmap(insert);
659
660 validate_mm(mm);
661
662 return 0;
663}
664
665/*
666 * If the vma has a ->close operation then the driver probably needs to release
667 * per-vma resources, so we don't attempt to merge those.
668 */
669static inline int is_mergeable_vma(struct vm_area_struct *vma,
670 struct file *file, unsigned long vm_flags)
671{
672 /* VM_CAN_NONLINEAR may get set later by f_op->mmap() */
673 if ((vma->vm_flags ^ vm_flags) & ~VM_CAN_NONLINEAR)
674 return 0;
675 if (vma->vm_file != file)
676 return 0;
677 if (vma->vm_ops && vma->vm_ops->close)
678 return 0;
679 return 1;
680}
681
682static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
683 struct anon_vma *anon_vma2,
684 struct vm_area_struct *vma)
685{
686 /*
687 * The list_is_singular() test is to avoid merging VMA cloned from
688 * parents. This can improve scalability caused by anon_vma lock.
689 */
690 if ((!anon_vma1 || !anon_vma2) && (!vma ||
691 list_is_singular(&vma->anon_vma_chain)))
692 return 1;
693 return anon_vma1 == anon_vma2;
694}
695
696/*
697 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
698 * in front of (at a lower virtual address and file offset than) the vma.
699 *
700 * We cannot merge two vmas if they have differently assigned (non-NULL)
701 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
702 *
703 * We don't check here for the merged mmap wrapping around the end of pagecache
704 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
705 * wrap, nor mmaps which cover the final page at index -1UL.
706 */
707static int
708can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
709 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
710{
711 if (is_mergeable_vma(vma, file, vm_flags) &&
712 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
713 if (vma->vm_pgoff == vm_pgoff)
714 return 1;
715 }
716 return 0;
717}
718
719/*
720 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
721 * beyond (at a higher virtual address and file offset than) the vma.
722 *
723 * We cannot merge two vmas if they have differently assigned (non-NULL)
724 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
725 */
726static int
727can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
728 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
729{
730 if (is_mergeable_vma(vma, file, vm_flags) &&
731 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
732 pgoff_t vm_pglen;
733 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
734 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
735 return 1;
736 }
737 return 0;
738}
739
740/*
741 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
742 * whether that can be merged with its predecessor or its successor.
743 * Or both (it neatly fills a hole).
744 *
745 * In most cases - when called for mmap, brk or mremap - [addr,end) is
746 * certain not to be mapped by the time vma_merge is called; but when
747 * called for mprotect, it is certain to be already mapped (either at
748 * an offset within prev, or at the start of next), and the flags of
749 * this area are about to be changed to vm_flags - and the no-change
750 * case has already been eliminated.
751 *
752 * The following mprotect cases have to be considered, where AAAA is
753 * the area passed down from mprotect_fixup, never extending beyond one
754 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
755 *
756 * AAAA AAAA AAAA AAAA
757 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
758 * cannot merge might become might become might become
759 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
760 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
761 * mremap move: PPPPNNNNNNNN 8
762 * AAAA
763 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
764 * might become case 1 below case 2 below case 3 below
765 *
766 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
767 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
768 */
769struct vm_area_struct *vma_merge(struct mm_struct *mm,
770 struct vm_area_struct *prev, unsigned long addr,
771 unsigned long end, unsigned long vm_flags,
772 struct anon_vma *anon_vma, struct file *file,
773 pgoff_t pgoff, struct mempolicy *policy)
774{
775 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
776 struct vm_area_struct *area, *next;
777 int err;
778
779 /*
780 * We later require that vma->vm_flags == vm_flags,
781 * so this tests vma->vm_flags & VM_SPECIAL, too.
782 */
783 if (vm_flags & VM_SPECIAL)
784 return NULL;
785
786 if (prev)
787 next = prev->vm_next;
788 else
789 next = mm->mmap;
790 area = next;
791 if (next && next->vm_end == end) /* cases 6, 7, 8 */
792 next = next->vm_next;
793
794 /*
795 * Can it merge with the predecessor?
796 */
797 if (prev && prev->vm_end == addr &&
798 mpol_equal(vma_policy(prev), policy) &&
799 can_vma_merge_after(prev, vm_flags,
800 anon_vma, file, pgoff)) {
801 /*
802 * OK, it can. Can we now merge in the successor as well?
803 */
804 if (next && end == next->vm_start &&
805 mpol_equal(policy, vma_policy(next)) &&
806 can_vma_merge_before(next, vm_flags,
807 anon_vma, file, pgoff+pglen) &&
808 is_mergeable_anon_vma(prev->anon_vma,
809 next->anon_vma, NULL)) {
810 /* cases 1, 6 */
811 err = vma_adjust(prev, prev->vm_start,
812 next->vm_end, prev->vm_pgoff, NULL);
813 } else /* cases 2, 5, 7 */
814 err = vma_adjust(prev, prev->vm_start,
815 end, prev->vm_pgoff, NULL);
816 if (err)
817 return NULL;
818 khugepaged_enter_vma_merge(prev);
819 return prev;
820 }
821
822 /*
823 * Can this new request be merged in front of next?
824 */
825 if (next && end == next->vm_start &&
826 mpol_equal(policy, vma_policy(next)) &&
827 can_vma_merge_before(next, vm_flags,
828 anon_vma, file, pgoff+pglen)) {
829 if (prev && addr < prev->vm_end) /* case 4 */
830 err = vma_adjust(prev, prev->vm_start,
831 addr, prev->vm_pgoff, NULL);
832 else /* cases 3, 8 */
833 err = vma_adjust(area, addr, next->vm_end,
834 next->vm_pgoff - pglen, NULL);
835 if (err)
836 return NULL;
837 khugepaged_enter_vma_merge(area);
838 return area;
839 }
840
841 return NULL;
842}
843
844/*
845 * Rough compatbility check to quickly see if it's even worth looking
846 * at sharing an anon_vma.
847 *
848 * They need to have the same vm_file, and the flags can only differ
849 * in things that mprotect may change.
850 *
851 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
852 * we can merge the two vma's. For example, we refuse to merge a vma if
853 * there is a vm_ops->close() function, because that indicates that the
854 * driver is doing some kind of reference counting. But that doesn't
855 * really matter for the anon_vma sharing case.
856 */
857static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
858{
859 return a->vm_end == b->vm_start &&
860 mpol_equal(vma_policy(a), vma_policy(b)) &&
861 a->vm_file == b->vm_file &&
862 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC)) &&
863 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
864}
865
866/*
867 * Do some basic sanity checking to see if we can re-use the anon_vma
868 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
869 * the same as 'old', the other will be the new one that is trying
870 * to share the anon_vma.
871 *
872 * NOTE! This runs with mm_sem held for reading, so it is possible that
873 * the anon_vma of 'old' is concurrently in the process of being set up
874 * by another page fault trying to merge _that_. But that's ok: if it
875 * is being set up, that automatically means that it will be a singleton
876 * acceptable for merging, so we can do all of this optimistically. But
877 * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
878 *
879 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
880 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
881 * is to return an anon_vma that is "complex" due to having gone through
882 * a fork).
883 *
884 * We also make sure that the two vma's are compatible (adjacent,
885 * and with the same memory policies). That's all stable, even with just
886 * a read lock on the mm_sem.
887 */
888static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
889{
890 if (anon_vma_compatible(a, b)) {
891 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma);
892
893 if (anon_vma && list_is_singular(&old->anon_vma_chain))
894 return anon_vma;
895 }
896 return NULL;
897}
898
899/*
900 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
901 * neighbouring vmas for a suitable anon_vma, before it goes off
902 * to allocate a new anon_vma. It checks because a repetitive
903 * sequence of mprotects and faults may otherwise lead to distinct
904 * anon_vmas being allocated, preventing vma merge in subsequent
905 * mprotect.
