Merge tag 'libnvdimm-for-5.7' of git://git.kernel.org/pub/scm/linux/kernel/git/nvdimm...
[linux-2.6-microblaze.git] / mm / mmap.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * mm/mmap.c
4  *
5  * Written by obz.
6  *
7  * Address space accounting code        <alan@lxorguk.ukuu.org.uk>
8  */
9
10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11
12 #include <linux/kernel.h>
13 #include <linux/slab.h>
14 #include <linux/backing-dev.h>
15 #include <linux/mm.h>
16 #include <linux/vmacache.h>
17 #include <linux/shm.h>
18 #include <linux/mman.h>
19 #include <linux/pagemap.h>
20 #include <linux/swap.h>
21 #include <linux/syscalls.h>
22 #include <linux/capability.h>
23 #include <linux/init.h>
24 #include <linux/file.h>
25 #include <linux/fs.h>
26 #include <linux/personality.h>
27 #include <linux/security.h>
28 #include <linux/hugetlb.h>
29 #include <linux/shmem_fs.h>
30 #include <linux/profile.h>
31 #include <linux/export.h>
32 #include <linux/mount.h>
33 #include <linux/mempolicy.h>
34 #include <linux/rmap.h>
35 #include <linux/mmu_notifier.h>
36 #include <linux/mmdebug.h>
37 #include <linux/perf_event.h>
38 #include <linux/audit.h>
39 #include <linux/khugepaged.h>
40 #include <linux/uprobes.h>
41 #include <linux/rbtree_augmented.h>
42 #include <linux/notifier.h>
43 #include <linux/memory.h>
44 #include <linux/printk.h>
45 #include <linux/userfaultfd_k.h>
46 #include <linux/moduleparam.h>
47 #include <linux/pkeys.h>
48 #include <linux/oom.h>
49 #include <linux/sched/mm.h>
50
51 #include <linux/uaccess.h>
52 #include <asm/cacheflush.h>
53 #include <asm/tlb.h>
54 #include <asm/mmu_context.h>
55
56 #define CREATE_TRACE_POINTS
57 #include <trace/events/mmap.h>
58
59 #include "internal.h"
60
61 #ifndef arch_mmap_check
62 #define arch_mmap_check(addr, len, flags)       (0)
63 #endif
64
65 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
66 const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
67 const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
68 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
69 #endif
70 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
71 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
72 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
73 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
74 #endif
75
76 static bool ignore_rlimit_data;
77 core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
78
79 static void unmap_region(struct mm_struct *mm,
80                 struct vm_area_struct *vma, struct vm_area_struct *prev,
81                 unsigned long start, unsigned long end);
82
83 /* description of effects of mapping type and prot in current implementation.
84  * this is due to the limited x86 page protection hardware.  The expected
85  * behavior is in parens:
86  *
87  * map_type     prot
88  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
89  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
90  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
91  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
92  *
93  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
94  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
95  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
96  */
97 pgprot_t protection_map[16] __ro_after_init = {
98         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
99         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
100 };
101
102 #ifndef CONFIG_ARCH_HAS_FILTER_PGPROT
103 static inline pgprot_t arch_filter_pgprot(pgprot_t prot)
104 {
105         return prot;
106 }
107 #endif
108
109 pgprot_t vm_get_page_prot(unsigned long vm_flags)
110 {
111         pgprot_t ret = __pgprot(pgprot_val(protection_map[vm_flags &
112                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
113                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
114
115         return arch_filter_pgprot(ret);
116 }
117 EXPORT_SYMBOL(vm_get_page_prot);
118
119 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
120 {
121         return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
122 }
123
124 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
125 void vma_set_page_prot(struct vm_area_struct *vma)
126 {
127         unsigned long vm_flags = vma->vm_flags;
128         pgprot_t vm_page_prot;
129
130         vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
131         if (vma_wants_writenotify(vma, vm_page_prot)) {
132                 vm_flags &= ~VM_SHARED;
133                 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
134         }
135         /* remove_protection_ptes reads vma->vm_page_prot without mmap_sem */
136         WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
137 }
138
139 /*
140  * Requires inode->i_mapping->i_mmap_rwsem
141  */
142 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
143                 struct file *file, struct address_space *mapping)
144 {
145         if (vma->vm_flags & VM_DENYWRITE)
146                 atomic_inc(&file_inode(file)->i_writecount);
147         if (vma->vm_flags & VM_SHARED)
148                 mapping_unmap_writable(mapping);
149
150         flush_dcache_mmap_lock(mapping);
151         vma_interval_tree_remove(vma, &mapping->i_mmap);
152         flush_dcache_mmap_unlock(mapping);
153 }
154
155 /*
156  * Unlink a file-based vm structure from its interval tree, to hide
157  * vma from rmap and vmtruncate before freeing its page tables.
