summaryrefslogtreecommitdiff
path: root/mm/mmap.c (plain)
blob: 6441ef827c2d5bc9df027fe41de1d4131aa90646
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
60const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
61const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
62int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
63#endif
64#ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
65const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
66const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
67int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
68#endif
69
70static bool ignore_rlimit_data;
71core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
72
73static 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 * On arm64, PROT_EXEC has the following behaviour for both MAP_SHARED and
92 * MAP_PRIVATE:
93 * r: (no) no
94 * w: (no) no
95 * x: (yes) yes
96 */
97pgprot_t protection_map[16] = {
98 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
99 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
100};
101
102pgprot_t vm_get_page_prot(unsigned long vm_flags)
103{
104 return __pgprot(pgprot_val(protection_map[vm_flags &
105 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
106 pgprot_val(arch_vm_get_page_prot(vm_flags)));
107}
108EXPORT_SYMBOL(vm_get_page_prot);
109
110static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
111{
112 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
113}
114
115/* Update vma->vm_page_prot to reflect vma->vm_flags. */
116void vma_set_page_prot(struct vm_area_struct *vma)
117{
118 unsigned long vm_flags = vma->vm_flags;
119 pgprot_t vm_page_prot;
120
121 vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
122 if (vma_wants_writenotify(vma, vm_page_prot)) {
123 vm_flags &= ~VM_SHARED;
124 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
125 }
126 /* remove_protection_ptes reads vma->vm_page_prot without mmap_sem */
127 WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
128}
129
130/*
131 * Requires inode->i_mapping->i_mmap_rwsem
132 */
133static void __remove_shared_vm_struct(struct vm_area_struct *vma,
134 struct file *file, struct address_space *mapping)
135{
136 if (vma->vm_flags & VM_DENYWRITE)
137 atomic_inc(&file_inode(file)->i_writecount);
138 if (vma->vm_flags & VM_SHARED)
139 mapping_unmap_writable(mapping);
140
141 flush_dcache_mmap_lock(mapping);
142 vma_interval_tree_remove(vma, &mapping->i_mmap);
143 flush_dcache_mmap_unlock(mapping);
144}
145
146/*
147 * Unlink a file-based vm structure from its interval tree, to hide
148 * vma from rmap and vmtruncate before freeing its page tables.
149 */
150void unlink_file_vma(struct vm_area_struct *vma)
151{
152 struct file *file = vma->vm_file;
153
154 if (file) {
155 struct address_space *mapping = file->f_mapping;
156 i_mmap_lock_write(mapping);
157 __remove_shared_vm_struct(vma, file, mapping);
158 i_mmap_unlock_write(mapping);
159 }
160}
161
162/*
163 * Close a vm structure and free it, returning the next.
164 */
165static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
166{
167 struct vm_area_struct *next = vma->vm_next;
168
169 might_sleep();
170 if (vma->vm_ops && vma->vm_ops->close)
171 vma->vm_ops->close(vma);
172 if (vma->vm_file)
173 fput(vma->vm_file);
174 mpol_put(vma_policy(vma));
175 kmem_cache_free(vm_area_cachep, vma);
176 return next;
177}
178
179static int do_brk(unsigned long addr, unsigned long len);
180
181SYSCALL_DEFINE1(brk, unsigned long, brk)
182{
183 unsigned long retval;
184 unsigned long newbrk, oldbrk;
185 struct mm_struct *mm = current->mm;
186 struct vm_area_struct *next;
187 unsigned long min_brk;
188 bool populate;
189
190 if (down_write_killable(&mm->mmap_sem))
191 return -EINTR;
192
193#ifdef CONFIG_COMPAT_BRK
194 /*
195 * CONFIG_COMPAT_BRK can still be overridden by setting
196 * randomize_va_space to 2, which will still cause mm->start_brk
197 * to be arbitrarily shifted
198 */
199 if (current->brk_randomized)
200 min_brk = mm->start_brk;
201 else
202 min_brk = mm->end_data;
203#else
204 min_brk = mm->start_brk;
205#endif
206 if (brk < min_brk)
207 goto out;
208
209 /*
210 * Check against rlimit here. If this check is done later after the test
211 * of oldbrk with newbrk then it can escape the test and let the data
212 * segment grow beyond its set limit the in case where the limit is
213 * not page aligned -Ram Gupta
214 */
215 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
216 mm->end_data, mm->start_data))
217 goto out;
218
219 newbrk = PAGE_ALIGN(brk);
220 oldbrk = PAGE_ALIGN(mm->brk);
221 if (oldbrk == newbrk)
222 goto set_brk;
223
224 /* Always allow shrinking brk. */
225 if (brk <= mm->brk) {
226 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
227 goto set_brk;
228 goto out;
229 }
230
231 /* Check against existing mmap mappings. */
232 next = find_vma(mm, oldbrk);
233 if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
234 goto out;
235
236 /* Ok, looks good - let it rip. */
237 if (do_brk(oldbrk, newbrk-oldbrk) < 0)
238 goto out;
239
240set_brk:
241 mm->brk = brk;
242 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
243 up_write(&mm->mmap_sem);
244 if (populate)
245 mm_populate(oldbrk, newbrk - oldbrk);
246 return brk;
247
248out:
249 retval = mm->brk;
250 up_write(&mm->mmap_sem);
251 return retval;
252}
253
254static long vma_compute_subtree_gap(struct vm_area_struct *vma)
255{
256 unsigned long max, prev_end, subtree_gap;
257
258 /*
259 * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
260 * allow two stack_guard_gaps between them here, and when choosing
261 * an unmapped area; whereas when expanding we only require one.
262 * That's a little inconsistent, but keeps the code here simpler.
263 */
264 max = vm_start_gap(vma);
265 if (vma->vm_prev) {
266 prev_end = vm_end_gap(vma->vm_prev);
267 if (max > prev_end)
268 max -= prev_end;
269 else
270 max = 0;
271 }
272 if (vma->vm_rb.rb_left) {
273 subtree_gap = rb_entry(vma->vm_rb.rb_left,
274 struct vm_area_struct, vm_rb)->rb_subtree_gap;
275 if (subtree_gap > max)
276 max = subtree_gap;
277 }
278 if (vma->vm_rb.rb_right) {
279 subtree_gap = rb_entry(vma->vm_rb.rb_right,
280 struct vm_area_struct, vm_rb)->rb_subtree_gap;
281 if (subtree_gap > max)
282 max = subtree_gap;
283 }
284 return max;
285}
286
287#ifdef CONFIG_DEBUG_VM_RB
288static int browse_rb(struct mm_struct *mm)
289{
290 struct rb_root *root = &mm->mm_rb;
291 int i = 0, j, bug = 0;
292 struct rb_node *nd, *pn = NULL;
293 unsigned long prev = 0, pend = 0;
294
295 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
296 struct vm_area_struct *vma;
297 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
298 if (vma->vm_start < prev) {
299 pr_emerg("vm_start %lx < prev %lx\n",
300 vma->vm_start, prev);
301 bug = 1;
302 }
303 if (vma->vm_start < pend) {
304 pr_emerg("vm_start %lx < pend %lx\n",
305 vma->vm_start, pend);
306 bug = 1;
307 }
308 if (vma->vm_start > vma->vm_end) {
309 pr_emerg("vm_start %lx > vm_end %lx\n",
310 vma->vm_start, vma->vm_end);
311 bug = 1;
312 }
313 spin_lock(&mm->page_table_lock);
314 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
315 pr_emerg("free gap %lx, correct %lx\n",
316 vma->rb_subtree_gap,
317 vma_compute_subtree_gap(vma));
318 bug = 1;
319 }
320 spin_unlock(&mm->page_table_lock);
321 i++;
322 pn = nd;
323 prev = vma->vm_start;
324 pend = vma->vm_end;
325 }
326 j = 0;
327 for (nd = pn; nd; nd = rb_prev(nd))
328 j++;
329 if (i != j) {
330 pr_emerg("backwards %d, forwards %d\n", j, i);
331 bug = 1;
332 }
333 return bug ? -1 : i;
334}
335
336static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
337{
338 struct rb_node *nd;
339
340 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
341 struct vm_area_struct *vma;
342 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
343 VM_BUG_ON_VMA(vma != ignore &&
344 vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
345 vma);
346 }
347}
348
349static void validate_mm(struct mm_struct *mm)
350{
351 int bug = 0;
352 int i = 0;
353 unsigned long highest_address = 0;
354 struct vm_area_struct *vma = mm->mmap;
355
356 while (vma) {
357 struct anon_vma *anon_vma = vma->anon_vma;
358 struct anon_vma_chain *avc;
359
360 if (anon_vma) {
361 anon_vma_lock_read(anon_vma);
362 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
363 anon_vma_interval_tree_verify(avc);
364 anon_vma_unlock_read(anon_vma);
365 }
366
367 highest_address = vm_end_gap(vma);
368 vma = vma->vm_next;
369 i++;
370 }
371 if (i != mm->map_count) {
372 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
373 bug = 1;
374 }
375 if (highest_address != mm->highest_vm_end) {
376 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
377 mm->highest_vm_end, highest_address);
378 bug = 1;
379 }
380 i = browse_rb(mm);
381 if (i != mm->map_count) {
382 if (i != -1)
383 pr_emerg("map_count %d rb %d\n", mm->map_count, i);
384 bug = 1;
385 }
386 VM_BUG_ON_MM(bug, mm);
387}
388#else
389#define validate_mm_rb(root, ignore) do { } while (0)
390#define validate_mm(mm) do { } while (0)
391#endif
392
393RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
394 unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
395
396/*
397 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
398 * vma->vm_prev->vm_end values changed, without modifying the vma's position
399 * in the rbtree.
400 */
401static void vma_gap_update(struct vm_area_struct *vma)
402{
403 /*
404 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
405 * function that does exacltly what we want.
406 */
407 vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
408}
409
410static inline void vma_rb_insert(struct vm_area_struct *vma,
411 struct rb_root *root)
412{
413 /* All rb_subtree_gap values must be consistent prior to insertion */
414 validate_mm_rb(root, NULL);
415
416 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
417}
418
419static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
420{
421 /*
422 * Note rb_erase_augmented is a fairly large inline function,
423 * so make sure we instantiate it only once with our desired
424 * augmented rbtree callbacks.
425 */
426 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
427}
428
429static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
430 struct rb_root *root,
431 struct vm_area_struct *ignore)
432{
433 /*
434 * All rb_subtree_gap values must be consistent prior to erase,
435 * with the possible exception of the "next" vma being erased if
436 * next->vm_start was reduced.
437 */
438 validate_mm_rb(root, ignore);
439
440 __vma_rb_erase(vma, root);
441}
442
443static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
444 struct rb_root *root)
445{
446 /*
447 * All rb_subtree_gap values must be consistent prior to erase,
448 * with the possible exception of the vma being erased.
449 */
450 validate_mm_rb(root, vma);
451
452 __vma_rb_erase(vma, root);
453}
454
455/*
456 * vma has some anon_vma assigned, and is already inserted on that
457 * anon_vma's interval trees.
458 *
459 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
460 * vma must be removed from the anon_vma's interval trees using
461 * anon_vma_interval_tree_pre_update_vma().
462 *
463 * After the update, the vma will be reinserted using
464 * anon_vma_interval_tree_post_update_vma().
465 *
466 * The entire update must be protected by exclusive mmap_sem and by
467 * the root anon_vma's mutex.
468 */
469static inline void
470anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
471{
472 struct anon_vma_chain *avc;
473
474 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
475 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
476}
477
478static inline void
479anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
480{
481 struct anon_vma_chain *avc;
482
483 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
484 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
485}
486
487static int find_vma_links(struct mm_struct *mm, unsigned long addr,
488 unsigned long end, struct vm_area_struct **pprev,
489 struct rb_node ***rb_link, struct rb_node **rb_parent)
490{
491 struct rb_node **__rb_link, *__rb_parent, *rb_prev;
492
493 __rb_link = &mm->mm_rb.rb_node;
494 rb_prev = __rb_parent = NULL;
495
496 while (*__rb_link) {
497 struct vm_area_struct *vma_tmp;
498
499 __rb_parent = *__rb_link;
500 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
501
502 if (vma_tmp->vm_end > addr) {
503 /* Fail if an existing vma overlaps the area */
504 if (vma_tmp->vm_start < end)
505 return -ENOMEM;
506 __rb_link = &__rb_parent->rb_left;
507 } else {
508 rb_prev = __rb_parent;
509 __rb_link = &__rb_parent->rb_right;
510 }
511 }
512
513 *pprev = NULL;
514 if (rb_prev)
515 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
516 *rb_link = __rb_link;
517 *rb_parent = __rb_parent;
518 return 0;
519}
520
521static unsigned long count_vma_pages_range(struct mm_struct *mm,
522 unsigned long addr, unsigned long end)
523{
524 unsigned long nr_pages = 0;
525 struct vm_area_struct *vma;
526
527 /* Find first overlaping mapping */
528 vma = find_vma_intersection(mm, addr, end);
529 if (!vma)
530 return 0;
531
532 nr_pages = (min(end, vma->vm_end) -
533 max(addr, vma->vm_start)) >> PAGE_SHIFT;
534
535 /* Iterate over the rest of the overlaps */
536 for (vma = vma->vm_next; vma; vma = vma->vm_next) {
537 unsigned long overlap_len;
538
539 if (vma->vm_start > end)
540 break;
541
542 overlap_len = min(end, vma->vm_end) - vma->vm_start;
543 nr_pages += overlap_len >> PAGE_SHIFT;
544 }
545
546 return nr_pages;
547}
548
549void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
550 struct rb_node **rb_link, struct rb_node *rb_parent)
551{
552 /* Update tracking information for the gap following the new vma. */
553 if (vma->vm_next)
554 vma_gap_update(vma->vm_next);
555 else
556 mm->highest_vm_end = vm_end_gap(vma);
557
558 /*
559 * vma->vm_prev wasn't known when we followed the rbtree to find the
560 * correct insertion point for that vma. As a result, we could not
561 * update the vma vm_rb parents rb_subtree_gap values on the way down.