906 */
907struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
908{
909 struct anon_vma *anon_vma;
910 struct vm_area_struct *near;
911
912 near = vma->vm_next;
913 if (!near)
914 goto try_prev;
915
916 anon_vma = reusable_anon_vma(near, vma, near);
917 if (anon_vma)
918 return anon_vma;
919try_prev:
920 near = vma->vm_prev;
921 if (!near)
922 goto none;
923
924 anon_vma = reusable_anon_vma(near, near, vma);
925 if (anon_vma)
926 return anon_vma;
927none:
928 /*
929 * There's no absolute need to look only at touching neighbours:
930 * we could search further afield for "compatible" anon_vmas.
931 * But it would probably just be a waste of time searching,
932 * or lead to too many vmas hanging off the same anon_vma.
933 * We're trying to allow mprotect remerging later on,
934 * not trying to minimize memory used for anon_vmas.
935 */
936 return NULL;
937}
938
939#ifdef CONFIG_PROC_FS
940void vm_stat_account(struct mm_struct *mm, unsigned long flags,
941 struct file *file, long pages)
942{
943 const unsigned long stack_flags
944 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
945
946 if (file) {
947 mm->shared_vm += pages;
948 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
949 mm->exec_vm += pages;
950 } else if (flags & stack_flags)
951 mm->stack_vm += pages;
952 if (flags & (VM_RESERVED|VM_IO))
953 mm->reserved_vm += pages;
954}
955#endif /* CONFIG_PROC_FS */
956
957/*
958 * If a hint addr is less than mmap_min_addr change hint to be as
959 * low as possible but still greater than mmap_min_addr
960 */
961static inline unsigned long round_hint_to_min(unsigned long hint)
962{
963 hint &= PAGE_MASK;
964 if (((void *)hint != NULL) &&
965 (hint < mmap_min_addr))
966 return PAGE_ALIGN(mmap_min_addr);
967 return hint;
968}
969
970/*
971 * The caller must hold down_write(¤t->mm->mmap_sem).
972 */
973
974unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
975 unsigned long len, unsigned long prot,
976 unsigned long flags, unsigned long pgoff)
977{
978 struct mm_struct * mm = current->mm;
979 struct inode *inode;
980 vm_flags_t vm_flags;
981
982 /*
983 * Does the application expect PROT_READ to imply PROT_EXEC?
984 *
985 * (the exception is when the underlying filesystem is noexec
986 * mounted, in which case we dont add PROT_EXEC.)
987 */
988 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
989 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
990 prot |= PROT_EXEC;
991
992 if (!len)
993 return -EINVAL;
994
995 if (!(flags & MAP_FIXED))
996 addr = round_hint_to_min(addr);
997
998 /* Careful about overflows.. */
999 len = PAGE_ALIGN(len);
1000 if (!len)
1001 return -ENOMEM;
1002
1003 /* offset overflow? */
1004 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1005 return -EOVERFLOW;
1006
1007 /* Too many mappings? */
1008 if (mm->map_count > sysctl_max_map_count)
1009 return -ENOMEM;
1010
1011 /* Obtain the address to map to. we verify (or select) it and ensure
1012 * that it represents a valid section of the address space.
1013 */
1014 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1015 if (addr & ~PAGE_MASK)
1016 return addr;
1017
1018 /* Do simple checking here so the lower-level routines won't have
1019 * to. we assume access permissions have been handled by the open
1020 * of the memory object, so we don't do any here.
1021 */
1022 vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1023 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1024
1025 if (flags & MAP_LOCKED)
1026 if (!can_do_mlock())
1027 return -EPERM;
1028
1029 /* mlock MCL_FUTURE? */
1030 if (vm_flags & VM_LOCKED) {
1031 unsigned long locked, lock_limit;
1032 locked = len >> PAGE_SHIFT;
1033 locked += mm->locked_vm;
1034 lock_limit = rlimit(RLIMIT_MEMLOCK);
1035 lock_limit >>= PAGE_SHIFT;
1036 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1037 return -EAGAIN;
1038 }
1039
1040 inode = file ? file->f_path.dentry->d_inode : NULL;
1041
1042 if (file) {
1043 switch (flags & MAP_TYPE) {
1044 case MAP_SHARED:
1045 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1046 return -EACCES;
1047
1048 /*
1049 * Make sure we don't allow writing to an append-only
1050 * file..
1051 */
1052 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1053 return -EACCES;
1054
1055 /*
1056 * Make sure there are no mandatory locks on the file.
1057 */
1058 if (locks_verify_locked(inode))
1059 return -EAGAIN;
1060
1061 vm_flags |= VM_SHARED | VM_MAYSHARE;
1062 if (!(file->f_mode & FMODE_WRITE))
1063 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1064
1065 /* fall through */
1066 case MAP_PRIVATE:
1067 if (!(file->f_mode & FMODE_READ))
1068 return -EACCES;
1069 if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1070 if (vm_flags & VM_EXEC)
1071 return -EPERM;
1072 vm_flags &= ~VM_MAYEXEC;
1073 }
1074
1075 if (!file->f_op || !file->f_op->mmap)
1076 return -ENODEV;
1077 break;
1078
1079 default:
1080 return -EINVAL;
1081 }
1082 } else {
1083 switch (flags & MAP_TYPE) {
1084 case MAP_SHARED:
1085 /*
1086 * Ignore pgoff.
1087 */
1088 pgoff = 0;
1089 vm_flags |= VM_SHARED | VM_MAYSHARE;
1090 break;
1091 case MAP_PRIVATE:
1092 /*
1093 * Set pgoff according to addr for anon_vma.
1094 */
1095 pgoff = addr >> PAGE_SHIFT;
1096 break;
1097 default:
1098 return -EINVAL;
1099 }
1100 }
1101
1102 return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1103}
1104
1105SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1106 unsigned long, prot, unsigned long, flags,
1107 unsigned long, fd, unsigned long, pgoff)
1108{
1109 struct file *file = NULL;
1110 unsigned long retval = -EBADF;
1111
1112 if (!(flags & MAP_ANONYMOUS)) {
1113 audit_mmap_fd(fd, flags);
1114 if (unlikely(flags & MAP_HUGETLB))
1115 return -EINVAL;
1116 file = fget(fd);
1117 if (!file)
1118 goto out;
1119 } else if (flags & MAP_HUGETLB) {
1120 struct user_struct *user = NULL;
1121 /*
1122 * VM_NORESERVE is used because the reservations will be
1123 * taken when vm_ops->mmap() is called
1124 * A dummy user value is used because we are not locking
1125 * memory so no accounting is necessary
1126 */
1127 file = hugetlb_file_setup(HUGETLB_ANON_FILE, addr, len,
1128 VM_NORESERVE, &user,
1129 HUGETLB_ANONHUGE_INODE);
1130 if (IS_ERR(file))
1131 return PTR_ERR(file);
1132 }
1133
1134 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1135
1136 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1137 if (file)
1138 fput(file);
1139out:
1140 return retval;
1141}
1142
1143#ifdef __ARCH_WANT_SYS_OLD_MMAP
1144struct mmap_arg_struct {
1145 unsigned long addr;
1146 unsigned long len;
1147 unsigned long prot;
1148 unsigned long flags;
1149 unsigned long fd;
1150 unsigned long offset;
1151};
1152
1153SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1154{
1155 struct mmap_arg_struct a;
1156
1157 if (copy_from_user(&a, arg, sizeof(a)))
1158 return -EFAULT;
1159 if (a.offset & ~PAGE_MASK)
1160 return -EINVAL;
1161
1162 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1163 a.offset >> PAGE_SHIFT);
1164}
1165#endif /* __ARCH_WANT_SYS_OLD_MMAP */
1166
1167/*
1168 * Some shared mappigns will want the pages marked read-only
1169 * to track write events. If so, we'll downgrade vm_page_prot
1170 * to the private version (using protection_map[] without the
1171 * VM_SHARED bit).
1172 */
1173int vma_wants_writenotify(struct vm_area_struct *vma)
1174{
1175 vm_flags_t vm_flags = vma->vm_flags;
1176
1177 /* If it was private or non-writable, the write bit is already clear */
1178 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1179 return 0;
1180
1181 /* The backer wishes to know when pages are first written to? */
1182 if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1183 return 1;
1184
1185 /* The open routine did something to the protections already? */
1186 if (pgprot_val(vma->vm_page_prot) !=
1187 pgprot_val(vm_get_page_prot(vm_flags)))
1188 return 0;
1189
1190 /* Specialty mapping? */
1191 if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1192 return 0;
1193
1194 /* Can the mapping track the dirty pages? */
1195 return vma->vm_file && vma->vm_file->f_mapping &&
1196 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1197}
1198
1199/*
1200 * We account for memory if it's a private writeable mapping,
1201 * not hugepages and VM_NORESERVE wasn't set.