158  */
159 void unlink_file_vma(struct vm_area_struct *vma)
160 {
161         struct file *file = vma->vm_file;
162
163         if (file) {
164                 struct address_space *mapping = file->f_mapping;
165                 i_mmap_lock_write(mapping);
166                 __remove_shared_vm_struct(vma, file, mapping);
167                 i_mmap_unlock_write(mapping);
168         }
169 }
170
171 /*
172  * Close a vm structure and free it, returning the next.
173  */
174 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
175 {
176         struct vm_area_struct *next = vma->vm_next;
177
178         might_sleep();
179         if (vma->vm_ops && vma->vm_ops->close)
180                 vma->vm_ops->close(vma);
181         if (vma->vm_file)
182                 fput(vma->vm_file);
183         mpol_put(vma_policy(vma));
184         vm_area_free(vma);
185         return next;
186 }
187
188 static int do_brk_flags(unsigned long addr, unsigned long request, unsigned long flags,
189                 struct list_head *uf);
190 SYSCALL_DEFINE1(brk, unsigned long, brk)
191 {
192         unsigned long retval;
193         unsigned long newbrk, oldbrk, origbrk;
194         struct mm_struct *mm = current->mm;
195         struct vm_area_struct *next;
196         unsigned long min_brk;
197         bool populate;
198         bool downgraded = false;
199         LIST_HEAD(uf);
200
201         if (down_write_killable(&mm->mmap_sem))
202                 return -EINTR;
203
204         origbrk = mm->brk;
205
206 #ifdef CONFIG_COMPAT_BRK
207         /*
208          * CONFIG_COMPAT_BRK can still be overridden by setting
209          * randomize_va_space to 2, which will still cause mm->start_brk
210          * to be arbitrarily shifted
211          */
212         if (current->brk_randomized)
213                 min_brk = mm->start_brk;
214         else
215                 min_brk = mm->end_data;
216 #else
217         min_brk = mm->start_brk;
218 #endif
219         if (brk < min_brk)
220                 goto out;
221
222         /*
223          * Check against rlimit here. If this check is done later after the test
224          * of oldbrk with newbrk then it can escape the test and let the data
225          * segment grow beyond its set limit the in case where the limit is
226          * not page aligned -Ram Gupta
227          */
228         if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
229                               mm->end_data, mm->start_data))
230                 goto out;
231
232         newbrk = PAGE_ALIGN(brk);
233         oldbrk = PAGE_ALIGN(mm->brk);
234         if (oldbrk == newbrk) {
235                 mm->brk = brk;
236                 goto success;
237         }
238
239         /*
240          * Always allow shrinking brk.
241          * __do_munmap() may downgrade mmap_sem to read.
242          */
243         if (brk <= mm->brk) {
244                 int ret;
245
246                 /*
247                  * mm->brk must to be protected by write mmap_sem so update it
248                  * before downgrading mmap_sem. When __do_munmap() fails,
249                  * mm->brk will be restored from origbrk.
