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