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