250                  */
251                 mm->brk = brk;
252                 ret = __do_munmap(mm, newbrk, oldbrk-newbrk, &uf, true);
253                 if (ret < 0) {
254                         mm->brk = origbrk;
255                         goto out;
256                 } else if (ret == 1) {
257                         downgraded = true;
258                 }
259                 goto success;
260         }
261
262         /* Check against existing mmap mappings. */
263         next = find_vma(mm, oldbrk);
264         if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
265                 goto out;
266
267         /* Ok, looks good - let it rip. */
268         if (do_brk_flags(oldbrk, newbrk-oldbrk, 0, &uf) < 0)
269                 goto out;
270         mm->brk = brk;
271
272 success:
273         populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
274         if (downgraded)
275                 up_read(&mm->mmap_sem);
276         else
277                 up_write(&mm->mmap_sem);
278         userfaultfd_unmap_complete(mm, &uf);
279         if (populate)
280                 mm_populate(oldbrk, newbrk - oldbrk);
281         return brk;
282
283 out:
284         retval = origbrk;
285         up_write(&mm->mmap_sem);
286         return retval;
287 }
288
289 static inline unsigned long vma_compute_gap(struct vm_area_struct *vma)
290 {
291         unsigned long gap, prev_end;
292
293         /*
294          * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
295          * allow two stack_guard_gaps between them here, and when choosing
296          * an unmapped area; whereas when expanding we only require one.
297          * That's a little inconsistent, but keeps the code here simpler.
298          */
299         gap = vm_start_gap(vma);
300         if (vma->vm_prev) {
301                 prev_end = vm_end_gap(vma->vm_prev);
302                 if (gap > prev_end)
303                         gap -= prev_end;
304                 else
305                         gap = 0;
306         }
307         return gap;
308 }
309
310 #ifdef CONFIG_DEBUG_VM_RB
311 static unsigned long vma_compute_subtree_gap(struct vm_area_struct *vma)
312 {
313         unsigned long max = vma_compute_gap(vma), subtree_gap;
314         if (vma->vm_rb.rb_left) {
315                 subtree_gap = rb_entry(vma->vm_rb.rb_left,
316                                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
317                 if (subtree_gap > max)
318                         max = subtree_gap;
319         }
320         if (vma->vm_rb.rb_right) {
321                 subtree_gap = rb_entry(vma->vm_rb.rb_right,
322                                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
323                 if (subtree_gap > max)
324                         max = subtree_gap;
325         }
326         return max;
327 }
328
329 static int browse_rb(struct mm_struct *mm)
330 {
331         struct rb_root *root = &mm->mm_rb;
332         int i = 0, j, bug = 0;
333         struct rb_node *nd, *pn = NULL;
334         unsigned long prev = 0, pend = 0;
335
336         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
337                 struct vm_area_struct *vma;
338                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
339                 if (vma->vm_start < prev) {
340                         pr_emerg("vm_start %lx < prev %lx\n",
341                                   vma->vm_start, prev);
342                         bug = 1;
343                 }
344                 if (vma->vm_start < pend) {
345                         pr_emerg("vm_start %lx < pend %lx\n",
346                                   vma->vm_start, pend);
347                         bug = 1;
348                 }
349                 if (vma->vm_start > vma->vm_end) {
350                         pr_emerg("vm_start %lx > vm_end %lx\n",
351                                   vma->vm_start, vma->vm_end);
352                         bug = 1;
353                 }
354                 spin_lock(&mm->page_table_lock);
355                 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
356                         pr_emerg("free gap %lx, correct %lx\n",
357                                vma->rb_subtree_gap,
358                                vma_compute_subtree_gap(vma));
359                         bug = 1;
360                 }
361                 spin_unlock(&mm->page_table_lock);
362                 i++;
363                 pn = nd;
364                 prev = vma->vm_start;
365                 pend = vma->vm_end;
366         }
367         j = 0;
368         for (nd = pn; nd; nd = rb_prev(nd))
369                 j++;
370         if (i != j) {
371                 pr_emerg("backwards %d, forwards %d\n", j, i);
372                 bug = 1;