562 * So, we first insert the vma with a zero rb_subtree_gap value
563 * (to be consistent with what we did on the way down), and then
564 * immediately update the gap to the correct value. Finally we
565 * rebalance the rbtree after all augmented values have been set.
566 */
567 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
568 vma->rb_subtree_gap = 0;
569 vma_gap_update(vma);
570 vma_rb_insert(vma, &mm->mm_rb);
571}
572
573static void __vma_link_file(struct vm_area_struct *vma)
574{
575 struct file *file;
576
577 file = vma->vm_file;
578 if (file) {
579 struct address_space *mapping = file->f_mapping;
580
581 if (vma->vm_flags & VM_DENYWRITE)
582 atomic_dec(&file_inode(file)->i_writecount);
583 if (vma->vm_flags & VM_SHARED)
584 atomic_inc(&mapping->i_mmap_writable);
585
586 flush_dcache_mmap_lock(mapping);
587 vma_interval_tree_insert(vma, &mapping->i_mmap);
588 flush_dcache_mmap_unlock(mapping);
589 }
590}
591
592static void
593__vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
594 struct vm_area_struct *prev, struct rb_node **rb_link,
595 struct rb_node *rb_parent)
596{
597 __vma_link_list(mm, vma, prev, rb_parent);
598 __vma_link_rb(mm, vma, rb_link, rb_parent);
599}
600
601static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
602 struct vm_area_struct *prev, struct rb_node **rb_link,
603 struct rb_node *rb_parent)
604{
605 struct address_space *mapping = NULL;
606
607 if (vma->vm_file) {
608 mapping = vma->vm_file->f_mapping;
609 i_mmap_lock_write(mapping);
610 }
611
612 __vma_link(mm, vma, prev, rb_link, rb_parent);
613 __vma_link_file(vma);
614
615 if (mapping)
616 i_mmap_unlock_write(mapping);
617
618 mm->map_count++;
619 validate_mm(mm);
620}
621
622/*
623 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
624 * mm's list and rbtree. It has already been inserted into the interval tree.
625 */
626static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
627{
628 struct vm_area_struct *prev;
629 struct rb_node **rb_link, *rb_parent;
630
631 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
632 &prev, &rb_link, &rb_parent))
633 BUG();
634 __vma_link(mm, vma, prev, rb_link, rb_parent);
635 mm->map_count++;
636}
637
638static __always_inline void __vma_unlink_common(struct mm_struct *mm,
639 struct vm_area_struct *vma,
640 struct vm_area_struct *prev,
641 bool has_prev,
642 struct vm_area_struct *ignore)
643{
644 struct vm_area_struct *next;
645
646 vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
647 next = vma->vm_next;
648 if (has_prev)
649 prev->vm_next = next;
650 else {
651 prev = vma->vm_prev;
652 if (prev)
653 prev->vm_next = next;
654 else
655 mm->mmap = next;
656 }
657 if (next)
658 next->vm_prev = prev;
659
660 /* Kill the cache */
661 vmacache_invalidate(mm);
662}
663
664static inline void __vma_unlink_prev(struct mm_struct *mm,
665 struct vm_area_struct *vma,
666 struct vm_area_struct *prev)
667{
668 __vma_unlink_common(mm, vma, prev, true, vma);
669}
670
671/*
672 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
673 * is already present in an i_mmap tree without adjusting the tree.
674 * The following helper function should be used when such adjustments
675 * are necessary. The "insert" vma (if any) is to be inserted
676 * before we drop the necessary locks.
677 */
678int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
679 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
680 struct vm_area_struct *expand)
681{
682 struct mm_struct *mm = vma->vm_mm;
683 struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
684 struct address_space *mapping = NULL;
685 struct rb_root *root = NULL;
686 struct anon_vma *anon_vma = NULL;
687 struct file *file = vma->vm_file;
688 bool start_changed = false, end_changed = false;
689 long adjust_next = 0;
690 int remove_next = 0;
691
692 if (next && !insert) {
693 struct vm_area_struct *exporter = NULL, *importer = NULL;
694
695 if (end >= next->vm_end) {
696 /*
697 * vma expands, overlapping all the next, and
698 * perhaps the one after too (mprotect case 6).
699 * The only other cases that gets here are
700 * case 1, case 7 and case 8.
701 */
702 if (next == expand) {
703 /*
704 * The only case where we don't expand "vma"
705 * and we expand "next" instead is case 8.
706 */
707 VM_WARN_ON(end != next->vm_end);
708 /*
709 * remove_next == 3 means we're
710 * removing "vma" and that to do so we
711 * swapped "vma" and "next".
712 */
713 remove_next = 3;
714 VM_WARN_ON(file != next->vm_file);
715 swap(vma, next);
716 } else {
717 VM_WARN_ON(expand != vma);
718 /*
719 * case 1, 6, 7, remove_next == 2 is case 6,
720 * remove_next == 1 is case 1 or 7.
721 */
722 remove_next = 1 + (end > next->vm_end);
723 VM_WARN_ON(remove_next == 2 &&
724 end != next->vm_next->vm_end);
725 VM_WARN_ON(remove_next == 1 &&
726 end != next->vm_end);
727 /* trim end to next, for case 6 first pass */
728 end = next->vm_end;
729 }
730
731 exporter = next;
732 importer = vma;
733
734 /*
735 * If next doesn't have anon_vma, import from vma after
736 * next, if the vma overlaps with it.
737 */
738 if (remove_next == 2 && !next->anon_vma)
739 exporter = next->vm_next;
740
741 } else if (end > next->vm_start) {
742 /*
743 * vma expands, overlapping part of the next:
744 * mprotect case 5 shifting the boundary up.
745 */
746 adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
747 exporter = next;
748 importer = vma;
749 VM_WARN_ON(expand != importer);
750 } else if (end < vma->vm_end) {
751 /*
752 * vma shrinks, and !insert tells it's not
753 * split_vma inserting another: so it must be
754 * mprotect case 4 shifting the boundary down.
755 */
756 adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT);
757 exporter = vma;
758 importer = next;
759 VM_WARN_ON(expand != importer);
760 }
761
762 /*
763 * Easily overlooked: when mprotect shifts the boundary,
764 * make sure the expanding vma has anon_vma set if the
765 * shrinking vma had, to cover any anon pages imported.
766 */
767 if (exporter && exporter->anon_vma && !importer->anon_vma) {
768 int error;
769
770 importer->anon_vma = exporter->anon_vma;
771 error = anon_vma_clone(importer, exporter);
772 if (error)
773 return error;
774 }
775 }
776again:
777 vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
778
779 if (file) {
780 mapping = file->f_mapping;
781 root = &mapping->i_mmap;
782 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
783
784 if (adjust_next)
785 uprobe_munmap(next, next->vm_start, next->vm_end);
786
787 i_mmap_lock_write(mapping);
788 if (insert) {
789 /*
790 * Put into interval tree now, so instantiated pages
791 * are visible to arm/parisc __flush_dcache_page
792 * throughout; but we cannot insert into address
793 * space until vma start or end is updated.
794 */
795 __vma_link_file(insert);
796 }
797 }
798
799 anon_vma = vma->anon_vma;
800 if (!anon_vma && adjust_next)
801 anon_vma = next->anon_vma;
802 if (anon_vma) {
803 VM_WARN_ON(adjust_next && next->anon_vma &&
804 anon_vma != next->anon_vma);
805 anon_vma_lock_write(anon_vma);
806 anon_vma_interval_tree_pre_update_vma(vma);
807 if (adjust_next)
808 anon_vma_interval_tree_pre_update_vma(next);
809 }
810
811 if (root) {
812 flush_dcache_mmap_lock(mapping);
813 vma_interval_tree_remove(vma, root);
814 if (adjust_next)
815 vma_interval_tree_remove(next, root);
816 }
817
818 if (start != vma->vm_start) {
819 vma->vm_start = start;
820 start_changed = true;
821 }
822 if (end != vma->vm_end) {
823 vma->vm_end = end;
824 end_changed = true;
825 }
826 vma->vm_pgoff = pgoff;
827 if (adjust_next) {
828 next->vm_start += adjust_next << PAGE_SHIFT;
829 next->vm_pgoff += adjust_next;
830 }
831
832 if (root) {
833 if (adjust_next)
834 vma_interval_tree_insert(next, root);
835 vma_interval_tree_insert(vma, root);
836 flush_dcache_mmap_unlock(mapping);
837 }
838
839 if (remove_next) {
840 /*
841 * vma_merge has merged next into vma, and needs
842 * us to remove next before dropping the locks.
843 */
844 if (remove_next != 3)
845 __vma_unlink_prev(mm, next, vma);
846 else
847 /*
848 * vma is not before next if they've been
849 * swapped.
850 *
851 * pre-swap() next->vm_start was reduced so
852 * tell validate_mm_rb to ignore pre-swap()
853 * "next" (which is stored in post-swap()
854 * "vma").
855 */
856 __vma_unlink_common(mm, next, NULL, false, vma);
857 if (file)
858 __remove_shared_vm_struct(next, file, mapping);
859 } else if (insert) {
860 /*
861 * split_vma has split insert from vma, and needs
862 * us to insert it before dropping the locks
863 * (it may either follow vma or precede it).
864 */
865 __insert_vm_struct(mm, insert);
866 } else {
867 if (start_changed)
868 vma_gap_update(vma);
869 if (end_changed) {
870 if (!next)
871 mm->highest_vm_end = vm_end_gap(vma);
872 else if (!adjust_next)
873 vma_gap_update(next);
874 }
875 }
876
877 if (anon_vma) {
878 anon_vma_interval_tree_post_update_vma(vma);
879 if (adjust_next)
880 anon_vma_interval_tree_post_update_vma(next);
881 anon_vma_unlock_write(anon_vma);
882 }
883 if (mapping)
884 i_mmap_unlock_write(mapping);
885
886 if (root) {
887 uprobe_mmap(vma);
888
889 if (adjust_next)
890 uprobe_mmap(next);
891 }
892
893 if (remove_next) {
894 if (file) {
895 uprobe_munmap(next, next->vm_start, next->vm_end);
896 fput(file);
897 }
898 if (next->anon_vma)
899 anon_vma_merge(vma, next);
900 mm->map_count--;
901 mpol_put(vma_policy(next));
902 kmem_cache_free(vm_area_cachep, next);
903 /*
904 * In mprotect's case 6 (see comments on vma_merge),
905 * we must remove another next too. It would clutter
906 * up the code too much to do both in one go.
907 */
908 if (remove_next != 3) {
909 /*
910 * If "next" was removed and vma->vm_end was
911 * expanded (up) over it, in turn
912 * "next->vm_prev->vm_end" changed and the
913 * "vma->vm_next" gap must be updated.
914 */
915 next = vma->vm_next;
916 } else {
917 /*
918 * For the scope of the comment "next" and
919 * "vma" considered pre-swap(): if "vma" was
920 * removed, next->vm_start was expanded (down)
921 * over it and the "next" gap must be updated.
922 * Because of the swap() the post-swap() "vma"
923 * actually points to pre-swap() "next"
924 * (post-swap() "next" as opposed is now a
925 * dangling pointer).
926 */
927 next = vma;
928 }
929 if (remove_next == 2) {
930 remove_next = 1;
931 end = next->vm_end;
932 goto again;
933 }
934 else if (next)
935 vma_gap_update(next);
936 else {
937 /*
938 * If remove_next == 2 we obviously can't
939 * reach this path.
940 *
941 * If remove_next == 3 we can't reach this
942 * path because pre-swap() next is always not
943 * NULL. pre-swap() "next" is not being
944 * removed and its next->vm_end is not altered
945 * (and furthermore "end" already matches
946 * next->vm_end in remove_next == 3).
947 *
948 * We reach this only in the remove_next == 1
949 * case if the "next" vma that was removed was
950 * the highest vma of the mm. However in such
951 * case next->vm_end == "end" and the extended
952 * "vma" has vma->vm_end == next->vm_end so
953 * mm->highest_vm_end doesn't need any update
954 * in remove_next == 1 case.
955 */
956 VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
957 }
958 }
959 if (insert && file)
960 uprobe_mmap(insert);
961
962 validate_mm(mm);
963
964 return 0;
965}
966
967/*
968 * If the vma has a ->close operation then the driver probably needs to release
969 * per-vma resources, so we don't attempt to merge those.
970 */
971static inline int is_mergeable_vma(struct vm_area_struct *vma,
972 struct file *file, unsigned long vm_flags,
973 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
974 const char __user *anon_name)
975{
976 /*
977 * VM_SOFTDIRTY should not prevent from VMA merging, if we
978 * match the flags but dirty bit -- the caller should mark
979 * merged VMA as dirty. If dirty bit won't be excluded from
980 * comparison, we increase pressue on the memory system forcing
981 * the kernel to generate new VMAs when old one could be
982 * extended instead.
983 */
984 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
985 return 0;
986 if (vma->vm_file != file)
987 return 0;
988 if (vma->vm_ops && vma->vm_ops->close)
989 return 0;
990 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
991 return 0;
992 if (vma_get_anon_name(vma) != anon_name)
993 return 0;
994 return 1;
995}
996
997static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
998 struct anon_vma *anon_vma2,
999 struct vm_area_struct *vma)
1000{
1001 /*
1002 * The list_is_singular() test is to avoid merging VMA cloned from
1003 * parents. This can improve scalability caused by anon_vma lock.
1004 */
1005 if ((!anon_vma1 || !anon_vma2) && (!vma ||
1006 list_is_singular(&vma->anon_vma_chain)))
1007 return 1;
1008 return anon_vma1 == anon_vma2;
1009}
1010
1011/*
1012 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1013 * in front of (at a lower virtual address and file offset than) the vma.