1202 */
1203static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1204{
1205 /*
1206 * hugetlb has its own accounting separate from the core VM
1207 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1208 */
1209 if (file && is_file_hugepages(file))
1210 return 0;
1211
1212 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1213}
1214
1215unsigned long mmap_region(struct file *file, unsigned long addr,
1216 unsigned long len, unsigned long flags,
1217 vm_flags_t vm_flags, unsigned long pgoff)
1218{
1219 struct mm_struct *mm = current->mm;
1220 struct vm_area_struct *vma, *prev;
1221 int correct_wcount = 0;
1222 int error;
1223 struct rb_node **rb_link, *rb_parent;
1224 unsigned long charged = 0;
1225 struct inode *inode = file ? file->f_path.dentry->d_inode : NULL;
1226
1227 /* Clear old maps */
1228 error = -ENOMEM;
1229munmap_back:
1230 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1231 if (vma && vma->vm_start < addr + len) {
1232 if (do_munmap(mm, addr, len))
1233 return -ENOMEM;
1234 goto munmap_back;
1235 }
1236
1237 /* Check against address space limit. */
1238 if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1239 return -ENOMEM;
1240
1241 /*
1242 * Set 'VM_NORESERVE' if we should not account for the
1243 * memory use of this mapping.
1244 */
1245 if ((flags & MAP_NORESERVE)) {
1246 /* We honor MAP_NORESERVE if allowed to overcommit */
1247 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1248 vm_flags |= VM_NORESERVE;
1249
1250 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1251 if (file && is_file_hugepages(file))
1252 vm_flags |= VM_NORESERVE;
1253 }
1254
1255 /*
1256 * Private writable mapping: check memory availability
1257 */
1258 if (accountable_mapping(file, vm_flags)) {
1259 charged = len >> PAGE_SHIFT;
1260 if (security_vm_enough_memory_mm(mm, charged))
1261 return -ENOMEM;
1262 vm_flags |= VM_ACCOUNT;
1263 }
1264
1265 /*
1266 * Can we just expand an old mapping?
1267 */
1268 vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1269 if (vma)
1270 goto out;
1271
1272 /*
1273 * Determine the object being mapped and call the appropriate
1274 * specific mapper. the address has already been validated, but
1275 * not unmapped, but the maps are removed from the list.
1276 */
1277 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1278 if (!vma) {
1279 error = -ENOMEM;
1280 goto unacct_error;
1281 }
1282
1283 vma->vm_mm = mm;
1284 vma->vm_start = addr;
1285 vma->vm_end = addr + len;
1286 vma->vm_flags = vm_flags;
1287 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1288 vma->vm_pgoff = pgoff;
1289 INIT_LIST_HEAD(&vma->anon_vma_chain);
1290
1291 error = -EINVAL; /* when rejecting VM_GROWSDOWN|VM_GROWSUP */
1292
1293 if (file) {
1294 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1295 goto free_vma;
1296 if (vm_flags & VM_DENYWRITE) {
1297 error = deny_write_access(file);
1298 if (error)
1299 goto free_vma;
1300 correct_wcount = 1;
1301 }
1302 vma->vm_file = file;
1303 get_file(file);
1304 error = file->f_op->mmap(file, vma);
1305 if (error)
1306 goto unmap_and_free_vma;
1307 if (vm_flags & VM_EXECUTABLE)
1308 added_exe_file_vma(mm);
1309
1310 /* Can addr have changed??
1311 *
1312 * Answer: Yes, several device drivers can do it in their
1313 * f_op->mmap method. -DaveM
1314 */
1315 addr = vma->vm_start;
1316 pgoff = vma->vm_pgoff;
1317 vm_flags = vma->vm_flags;
1318 } else if (vm_flags & VM_SHARED) {
1319 if (unlikely(vm_flags & (VM_GROWSDOWN|VM_GROWSUP)))
1320 goto free_vma;
1321 error = shmem_zero_setup(vma);
1322 if (error)
1323 goto free_vma;
1324 }
1325
1326 if (vma_wants_writenotify(vma)) {
1327 pgprot_t pprot = vma->vm_page_prot;
1328
1329 /* Can vma->vm_page_prot have changed??
1330 *
1331 * Answer: Yes, drivers may have changed it in their
1332 * f_op->mmap method.
1333 *
1334 * Ensures that vmas marked as uncached stay that way.
1335 */
1336 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1337 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1338 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1339 }
1340
1341 vma_link(mm, vma, prev, rb_link, rb_parent);
1342 file = vma->vm_file;
1343
1344 /* Once vma denies write, undo our temporary denial count */
1345 if (correct_wcount)
1346 atomic_inc(&inode->i_writecount);
1347out:
1348 perf_event_mmap(vma);
1349
1350 mm->total_vm += len >> PAGE_SHIFT;
1351 vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1352 if (vm_flags & VM_LOCKED) {
1353 if (!mlock_vma_pages_range(vma, addr, addr + len))
1354 mm->locked_vm += (len >> PAGE_SHIFT);
1355 } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1356 make_pages_present(addr, addr + len);
1357
1358 if (file)
1359 uprobe_mmap(vma);
1360
1361 return addr;
1362
1363unmap_and_free_vma:
1364 if (correct_wcount)
1365 atomic_inc(&inode->i_writecount);
1366 vma->vm_file = NULL;
1367 fput(file);
1368
1369 /* Undo any partial mapping done by a device driver. */
1370 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1371 charged = 0;
1372free_vma:
1373 kmem_cache_free(vm_area_cachep, vma);
1374unacct_error:
1375 if (charged)
1376 vm_unacct_memory(charged);
1377 return error;
1378}
1379
1380/* Get an address range which is currently unmapped.
1381 * For shmat() with addr=0.
1382 *
1383 * Ugly calling convention alert:
1384 * Return value with the low bits set means error value,
1385 * ie
1386 * if (ret & ~PAGE_MASK)
1387 * error = ret;
1388 *
1389 * This function "knows" that -ENOMEM has the bits set.
1390 */
1391#ifndef HAVE_ARCH_UNMAPPED_AREA
1392unsigned long
1393arch_get_unmapped_area(struct file *filp, unsigned long addr,
1394 unsigned long len, unsigned long pgoff, unsigned long flags)
1395{
1396 struct mm_struct *mm = current->mm;
1397 struct vm_area_struct *vma;
1398 unsigned long start_addr;
1399
1400 if (len > TASK_SIZE)
1401 return -ENOMEM;
1402
1403 if (flags & MAP_FIXED)
1404 return addr;
1405
1406 if (addr) {
1407 addr = PAGE_ALIGN(addr);
1408 vma = find_vma(mm, addr);
1409 if (TASK_SIZE - len >= addr &&
1410 (!vma || addr + len <= vma->vm_start))
1411 return addr;
1412 }
1413 if (len > mm->cached_hole_size) {
1414 start_addr = addr = mm->free_area_cache;
1415 } else {
1416 start_addr = addr = TASK_UNMAPPED_BASE;
1417 mm->cached_hole_size = 0;
1418 }
1419
1420full_search:
1421 for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1422 /* At this point: (!vma || addr < vma->vm_end). */
1423 if (TASK_SIZE - len < addr) {
1424 /*
1425 * Start a new search - just in case we missed
1426 * some holes.
1427 */
1428 if (start_addr != TASK_UNMAPPED_BASE) {
1429 addr = TASK_UNMAPPED_BASE;
1430 start_addr = addr;
1431 mm->cached_hole_size = 0;
1432 goto full_search;
1433 }
1434 return -ENOMEM;
1435 }
1436 if (!vma || addr + len <= vma->vm_start) {
1437 /*
1438 * Remember the place where we stopped the search:
1439 */
1440 mm->free_area_cache = addr + len;
1441 return addr;
1442 }
1443 if (addr + mm->cached_hole_size < vma->vm_start)
1444 mm->cached_hole_size = vma->vm_start - addr;
1445 addr = vma->vm_end;
1446 }
1447}
1448#endif
1449
1450void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1451{
1452 /*
1453 * Is this a new hole at the lowest possible address?