373         }
374         return bug ? -1 : i;
375 }
376
377 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
378 {
379         struct rb_node *nd;
380
381         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
382                 struct vm_area_struct *vma;
383                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
384                 VM_BUG_ON_VMA(vma != ignore &&
385                         vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
386                         vma);
387         }
388 }
389
390 static void validate_mm(struct mm_struct *mm)
391 {
392         int bug = 0;
393         int i = 0;
394         unsigned long highest_address = 0;
395         struct vm_area_struct *vma = mm->mmap;
396
397         while (vma) {
398                 struct anon_vma *anon_vma = vma->anon_vma;
399                 struct anon_vma_chain *avc;
400
401                 if (anon_vma) {
402                         anon_vma_lock_read(anon_vma);
403                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
404                                 anon_vma_interval_tree_verify(avc);
405                         anon_vma_unlock_read(anon_vma);
406                 }
407
408                 highest_address = vm_end_gap(vma);
409                 vma = vma->vm_next;
410                 i++;
411         }
412         if (i != mm->map_count) {
413                 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
414                 bug = 1;
415         }
416         if (highest_address != mm->highest_vm_end) {
417                 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
418                           mm->highest_vm_end, highest_address);
419                 bug = 1;
420         }
421         i = browse_rb(mm);
422         if (i != mm->map_count) {
423                 if (i != -1)
424                         pr_emerg("map_count %d rb %d\n", mm->map_count, i);
425                 bug = 1;
426         }
427         VM_BUG_ON_MM(bug, mm);
428 }
429 #else
430 #define validate_mm_rb(root, ignore) do { } while (0)
431 #define validate_mm(mm) do { } while (0)
432 #endif
433
434 RB_DECLARE_CALLBACKS_MAX(static, vma_gap_callbacks,
435                          struct vm_area_struct, vm_rb,
436                          unsigned long, rb_subtree_gap, vma_compute_gap)
437
438 /*
439  * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
440  * vma->vm_prev->vm_end values changed, without modifying the vma's position
441  * in the rbtree.
442  */
443 static void vma_gap_update(struct vm_area_struct *vma)
444 {
445         /*
446          * As it turns out, RB_DECLARE_CALLBACKS_MAX() already created
447          * a callback function that does exactly what we want.
448          */
449         vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
450 }
451
452 static inline void vma_rb_insert(struct vm_area_struct *vma,
453                                  struct rb_root *root)
454 {
455         /* All rb_subtree_gap values must be consistent prior to insertion */
456         validate_mm_rb(root, NULL);
457
458         rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
459 }
460
461 static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
462 {
463         /*
464          * Note rb_erase_augmented is a fairly large inline function,
465          * so make sure we instantiate it only once with our desired
466          * augmented rbtree callbacks.
467          */
468         rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
469 }
470
471 static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
472                                                 struct rb_root *root,
473                                                 struct vm_area_struct *ignore)
474 {
475         /*
476          * All rb_subtree_gap values must be consistent prior to erase,
477          * with the possible exception of the "next" vma being erased if
478          * next->vm_start was reduced.
479          */
480         validate_mm_rb(root, ignore);
481
482         __vma_rb_erase(vma, root);
483 }
484
485 static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
486                                          struct rb_root *root)
487 {
488         /*
489          * All rb_subtree_gap values must be consistent prior to erase,
490          * with the possible exception of the vma being erased.
491          */
492         validate_mm_rb(root, vma);
493
494         __vma_rb_erase(vma, root);
495 }
496
497 /*
498  * vma has some anon_vma assigned, and is already inserted on that
499  * anon_vma's interval trees.
500  *
501  * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
502  * vma must be removed from the anon_vma's interval trees using
503  * anon_vma_interval_tree_pre_update_vma().