1014 *
1015 * We cannot merge two vmas if they have differently assigned (non-NULL)
1016 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1017 *
1018 * We don't check here for the merged mmap wrapping around the end of pagecache
1019 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
1020 * wrap, nor mmaps which cover the final page at index -1UL.
1021 */
1022static int
1023can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1024 struct anon_vma *anon_vma, struct file *file,
1025 pgoff_t vm_pgoff,
1026 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
1027 const char __user *anon_name)
1028{
1029 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name) &&
1030 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1031 if (vma->vm_pgoff == vm_pgoff)
1032 return 1;
1033 }
1034 return 0;
1035}
1036
1037/*
1038 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1039 * beyond (at a higher virtual address and file offset than) the vma.
1040 *
1041 * We cannot merge two vmas if they have differently assigned (non-NULL)
1042 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1043 */
1044static int
1045can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1046 struct anon_vma *anon_vma, struct file *file,
1047 pgoff_t vm_pgoff,
1048 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
1049 const char __user *anon_name)
1050{
1051 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name) &&
1052 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1053 pgoff_t vm_pglen;
1054 vm_pglen = vma_pages(vma);
1055 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1056 return 1;
1057 }
1058 return 0;
1059}
1060
1061/*
1062 * Given a mapping request (addr,end,vm_flags,file,pgoff,anon_name),
1063 * figure out whether that can be merged with its predecessor or its
1064 * successor. Or both (it neatly fills a hole).
1065 *
1066 * In most cases - when called for mmap, brk or mremap - [addr,end) is
1067 * certain not to be mapped by the time vma_merge is called; but when
1068 * called for mprotect, it is certain to be already mapped (either at
1069 * an offset within prev, or at the start of next), and the flags of
1070 * this area are about to be changed to vm_flags - and the no-change
1071 * case has already been eliminated.
1072 *
1073 * The following mprotect cases have to be considered, where AAAA is
1074 * the area passed down from mprotect_fixup, never extending beyond one
1075 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1076 *
1077 * AAAA AAAA AAAA AAAA
1078 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
1079 * cannot merge might become might become might become
1080 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
1081 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
1082 * mremap move: PPPPXXXXXXXX 8
1083 * AAAA
1084 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
1085 * might become case 1 below case 2 below case 3 below
1086 *
1087 * It is important for case 8 that the the vma NNNN overlapping the
1088 * region AAAA is never going to extended over XXXX. Instead XXXX must
1089 * be extended in region AAAA and NNNN must be removed. This way in
1090 * all cases where vma_merge succeeds, the moment vma_adjust drops the
1091 * rmap_locks, the properties of the merged vma will be already
1092 * correct for the whole merged range. Some of those properties like
1093 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1094 * be correct for the whole merged range immediately after the
1095 * rmap_locks are released. Otherwise if XXXX would be removed and
1096 * NNNN would be extended over the XXXX range, remove_migration_ptes
1097 * or other rmap walkers (if working on addresses beyond the "end"
1098 * parameter) may establish ptes with the wrong permissions of NNNN
1099 * instead of the right permissions of XXXX.
1100 */
1101struct vm_area_struct *vma_merge(struct mm_struct *mm,
1102 struct vm_area_struct *prev, unsigned long addr,
1103 unsigned long end, unsigned long vm_flags,
1104 struct anon_vma *anon_vma, struct file *file,
1105 pgoff_t pgoff, struct mempolicy *policy,
1106 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
1107 const char __user *anon_name)
1108{
1109 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1110 struct vm_area_struct *area, *next;
1111 int err;
1112
1113 /*
1114 * We later require that vma->vm_flags == vm_flags,
1115 * so this tests vma->vm_flags & VM_SPECIAL, too.
1116 */
1117 if (vm_flags & VM_SPECIAL)
1118 return NULL;
1119
1120 if (prev)
1121 next = prev->vm_next;
1122 else
1123 next = mm->mmap;
1124 area = next;
1125 if (area && area->vm_end == end) /* cases 6, 7, 8 */
1126 next = next->vm_next;
1127
1128 /* verify some invariant that must be enforced by the caller */
1129 VM_WARN_ON(prev && addr <= prev->vm_start);
1130 VM_WARN_ON(area && end > area->vm_end);
1131 VM_WARN_ON(addr >= end);
1132
1133 /*
1134 * Can it merge with the predecessor?
1135 */
1136 if (prev && prev->vm_end == addr &&
1137 mpol_equal(vma_policy(prev), policy) &&
1138 can_vma_merge_after(prev, vm_flags,
1139 anon_vma, file, pgoff,
1140 vm_userfaultfd_ctx,
1141 anon_name)) {
1142 /*
1143 * OK, it can. Can we now merge in the successor as well?
1144 */
1145 if (next && end == next->vm_start &&
1146 mpol_equal(policy, vma_policy(next)) &&
1147 can_vma_merge_before(next, vm_flags,
1148 anon_vma, file,
1149 pgoff+pglen,
1150 vm_userfaultfd_ctx,
1151 anon_name) &&
1152 is_mergeable_anon_vma(prev->anon_vma,
1153 next->anon_vma, NULL)) {
1154 /* cases 1, 6 */
1155 err = __vma_adjust(prev, prev->vm_start,
1156 next->vm_end, prev->vm_pgoff, NULL,
1157 prev);
1158 } else /* cases 2, 5, 7 */
1159 err = __vma_adjust(prev, prev->vm_start,
1160 end, prev->vm_pgoff, NULL, prev);
1161 if (err)
1162 return NULL;
1163 khugepaged_enter_vma_merge(prev, vm_flags);
1164 return prev;
1165 }
1166
1167 /*
1168 * Can this new request be merged in front of next?
1169 */
1170 if (next && end == next->vm_start &&
1171 mpol_equal(policy, vma_policy(next)) &&
1172 can_vma_merge_before(next, vm_flags,
1173 anon_vma, file, pgoff+pglen,
1174 vm_userfaultfd_ctx,
1175 anon_name)) {
1176 if (prev && addr < prev->vm_end) /* case 4 */
1177 err = __vma_adjust(prev, prev->vm_start,
1178 addr, prev->vm_pgoff, NULL, next);
1179 else { /* cases 3, 8 */
1180 err = __vma_adjust(area, addr, next->vm_end,
1181 next->vm_pgoff - pglen, NULL, next);
1182 /*
1183 * In case 3 area is already equal to next and
1184 * this is a noop, but in case 8 "area" has
1185 * been removed and next was expanded over it.
1186 */
1187 area = next;
1188 }
1189 if (err)
1190 return NULL;
1191 khugepaged_enter_vma_merge(area, vm_flags);
1192 return area;
1193 }
1194
1195 return NULL;
1196}
1197
1198/*
1199 * Rough compatbility check to quickly see if it's even worth looking
1200 * at sharing an anon_vma.
1201 *
1202 * They need to have the same vm_file, and the flags can only differ
1203 * in things that mprotect may change.
1204 *
1205 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1206 * we can merge the two vma's. For example, we refuse to merge a vma if
1207 * there is a vm_ops->close() function, because that indicates that the
1208 * driver is doing some kind of reference counting. But that doesn't
1209 * really matter for the anon_vma sharing case.
1210 */
1211static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1212{
1213 return a->vm_end == b->vm_start &&
1214 mpol_equal(vma_policy(a), vma_policy(b)) &&
1215 a->vm_file == b->vm_file &&
1216 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) &&
1217 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1218}
1219
1220/*
1221 * Do some basic sanity checking to see if we can re-use the anon_vma
1222 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1223 * the same as 'old', the other will be the new one that is trying
1224 * to share the anon_vma.
1225 *
1226 * NOTE! This runs with mm_sem held for reading, so it is possible that
1227 * the anon_vma of 'old' is concurrently in the process of being set up
1228 * by another page fault trying to merge _that_. But that's ok: if it
1229 * is being set up, that automatically means that it will be a singleton
1230 * acceptable for merging, so we can do all of this optimistically. But
1231 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1232 *
1233 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1234 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1235 * is to return an anon_vma that is "complex" due to having gone through
1236 * a fork).
1237 *
1238 * We also make sure that the two vma's are compatible (adjacent,
1239 * and with the same memory policies). That's all stable, even with just
1240 * a read lock on the mm_sem.
1241 */
1242static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1243{
1244 if (anon_vma_compatible(a, b)) {
1245 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1246
1247 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1248 return anon_vma;
1249 }
1250 return NULL;
1251}
1252
1253/*
1254 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1255 * neighbouring vmas for a suitable anon_vma, before it goes off
1256 * to allocate a new anon_vma. It checks because a repetitive
1257 * sequence of mprotects and faults may otherwise lead to distinct
1258 * anon_vmas being allocated, preventing vma merge in subsequent
1259 * mprotect.
1260 */
1261struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1262{
1263 struct anon_vma *anon_vma;
1264 struct vm_area_struct *near;
1265
1266 near = vma->vm_next;
1267 if (!near)
1268 goto try_prev;
1269
1270 anon_vma = reusable_anon_vma(near, vma, near);
1271 if (anon_vma)
1272 return anon_vma;
1273try_prev:
1274 near = vma->vm_prev;
1275 if (!near)
1276 goto none;
1277
1278 anon_vma = reusable_anon_vma(near, near, vma);
1279 if (anon_vma)
1280 return anon_vma;
1281none:
1282 /*
1283 * There's no absolute need to look only at touching neighbours:
1284 * we could search further afield for "compatible" anon_vmas.
1285 * But it would probably just be a waste of time searching,
1286 * or lead to too many vmas hanging off the same anon_vma.
1287 * We're trying to allow mprotect remerging later on,
1288 * not trying to minimize memory used for anon_vmas.
1289 */
1290 return NULL;
1291}
1292
1293/*
1294 * If a hint addr is less than mmap_min_addr change hint to be as
1295 * low as possible but still greater than mmap_min_addr
1296 */
1297static inline unsigned long round_hint_to_min(unsigned long hint)
1298{
1299 hint &= PAGE_MASK;
1300 if (((void *)hint != NULL) &&
1301 (hint < mmap_min_addr))
1302 return PAGE_ALIGN(mmap_min_addr);
1303 return hint;
1304}
1305
1306static inline int mlock_future_check(struct mm_struct *mm,
1307 unsigned long flags,
1308 unsigned long len)
1309{
1310 unsigned long locked, lock_limit;
1311
1312 /* mlock MCL_FUTURE? */
1313 if (flags & VM_LOCKED) {
1314 locked = len >> PAGE_SHIFT;
1315 locked += mm->locked_vm;
1316 lock_limit = rlimit(RLIMIT_MEMLOCK);
1317 lock_limit >>= PAGE_SHIFT;
1318 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1319 return -EAGAIN;
1320 }
1321 return 0;
1322}
1323
1324static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1325{
1326 if (S_ISREG(inode->i_mode))
1327 return MAX_LFS_FILESIZE;
1328
1329 if (S_ISBLK(inode->i_mode))
1330 return MAX_LFS_FILESIZE;
1331
1332 /* Special "we do even unsigned file positions" case */
1333 if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1334 return 0;
1335
1336 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1337 return ULONG_MAX;
1338}
1339
1340static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1341 unsigned long pgoff, unsigned long len)
1342{
1343 u64 maxsize = file_mmap_size_max(file, inode);
1344
1345 if (maxsize && len > maxsize)
1346 return false;
1347 maxsize -= len;
1348 if (pgoff > maxsize >> PAGE_SHIFT)
1349 return false;
1350 return true;
1351}
1352
1353/*
1354 * The caller must hold down_write(&current->mm->mmap_sem).
1355 */
1356unsigned long do_mmap(struct file *file, unsigned long addr,
1357 unsigned long len, unsigned long prot,
1358 unsigned long flags, vm_flags_t vm_flags,
1359 unsigned long pgoff, unsigned long *populate)
1360{
1361 struct mm_struct *mm = current->mm;
1362 int pkey = 0;
1363
1364 *populate = 0;
1365
1366 if (!len)
1367 return -EINVAL;
1368
1369 /*
1370 * Does the application expect PROT_READ to imply PROT_EXEC?
1371 *
1372 * (the exception is when the underlying filesystem is noexec
1373 * mounted, in which case we dont add PROT_EXEC.)
1374 */
1375 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1376 if (!(file && path_noexec(&file->f_path)))
1377 prot |= PROT_EXEC;
1378
1379 if (!(flags & MAP_FIXED))
1380 addr = round_hint_to_min(addr);
1381
1382 /* Careful about overflows.. */
1383 len = PAGE_ALIGN(len);
1384 if (!len)
1385 return -ENOMEM;
1386
1387 /* offset overflow? */
1388 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1389 return -EOVERFLOW;
1390
1391 /* Too many mappings? */
1392 if (mm->map_count > sysctl_max_map_count)
1393 return -ENOMEM;
1394
1395 /* Obtain the address to map to. we verify (or select) it and ensure
1396 * that it represents a valid section of the address space.
1397 */
1398 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1399 if (offset_in_page(addr))
1400 return addr;
1401
1402 if (prot == PROT_EXEC) {
1403 pkey = execute_only_pkey(mm);
1404 if (pkey < 0)
1405 pkey = 0;
1406 }
1407
1408 /* Do simple checking here so the lower-level routines won't have
1409 * to. we assume access permissions have been handled by the open
1410 * of the memory object, so we don't do any here.
1411 */
1412 vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1413 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1414
1415 if (flags & MAP_LOCKED)
1416 if (!can_do_mlock())
1417 return -EPERM;
1418
1419 if (mlock_future_check(mm, vm_flags, len))
1420 return -EAGAIN;
1421
1422 if (file) {
1423 struct inode *inode = file_inode(file);
1424
1425 if (!file_mmap_ok(file, inode, pgoff, len))
1426 return -EOVERFLOW;
1427
1428 switch (flags & MAP_TYPE) {
1429 case MAP_SHARED:
1430 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1431 return -EACCES;
1432
1433 /*
1434 * Make sure we don't allow writing to an append-only
1435 * file..