1454 */
1455 if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache)
1456 mm->free_area_cache = addr;
1457}
1458
1459/*
1460 * This mmap-allocator allocates new areas top-down from below the
1461 * stack's low limit (the base):
1462 */
1463#ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1464unsigned long
1465arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1466 const unsigned long len, const unsigned long pgoff,
1467 const unsigned long flags)
1468{
1469 struct vm_area_struct *vma;
1470 struct mm_struct *mm = current->mm;
1471 unsigned long addr = addr0, start_addr;
1472
1473 /* requested length too big for entire address space */
1474 if (len > TASK_SIZE)
1475 return -ENOMEM;
1476
1477 if (flags & MAP_FIXED)
1478 return addr;
1479
1480 /* requesting a specific address */
1481 if (addr) {
1482 addr = PAGE_ALIGN(addr);
1483 vma = find_vma(mm, addr);
1484 if (TASK_SIZE - len >= addr &&
1485 (!vma || addr + len <= vma->vm_start))
1486 return addr;
1487 }
1488
1489 /* check if free_area_cache is useful for us */
1490 if (len <= mm->cached_hole_size) {
1491 mm->cached_hole_size = 0;
1492 mm->free_area_cache = mm->mmap_base;
1493 }
1494
1495try_again:
1496 /* either no address requested or can't fit in requested address hole */
1497 start_addr = addr = mm->free_area_cache;
1498
1499 if (addr < len)
1500 goto fail;
1501
1502 addr -= len;
1503 do {
1504 /*
1505 * Lookup failure means no vma is above this address,
1506 * else if new region fits below vma->vm_start,
1507 * return with success:
1508 */
1509 vma = find_vma(mm, addr);
1510 if (!vma || addr+len <= vma->vm_start)
1511 /* remember the address as a hint for next time */
1512 return (mm->free_area_cache = addr);
1513
1514 /* remember the largest hole we saw so far */
1515 if (addr + mm->cached_hole_size < vma->vm_start)
1516 mm->cached_hole_size = vma->vm_start - addr;
1517
1518 /* try just below the current vma->vm_start */
1519 addr = vma->vm_start-len;
1520 } while (len < vma->vm_start);
1521
1522fail:
1523 /*
1524 * if hint left us with no space for the requested
1525 * mapping then try again:
1526 *
1527 * Note: this is different with the case of bottomup
1528 * which does the fully line-search, but we use find_vma
1529 * here that causes some holes skipped.
1530 */
1531 if (start_addr != mm->mmap_base) {
1532 mm->free_area_cache = mm->mmap_base;
1533 mm->cached_hole_size = 0;
1534 goto try_again;
1535 }
1536
1537 /*
1538 * A failed mmap() very likely causes application failure,
1539 * so fall back to the bottom-up function here. This scenario
1540 * can happen with large stack limits and large mmap()
1541 * allocations.
1542 */
1543 mm->cached_hole_size = ~0UL;
1544 mm->free_area_cache = TASK_UNMAPPED_BASE;
1545 addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1546 /*
1547 * Restore the topdown base:
1548 */
1549 mm->free_area_cache = mm->mmap_base;
1550 mm->cached_hole_size = ~0UL;
1551
1552 return addr;
1553}
1554#endif
1555
1556void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1557{
1558 /*
1559 * Is this a new hole at the highest possible address?
1560 */
1561 if (addr > mm->free_area_cache)
1562 mm->free_area_cache = addr;
1563
1564 /* dont allow allocations above current base */
1565 if (mm->free_area_cache > mm->mmap_base)
1566 mm->free_area_cache = mm->mmap_base;
1567}
1568
1569unsigned long
1570get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1571 unsigned long pgoff, unsigned long flags)
1572{
1573 unsigned long (*get_area)(struct file *, unsigned long,
1574 unsigned long, unsigned long, unsigned long);
1575
1576 unsigned long error = arch_mmap_check(addr, len, flags);
1577 if (error)
1578 return error;
1579
1580 /* Careful about overflows.. */
1581 if (len > TASK_SIZE)
1582 return -ENOMEM;
1583
1584 get_area = current->mm->get_unmapped_area;
1585 if (file && file->f_op && file->f_op->get_unmapped_area)
1586 get_area = file->f_op->get_unmapped_area;
1587 addr = get_area(file, addr, len, pgoff, flags);
1588 if (IS_ERR_VALUE(addr))
1589 return addr;
1590
1591 if (addr > TASK_SIZE - len)
1592 return -ENOMEM;
1593 if (addr & ~PAGE_MASK)
1594 return -EINVAL;
1595
1596 addr = arch_rebalance_pgtables(addr, len);
1597 error = security_mmap_addr(addr);
1598 return error ? error : addr;
1599}
1600
1601EXPORT_SYMBOL(get_unmapped_area);
1602
1603/* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
1604struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1605{
1606 struct vm_area_struct *vma = NULL;
1607
1608 if (WARN_ON_ONCE(!mm)) /* Remove this in linux-3.6 */
1609 return NULL;
1610
1611 /* Check the cache first. */
1612 /* (Cache hit rate is typically around 35%.) */
1613 vma = mm->mmap_cache;
1614 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1615 struct rb_node *rb_node;
1616
1617 rb_node = mm->mm_rb.rb_node;
1618 vma = NULL;
1619
1620 while (rb_node) {
1621 struct vm_area_struct *vma_tmp;
1622
1623 vma_tmp = rb_entry(rb_node,
1624 struct vm_area_struct, vm_rb);
1625
1626 if (vma_tmp->vm_end > addr) {
1627 vma = vma_tmp;
1628 if (vma_tmp->vm_start <= addr)
1629 break;
1630 rb_node = rb_node->rb_left;
1631 } else
1632 rb_node = rb_node->rb_right;
1633 }
1634 if (vma)
1635 mm->mmap_cache = vma;
1636 }
1637 return vma;
1638}
1639
1640EXPORT_SYMBOL(find_vma);
1641
1642/*
1643 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
1644 */
1645struct vm_area_struct *
1646find_vma_prev(struct mm_struct *mm, unsigned long addr,
1647 struct vm_area_struct **pprev)
1648{
1649 struct vm_area_struct *vma;
1650
1651 vma = find_vma(mm, addr);
1652 if (vma) {
1653 *pprev = vma->vm_prev;
1654 } else {
1655 struct rb_node *rb_node = mm->mm_rb.rb_node;
1656 *pprev = NULL;
1657 while (rb_node) {
1658 *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1659 rb_node = rb_node->rb_right;
1660 }
1661 }
1662 return vma;
1663}
1664
1665/*
1666 * Verify that the stack growth is acceptable and
1667 * update accounting. This is shared with both the
1668 * grow-up and grow-down cases.
1669 */
1670static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1671{
1672 struct mm_struct *mm = vma->vm_mm;
1673 struct rlimit *rlim = current->signal->rlim;
1674 unsigned long new_start;
1675
1676 /* address space limit tests */
1677 if (!may_expand_vm(mm, grow))
1678 return -ENOMEM;
1679
1680 /* Stack limit test */
1681 if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
1682 return -ENOMEM;
1683
1684 /* mlock limit tests */
1685 if (vma->vm_flags & VM_LOCKED) {
1686 unsigned long locked;
1687 unsigned long limit;
1688 locked = mm->locked_vm + grow;
1689 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
1690 limit >>= PAGE_SHIFT;
1691 if (locked > limit && !capable(CAP_IPC_LOCK))
1692 return -ENOMEM;
1693 }
1694
1695 /* Check to ensure the stack will not grow into a hugetlb-only region */
1696 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1697 vma->vm_end - size;
1698 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1699 return -EFAULT;
1700
1701 /*
1702 * Overcommit.. This must be the final test, as it will
1703 * update security statistics.