504  *
505  * After the update, the vma will be reinserted using
506  * anon_vma_interval_tree_post_update_vma().
507  *
508  * The entire update must be protected by exclusive mmap_sem and by
509  * the root anon_vma's mutex.
510  */
511 static inline void
512 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
513 {
514         struct anon_vma_chain *avc;
515
516         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
517                 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
518 }
519
520 static inline void
521 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
522 {
523         struct anon_vma_chain *avc;
524
525         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
526                 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
527 }
528
529 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
530                 unsigned long end, struct vm_area_struct **pprev,
531                 struct rb_node ***rb_link, struct rb_node **rb_parent)
532 {
533         struct rb_node **__rb_link, *__rb_parent, *rb_prev;
534
535         __rb_link = &mm->mm_rb.rb_node;
536         rb_prev = __rb_parent = NULL;
537
538         while (*__rb_link) {
539                 struct vm_area_struct *vma_tmp;
540
541                 __rb_parent = *__rb_link;
542                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
543
544                 if (vma_tmp->vm_end > addr) {
545                         /* Fail if an existing vma overlaps the area */
546                         if (vma_tmp->vm_start < end)
547                                 return -ENOMEM;
548                         __rb_link = &__rb_parent->rb_left;
549                 } else {
550                         rb_prev = __rb_parent;
551                         __rb_link = &__rb_parent->rb_right;
552                 }
553         }
554
555         *pprev = NULL;
556         if (rb_prev)
557                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
558         *rb_link = __rb_link;
559         *rb_parent = __rb_parent;
560         return 0;
561 }
562
563 static unsigned long count_vma_pages_range(struct mm_struct *mm,
564                 unsigned long addr, unsigned long end)
565 {
566         unsigned long nr_pages = 0;
567         struct vm_area_struct *vma;
568
569         /* Find first overlaping mapping */
570         vma = find_vma_intersection(mm, addr, end);
571         if (!vma)
572                 return 0;
573
574         nr_pages = (min(end, vma->vm_end) -
575                 max(addr, vma->vm_start)) >> PAGE_SHIFT;
576
577         /* Iterate over the rest of the overlaps */
578         for (vma = vma->vm_next; vma; vma = vma->vm_next) {
579                 unsigned long overlap_len;
580
581                 if (vma->vm_start > end)
582                         break;
583
584                 overlap_len = min(end, vma->vm_end) - vma->vm_start;
585                 nr_pages += overlap_len >> PAGE_SHIFT;
586         }
587
588         return nr_pages;
589 }
590
591 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
592                 struct rb_node **rb_link, struct rb_node *rb_parent)
593 {
594         /* Update tracking information for the gap following the new vma. */
595         if (vma->vm_next)
596                 vma_gap_update(vma->vm_next);
597         else
598                 mm->highest_vm_end = vm_end_gap(vma);
599
600         /*
601          * vma->vm_prev wasn't known when we followed the rbtree to find the
602          * correct insertion point for that vma. As a result, we could not
603          * update the vma vm_rb parents rb_subtree_gap values on the way down.
604          * So, we first insert the vma with a zero rb_subtree_gap value
605          * (to be consistent with what we did on the way down), and then
606          * immediately update the gap to the correct value. Finally we
607          * rebalance the rbtree after all augmented values have been set.