1436 */
1437 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1438 return -EACCES;
1439
1440 /*
1441 * Make sure there are no mandatory locks on the file.
1442 */
1443 if (locks_verify_locked(file))
1444 return -EAGAIN;
1445
1446 vm_flags |= VM_SHARED | VM_MAYSHARE;
1447 if (!(file->f_mode & FMODE_WRITE))
1448 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1449
1450 /* fall through */
1451 case MAP_PRIVATE:
1452 if (!(file->f_mode & FMODE_READ))
1453 return -EACCES;
1454 if (path_noexec(&file->f_path)) {
1455 if (vm_flags & VM_EXEC)
1456 return -EPERM;
1457 vm_flags &= ~VM_MAYEXEC;
1458 }
1459
1460 if (!file->f_op->mmap)
1461 return -ENODEV;
1462 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1463 return -EINVAL;
1464 break;
1465
1466 default:
1467 return -EINVAL;
1468 }
1469 } else {
1470 switch (flags & MAP_TYPE) {
1471 case MAP_SHARED:
1472 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1473 return -EINVAL;
1474 /*
1475 * Ignore pgoff.
1476 */
1477 pgoff = 0;
1478 vm_flags |= VM_SHARED | VM_MAYSHARE;
1479 break;
1480 case MAP_PRIVATE:
1481 /*
1482 * Set pgoff according to addr for anon_vma.
1483 */
1484 pgoff = addr >> PAGE_SHIFT;
1485 break;
1486 default:
1487 return -EINVAL;
1488 }
1489 }
1490
1491 /*
1492 * Set 'VM_NORESERVE' if we should not account for the
1493 * memory use of this mapping.
1494 */
1495 if (flags & MAP_NORESERVE) {
1496 /* We honor MAP_NORESERVE if allowed to overcommit */
1497 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1498 vm_flags |= VM_NORESERVE;
1499
1500 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1501 if (file && is_file_hugepages(file))
1502 vm_flags |= VM_NORESERVE;
1503 }
1504
1505 addr = mmap_region(file, addr, len, vm_flags, pgoff);
1506 if (!IS_ERR_VALUE(addr) &&
1507 ((vm_flags & VM_LOCKED) ||
1508 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1509 *populate = len;
1510 return addr;
1511}
1512
1513SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1514 unsigned long, prot, unsigned long, flags,
1515 unsigned long, fd, unsigned long, pgoff)
1516{
1517 struct file *file = NULL;
1518 unsigned long retval;
1519
1520 if (!(flags & MAP_ANONYMOUS)) {
1521 audit_mmap_fd(fd, flags);
1522 file = fget(fd);
1523 if (!file)
1524 return -EBADF;
1525 if (is_file_hugepages(file))
1526 len = ALIGN(len, huge_page_size(hstate_file(file)));
1527 retval = -EINVAL;
1528 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1529 goto out_fput;
1530 } else if (flags & MAP_HUGETLB) {
1531 struct user_struct *user = NULL;
1532 struct hstate *hs;
1533
1534 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & SHM_HUGE_MASK);
1535 if (!hs)
1536 return -EINVAL;
1537
1538 len = ALIGN(len, huge_page_size(hs));
1539 /*
1540 * VM_NORESERVE is used because the reservations will be
1541 * taken when vm_ops->mmap() is called
1542 * A dummy user value is used because we are not locking
1543 * memory so no accounting is necessary
1544 */
1545 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1546 VM_NORESERVE,
1547 &user, HUGETLB_ANONHUGE_INODE,
1548 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1549 if (IS_ERR(file))
1550 return PTR_ERR(file);
1551 }
1552
1553 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1554
1555 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1556out_fput:
1557 if (file)
1558 fput(file);
1559 return retval;
1560}
1561
1562#ifdef __ARCH_WANT_SYS_OLD_MMAP
1563struct mmap_arg_struct {
1564 unsigned long addr;
1565 unsigned long len;
1566 unsigned long prot;
1567 unsigned long flags;
1568 unsigned long fd;
1569 unsigned long offset;
1570};
1571
1572SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1573{
1574 struct mmap_arg_struct a;
1575
1576 if (copy_from_user(&a, arg, sizeof(a)))
1577 return -EFAULT;
1578 if (offset_in_page(a.offset))
1579 return -EINVAL;
1580
1581 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1582 a.offset >> PAGE_SHIFT);
1583}
1584#endif /* __ARCH_WANT_SYS_OLD_MMAP */
1585
1586/*
1587 * Some shared mappigns will want the pages marked read-only
1588 * to track write events. If so, we'll downgrade vm_page_prot
1589 * to the private version (using protection_map[] without the
1590 * VM_SHARED bit).
1591 */
1592int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1593{
1594 vm_flags_t vm_flags = vma->vm_flags;
1595 const struct vm_operations_struct *vm_ops = vma->vm_ops;
1596
1597 /* If it was private or non-writable, the write bit is already clear */
1598 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1599 return 0;
1600
1601 /* The backer wishes to know when pages are first written to? */
1602 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1603 return 1;
1604
1605 /* The open routine did something to the protections that pgprot_modify
1606 * won't preserve? */
1607 if (pgprot_val(vm_page_prot) !=
1608 pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1609 return 0;
1610
1611 /* Do we need to track softdirty? */
1612 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1613 return 1;
1614
1615 /* Specialty mapping? */
1616 if (vm_flags & VM_PFNMAP)
1617 return 0;
1618
1619 /* Can the mapping track the dirty pages? */
1620 return vma->vm_file && vma->vm_file->f_mapping &&
1621 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1622}
1623
1624/*
1625 * We account for memory if it's a private writeable mapping,
1626 * not hugepages and VM_NORESERVE wasn't set.
1627 */
1628static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1629{
1630 /*
1631 * hugetlb has its own accounting separate from the core VM
1632 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1633 */
1634 if (file && is_file_hugepages(file))
1635 return 0;
1636
1637 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1638}
1639
1640unsigned long mmap_region(struct file *file, unsigned long addr,
1641 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff)
1642{
1643 struct mm_struct *mm = current->mm;
1644 struct vm_area_struct *vma, *prev;
1645 int error;
1646 struct rb_node **rb_link, *rb_parent;
1647 unsigned long charged = 0;
1648
1649 /* Check against address space limit. */
1650 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1651 unsigned long nr_pages;
1652
1653 /*
1654 * MAP_FIXED may remove pages of mappings that intersects with
1655 * requested mapping. Account for the pages it would unmap.
1656 */
1657 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1658
1659 if (!may_expand_vm(mm, vm_flags,
1660 (len >> PAGE_SHIFT) - nr_pages))
1661 return -ENOMEM;
1662 }
1663
1664 /* Clear old maps */
1665 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
1666 &rb_parent)) {
1667 if (do_munmap(mm, addr, len))
1668 return -ENOMEM;
1669 }
1670
1671 /*
1672 * Private writable mapping: check memory availability
1673 */
1674 if (accountable_mapping(file, vm_flags)) {
1675 charged = len >> PAGE_SHIFT;
1676 if (security_vm_enough_memory_mm(mm, charged))
1677 return -ENOMEM;
1678 vm_flags |= VM_ACCOUNT;
1679 }
1680
1681 /*
1682 * Can we just expand an old mapping?
1683 */
1684 vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1685 NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX, NULL);
1686 if (vma)
1687 goto out;
1688
1689 /*
1690 * Determine the object being mapped and call the appropriate
1691 * specific mapper. the address has already been validated, but
1692 * not unmapped, but the maps are removed from the list.
1693 */
1694 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1695 if (!vma) {
1696 error = -ENOMEM;
1697 goto unacct_error;
1698 }
1699
1700 vma->vm_mm = mm;
1701 vma->vm_start = addr;
1702 vma->vm_end = addr + len;
1703 vma->vm_flags = vm_flags;
1704 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1705 vma->vm_pgoff = pgoff;
1706 INIT_LIST_HEAD(&vma->anon_vma_chain);
1707
1708 if (file) {
1709 if (vm_flags & VM_DENYWRITE) {
1710 error = deny_write_access(file);
1711 if (error)
1712 goto free_vma;
1713 }
1714 if (vm_flags & VM_SHARED) {
1715 error = mapping_map_writable(file->f_mapping);
1716 if (error)
1717 goto allow_write_and_free_vma;
1718 }
1719
1720 /* ->mmap() can change vma->vm_file, but must guarantee that
1721 * vma_link() below can deny write-access if VM_DENYWRITE is set
1722 * and map writably if VM_SHARED is set. This usually means the
1723 * new file must not have been exposed to user-space, yet.
1724 */
1725 vma->vm_file = get_file(file);
1726 error = file->f_op->mmap(file, vma);
1727 if (error)
1728 goto unmap_and_free_vma;
1729
1730 /* Can addr have changed??
1731 *
1732 * Answer: Yes, several device drivers can do it in their
1733 * f_op->mmap method. -DaveM
1734 * Bug: If addr is changed, prev, rb_link, rb_parent should
1735 * be updated for vma_link()
1736 */
1737 WARN_ON_ONCE(addr != vma->vm_start);
1738
1739 addr = vma->vm_start;
1740 vm_flags = vma->vm_flags;
1741 } else if (vm_flags & VM_SHARED) {
1742 error = shmem_zero_setup(vma);
1743 if (error)
1744 goto free_vma;
1745 }
1746
1747 vma_link(mm, vma, prev, rb_link, rb_parent);
1748 /* Once vma denies write, undo our temporary denial count */
1749 if (file) {
1750 if (vm_flags & VM_SHARED)
1751 mapping_unmap_writable(file->f_mapping);
1752 if (vm_flags & VM_DENYWRITE)
1753 allow_write_access(file);
1754 }
1755 file = vma->vm_file;
1756out:
1757 perf_event_mmap(vma);
1758
1759 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1760 if (vm_flags & VM_LOCKED) {
1761 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1762 vma == get_gate_vma(current->mm)))
1763 mm->locked_vm += (len >> PAGE_SHIFT);
1764 else
1765 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1766 }
1767
1768 if (file)
1769 uprobe_mmap(vma);
1770
1771 /*
1772 * New (or expanded) vma always get soft dirty status.
1773 * Otherwise user-space soft-dirty page tracker won't
1774 * be able to distinguish situation when vma area unmapped,
1775 * then new mapped in-place (which must be aimed as
1776 * a completely new data area).
1777 */
1778 vma->vm_flags |= VM_SOFTDIRTY;
1779
1780 vma_set_page_prot(vma);
1781
1782 return addr;
1783
1784unmap_and_free_vma:
1785 vma->vm_file = NULL;
1786 fput(file);
1787
1788 /* Undo any partial mapping done by a device driver. */
1789 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1790 charged = 0;
1791 if (vm_flags & VM_SHARED)
1792 mapping_unmap_writable(file->f_mapping);
1793allow_write_and_free_vma:
1794 if (vm_flags & VM_DENYWRITE)
1795 allow_write_access(file);
1796free_vma:
1797 kmem_cache_free(vm_area_cachep, vma);
1798unacct_error:
1799 if (charged)
1800 vm_unacct_memory(charged);
1801 return error;
1802}
1803
1804unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1805{
1806 /*
1807 * We implement the search by looking for an rbtree node that
1808 * immediately follows a suitable gap. That is,
1809 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1810 * - gap_end = vma->vm_start >= info->low_limit + length;
1811 * - gap_end - gap_start >= length
1812 */
1813
1814 struct mm_struct *mm = current->mm;
1815 struct vm_area_struct *vma;
1816 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1817
1818 /* Adjust search length to account for worst case alignment overhead */
1819 length = info->length + info->align_mask;
1820 if (length < info->length)
1821 return -ENOMEM;
1822
1823 /* Adjust search limits by the desired length */
1824 if (info->high_limit < length)
1825 return -ENOMEM;
1826 high_limit = info->high_limit - length;
1827
1828 if (info->low_limit > high_limit)
1829 return -ENOMEM;
1830 low_limit = info->low_limit + length;
1831
1832 /* Check if rbtree root looks promising */
1833 if (RB_EMPTY_ROOT(&mm->mm_rb))
1834 goto check_highest;
1835 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1836 if (vma->rb_subtree_gap < length)
1837 goto check_highest;
1838
1839 while (true) {
1840 /* Visit left subtree if it looks promising */
1841 gap_end = vm_start_gap(vma);
1842 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1843 struct vm_area_struct *left =
1844 rb_entry(vma->vm_rb.rb_left,
1845 struct vm_area_struct, vm_rb);
1846 if (left->rb_subtree_gap >= length) {
1847 vma = left;
1848 continue;
1849 }
1850 }
1851
1852 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1853check_current:
1854 /* Check if current node has a suitable gap */
1855 if (gap_start > high_limit)
1856 return -ENOMEM;
1857 if (gap_end >= low_limit &&
1858 gap_end > gap_start && gap_end - gap_start >= length)
1859 goto found;
1860
1861 /* Visit right subtree if it looks promising */
1862 if (vma->vm_rb.rb_right) {
1863 struct vm_area_struct *right =
1864 rb_entry(vma->vm_rb.rb_right,
1865 struct vm_area_struct, vm_rb);
1866 if (right->rb_subtree_gap >= length) {
1867 vma = right;
1868 continue;
1869 }
1870 }
1871
1872 /* Go back up the rbtree to find next candidate node */
1873 while (true) {
1874 struct rb_node *prev = &vma->vm_rb;
1875 if (!rb_parent(prev))
1876 goto check_highest;
1877 vma = rb_entry(rb_parent(prev),
1878 struct vm_area_struct, vm_rb);
1879 if (prev == vma->vm_rb.rb_left) {
1880 gap_start = vm_end_gap(vma->vm_prev);
1881 gap_end = vm_start_gap(vma);
1882 goto check_current;
1883 }
1884 }
1885 }
1886
1887check_highest:
1888 /* Check highest gap, which does not precede any rbtree node */
1889 gap_start = mm->highest_vm_end;
1890 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */
1891 if (gap_start > high_limit)
1892 return -ENOMEM;
1893
1894found:
1895 /* We found a suitable gap. Clip it with the original low_limit. */
1896 if (gap_start < info->low_limit)
1897 gap_start = info->low_limit;
1898
1899 /* Adjust gap address to the desired alignment */
1900 gap_start += (info->align_offset - gap_start) & info->align_mask;
1901
1902 VM_BUG_ON(gap_start + info->length > info->high_limit);
1903 VM_BUG_ON(gap_start + info->length > gap_end);
1904 return gap_start;
1905}
1906
1907unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1908{
1909 struct mm_struct *mm = current->mm;
1910 struct vm_area_struct *vma;
1911 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1912
1913 /* Adjust search length to account for worst case alignment overhead */
1914 length = info->length + info->align_mask;
1915 if (length < info->length)
1916 return -ENOMEM;
1917
1918 /*
1919 * Adjust search limits by the desired length.