1704 */
1705 if (security_vm_enough_memory_mm(mm, grow))
1706 return -ENOMEM;
1707
1708 /* Ok, everything looks good - let it rip */
1709 mm->total_vm += grow;
1710 if (vma->vm_flags & VM_LOCKED)
1711 mm->locked_vm += grow;
1712 vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1713 return 0;
1714}
1715
1716#if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1717/*
1718 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1719 * vma is the last one with address > vma->vm_end. Have to extend vma.
1720 */
1721int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1722{
1723 int error;
1724
1725 if (!(vma->vm_flags & VM_GROWSUP))
1726 return -EFAULT;
1727
1728 /*
1729 * We must make sure the anon_vma is allocated
1730 * so that the anon_vma locking is not a noop.
1731 */
1732 if (unlikely(anon_vma_prepare(vma)))
1733 return -ENOMEM;
1734 vma_lock_anon_vma(vma);
1735
1736 /*
1737 * vma->vm_start/vm_end cannot change under us because the caller
1738 * is required to hold the mmap_sem in read mode. We need the
1739 * anon_vma lock to serialize against concurrent expand_stacks.
1740 * Also guard against wrapping around to address 0.
1741 */
1742 if (address < PAGE_ALIGN(address+4))
1743 address = PAGE_ALIGN(address+4);
1744 else {
1745 vma_unlock_anon_vma(vma);
1746 return -ENOMEM;
1747 }
1748 error = 0;
1749
1750 /* Somebody else might have raced and expanded it already */
1751 if (address > vma->vm_end) {
1752 unsigned long size, grow;
1753
1754 size = address - vma->vm_start;
1755 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1756
1757 error = -ENOMEM;
1758 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
1759 error = acct_stack_growth(vma, size, grow);
1760 if (!error) {
1761 vma->vm_end = address;
1762 perf_event_mmap(vma);
1763 }
1764 }
1765 }
1766 vma_unlock_anon_vma(vma);
1767 khugepaged_enter_vma_merge(vma);
1768 return error;
1769}
1770#endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1771
1772/*
1773 * vma is the first one with address < vma->vm_start. Have to extend vma.
1774 */
1775int expand_downwards(struct vm_area_struct *vma,
1776 unsigned long address)
1777{
1778 int error;
1779
1780 /*
1781 * We must make sure the anon_vma is allocated
1782 * so that the anon_vma locking is not a noop.
1783 */
1784 if (unlikely(anon_vma_prepare(vma)))
1785 return -ENOMEM;
1786
1787 address &= PAGE_MASK;
1788 error = security_mmap_addr(address);
1789 if (error)
1790 return error;
1791
1792 vma_lock_anon_vma(vma);
1793
1794 /*
1795 * vma->vm_start/vm_end cannot change under us because the caller
1796 * is required to hold the mmap_sem in read mode. We need the
1797 * anon_vma lock to serialize against concurrent expand_stacks.
1798 */
1799
1800 /* Somebody else might have raced and expanded it already */
1801 if (address < vma->vm_start) {
1802 unsigned long size, grow;
1803
1804 size = vma->vm_end - address;
1805 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1806
1807 error = -ENOMEM;
1808 if (grow <= vma->vm_pgoff) {
1809 error = acct_stack_growth(vma, size, grow);
1810 if (!error) {
1811 vma->vm_start = address;
1812 vma->vm_pgoff -= grow;
1813 perf_event_mmap(vma);
1814 }
1815 }
1816 }
1817 vma_unlock_anon_vma(vma);
1818 khugepaged_enter_vma_merge(vma);
1819 return error;
1820}
1821
1822#ifdef CONFIG_STACK_GROWSUP
1823int expand_stack(struct vm_area_struct *vma, unsigned long address)
1824{
1825 return expand_upwards(vma, address);
1826}
1827
1828struct vm_area_struct *
1829find_extend_vma(struct mm_struct *mm, unsigned long addr)
1830{
1831 struct vm_area_struct *vma, *prev;
1832
1833 addr &= PAGE_MASK;
1834 vma = find_vma_prev(mm, addr, &prev);
1835 if (vma && (vma->vm_start <= addr))
1836 return vma;
1837 if (!prev || expand_stack(prev, addr))
1838 return NULL;
1839 if (prev->vm_flags & VM_LOCKED) {
1840 mlock_vma_pages_range(prev, addr, prev->vm_end);
1841 }
1842 return prev;
1843}
1844#else
1845int expand_stack(struct vm_area_struct *vma, unsigned long address)
1846{
1847 return expand_downwards(vma, address);
1848}
1849
1850struct vm_area_struct *
1851find_extend_vma(struct mm_struct * mm, unsigned long addr)
1852{
1853 struct vm_area_struct * vma;
1854 unsigned long start;
1855
1856 addr &= PAGE_MASK;
1857 vma = find_vma(mm,addr);
1858 if (!vma)
1859 return NULL;
1860 if (vma->vm_start <= addr)
1861 return vma;
1862 if (!(vma->vm_flags & VM_GROWSDOWN))
1863 return NULL;
1864 start = vma->vm_start;
1865 if (expand_stack(vma, addr))
1866 return NULL;
1867 if (vma->vm_flags & VM_LOCKED) {
1868 mlock_vma_pages_range(vma, addr, start);
1869 }
1870 return vma;
1871}
1872#endif
1873
1874/*
1875 * Ok - we have the memory areas we should free on the vma list,
1876 * so release them, and do the vma updates.
1877 *
1878 * Called with the mm semaphore held.
1879 */
1880static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1881{
1882 unsigned long nr_accounted = 0;
1883
1884 /* Update high watermark before we lower total_vm */
1885 update_hiwater_vm(mm);
1886 do {
1887 long nrpages = vma_pages(vma);
1888
1889 if (vma->vm_flags & VM_ACCOUNT)
1890 nr_accounted += nrpages;
1891 mm->total_vm -= nrpages;
1892 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1893 vma = remove_vma(vma);
1894 } while (vma);
1895 vm_unacct_memory(nr_accounted);
1896 validate_mm(mm);
1897}
1898
1899/*
1900 * Get rid of page table information in the indicated region.
1901 *
1902 * Called with the mm semaphore held.
1903 */
1904static void unmap_region(struct mm_struct *mm,
1905 struct vm_area_struct *vma, struct vm_area_struct *prev,
1906 unsigned long start, unsigned long end)
1907{
1908 struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1909 struct mmu_gather tlb;
1910
1911 lru_add_drain();
1912 tlb_gather_mmu(&tlb, mm, 0);
1913 update_hiwater_rss(mm);
1914 unmap_vmas(&tlb, vma, start, end);
1915 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
1916 next ? next->vm_start : 0);
1917 tlb_finish_mmu(&tlb, start, end);
1918}
1919
1920/*
1921 * Create a list of vma's touched by the unmap, removing them from the mm's
1922 * vma list as we go..
1923 */
1924static void
1925detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1926 struct vm_area_struct *prev, unsigned long end)
1927{
1928 struct vm_area_struct **insertion_point;
1929 struct vm_area_struct *tail_vma = NULL;
1930 unsigned long addr;
1931
1932 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1933 vma->vm_prev = NULL;
1934 do {
1935 rb_erase(&vma->vm_rb, &mm->mm_rb);
1936 mm->map_count--;
1937 tail_vma = vma;
1938 vma = vma->vm_next;
1939 } while (vma && vma->vm_start < end);
1940 *insertion_point = vma;
1941 if (vma)
1942 vma->vm_prev = prev;
1943 tail_vma->vm_next = NULL;
1944 if (mm->unmap_area == arch_unmap_area)
1945 addr = prev ? prev->vm_end : mm->mmap_base;
1946 else
1947 addr = vma ? vma->vm_start : mm->mmap_base;
1948 mm->unmap_area(mm, addr);
1949 mm->mmap_cache = NULL; /* Kill the cache. */
1950}
1951
1952/*
1953 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
1954 * munmap path where it doesn't make sense to fail.