608          */
609         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
610         vma->rb_subtree_gap = 0;
611         vma_gap_update(vma);
612         vma_rb_insert(vma, &mm->mm_rb);
613 }
614
615 static void __vma_link_file(struct vm_area_struct *vma)
616 {
617         struct file *file;
618
619         file = vma->vm_file;
620         if (file) {
621                 struct address_space *mapping = file->f_mapping;
622
623                 if (vma->vm_flags & VM_DENYWRITE)
624                         atomic_dec(&file_inode(file)->i_writecount);
625                 if (vma->vm_flags & VM_SHARED)
626                         atomic_inc(&mapping->i_mmap_writable);
627
628                 flush_dcache_mmap_lock(mapping);
629                 vma_interval_tree_insert(vma, &mapping->i_mmap);
630                 flush_dcache_mmap_unlock(mapping);
631         }
632 }
633
634 static void
635 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
636         struct vm_area_struct *prev, struct rb_node **rb_link,
637         struct rb_node *rb_parent)
638 {
639         __vma_link_list(mm, vma, prev);
640         __vma_link_rb(mm, vma, rb_link, rb_parent);
641 }
642
643 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
644                         struct vm_area_struct *prev, struct rb_node **rb_link,
645                         struct rb_node *rb_parent)
646 {
647         struct address_space *mapping = NULL;
648
649         if (vma->vm_file) {
650                 mapping = vma->vm_file->f_mapping;
651                 i_mmap_lock_write(mapping);
652         }
653
654         __vma_link(mm, vma, prev, rb_link, rb_parent);
655         __vma_link_file(vma);
656
657         if (mapping)
658                 i_mmap_unlock_write(mapping);
659
660         mm->map_count++;
661         validate_mm(mm);
662 }
663
664 /*
665  * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
666  * mm's list and rbtree.  It has already been inserted into the interval tree.
667  */
668 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
669 {
670         struct vm_area_struct *prev;
671         struct rb_node **rb_link, *rb_parent;
672
673         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
674                            &prev, &rb_link, &rb_parent))
675                 BUG();
676         __vma_link(mm, vma, prev, rb_link, rb_parent);
677         mm->map_count++;
678 }
679
680 static __always_inline void __vma_unlink_common(struct mm_struct *mm,
681                                                 struct vm_area_struct *vma,
682                                                 struct vm_area_struct *ignore)
683 {
684         vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
685         __vma_unlink_list(mm, vma);
686         /* Kill the cache */
687         vmacache_invalidate(mm);
688 }
689
690 /*
691  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
692  * is already present in an i_mmap tree without adjusting the tree.
693  * The following helper function should be used when such adjustments
694  * are necessary.  The "insert" vma (if any) is to be inserted
695  * before we drop the necessary locks.
696  */
697 int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
698         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
699         struct vm_area_struct *expand)
700 {
701         struct mm_struct *mm = vma->vm_mm;
702         struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
703         struct address_space *mapping = NULL;
704         struct rb_root_cached *root = NULL;
705         struct anon_vma *anon_vma = NULL;
706         struct file *file = vma->vm_file;
707         bool start_changed = false, end_changed = false;
708         long adjust_next = 0;
709         int remove_next = 0;
710
711         if (next && !insert) {
712                 struct vm_area_struct *exporter = NULL, *importer = NULL;
713
714                 if (end >= next->vm_end) {
715                         /*
716                          * vma expands, overlapping all the next, and
717                          * perhaps the one after too (mprotect case 6).
718                          * The only other cases that gets here are
719                          * case 1, case 7 and case 8.
720                          */
721                         if (next == expand) {
722                                 /*
723                                  * The only case where we don't expand "vma"
724                                  * and we expand "next" instead is case 8.
725                                  */
726                                 VM_WARN_ON(end != next->vm_end);
727                                 /*
728                                  * remove_next == 3 means we're
729                                  * removing "vma" and that to do so we
730                                  * swapped "vma" and "next".
731                                  */
732                                 remove_next = 3;
733                                 VM_WARN_ON(file != next->vm_file);
734                                 swap(vma, next);
735                         } else {
736                                 VM_WARN_ON(expand != vma);
737                                 /*
738                                  * case 1, 6, 7, remove_next == 2 is case 6,
739                                  * remove_next == 1 is case 1 or 7.