1920 * See implementation comment at top of unmapped_area().
1921 */
1922 gap_end = info->high_limit;
1923 if (gap_end < length)
1924 return -ENOMEM;
1925 high_limit = gap_end - length;
1926
1927 if (info->low_limit > high_limit)
1928 return -ENOMEM;
1929 low_limit = info->low_limit + length;
1930
1931 /* Check highest gap, which does not precede any rbtree node */
1932 gap_start = mm->highest_vm_end;
1933 if (gap_start <= high_limit)
1934 goto found_highest;
1935
1936 /* Check if rbtree root looks promising */
1937 if (RB_EMPTY_ROOT(&mm->mm_rb))
1938 return -ENOMEM;
1939 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1940 if (vma->rb_subtree_gap < length)
1941 return -ENOMEM;
1942
1943 while (true) {
1944 /* Visit right subtree if it looks promising */
1945 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1946 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1947 struct vm_area_struct *right =
1948 rb_entry(vma->vm_rb.rb_right,
1949 struct vm_area_struct, vm_rb);
1950 if (right->rb_subtree_gap >= length) {
1951 vma = right;
1952 continue;
1953 }
1954 }
1955
1956check_current:
1957 /* Check if current node has a suitable gap */
1958 gap_end = vm_start_gap(vma);
1959 if (gap_end < low_limit)
1960 return -ENOMEM;
1961 if (gap_start <= high_limit &&
1962 gap_end > gap_start && gap_end - gap_start >= length)
1963 goto found;
1964
1965 /* Visit left subtree if it looks promising */
1966 if (vma->vm_rb.rb_left) {
1967 struct vm_area_struct *left =
1968 rb_entry(vma->vm_rb.rb_left,
1969 struct vm_area_struct, vm_rb);
1970 if (left->rb_subtree_gap >= length) {
1971 vma = left;
1972 continue;
1973 }
1974 }
1975
1976 /* Go back up the rbtree to find next candidate node */
1977 while (true) {
1978 struct rb_node *prev = &vma->vm_rb;
1979 if (!rb_parent(prev))
1980 return -ENOMEM;
1981 vma = rb_entry(rb_parent(prev),
1982 struct vm_area_struct, vm_rb);
1983 if (prev == vma->vm_rb.rb_right) {
1984 gap_start = vma->vm_prev ?
1985 vm_end_gap(vma->vm_prev) : 0;
1986 goto check_current;
1987 }
1988 }
1989 }
1990
1991found:
1992 /* We found a suitable gap. Clip it with the original high_limit. */
1993 if (gap_end > info->high_limit)
1994 gap_end = info->high_limit;
1995
1996found_highest:
1997 /* Compute highest gap address at the desired alignment */
1998 gap_end -= info->length;
1999 gap_end -= (gap_end - info->align_offset) & info->align_mask;
2000
2001 VM_BUG_ON(gap_end < info->low_limit);
2002 VM_BUG_ON(gap_end < gap_start);
2003 return gap_end;
2004}
2005
2006/* Get an address range which is currently unmapped.
2007 * For shmat() with addr=0.
2008 *
2009 * Ugly calling convention alert:
2010 * Return value with the low bits set means error value,
2011 * ie
2012 * if (ret & ~PAGE_MASK)
2013 * error = ret;
2014 *
2015 * This function "knows" that -ENOMEM has the bits set.
2016 */
2017#ifndef HAVE_ARCH_UNMAPPED_AREA
2018unsigned long
2019arch_get_unmapped_area(struct file *filp, unsigned long addr,
2020 unsigned long len, unsigned long pgoff, unsigned long flags)
2021{
2022 struct mm_struct *mm = current->mm;
2023 struct vm_area_struct *vma, *prev;
2024 struct vm_unmapped_area_info info;
2025
2026 if (len > TASK_SIZE - mmap_min_addr)
2027 return -ENOMEM;
2028
2029 if (flags & MAP_FIXED)
2030 return addr;
2031
2032 if (addr) {
2033 addr = PAGE_ALIGN(addr);
2034 vma = find_vma_prev(mm, addr, &prev);
2035 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
2036 (!vma || addr + len <= vm_start_gap(vma)) &&
2037 (!prev || addr >= vm_end_gap(prev)))
2038 return addr;
2039 }
2040
2041 info.flags = 0;
2042 info.length = len;
2043 info.low_limit = mm->mmap_base;
2044 info.high_limit = TASK_SIZE;
2045 info.align_mask = 0;
2046 return vm_unmapped_area(&info);
2047}
2048#endif
2049
2050/*
2051 * This mmap-allocator allocates new areas top-down from below the
2052 * stack's low limit (the base):
2053 */
2054#ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2055unsigned long
2056arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
2057 const unsigned long len, const unsigned long pgoff,
2058 const unsigned long flags)
2059{
2060 struct vm_area_struct *vma, *prev;
2061 struct mm_struct *mm = current->mm;
2062 unsigned long addr = addr0;
2063 struct vm_unmapped_area_info info;
2064
2065 /* requested length too big for entire address space */
2066 if (len > TASK_SIZE - mmap_min_addr)
2067 return -ENOMEM;
2068
2069 if (flags & MAP_FIXED)
2070 return addr;
2071
2072 /* requesting a specific address */
2073 if (addr) {
2074 addr = PAGE_ALIGN(addr);
2075 vma = find_vma_prev(mm, addr, &prev);
2076 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
2077 (!vma || addr + len <= vm_start_gap(vma)) &&
2078 (!prev || addr >= vm_end_gap(prev)))
2079 return addr;
2080 }
2081
2082 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2083 info.length = len;
2084 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2085 info.high_limit = mm->mmap_base;
2086 info.align_mask = 0;
2087 addr = vm_unmapped_area(&info);
2088
2089 /*
2090 * A failed mmap() very likely causes application failure,
2091 * so fall back to the bottom-up function here. This scenario
2092 * can happen with large stack limits and large mmap()
2093 * allocations.
2094 */
2095 if (offset_in_page(addr)) {
2096 VM_BUG_ON(addr != -ENOMEM);
2097 info.flags = 0;
2098 info.low_limit = TASK_UNMAPPED_BASE;
2099 info.high_limit = TASK_SIZE;
2100 addr = vm_unmapped_area(&info);
2101 }
2102
2103 return addr;
2104}
2105#endif
2106
2107unsigned long
2108get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2109 unsigned long pgoff, unsigned long flags)
2110{
2111 unsigned long (*get_area)(struct file *, unsigned long,
2112 unsigned long, unsigned long, unsigned long);
2113
2114 unsigned long error = arch_mmap_check(addr, len, flags);
2115 if (error)
2116 return error;
2117
2118 /* Careful about overflows.. */
2119 if (len > TASK_SIZE)
2120 return -ENOMEM;
2121
2122 get_area = current->mm->get_unmapped_area;
2123 if (file) {
2124 if (file->f_op->get_unmapped_area)
2125 get_area = file->f_op->get_unmapped_area;
2126 } else if (flags & MAP_SHARED) {
2127 /*
2128 * mmap_region() will call shmem_zero_setup() to create a file,
2129 * so use shmem's get_unmapped_area in case it can be huge.
2130 * do_mmap_pgoff() will clear pgoff, so match alignment.
2131 */
2132 pgoff = 0;
2133 get_area = shmem_get_unmapped_area;
2134 }
2135
2136 addr = get_area(file, addr, len, pgoff, flags);
2137 if (IS_ERR_VALUE(addr))
2138 return addr;
2139
2140 if (addr > TASK_SIZE - len)
2141 return -ENOMEM;
2142 if (offset_in_page(addr))
2143 return -EINVAL;
2144
2145 error = security_mmap_addr(addr);
2146 return error ? error : addr;
2147}
2148
2149EXPORT_SYMBOL(get_unmapped_area);
2150
2151/* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
2152struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2153{
2154 struct rb_node *rb_node;
2155 struct vm_area_struct *vma;
2156
2157 /* Check the cache first. */
2158 vma = vmacache_find(mm, addr);
2159 if (likely(vma))
2160 return vma;
2161
2162 rb_node = mm->mm_rb.rb_node;
2163
2164 while (rb_node) {
2165 struct vm_area_struct *tmp;
2166
2167 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2168
2169 if (tmp->vm_end > addr) {
2170 vma = tmp;
2171 if (tmp->vm_start <= addr)
2172 break;
2173 rb_node = rb_node->rb_left;
2174 } else
2175 rb_node = rb_node->rb_right;
2176 }
2177
2178 if (vma)
2179 vmacache_update(addr, vma);
2180 return vma;
2181}
2182
2183EXPORT_SYMBOL(find_vma);
2184
2185/*
2186 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2187 */
2188struct vm_area_struct *
2189find_vma_prev(struct mm_struct *mm, unsigned long addr,
2190 struct vm_area_struct **pprev)
2191{
2192 struct vm_area_struct *vma;
2193
2194 vma = find_vma(mm, addr);
2195 if (vma) {
2196 *pprev = vma->vm_prev;
2197 } else {
2198 struct rb_node *rb_node = mm->mm_rb.rb_node;
2199 *pprev = NULL;
2200 while (rb_node) {
2201 *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2202 rb_node = rb_node->rb_right;
2203 }
2204 }
2205 return vma;
2206}
2207
2208/*
2209 * Verify that the stack growth is acceptable and
2210 * update accounting. This is shared with both the
2211 * grow-up and grow-down cases.
2212 */
2213static int acct_stack_growth(struct vm_area_struct *vma,
2214 unsigned long size, unsigned long grow)
2215{
2216 struct mm_struct *mm = vma->vm_mm;
2217 struct rlimit *rlim = current->signal->rlim;
2218 unsigned long new_start;
2219
2220 /* address space limit tests */
2221 if (!may_expand_vm(mm, vma->vm_flags, grow))
2222 return -ENOMEM;
2223
2224 /* Stack limit test */
2225 if (size > READ_ONCE(rlim[RLIMIT_STACK].rlim_cur))
2226 return -ENOMEM;
2227
2228 /* mlock limit tests */
2229 if (vma->vm_flags & VM_LOCKED) {
2230 unsigned long locked;
2231 unsigned long limit;
2232 locked = mm->locked_vm + grow;
2233 limit = READ_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
2234 limit >>= PAGE_SHIFT;
2235 if (locked > limit && !capable(CAP_IPC_LOCK))
2236 return -ENOMEM;
2237 }
2238
2239 /* Check to ensure the stack will not grow into a hugetlb-only region */
2240 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2241 vma->vm_end - size;
2242 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2243 return -EFAULT;
2244
2245 /*
2246 * Overcommit.. This must be the final test, as it will
2247 * update security statistics.
2248 */
2249 if (security_vm_enough_memory_mm(mm, grow))
2250 return -ENOMEM;
2251
2252 return 0;
2253}
2254
2255#if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2256/*
2257 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2258 * vma is the last one with address > vma->vm_end. Have to extend vma.
2259 */
2260int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2261{
2262 struct mm_struct *mm = vma->vm_mm;
2263 struct vm_area_struct *next;
2264 unsigned long gap_addr;
2265 int error = 0;
2266
2267 if (!(vma->vm_flags & VM_GROWSUP))
2268 return -EFAULT;
2269
2270 /* Guard against exceeding limits of the address space. */
2271 address &= PAGE_MASK;
2272 if (address >= (TASK_SIZE & PAGE_MASK))
2273 return -ENOMEM;
2274 address += PAGE_SIZE;
2275
2276 /* Enforce stack_guard_gap */
2277 gap_addr = address + stack_guard_gap;
2278
2279 /* Guard against overflow */
2280 if (gap_addr < address || gap_addr > TASK_SIZE)
2281 gap_addr = TASK_SIZE;
2282
2283 next = vma->vm_next;
2284 if (next && next->vm_start < gap_addr &&
2285 (next->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2286 if (!(next->vm_flags & VM_GROWSUP))
2287 return -ENOMEM;
2288 /* Check that both stack segments have the same anon_vma? */
2289 }
2290
2291 /* We must make sure the anon_vma is allocated. */
2292 if (unlikely(anon_vma_prepare(vma)))
2293 return -ENOMEM;
2294
2295 /*
2296 * vma->vm_start/vm_end cannot change under us because the caller
2297 * is required to hold the mmap_sem in read mode. We need the
2298 * anon_vma lock to serialize against concurrent expand_stacks.