1955 */
1956static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1957 unsigned long addr, int new_below)
1958{
1959 struct mempolicy *pol;
1960 struct vm_area_struct *new;
1961 int err = -ENOMEM;
1962
1963 if (is_vm_hugetlb_page(vma) && (addr &
1964 ~(huge_page_mask(hstate_vma(vma)))))
1965 return -EINVAL;
1966
1967 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1968 if (!new)
1969 goto out_err;
1970
1971 /* most fields are the same, copy all, and then fixup */
1972 *new = *vma;
1973
1974 INIT_LIST_HEAD(&new->anon_vma_chain);
1975
1976 if (new_below)
1977 new->vm_end = addr;
1978 else {
1979 new->vm_start = addr;
1980 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1981 }
1982
1983 pol = mpol_dup(vma_policy(vma));
1984 if (IS_ERR(pol)) {
1985 err = PTR_ERR(pol);
1986 goto out_free_vma;
1987 }
1988 vma_set_policy(new, pol);
1989
1990 if (anon_vma_clone(new, vma))
1991 goto out_free_mpol;
1992
1993 if (new->vm_file) {
1994 get_file(new->vm_file);
1995 if (vma->vm_flags & VM_EXECUTABLE)
1996 added_exe_file_vma(mm);
1997 }
1998
1999 if (new->vm_ops && new->vm_ops->open)
2000 new->vm_ops->open(new);
2001
2002 if (new_below)
2003 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2004 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2005 else
2006 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2007
2008 /* Success. */
2009 if (!err)
2010 return 0;
2011
2012 /* Clean everything up if vma_adjust failed. */
2013 if (new->vm_ops && new->vm_ops->close)
2014 new->vm_ops->close(new);
2015 if (new->vm_file) {
2016 if (vma->vm_flags & VM_EXECUTABLE)
2017 removed_exe_file_vma(mm);
2018 fput(new->vm_file);
2019 }
2020 unlink_anon_vmas(new);
2021 out_free_mpol:
2022 mpol_put(pol);
2023 out_free_vma:
2024 kmem_cache_free(vm_area_cachep, new);
2025 out_err:
2026 return err;
2027}
2028
2029/*
2030 * Split a vma into two pieces at address 'addr', a new vma is allocated
2031 * either for the first part or the tail.
2032 */
2033int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2034 unsigned long addr, int new_below)
2035{
2036 if (mm->map_count >= sysctl_max_map_count)
2037 return -ENOMEM;
2038
2039 return __split_vma(mm, vma, addr, new_below);
2040}
2041
2042/* Munmap is split into 2 main parts -- this part which finds
2043 * what needs doing, and the areas themselves, which do the
2044 * work. This now handles partial unmappings.
2045 * Jeremy Fitzhardinge <jeremy@goop.org>
2046 */
2047int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2048{
2049 unsigned long end;
2050 struct vm_area_struct *vma, *prev, *last;
2051
2052 if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2053 return -EINVAL;
2054
2055 if ((len = PAGE_ALIGN(len)) == 0)
2056 return -EINVAL;
2057
2058 /* Find the first overlapping VMA */
2059 vma = find_vma(mm, start);
2060 if (!vma)
2061 return 0;
2062 prev = vma->vm_prev;
2063 /* we have start < vma->vm_end */
2064
2065 /* if it doesn't overlap, we have nothing.. */
2066 end = start + len;
2067 if (vma->vm_start >= end)
2068 return 0;
2069
2070 /*
2071 * If we need to split any vma, do it now to save pain later.
2072 *
2073 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2074 * unmapped vm_area_struct will remain in use: so lower split_vma
2075 * places tmp vma above, and higher split_vma places tmp vma below.
2076 */
2077 if (start > vma->vm_start) {
2078 int error;
2079
2080 /*
2081 * Make sure that map_count on return from munmap() will
2082 * not exceed its limit; but let map_count go just above
2083 * its limit temporarily, to help free resources as expected.
2084 */
2085 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2086 return -ENOMEM;
2087
2088 error = __split_vma(mm, vma, start, 0);
2089 if (error)
2090 return error;
2091 prev = vma;
2092 }
2093
2094 /* Does it split the last one? */
2095 last = find_vma(mm, end);
2096 if (last && end > last->vm_start) {
2097 int error = __split_vma(mm, last, end, 1);
2098 if (error)
2099 return error;
2100 }
2101 vma = prev? prev->vm_next: mm->mmap;
2102
2103 /*
2104 * unlock any mlock()ed ranges before detaching vmas
2105 */
2106 if (mm->locked_vm) {
2107 struct vm_area_struct *tmp = vma;
2108 while (tmp && tmp->vm_start < end) {
2109 if (tmp->vm_flags & VM_LOCKED) {
2110 mm->locked_vm -= vma_pages(tmp);
2111 munlock_vma_pages_all(tmp);
2112 }
2113 tmp = tmp->vm_next;
2114 }
2115 }
2116
2117 /*
2118 * Remove the vma's, and unmap the actual pages
2119 */
2120 detach_vmas_to_be_unmapped(mm, vma, prev, end);
2121 unmap_region(mm, vma, prev, start, end);
2122
2123 /* Fix up all other VM information */
2124 remove_vma_list(mm, vma);
2125
2126 return 0;
2127}
2128
2129int vm_munmap(unsigned long start, size_t len)
2130{
2131 int ret;
2132 struct mm_struct *mm = current->mm;
2133
2134 down_write(&mm->mmap_sem);
2135 ret = do_munmap(mm, start, len);
2136 up_write(&mm->mmap_sem);
2137 return ret;
2138}
2139EXPORT_SYMBOL(vm_munmap);
2140
2141SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2142{
2143 profile_munmap(addr);
2144 return vm_munmap(addr, len);
2145}
2146
2147static inline void verify_mm_writelocked(struct mm_struct *mm)
2148{
2149#ifdef CONFIG_DEBUG_VM
2150 if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2151 WARN_ON(1);
2152 up_read(&mm->mmap_sem);
2153 }
2154#endif
2155}
2156
2157/*
2158 * this is really a simplified "do_mmap". it only handles
2159 * anonymous maps. eventually we may be able to do some
2160 * brk-specific accounting here.
2161 */
2162static unsigned long do_brk(unsigned long addr, unsigned long len)
2163{
2164 struct mm_struct * mm = current->mm;
2165 struct vm_area_struct * vma, * prev;
2166 unsigned long flags;
2167 struct rb_node ** rb_link, * rb_parent;
2168 pgoff_t pgoff = addr >> PAGE_SHIFT;
2169 int error;
2170
2171 len = PAGE_ALIGN(len);
2172 if (!len)
2173 return addr;
2174
2175 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2176
2177 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2178 if (error & ~PAGE_MASK)
2179 return error;
2180
2181 /*
2182 * mlock MCL_FUTURE?
2183 */
2184 if (mm->def_flags & VM_LOCKED) {
2185 unsigned long locked, lock_limit;
2186 locked = len >> PAGE_SHIFT;
2187 locked += mm->locked_vm;
2188 lock_limit = rlimit(RLIMIT_MEMLOCK);
2189 lock_limit >>= PAGE_SHIFT;
2190 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2191 return -EAGAIN;
2192 }
2193
2194 /*
2195 * mm->mmap_sem is required to protect against another thread
2196 * changing the mappings in case we sleep.