740                                  */
741                                 remove_next = 1 + (end > next->vm_end);
742                                 VM_WARN_ON(remove_next == 2 &&
743                                            end != next->vm_next->vm_end);
744                                 /* trim end to next, for case 6 first pass */
745                                 end = next->vm_end;
746                         }
747
748                         exporter = next;
749                         importer = vma;
750
751                         /*
752                          * If next doesn't have anon_vma, import from vma after
753                          * next, if the vma overlaps with it.
754                          */
755                         if (remove_next == 2 && !next->anon_vma)
756                                 exporter = next->vm_next;
757
758                 } else if (end > next->vm_start) {
759                         /*
760                          * vma expands, overlapping part of the next:
761                          * mprotect case 5 shifting the boundary up.
762                          */
763                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
764                         exporter = next;
765                         importer = vma;
766                         VM_WARN_ON(expand != importer);
767                 } else if (end < vma->vm_end) {
768                         /*
769                          * vma shrinks, and !insert tells it's not
770                          * split_vma inserting another: so it must be
771                          * mprotect case 4 shifting the boundary down.
772                          */
773                         adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT);
774                         exporter = vma;
775                         importer = next;
776                         VM_WARN_ON(expand != importer);
777                 }
778
779                 /*
780                  * Easily overlooked: when mprotect shifts the boundary,
781                  * make sure the expanding vma has anon_vma set if the
782                  * shrinking vma had, to cover any anon pages imported.
783                  */
784                 if (exporter && exporter->anon_vma && !importer->anon_vma) {
785                         int error;
786
787                         importer->anon_vma = exporter->anon_vma;
788                         error = anon_vma_clone(importer, exporter);
789                         if (error)
790                                 return error;
791                 }
792         }
793 again:
794         vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
795
796         if (file) {
797                 mapping = file->f_mapping;
798                 root = &mapping->i_mmap;
799                 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
800
801                 if (adjust_next)
802                         uprobe_munmap(next, next->vm_start, next->vm_end);
803
804                 i_mmap_lock_write(mapping);
805                 if (insert) {
806                         /*
807                          * Put into interval tree now, so instantiated pages
808                          * are visible to arm/parisc __flush_dcache_page
809                          * throughout; but we cannot insert into address
810                          * space until vma start or end is updated.
811                          */
812                         __vma_link_file(insert);
813                 }
814         }
815
816         anon_vma = vma->anon_vma;
817         if (!anon_vma && adjust_next)
818                 anon_vma = next->anon_vma;
819         if (anon_vma) {
820                 VM_WARN_ON(adjust_next && next->anon_vma &&
821                            anon_vma != next->anon_vma);
822                 anon_vma_lock_write(anon_vma);
823                 anon_vma_interval_tree_pre_update_vma(vma);
824                 if (adjust_next)
825                         anon_vma_interval_tree_pre_update_vma(next);
826         }
827
828         if (root) {
829                 flush_dcache_mmap_lock(mapping);
830                 vma_interval_tree_remove(vma, root);
831                 if (adjust_next)
832                         vma_interval_tree_remove(next, root);
833         }
834
835         if (start != vma->vm_start) {
836                 vma->vm_start = start;
837                 start_changed = true;
838         }
839         if (end != vma->vm_end) {
840                 vma->vm_end = end;
841                 end_changed = true;
842         }
843         vma->vm_pgoff = pgoff;
844         if (adjust_next) {
845                 next->vm_start += adjust_next << PAGE_SHIFT;
846                 next->vm_pgoff += adjust_next;
847         }
848
849         if (root) {
850                 if (adjust_next)
851                         vma_interval_tree_insert(next, root);
852                 vma_interval_tree_insert(vma, root);
853                 flush_dcache_mmap_unlock(mapping);
854         }
855
856         if (remove_next) {
857                 /*
858                  * vma_merge has merged next into vma, and needs
859                  * us to remove next before dropping the locks.
860                  */
861                 if (remove_next != 3)
862                         __vma_unlink_common(mm, next, next);
863                 else
864                         /*
865                          * vma is not before next if they've been
866                          * swapped.