2299 */
2300 anon_vma_lock_write(vma->anon_vma);
2301
2302 /* Somebody else might have raced and expanded it already */
2303 if (address > vma->vm_end) {
2304 unsigned long size, grow;
2305
2306 size = address - vma->vm_start;
2307 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2308
2309 error = -ENOMEM;
2310 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2311 error = acct_stack_growth(vma, size, grow);
2312 if (!error) {
2313 /*
2314 * vma_gap_update() doesn't support concurrent
2315 * updates, but we only hold a shared mmap_sem
2316 * lock here, so we need to protect against
2317 * concurrent vma expansions.
2318 * anon_vma_lock_write() doesn't help here, as
2319 * we don't guarantee that all growable vmas
2320 * in a mm share the same root anon vma.
2321 * So, we reuse mm->page_table_lock to guard
2322 * against concurrent vma expansions.
2323 */
2324 spin_lock(&mm->page_table_lock);
2325 if (vma->vm_flags & VM_LOCKED)
2326 mm->locked_vm += grow;
2327 vm_stat_account(mm, vma->vm_flags, grow);
2328 anon_vma_interval_tree_pre_update_vma(vma);
2329 vma->vm_end = address;
2330 anon_vma_interval_tree_post_update_vma(vma);
2331 if (vma->vm_next)
2332 vma_gap_update(vma->vm_next);
2333 else
2334 mm->highest_vm_end = vm_end_gap(vma);
2335 spin_unlock(&mm->page_table_lock);
2336
2337 perf_event_mmap(vma);
2338 }
2339 }
2340 }
2341 anon_vma_unlock_write(vma->anon_vma);
2342 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2343 validate_mm(mm);
2344 return error;
2345}
2346#endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2347
2348/*
2349 * vma is the first one with address < vma->vm_start. Have to extend vma.
2350 */
2351int expand_downwards(struct vm_area_struct *vma,
2352 unsigned long address)
2353{
2354 struct mm_struct *mm = vma->vm_mm;
2355 struct vm_area_struct *prev;
2356 unsigned long gap_addr;
2357 int error = 0;
2358
2359 address &= PAGE_MASK;
2360 if (address < mmap_min_addr)
2361 return -EPERM;
2362
2363 /* Enforce stack_guard_gap */
2364 gap_addr = address - stack_guard_gap;
2365 if (gap_addr > address)
2366 return -ENOMEM;
2367 prev = vma->vm_prev;
2368 if (prev && prev->vm_end > gap_addr &&
2369 (prev->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2370 if (!(prev->vm_flags & VM_GROWSDOWN))
2371 return -ENOMEM;
2372 /* Check that both stack segments have the same anon_vma? */
2373 }
2374
2375 /* We must make sure the anon_vma is allocated. */
2376 if (unlikely(anon_vma_prepare(vma)))
2377 return -ENOMEM;
2378
2379 /*
2380 * vma->vm_start/vm_end cannot change under us because the caller
2381 * is required to hold the mmap_sem in read mode. We need the
2382 * anon_vma lock to serialize against concurrent expand_stacks.
2383 */
2384 anon_vma_lock_write(vma->anon_vma);
2385
2386 /* Somebody else might have raced and expanded it already */
2387 if (address < vma->vm_start) {
2388 unsigned long size, grow;
2389
2390 size = vma->vm_end - address;
2391 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2392
2393 error = -ENOMEM;
2394 if (grow <= vma->vm_pgoff) {
2395 error = acct_stack_growth(vma, size, grow);
2396 if (!error) {
2397 /*
2398 * vma_gap_update() doesn't support concurrent
2399 * updates, but we only hold a shared mmap_sem
2400 * lock here, so we need to protect against
2401 * concurrent vma expansions.
2402 * anon_vma_lock_write() doesn't help here, as
2403 * we don't guarantee that all growable vmas
2404 * in a mm share the same root anon vma.
2405 * So, we reuse mm->page_table_lock to guard
2406 * against concurrent vma expansions.
2407 */
2408 spin_lock(&mm->page_table_lock);
2409 if (vma->vm_flags & VM_LOCKED)
2410 mm->locked_vm += grow;
2411 vm_stat_account(mm, vma->vm_flags, grow);
2412 anon_vma_interval_tree_pre_update_vma(vma);
2413 vma->vm_start = address;
2414 vma->vm_pgoff -= grow;
2415 anon_vma_interval_tree_post_update_vma(vma);
2416 vma_gap_update(vma);
2417 spin_unlock(&mm->page_table_lock);
2418
2419 perf_event_mmap(vma);
2420 }
2421 }
2422 }
2423 anon_vma_unlock_write(vma->anon_vma);
2424 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2425 validate_mm(mm);
2426 return error;
2427}
2428
2429/* enforced gap between the expanding stack and other mappings. */
2430unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2431
2432static int __init cmdline_parse_stack_guard_gap(char *p)
2433{
2434 unsigned long val;
2435 char *endptr;
2436
2437 val = simple_strtoul(p, &endptr, 10);
2438 if (!*endptr)
2439 stack_guard_gap = val << PAGE_SHIFT;
2440
2441 return 0;
2442}
2443__setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2444
2445#ifdef CONFIG_STACK_GROWSUP
2446int expand_stack(struct vm_area_struct *vma, unsigned long address)
2447{
2448 return expand_upwards(vma, address);
2449}
2450
2451struct vm_area_struct *
2452find_extend_vma(struct mm_struct *mm, unsigned long addr)
2453{
2454 struct vm_area_struct *vma, *prev;
2455
2456 addr &= PAGE_MASK;
2457 vma = find_vma_prev(mm, addr, &prev);
2458 if (vma && (vma->vm_start <= addr))
2459 return vma;
2460 if (!prev || expand_stack(prev, addr))
2461 return NULL;
2462 if (prev->vm_flags & VM_LOCKED)
2463 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2464 return prev;
2465}
2466#else
2467int expand_stack(struct vm_area_struct *vma, unsigned long address)
2468{
2469 return expand_downwards(vma, address);
2470}
2471
2472struct vm_area_struct *
2473find_extend_vma(struct mm_struct *mm, unsigned long addr)
2474{
2475 struct vm_area_struct *vma;
2476 unsigned long start;
2477
2478 addr &= PAGE_MASK;
2479 vma = find_vma(mm, addr);
2480 if (!vma)
2481 return NULL;
2482 if (vma->vm_start <= addr)
2483 return vma;
2484 if (!(vma->vm_flags & VM_GROWSDOWN))
2485 return NULL;
2486 start = vma->vm_start;
2487 if (expand_stack(vma, addr))
2488 return NULL;
2489 if (vma->vm_flags & VM_LOCKED)
2490 populate_vma_page_range(vma, addr, start, NULL);
2491 return vma;
2492}
2493#endif
2494
2495EXPORT_SYMBOL_GPL(find_extend_vma);
2496
2497/*
2498 * Ok - we have the memory areas we should free on the vma list,
2499 * so release them, and do the vma updates.
2500 *
2501 * Called with the mm semaphore held.
2502 */
2503static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2504{
2505 unsigned long nr_accounted = 0;
2506
2507 /* Update high watermark before we lower total_vm */
2508 update_hiwater_vm(mm);
2509 do {
2510 long nrpages = vma_pages(vma);
2511
2512 if (vma->vm_flags & VM_ACCOUNT)
2513 nr_accounted += nrpages;
2514 vm_stat_account(mm, vma->vm_flags, -nrpages);
2515 vma = remove_vma(vma);
2516 } while (vma);
2517 vm_unacct_memory(nr_accounted);
2518 validate_mm(mm);
2519}
2520
2521/*
2522 * Get rid of page table information in the indicated region.
2523 *
2524 * Called with the mm semaphore held.
2525 */
2526static void unmap_region(struct mm_struct *mm,
2527 struct vm_area_struct *vma, struct vm_area_struct *prev,
2528 unsigned long start, unsigned long end)
2529{
2530 struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap;
2531 struct mmu_gather tlb;
2532
2533 lru_add_drain();
2534 tlb_gather_mmu(&tlb, mm, start, end);
2535 update_hiwater_rss(mm);
2536 unmap_vmas(&tlb, vma, start, end);
2537 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2538 next ? next->vm_start : USER_PGTABLES_CEILING);
2539 tlb_finish_mmu(&tlb, start, end);
2540}
2541
2542/*
2543 * Create a list of vma's touched by the unmap, removing them from the mm's
2544 * vma list as we go..
2545 */
2546static void
2547detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2548 struct vm_area_struct *prev, unsigned long end)
2549{
2550 struct vm_area_struct **insertion_point;
2551 struct vm_area_struct *tail_vma = NULL;
2552
2553 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2554 vma->vm_prev = NULL;
2555 do {
2556 vma_rb_erase(vma, &mm->mm_rb);
2557 mm->map_count--;
2558 tail_vma = vma;
2559 vma = vma->vm_next;
2560 } while (vma && vma->vm_start < end);
2561 *insertion_point = vma;
2562 if (vma) {
2563 vma->vm_prev = prev;
2564 vma_gap_update(vma);
2565 } else
2566 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2567 tail_vma->vm_next = NULL;
2568
2569 /* Kill the cache */
2570 vmacache_invalidate(mm);
2571}
2572
2573/*
2574 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
2575 * munmap path where it doesn't make sense to fail.
2576 */
2577static int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2578 unsigned long addr, int new_below)
2579{
2580 struct vm_area_struct *new;
2581 int err;
2582
2583 if (vma->vm_ops && vma->vm_ops->split) {
2584 err = vma->vm_ops->split(vma, addr);
2585 if (err)
2586 return err;
2587 }
2588
2589 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2590 if (!new)
2591 return -ENOMEM;
2592
2593 /* most fields are the same, copy all, and then fixup */
2594 *new = *vma;
2595
2596 INIT_LIST_HEAD(&new->anon_vma_chain);
2597
2598 if (new_below)
2599 new->vm_end = addr;
2600 else {
2601 new->vm_start = addr;
2602 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2603 }
2604
2605 err = vma_dup_policy(vma, new);
2606 if (err)
2607 goto out_free_vma;
2608
2609 err = anon_vma_clone(new, vma);
2610 if (err)
2611 goto out_free_mpol;
2612
2613 if (new->vm_file)
2614 get_file(new->vm_file);
2615
2616 if (new->vm_ops && new->vm_ops->open)
2617 new->vm_ops->open(new);
2618
2619 if (new_below)
2620 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2621 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2622 else
2623 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2624
2625 /* Success. */
2626 if (!err)
2627 return 0;
2628
2629 /* Clean everything up if vma_adjust failed. */
2630 if (new->vm_ops && new->vm_ops->close)
2631 new->vm_ops->close(new);
2632 if (new->vm_file)
2633 fput(new->vm_file);
2634 unlink_anon_vmas(new);
2635 out_free_mpol:
2636 mpol_put(vma_policy(new));
2637 out_free_vma:
2638 kmem_cache_free(vm_area_cachep, new);
2639 return err;
2640}
2641
2642/*
2643 * Split a vma into two pieces at address 'addr', a new vma is allocated
2644 * either for the first part or the tail.
2645 */
2646int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2647 unsigned long addr, int new_below)
2648{
2649 if (mm->map_count >= sysctl_max_map_count)
2650 return -ENOMEM;
2651
2652 return __split_vma(mm, vma, addr, new_below);
2653}
2654
2655/* Munmap is split into 2 main parts -- this part which finds
2656 * what needs doing, and the areas themselves, which do the
2657 * work. This now handles partial unmappings.
2658 * Jeremy Fitzhardinge <jeremy@goop.org>
2659 */
2660int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2661{
2662 unsigned long end;
2663 struct vm_area_struct *vma, *prev, *last;
2664
2665 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2666 return -EINVAL;
2667
2668 len = PAGE_ALIGN(len);
2669 if (len == 0)
2670 return -EINVAL;
2671
2672 /* Find the first overlapping VMA */
2673 vma = find_vma(mm, start);
2674 if (!vma)
2675 return 0;
2676 prev = vma->vm_prev;
2677 /* we have start < vma->vm_end */
2678
2679 /* if it doesn't overlap, we have nothing.. */
2680 end = start + len;
2681 if (vma->vm_start >= end)
2682 return 0;
2683
2684 /*
2685 * If we need to split any vma, do it now to save pain later.
2686 *
2687 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2688 * unmapped vm_area_struct will remain in use: so lower split_vma
2689 * places tmp vma above, and higher split_vma places tmp vma below.
2690 */
2691 if (start > vma->vm_start) {
2692 int error;
2693
2694 /*
2695 * Make sure that map_count on return from munmap() will
2696 * not exceed its limit; but let map_count go just above
2697 * its limit temporarily, to help free resources as expected.
2698 */
2699 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2700 return -ENOMEM;
2701
2702 error = __split_vma(mm, vma, start, 0);
2703 if (error)
2704 return error;
2705 prev = vma;
2706 }
2707
2708 /* Does it split the last one? */
2709 last = find_vma(mm, end);
2710 if (last && end > last->vm_start) {
2711 int error = __split_vma(mm, last, end, 1);
2712 if (error)
2713 return error;
2714 }
2715 vma = prev ? prev->vm_next : mm->mmap;
2716
2717 /*
2718 * unlock any mlock()ed ranges before detaching vmas
2719 */
2720 if (mm->locked_vm) {
2721 struct vm_area_struct *tmp = vma;
2722 while (tmp && tmp->vm_start < end) {
2723 if (tmp->vm_flags & VM_LOCKED) {
2724 mm->locked_vm -= vma_pages(tmp);
2725 munlock_vma_pages_all(tmp);
2726 }
2727 tmp = tmp->vm_next;
2728 }
2729 }
2730
2731 /*
2732 * Remove the vma's, and unmap the actual pages
2733 */
2734 detach_vmas_to_be_unmapped(mm, vma, prev, end);
2735 unmap_region(mm, vma, prev, start, end);
2736
2737 arch_unmap(mm, vma, start, end);
2738
2739 /* Fix up all other VM information */
2740 remove_vma_list(mm, vma);
2741
2742 return 0;
2743}
2744EXPORT_SYMBOL(do_munmap);
2745
2746int vm_munmap(unsigned long start, size_t len)
2747{
2748 int ret;
2749 struct mm_struct *mm = current->mm;
2750
2751 if (down_write_killable(&mm->mmap_sem))
2752 return -EINTR;
2753
2754 ret = do_munmap(mm, start, len);
2755 up_write(&mm->mmap_sem);
2756 return ret;
2757}
2758EXPORT_SYMBOL(vm_munmap);
2759
2760SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2761{
2762 int ret;
2763 struct mm_struct *mm = current->mm;
2764
2765 profile_munmap(addr);
2766 if (down_write_killable(&mm->mmap_sem))
2767 return -EINTR;
2768 ret = do_munmap(mm, addr, len);
2769 up_write(&mm->mmap_sem);
2770 return ret;
2771}
2772
2773
2774/*
2775 * Emulation of deprecated remap_file_pages() syscall.