2197 */
2198 verify_mm_writelocked(mm);
2199
2200 /*
2201 * Clear old maps. this also does some error checking for us
2202 */
2203 munmap_back:
2204 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2205 if (vma && vma->vm_start < addr + len) {
2206 if (do_munmap(mm, addr, len))
2207 return -ENOMEM;
2208 goto munmap_back;
2209 }
2210
2211 /* Check against address space limits *after* clearing old maps... */
2212 if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2213 return -ENOMEM;
2214
2215 if (mm->map_count > sysctl_max_map_count)
2216 return -ENOMEM;
2217
2218 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2219 return -ENOMEM;
2220
2221 /* Can we just expand an old private anonymous mapping? */
2222 vma = vma_merge(mm, prev, addr, addr + len, flags,
2223 NULL, NULL, pgoff, NULL);
2224 if (vma)
2225 goto out;
2226
2227 /*
2228 * create a vma struct for an anonymous mapping
2229 */
2230 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2231 if (!vma) {
2232 vm_unacct_memory(len >> PAGE_SHIFT);
2233 return -ENOMEM;
2234 }
2235
2236 INIT_LIST_HEAD(&vma->anon_vma_chain);
2237 vma->vm_mm = mm;
2238 vma->vm_start = addr;
2239 vma->vm_end = addr + len;
2240 vma->vm_pgoff = pgoff;
2241 vma->vm_flags = flags;
2242 vma->vm_page_prot = vm_get_page_prot(flags);
2243 vma_link(mm, vma, prev, rb_link, rb_parent);
2244out:
2245 perf_event_mmap(vma);
2246 mm->total_vm += len >> PAGE_SHIFT;
2247 if (flags & VM_LOCKED) {
2248 if (!mlock_vma_pages_range(vma, addr, addr + len))
2249 mm->locked_vm += (len >> PAGE_SHIFT);
2250 }
2251 return addr;
2252}
2253
2254unsigned long vm_brk(unsigned long addr, unsigned long len)
2255{
2256 struct mm_struct *mm = current->mm;
2257 unsigned long ret;
2258
2259 down_write(&mm->mmap_sem);
2260 ret = do_brk(addr, len);
2261 up_write(&mm->mmap_sem);
2262 return ret;
2263}
2264EXPORT_SYMBOL(vm_brk);
2265
2266/* Release all mmaps. */
2267void exit_mmap(struct mm_struct *mm)
2268{
2269 struct mmu_gather tlb;
2270 struct vm_area_struct *vma;
2271 unsigned long nr_accounted = 0;
2272
2273 /* mm's last user has gone, and its about to be pulled down */
2274 mmu_notifier_release(mm);
2275
2276 if (mm->locked_vm) {
2277 vma = mm->mmap;
2278 while (vma) {
2279 if (vma->vm_flags & VM_LOCKED)
2280 munlock_vma_pages_all(vma);
2281 vma = vma->vm_next;
2282 }
2283 }
2284
2285 arch_exit_mmap(mm);
2286
2287 vma = mm->mmap;
2288 if (!vma) /* Can happen if dup_mmap() received an OOM */
2289 return;
2290
2291 lru_add_drain();
2292 flush_cache_mm(mm);
2293 tlb_gather_mmu(&tlb, mm, 1);
2294 /* update_hiwater_rss(mm) here? but nobody should be looking */
2295 /* Use -1 here to ensure all VMAs in the mm are unmapped */
2296 unmap_vmas(&tlb, vma, 0, -1);
2297
2298 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
2299 tlb_finish_mmu(&tlb, 0, -1);
2300
2301 /*
2302 * Walk the list again, actually closing and freeing it,
2303 * with preemption enabled, without holding any MM locks.
2304 */
2305 while (vma) {
2306 if (vma->vm_flags & VM_ACCOUNT)
2307 nr_accounted += vma_pages(vma);
2308 vma = remove_vma(vma);
2309 }
2310 vm_unacct_memory(nr_accounted);
2311
2312 BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2313}
2314
2315/* Insert vm structure into process list sorted by address
2316 * and into the inode's i_mmap tree. If vm_file is non-NULL
2317 * then i_mmap_mutex is taken here.
2318 */
2319int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2320{
2321 struct vm_area_struct * __vma, * prev;
2322 struct rb_node ** rb_link, * rb_parent;
2323
2324 /*
2325 * The vm_pgoff of a purely anonymous vma should be irrelevant
2326 * until its first write fault, when page's anon_vma and index
2327 * are set. But now set the vm_pgoff it will almost certainly
2328 * end up with (unless mremap moves it elsewhere before that
2329 * first wfault), so /proc/pid/maps tells a consistent story.
2330 *
2331 * By setting it to reflect the virtual start address of the
2332 * vma, merges and splits can happen in a seamless way, just
2333 * using the existing file pgoff checks and manipulations.
2334 * Similarly in do_mmap_pgoff and in do_brk.
2335 */
2336 if (!vma->vm_file) {
2337 BUG_ON(vma->anon_vma);
2338 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2339 }
2340 __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2341 if (__vma && __vma->vm_start < vma->vm_end)
2342 return -ENOMEM;
2343 if ((vma->vm_flags & VM_ACCOUNT) &&
2344 security_vm_enough_memory_mm(mm, vma_pages(vma)))
2345 return -ENOMEM;
2346
2347 if (vma->vm_file && uprobe_mmap(vma))
2348 return -EINVAL;
2349
2350 vma_link(mm, vma, prev, rb_link, rb_parent);
2351 return 0;
2352}
2353
2354/*
2355 * Copy the vma structure to a new location in the same mm,
2356 * prior to moving page table entries, to effect an mremap move.
2357 */
2358struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2359 unsigned long addr, unsigned long len, pgoff_t pgoff)
2360{
2361 struct vm_area_struct *vma = *vmap;
2362 unsigned long vma_start = vma->vm_start;
2363 struct mm_struct *mm = vma->vm_mm;
2364 struct vm_area_struct *new_vma, *prev;
2365 struct rb_node **rb_link, *rb_parent;
2366 struct mempolicy *pol;
2367 bool faulted_in_anon_vma = true;
2368
2369 /*
2370 * If anonymous vma has not yet been faulted, update new pgoff
2371 * to match new location, to increase its chance of merging.
2372 */
2373 if (unlikely(!vma->vm_file && !vma->anon_vma)) {
2374 pgoff = addr >> PAGE_SHIFT;
2375 faulted_in_anon_vma = false;
2376 }
2377
2378 find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2379 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2380 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2381 if (new_vma) {
2382 /*
2383 * Source vma may have been merged into new_vma
2384 */
2385 if (unlikely(vma_start >= new_vma->vm_start &&
2386 vma_start < new_vma->vm_end)) {
2387 /*
2388 * The only way we can get a vma_merge with
2389 * self during an mremap is if the vma hasn't
2390 * been faulted in yet and we were allowed to
2391 * reset the dst vma->vm_pgoff to the
2392 * destination address of the mremap to allow
2393 * the merge to happen. mremap must change the
2394 * vm_pgoff linearity between src and dst vmas
2395 * (in turn preventing a vma_merge) to be
2396 * safe. It is only safe to keep the vm_pgoff
2397 * linear if there are no pages mapped yet.
2398 */
2399 VM_BUG_ON(faulted_in_anon_vma);
2400 *vmap = new_vma;
2401 } else
2402 anon_vma_moveto_tail(new_vma);
2403 } else {
2404 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2405 if (new_vma) {
2406 *new_vma = *vma;
2407 pol = mpol_dup(vma_policy(vma));
2408 if (IS_ERR(pol))
2409 goto out_free_vma;
2410 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2411 if (anon_vma_clone(new_vma, vma))
2412 goto out_free_mempol;
2413 vma_set_policy(new_vma, pol);
2414 new_vma->vm_start = addr;
2415 new_vma->vm_end = addr + len;
2416 new_vma->vm_pgoff = pgoff;
2417 if (new_vma->vm_file) {
2418 get_file(new_vma->vm_file);
2419
2420 if (uprobe_mmap(new_vma))
2421 goto out_free_mempol;
2422
2423 if (vma->vm_flags & VM_EXECUTABLE)
2424 added_exe_file_vma(mm);
2425 }
2426 if (new_vma->vm_ops && new_vma->vm_ops->open)
2427 new_vma->vm_ops->open(new_vma);
2428 vma_link(mm, new_vma, prev, rb_link, rb_parent);
2429 }
2430 }
2431 return new_vma;
2432
2433 out_free_mempol:
2434 mpol_put(pol);
2435 out_free_vma:
2436 kmem_cache_free(vm_area_cachep, new_vma);
2437 return NULL;
2438}
2439
2440/*
2441 * Return true if the calling process may expand its vm space by the passed
2442 * number of pages
2443 */
2444int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2445{
2446 unsigned long cur = mm->total_vm; /* pages */
2447 unsigned long lim;
2448
2449 lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2450
2451 if (cur + npages > lim)
2452 return 0;
2453 return 1;
2454}
2455
2456
2457static int special_mapping_fault(struct vm_area_struct *vma,
2458 struct vm_fault *vmf)
2459{
2460 pgoff_t pgoff;
2461 struct page **pages;
2462
2463 /*
2464 * special mappings have no vm_file, and in that case, the mm
2465 * uses vm_pgoff internally. So we have to subtract it from here.