2776 */
2777SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2778 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2779{
2780
2781 struct mm_struct *mm = current->mm;
2782 struct vm_area_struct *vma;
2783 unsigned long populate = 0;
2784 unsigned long ret = -EINVAL;
2785 struct file *file;
2786
2787 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.txt.\n",
2788 current->comm, current->pid);
2789
2790 if (prot)
2791 return ret;
2792 start = start & PAGE_MASK;
2793 size = size & PAGE_MASK;
2794
2795 if (start + size <= start)
2796 return ret;
2797
2798 /* Does pgoff wrap? */
2799 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2800 return ret;
2801
2802 if (down_write_killable(&mm->mmap_sem))
2803 return -EINTR;
2804
2805 vma = find_vma(mm, start);
2806
2807 if (!vma || !(vma->vm_flags & VM_SHARED))
2808 goto out;
2809
2810 if (start < vma->vm_start)
2811 goto out;
2812
2813 if (start + size > vma->vm_end) {
2814 struct vm_area_struct *next;
2815
2816 for (next = vma->vm_next; next; next = next->vm_next) {
2817 /* hole between vmas ? */
2818 if (next->vm_start != next->vm_prev->vm_end)
2819 goto out;
2820
2821 if (next->vm_file != vma->vm_file)
2822 goto out;
2823
2824 if (next->vm_flags != vma->vm_flags)
2825 goto out;
2826
2827 if (start + size <= next->vm_end)
2828 break;
2829 }
2830
2831 if (!next)
2832 goto out;
2833 }
2834
2835 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2836 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2837 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2838
2839 flags &= MAP_NONBLOCK;
2840 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2841 if (vma->vm_flags & VM_LOCKED) {
2842 struct vm_area_struct *tmp;
2843 flags |= MAP_LOCKED;
2844
2845 /* drop PG_Mlocked flag for over-mapped range */
2846 for (tmp = vma; tmp->vm_start >= start + size;
2847 tmp = tmp->vm_next) {
2848 /*
2849 * Split pmd and munlock page on the border
2850 * of the range.
2851 */
2852 vma_adjust_trans_huge(tmp, start, start + size, 0);
2853
2854 munlock_vma_pages_range(tmp,
2855 max(tmp->vm_start, start),
2856 min(tmp->vm_end, start + size));
2857 }
2858 }
2859
2860 file = get_file(vma->vm_file);
2861 ret = do_mmap_pgoff(vma->vm_file, start, size,
2862 prot, flags, pgoff, &populate);
2863 fput(file);
2864out:
2865 up_write(&mm->mmap_sem);
2866 if (populate)
2867 mm_populate(ret, populate);
2868 if (!IS_ERR_VALUE(ret))
2869 ret = 0;
2870 return ret;
2871}
2872
2873static inline void verify_mm_writelocked(struct mm_struct *mm)
2874{
2875#ifdef CONFIG_DEBUG_VM
2876 if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2877 WARN_ON(1);
2878 up_read(&mm->mmap_sem);
2879 }
2880#endif
2881}
2882
2883/*
2884 * this is really a simplified "do_mmap". it only handles
2885 * anonymous maps. eventually we may be able to do some
2886 * brk-specific accounting here.
2887 */
2888static int do_brk(unsigned long addr, unsigned long len)
2889{
2890 struct mm_struct *mm = current->mm;
2891 struct vm_area_struct *vma, *prev;
2892 unsigned long flags;
2893 struct rb_node **rb_link, *rb_parent;
2894 pgoff_t pgoff = addr >> PAGE_SHIFT;
2895 int error;
2896
2897 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2898
2899 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2900 if (offset_in_page(error))
2901 return error;
2902
2903 error = mlock_future_check(mm, mm->def_flags, len);
2904 if (error)
2905 return error;
2906
2907 /*
2908 * mm->mmap_sem is required to protect against another thread
2909 * changing the mappings in case we sleep.
2910 */
2911 verify_mm_writelocked(mm);
2912
2913 /*
2914 * Clear old maps. this also does some error checking for us
2915 */
2916 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
2917 &rb_parent)) {
2918 if (do_munmap(mm, addr, len))
2919 return -ENOMEM;
2920 }
2921
2922 /* Check against address space limits *after* clearing old maps... */
2923 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
2924 return -ENOMEM;
2925
2926 if (mm->map_count > sysctl_max_map_count)
2927 return -ENOMEM;
2928
2929 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2930 return -ENOMEM;
2931
2932 /* Can we just expand an old private anonymous mapping? */
2933 vma = vma_merge(mm, prev, addr, addr + len, flags,
2934 NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX, NULL);
2935 if (vma)
2936 goto out;
2937
2938 /*
2939 * create a vma struct for an anonymous mapping
2940 */
2941 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2942 if (!vma) {
2943 vm_unacct_memory(len >> PAGE_SHIFT);
2944 return -ENOMEM;
2945 }
2946
2947 INIT_LIST_HEAD(&vma->anon_vma_chain);
2948 vma->vm_mm = mm;
2949 vma->vm_start = addr;
2950 vma->vm_end = addr + len;
2951 vma->vm_pgoff = pgoff;
2952 vma->vm_flags = flags;
2953 vma->vm_page_prot = vm_get_page_prot(flags);
2954 vma_link(mm, vma, prev, rb_link, rb_parent);
2955out:
2956 perf_event_mmap(vma);
2957 mm->total_vm += len >> PAGE_SHIFT;
2958 mm->data_vm += len >> PAGE_SHIFT;
2959 if (flags & VM_LOCKED)
2960 mm->locked_vm += (len >> PAGE_SHIFT);
2961 vma->vm_flags |= VM_SOFTDIRTY;
2962 return 0;
2963}
2964
2965int vm_brk(unsigned long addr, unsigned long request)
2966{
2967 struct mm_struct *mm = current->mm;
2968 unsigned long len;
2969 int ret;
2970 bool populate;
2971
2972 len = PAGE_ALIGN(request);
2973 if (len < request)
2974 return -ENOMEM;
2975 if (!len)
2976 return 0;
2977
2978 if (down_write_killable(&mm->mmap_sem))
2979 return -EINTR;
2980
2981 ret = do_brk(addr, len);
2982 populate = ((mm->def_flags & VM_LOCKED) != 0);
2983 up_write(&mm->mmap_sem);
2984 if (populate && !ret)
2985 mm_populate(addr, len);
2986 return ret;
2987}
2988EXPORT_SYMBOL(vm_brk);
2989
2990/* Release all mmaps. */
2991void exit_mmap(struct mm_struct *mm)
2992{
2993 struct mmu_gather tlb;
2994 struct vm_area_struct *vma;
2995 unsigned long nr_accounted = 0;
2996
2997 /* mm's last user has gone, and its about to be pulled down */
2998 mmu_notifier_release(mm);
2999
3000 if (mm->locked_vm) {
3001 vma = mm->mmap;
3002 while (vma) {
3003 if (vma->vm_flags & VM_LOCKED)
3004 munlock_vma_pages_all(vma);
3005 vma = vma->vm_next;
3006 }
3007 }
3008
3009 arch_exit_mmap(mm);
3010
3011 vma = mm->mmap;
3012 if (!vma) /* Can happen if dup_mmap() received an OOM */
3013 return;
3014
3015 lru_add_drain();
3016 flush_cache_mm(mm);
3017 tlb_gather_mmu(&tlb, mm, 0, -1);
3018 /* update_hiwater_rss(mm) here? but nobody should be looking */
3019 /* Use -1 here to ensure all VMAs in the mm are unmapped */
3020 unmap_vmas(&tlb, vma, 0, -1);
3021
3022 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3023 tlb_finish_mmu(&tlb, 0, -1);
3024
3025 /*
3026 * Walk the list again, actually closing and freeing it,
3027 * with preemption enabled, without holding any MM locks.
3028 */
3029 while (vma) {
3030 if (vma->vm_flags & VM_ACCOUNT)
3031 nr_accounted += vma_pages(vma);
3032 vma = remove_vma(vma);
3033 }
3034 vm_unacct_memory(nr_accounted);
3035}
3036
3037/* Insert vm structure into process list sorted by address
3038 * and into the inode's i_mmap tree. If vm_file is non-NULL
3039 * then i_mmap_rwsem is taken here.
3040 */
3041int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3042{
3043 struct vm_area_struct *prev;
3044 struct rb_node **rb_link, *rb_parent;
3045
3046 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3047 &prev, &rb_link, &rb_parent))
3048 return -ENOMEM;
3049 if ((vma->vm_flags & VM_ACCOUNT) &&
3050 security_vm_enough_memory_mm(mm, vma_pages(vma)))
3051 return -ENOMEM;
3052
3053 /*
3054 * The vm_pgoff of a purely anonymous vma should be irrelevant
3055 * until its first write fault, when page's anon_vma and index
3056 * are set. But now set the vm_pgoff it will almost certainly
3057 * end up with (unless mremap moves it elsewhere before that
3058 * first wfault), so /proc/pid/maps tells a consistent story.
3059 *
3060 * By setting it to reflect the virtual start address of the
3061 * vma, merges and splits can happen in a seamless way, just
3062 * using the existing file pgoff checks and manipulations.
3063 * Similarly in do_mmap_pgoff and in do_brk.
3064 */
3065 if (vma_is_anonymous(vma)) {
3066 BUG_ON(vma->anon_vma);
3067 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3068 }
3069
3070 vma_link(mm, vma, prev, rb_link, rb_parent);
3071 return 0;
3072}
3073
3074/*
3075 * Copy the vma structure to a new location in the same mm,
3076 * prior to moving page table entries, to effect an mremap move.
3077 */
3078struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3079 unsigned long addr, unsigned long len, pgoff_t pgoff,
3080 bool *need_rmap_locks)
3081{
3082 struct vm_area_struct *vma = *vmap;
3083 unsigned long vma_start = vma->vm_start;
3084 struct mm_struct *mm = vma->vm_mm;
3085 struct vm_area_struct *new_vma, *prev;
3086 struct rb_node **rb_link, *rb_parent;
3087 bool faulted_in_anon_vma = true;
3088
3089 /*
3090 * If anonymous vma has not yet been faulted, update new pgoff
3091 * to match new location, to increase its chance of merging.
3092 */
3093 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3094 pgoff = addr >> PAGE_SHIFT;
3095 faulted_in_anon_vma = false;
3096 }
3097
3098 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3099 return NULL; /* should never get here */
3100 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3101 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3102 vma->vm_userfaultfd_ctx, vma_get_anon_name(vma));
3103 if (new_vma) {
3104 /*
3105 * Source vma may have been merged into new_vma
3106 */
3107 if (unlikely(vma_start >= new_vma->vm_start &&
3108 vma_start < new_vma->vm_end)) {
3109 /*
3110 * The only way we can get a vma_merge with
3111 * self during an mremap is if the vma hasn't
3112 * been faulted in yet and we were allowed to
3113 * reset the dst vma->vm_pgoff to the
3114 * destination address of the mremap to allow
3115 * the merge to happen. mremap must change the
3116 * vm_pgoff linearity between src and dst vmas
3117 * (in turn preventing a vma_merge) to be
3118 * safe. It is only safe to keep the vm_pgoff
3119 * linear if there are no pages mapped yet.
3120 */
3121 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3122 *vmap = vma = new_vma;
3123 }
3124 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3125 } else {
3126 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
3127 if (!new_vma)
3128 goto out;
3129 *new_vma = *vma;
3130 new_vma->vm_start = addr;
3131 new_vma->vm_end = addr + len;
3132 new_vma->vm_pgoff = pgoff;
3133 if (vma_dup_policy(vma, new_vma))
3134 goto out_free_vma;
3135 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
3136 if (anon_vma_clone(new_vma, vma))
3137 goto out_free_mempol;
3138 if (new_vma->vm_file)
3139 get_file(new_vma->vm_file);
3140 if (new_vma->vm_ops && new_vma->vm_ops->open)
3141 new_vma->vm_ops->open(new_vma);
3142 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3143 *need_rmap_locks = false;
3144 }
3145 return new_vma;
3146
3147out_free_mempol:
3148 mpol_put(vma_policy(new_vma));
3149out_free_vma:
3150 kmem_cache_free(vm_area_cachep, new_vma);
3151out:
3152 return NULL;
3153}
3154
3155/*
3156 * Return true if the calling process may expand its vm space by the passed
3157 * number of pages
3158 */
3159bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3160{
3161 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3162 return false;
3163
3164 if (is_data_mapping(flags) &&
3165 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3166 /* Workaround for Valgrind */
3167 if (rlimit(RLIMIT_DATA) == 0 &&
3168 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3169 return true;
3170 if (!ignore_rlimit_data) {
3171 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits or use boot option ignore_rlimit_data.\n",
3172 current->comm, current->pid,
3173 (mm->data_vm + npages) << PAGE_SHIFT,
3174 rlimit(RLIMIT_DATA));
3175 return false;
3176 }
3177 }
3178
3179 return true;
3180}
3181
3182void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3183{
3184 mm->total_vm += npages;
3185
3186 if (is_exec_mapping(flags))
3187 mm->exec_vm += npages;
3188 else if (is_stack_mapping(flags))
3189 mm->stack_vm += npages;
3190 else if (is_data_mapping(flags))
3191 mm->data_vm += npages;
3192}
3193
3194static int special_mapping_fault(struct vm_area_struct *vma,
3195 struct vm_fault *vmf);
3196
3197/*
3198 * Having a close hook prevents vma merging regardless of flags.
3199 */
3200static void special_mapping_close(struct vm_area_struct *vma)
3201{
3202}
3203
3204static const char *special_mapping_name(struct vm_area_struct *vma)
3205{
3206 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3207}
3208
3209static int special_mapping_mremap(struct vm_area_struct *new_vma)
3210{
3211 struct vm_special_mapping *sm = new_vma->vm_private_data;
3212
3213 if (sm->mremap)
3214 return sm->mremap(sm, new_vma);
3215 return 0;
3216}
3217
3218static const struct vm_operations_struct special_mapping_vmops = {
3219 .close = special_mapping_close,
3220 .fault = special_mapping_fault,
3221 .mremap = special_mapping_mremap,
3222 .name = special_mapping_name,
3223};
3224
3225static const struct vm_operations_struct legacy_special_mapping_vmops = {
3226 .close = special_mapping_close,
3227 .fault = special_mapping_fault,
3228};
3229
3230static int special_mapping_fault(struct vm_area_struct *vma,
3231 struct vm_fault *vmf)
3232{
3233 pgoff_t pgoff;
3234 struct page **pages;
3235
3236 if (vma->vm_ops == &legacy_special_mapping_vmops) {
3237 pages = vma->vm_private_data;
3238 } else {
3239 struct vm_special_mapping *sm = vma->vm_private_data;
3240
3241 if (sm->fault)
3242 return sm->fault(sm, vma, vmf);
3243
3244 pages = sm->pages;
3245 }
3246
3247 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3248 pgoff--;
3249
3250 if (*pages) {
3251 struct page *page = *pages;
3252 get_page(page);
3253 vmf->page = page;
3254 return 0;
3255 }
3256
3257 return VM_FAULT_SIGBUS;
3258}
3259
3260static struct vm_area_struct *__install_special_mapping(
3261 struct mm_struct *mm,
3262 unsigned long addr, unsigned long len,
3263 unsigned long vm_flags, void *priv,
3264 const struct vm_operations_struct *ops)
3265{
3266 int ret;
3267 struct vm_area_struct *vma;
3268
3269 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
3270 if (unlikely(vma == NULL))
3271 return ERR_PTR(-ENOMEM);
3272
3273 INIT_LIST_HEAD(&vma->anon_vma_chain);
3274 vma->vm_mm = mm;
3275 vma->vm_start = addr;
3276 vma->vm_end = addr + len;
3277
3278 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3279 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3280
3281 vma->vm_ops = ops;
3282 vma->vm_private_data = priv;
3283
3284 ret = insert_vm_struct(mm, vma);
3285 if (ret)
3286 goto out;
3287
3288 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3289
3290 perf_event_mmap(vma);
3291
3292 return vma;
3293
3294out:
3295 kmem_cache_free(vm_area_cachep, vma);
3296 return ERR_PTR(ret);
3297}
3298
3299bool vma_is_special_mapping(const struct vm_area_struct *vma,
3300 const struct vm_special_mapping *sm)
3301{
3302 return vma->vm_private_data == sm &&
3303 (vma->vm_ops == &special_mapping_vmops ||
3304 vma->vm_ops == &legacy_special_mapping_vmops);
3305}
3306
3307/*
3308 * Called with mm->mmap_sem held for writing.
3309 * Insert a new vma covering the given region, with the given flags.
3310 * Its pages are supplied by the given array of struct page *.
3311 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3312 * The region past the last page supplied will always produce SIGBUS.
3313 * The array pointer and the pages it points to are assumed to stay alive
3314 * for as long as this mapping might exist.
3315 */
3316struct vm_area_struct *_install_special_mapping(
3317 struct mm_struct *mm,
3318 unsigned long addr, unsigned long len,
3319 unsigned long vm_flags, const struct vm_special_mapping *spec)
3320{
3321 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3322 &special_mapping_vmops);
3323}
3324
3325int install_special_mapping(struct mm_struct *mm,
3326 unsigned long addr, unsigned long len,
3327 unsigned long vm_flags, struct page **pages)
3328{
3329 struct vm_area_struct *vma = __install_special_mapping(
3330 mm, addr, len, vm_flags, (void *)pages,
3331 &legacy_special_mapping_vmops);
3332
3333 return PTR_ERR_OR_ZERO(vma);
3334}
3335
3336static DEFINE_MUTEX(mm_all_locks_mutex);
3337
3338static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3339{
3340 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3341 /*
3342 * The LSB of head.next can't change from under us
3343 * because we hold the mm_all_locks_mutex.
3344 */
3345 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
3346 /*
3347 * We can safely modify head.next after taking the
3348 * anon_vma->root->rwsem. If some other vma in this mm shares
3349 * the same anon_vma we won't take it again.
3350 *
3351 * No need of atomic instructions here, head.next
3352 * can't change from under us thanks to the
3353 * anon_vma->root->rwsem.
3354 */
3355 if (__test_and_set_bit(0, (unsigned long *)
3356 &anon_vma->root->rb_root.rb_node))
3357 BUG();
3358 }
3359}
3360
3361static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3362{
3363 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3364 /*
3365 * AS_MM_ALL_LOCKS can't change from under us because
3366 * we hold the mm_all_locks_mutex.
3367 *
3368 * Operations on ->flags have to be atomic because
3369 * even if AS_MM_ALL_LOCKS is stable thanks to the
3370 * mm_all_locks_mutex, there may be other cpus
3371 * changing other bitflags in parallel to us.
3372 */
3373 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3374 BUG();
3375 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem);
3376 }
3377}
3378
3379/*
3380 * This operation locks against the VM for all pte/vma/mm related
3381 * operations that could ever happen on a certain mm. This includes
3382 * vmtruncate, try_to_unmap, and all page faults.
3383 *
3384 * The caller must take the mmap_sem in write mode before calling
3385 * mm_take_all_locks(). The caller isn't allowed to release the
3386 * mmap_sem until mm_drop_all_locks() returns.
3387 *
3388 * mmap_sem in write mode is required in order to block all operations
3389 * that could modify pagetables and free pages without need of
3390 * altering the vma layout. It's also needed in write mode to avoid new
3391 * anon_vmas to be associated with existing vmas.
3392 *
3393 * A single task can't take more than one mm_take_all_locks() in a row
3394 * or it would deadlock.
3395 *
3396 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3397 * mapping->flags avoid to take the same lock twice, if more than one
3398 * vma in this mm is backed by the same anon_vma or address_space.
3399 *
3400 * We take locks in following order, accordingly to comment at beginning
3401 * of mm/rmap.c:
3402 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3403 * hugetlb mapping);
3404 * - all i_mmap_rwsem locks;
3405 * - all anon_vma->rwseml
3406 *
3407 * We can take all locks within these types randomly because the VM code
3408 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3409 * mm_all_locks_mutex.
3410 *
3411 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3412 * that may have to take thousand of locks.
3413 *
3414 * mm_take_all_locks() can fail if it's interrupted by signals.
3415 */
3416int mm_take_all_locks(struct mm_struct *mm)
3417{
3418 struct vm_area_struct *vma;
3419 struct anon_vma_chain *avc;
3420
3421 BUG_ON(down_read_trylock(&mm->mmap_sem));
3422
3423 mutex_lock(&mm_all_locks_mutex);
3424
3425 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3426 if (signal_pending(current))
3427 goto out_unlock;
3428 if (vma->vm_file && vma->vm_file->f_mapping &&
3429 is_vm_hugetlb_page(vma))
3430 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3431 }
3432
3433 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3434 if (signal_pending(current))
3435 goto out_unlock;
3436 if (vma->vm_file && vma->vm_file->f_mapping &&
3437 !is_vm_hugetlb_page(vma))
3438 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3439 }
3440
3441 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3442 if (signal_pending(current))
3443 goto out_unlock;
3444 if (vma->anon_vma)
3445 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3446 vm_lock_anon_vma(mm, avc->anon_vma);
3447 }
3448
3449 return 0;
3450
3451out_unlock:
3452 mm_drop_all_locks(mm);
3453 return -EINTR;
3454}
3455
3456static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3457{
3458 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3459 /*
3460 * The LSB of head.next can't change to 0 from under
3461 * us because we hold the mm_all_locks_mutex.
3462 *
3463 * We must however clear the bitflag before unlocking
3464 * the vma so the users using the anon_vma->rb_root will
3465 * never see our bitflag.
3466 *
3467 * No need of atomic instructions here, head.next
3468 * can't change from under us until we release the
3469 * anon_vma->root->rwsem.
3470 */
3471 if (!__test_and_clear_bit(0, (unsigned long *)
3472 &anon_vma->root->rb_root.rb_node))
3473 BUG();
3474 anon_vma_unlock_write(anon_vma);
3475 }
3476}
3477
3478static void vm_unlock_mapping(struct address_space *mapping)
3479{
3480 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3481 /*
3482 * AS_MM_ALL_LOCKS can't change to 0 from under us
3483 * because we hold the mm_all_locks_mutex.
3484 */
3485 i_mmap_unlock_write(mapping);
3486 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3487 &mapping->flags))
3488 BUG();
3489 }
3490}
3491
3492/*
3493 * The mmap_sem cannot be released by the caller until
3494 * mm_drop_all_locks() returns.
3495 */
3496void mm_drop_all_locks(struct mm_struct *mm)
3497{
3498 struct vm_area_struct *vma;
3499 struct anon_vma_chain *avc;
3500
3501 BUG_ON(down_read_trylock(&mm->mmap_sem));
3502 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3503
3504 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3505 if (vma->anon_vma)
3506 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3507 vm_unlock_anon_vma(avc->anon_vma);
3508 if (vma->vm_file && vma->vm_file->f_mapping)
3509 vm_unlock_mapping(vma->vm_file->f_mapping);
3510 }
3511
3512 mutex_unlock(&mm_all_locks_mutex);
3513}
3514
3515/*
3516 * initialise the VMA slab
3517 */
3518void __init mmap_init(void)
3519{
3520 int ret;
3521
3522 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3523 VM_BUG_ON(ret);
3524}
3525
3526/*
3527 * Initialise sysctl_user_reserve_kbytes.
3528 *
3529 * This is intended to prevent a user from starting a single memory hogging
3530 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3531 * mode.
3532 *
3533 * The default value is min(3% of free memory, 128MB)
3534 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3535 */
3536static int init_user_reserve(void)
3537{
3538 unsigned long free_kbytes;
3539
3540 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3541
3542 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3543 return 0;
3544}
3545subsys_initcall(init_user_reserve);
3546
3547/*
3548 * Initialise sysctl_admin_reserve_kbytes.
3549 *
3550 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3551 * to log in and kill a memory hogging process.
3552 *
3553 * Systems with more than 256MB will reserve 8MB, enough to recover
3554 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3555 * only reserve 3% of free pages by default.
3556 */
3557static int init_admin_reserve(void)
3558{
3559 unsigned long free_kbytes;
3560
3561 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3562
3563 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3564 return 0;
3565}
3566subsys_initcall(init_admin_reserve);
3567
3568/*
3569 * Reinititalise user and admin reserves if memory is added or removed.
3570 *
3571 * The default user reserve max is 128MB, and the default max for the
3572 * admin reserve is 8MB. These are usually, but not always, enough to
3573 * enable recovery from a memory hogging process using login/sshd, a shell,
3574 * and tools like top. It may make sense to increase or even disable the
3575 * reserve depending on the existence of swap or variations in the recovery
3576 * tools. So, the admin may have changed them.
3577 *
3578 * If memory is added and the reserves have been eliminated or increased above
3579 * the default max, then we'll trust the admin.
3580 *
3581 * If memory is removed and there isn't enough free memory, then we
3582 * need to reset the reserves.
3583 *
3584 * Otherwise keep the reserve set by the admin.
3585 */
3586static int reserve_mem_notifier(struct notifier_block *nb,
3587 unsigned long action, void *data)
3588{
3589 unsigned long tmp, free_kbytes;
3590
3591 switch (action) {
3592 case MEM_ONLINE:
3593 /* Default max is 128MB. Leave alone if modified by operator. */
3594 tmp = sysctl_user_reserve_kbytes;
3595 if (0 < tmp && tmp < (1UL << 17))
3596 init_user_reserve();
3597
3598 /* Default max is 8MB. Leave alone if modified by operator. */
3599 tmp = sysctl_admin_reserve_kbytes;
3600 if (0 < tmp && tmp < (1UL << 13))
3601 init_admin_reserve();
3602
3603 break;
3604 case MEM_OFFLINE:
3605 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3606
3607 if (sysctl_user_reserve_kbytes > free_kbytes) {
3608 init_user_reserve();
3609 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3610 sysctl_user_reserve_kbytes);
3611 }
3612
3613 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3614 init_admin_reserve();
3615 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3616 sysctl_admin_reserve_kbytes);
3617 }
3618 break;
3619 default:
3620 break;
3621 }
3622 return NOTIFY_OK;
3623}
3624
3625static struct notifier_block reserve_mem_nb = {
3626 .notifier_call = reserve_mem_notifier,
3627};
3628
3629static int __meminit init_reserve_notifier(void)
3630{
3631 if (register_hotmemory_notifier(&reserve_mem_nb))
3632 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3633
3634 return 0;
3635}
3636subsys_initcall(init_reserve_notifier);
3637
generated by cgit at 2024-05-18 15:54:12 (GMT)