2466 * We are allowed to do this because we are the mm; do not copy
2467 * this code into drivers!
2468 */
2469 pgoff = vmf->pgoff - vma->vm_pgoff;
2470
2471 for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2472 pgoff--;
2473
2474 if (*pages) {
2475 struct page *page = *pages;
2476 get_page(page);
2477 vmf->page = page;
2478 return 0;
2479 }
2480
2481 return VM_FAULT_SIGBUS;
2482}
2483
2484/*
2485 * Having a close hook prevents vma merging regardless of flags.
2486 */
2487static void special_mapping_close(struct vm_area_struct *vma)
2488{
2489}
2490
2491static const struct vm_operations_struct special_mapping_vmops = {
2492 .close = special_mapping_close,
2493 .fault = special_mapping_fault,
2494};
2495
2496/*
2497 * Called with mm->mmap_sem held for writing.
2498 * Insert a new vma covering the given region, with the given flags.
2499 * Its pages are supplied by the given array of struct page *.
2500 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2501 * The region past the last page supplied will always produce SIGBUS.
2502 * The array pointer and the pages it points to are assumed to stay alive
2503 * for as long as this mapping might exist.
2504 */
2505int install_special_mapping(struct mm_struct *mm,
2506 unsigned long addr, unsigned long len,
2507 unsigned long vm_flags, struct page **pages)
2508{
2509 int ret;
2510 struct vm_area_struct *vma;
2511
2512 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2513 if (unlikely(vma == NULL))
2514 return -ENOMEM;
2515
2516 INIT_LIST_HEAD(&vma->anon_vma_chain);
2517 vma->vm_mm = mm;
2518 vma->vm_start = addr;
2519 vma->vm_end = addr + len;
2520
2521 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2522 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2523
2524 vma->vm_ops = &special_mapping_vmops;
2525 vma->vm_private_data = pages;
2526
2527 ret = insert_vm_struct(mm, vma);
2528 if (ret)
2529 goto out;
2530
2531 mm->total_vm += len >> PAGE_SHIFT;
2532
2533 perf_event_mmap(vma);
2534
2535 return 0;
2536
2537out:
2538 kmem_cache_free(vm_area_cachep, vma);
2539 return ret;
2540}
2541
2542static DEFINE_MUTEX(mm_all_locks_mutex);
2543
2544static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2545{
2546 if (!test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2547 /*
2548 * The LSB of head.next can't change from under us
2549 * because we hold the mm_all_locks_mutex.
2550 */
2551 mutex_lock_nest_lock(&anon_vma->root->mutex, &mm->mmap_sem);
2552 /*
2553 * We can safely modify head.next after taking the
2554 * anon_vma->root->mutex. If some other vma in this mm shares
2555 * the same anon_vma we won't take it again.
2556 *
2557 * No need of atomic instructions here, head.next
2558 * can't change from under us thanks to the
2559 * anon_vma->root->mutex.
2560 */
2561 if (__test_and_set_bit(0, (unsigned long *)
2562 &anon_vma->root->head.next))
2563 BUG();
2564 }
2565}
2566
2567static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2568{
2569 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2570 /*
2571 * AS_MM_ALL_LOCKS can't change from under us because
2572 * we hold the mm_all_locks_mutex.
2573 *
2574 * Operations on ->flags have to be atomic because
2575 * even if AS_MM_ALL_LOCKS is stable thanks to the
2576 * mm_all_locks_mutex, there may be other cpus
2577 * changing other bitflags in parallel to us.
2578 */
2579 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2580 BUG();
2581 mutex_lock_nest_lock(&mapping->i_mmap_mutex, &mm->mmap_sem);
2582 }
2583}
2584
2585/*
2586 * This operation locks against the VM for all pte/vma/mm related
2587 * operations that could ever happen on a certain mm. This includes
2588 * vmtruncate, try_to_unmap, and all page faults.
2589 *
2590 * The caller must take the mmap_sem in write mode before calling
2591 * mm_take_all_locks(). The caller isn't allowed to release the
2592 * mmap_sem until mm_drop_all_locks() returns.
2593 *
2594 * mmap_sem in write mode is required in order to block all operations
2595 * that could modify pagetables and free pages without need of
2596 * altering the vma layout (for example populate_range() with
2597 * nonlinear vmas). It's also needed in write mode to avoid new
2598 * anon_vmas to be associated with existing vmas.
2599 *
2600 * A single task can't take more than one mm_take_all_locks() in a row
2601 * or it would deadlock.
2602 *
2603 * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2604 * mapping->flags avoid to take the same lock twice, if more than one
2605 * vma in this mm is backed by the same anon_vma or address_space.
2606 *
2607 * We can take all the locks in random order because the VM code
2608 * taking i_mmap_mutex or anon_vma->mutex outside the mmap_sem never
2609 * takes more than one of them in a row. Secondly we're protected
2610 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2611 *
2612 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2613 * that may have to take thousand of locks.
2614 *
2615 * mm_take_all_locks() can fail if it's interrupted by signals.
2616 */
2617int mm_take_all_locks(struct mm_struct *mm)
2618{
2619 struct vm_area_struct *vma;
2620 struct anon_vma_chain *avc;
2621
2622 BUG_ON(down_read_trylock(&mm->mmap_sem));
2623
2624 mutex_lock(&mm_all_locks_mutex);
2625
2626 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2627 if (signal_pending(current))
2628 goto out_unlock;
2629 if (vma->vm_file && vma->vm_file->f_mapping)
2630 vm_lock_mapping(mm, vma->vm_file->f_mapping);
2631 }
2632
2633 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2634 if (signal_pending(current))
2635 goto out_unlock;
2636 if (vma->anon_vma)
2637 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2638 vm_lock_anon_vma(mm, avc->anon_vma);
2639 }
2640
2641 return 0;
2642
2643out_unlock:
2644 mm_drop_all_locks(mm);
2645 return -EINTR;
2646}
2647
2648static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2649{
2650 if (test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2651 /*
2652 * The LSB of head.next can't change to 0 from under
2653 * us because we hold the mm_all_locks_mutex.
2654 *
2655 * We must however clear the bitflag before unlocking
2656 * the vma so the users using the anon_vma->head will
2657 * never see our bitflag.
2658 *
2659 * No need of atomic instructions here, head.next
2660 * can't change from under us until we release the
2661 * anon_vma->root->mutex.
2662 */
2663 if (!__test_and_clear_bit(0, (unsigned long *)
2664 &anon_vma->root->head.next))
2665 BUG();
2666 anon_vma_unlock(anon_vma);
2667 }
2668}
2669
2670static void vm_unlock_mapping(struct address_space *mapping)
2671{
2672 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2673 /*
2674 * AS_MM_ALL_LOCKS can't change to 0 from under us
2675 * because we hold the mm_all_locks_mutex.
2676 */
2677 mutex_unlock(&mapping->i_mmap_mutex);
2678 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2679 &mapping->flags))
2680 BUG();
2681 }
2682}
2683
2684/*
2685 * The mmap_sem cannot be released by the caller until
2686 * mm_drop_all_locks() returns.
2687 */
2688void mm_drop_all_locks(struct mm_struct *mm)
2689{
2690 struct vm_area_struct *vma;
2691 struct anon_vma_chain *avc;
2692
2693 BUG_ON(down_read_trylock(&mm->mmap_sem));
2694 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2695
2696 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2697 if (vma->anon_vma)
2698 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2699 vm_unlock_anon_vma(avc->anon_vma);
2700 if (vma->vm_file && vma->vm_file->f_mapping)
2701 vm_unlock_mapping(vma->vm_file->f_mapping);
2702 }
2703
2704 mutex_unlock(&mm_all_locks_mutex);
2705}
2706
2707/*
2708 * initialise the VMA slab
2709 */
2710void __init mmap_init(void)
2711{
2712 int ret;
2713
2714 ret = percpu_counter_init(&vm_committed_as, 0);
2715 VM_BUG_ON(ret);
2716}