867                          *
868                          * pre-swap() next->vm_start was reduced so
869                          * tell validate_mm_rb to ignore pre-swap()
870                          * "next" (which is stored in post-swap()
871                          * "vma").
872                          */
873                         __vma_unlink_common(mm, next, vma);
874                 if (file)
875                         __remove_shared_vm_struct(next, file, mapping);
876         } else if (insert) {
877                 /*
878                  * split_vma has split insert from vma, and needs
879                  * us to insert it before dropping the locks
880                  * (it may either follow vma or precede it).
881                  */
882                 __insert_vm_struct(mm, insert);
883         } else {
884                 if (start_changed)
885                         vma_gap_update(vma);
886                 if (end_changed) {
887                         if (!next)
888                                 mm->highest_vm_end = vm_end_gap(vma);
889                         else if (!adjust_next)
890                                 vma_gap_update(next);
891                 }
892         }
893
894         if (anon_vma) {
895                 anon_vma_interval_tree_post_update_vma(vma);
896                 if (adjust_next)
897                         anon_vma_interval_tree_post_update_vma(next);
898                 anon_vma_unlock_write(anon_vma);
899         }
900         if (mapping)
901                 i_mmap_unlock_write(mapping);
902
903         if (root) {
904                 uprobe_mmap(vma);
905
906                 if (adjust_next)
907                         uprobe_mmap(next);
908         }
909
910         if (remove_next) {
911                 if (file) {
912                         uprobe_munmap(next, next->vm_start, next->vm_end);
913                         fput(file);
914                 }
915                 if (next->anon_vma)
916                         anon_vma_merge(vma, next);
917                 mm->map_count--;
918                 mpol_put(vma_policy(next));
919                 vm_area_free(next);
920                 /*
921                  * In mprotect's case 6 (see comments on vma_merge),
922                  * we must remove another next too. It would clutter
923                  * up the code too much to do both in one go.
924                  */
925                 if (remove_next != 3) {
926                         /*
927                          * If "next" was removed and vma->vm_end was
928                          * expanded (up) over it, in turn
929                          * "next->vm_prev->vm_end" changed and the
930                          * "vma->vm_next" gap must be updated.
931                          */
932                         next = vma->vm_next;
933                 } else {
934                         /*
935                          * For the scope of the comment "next" and
936                          * "vma" considered pre-swap(): if "vma" was
937                          * removed, next->vm_start was expanded (down)
938                          * over it and the "next" gap must be updated.
939                          * Because of the swap() the post-swap() "vma"
940                          * actually points to pre-swap() "next"
941                          * (post-swap() "next" as opposed is now a
942                          * dangling pointer).
943                          */
944                         next = vma;
945                 }
946                 if (remove_next == 2) {
947                         remove_next = 1;
948                         end = next->vm_end;
949                         goto again;
950                 }
951                 else if (next)
952                         vma_gap_update(next);
953                 else {
954                         /*
955                          * If remove_next == 2 we obviously can't
956                          * reach this path.
957                          *
958                          * If remove_next == 3 we can't reach this
959                          * path because pre-swap() next is always not
960                          * NULL. pre-swap() "next" is not being
961                          * removed and its next->vm_end is not altered
962                          * (and furthermore "end" already matches
963                          * next->vm_end in remove_next == 3).
964                          *
965                          * We reach this only in the remove_next == 1
966                          * case if the "next" vma that was removed was
967                          * the highest vma of the mm. However in such
968                          * case next->vm_end == "end" and the extended
969                          * "vma" has vma->vm_end == next->vm_end so
970                          * mm->highest_vm_end doesn't need any update
971                          * in remove_next == 1 case.
972                          */
973                         VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
974                 }
975         }
976         if (insert && file)
977                 uprobe_mmap(insert);
978
979         validate_mm(mm);
980
981         return 0;
982 }
983
984 /*
985  * If the vma has a ->close operation then the driver probably needs to release
986  * per-vma resources, so we don't attempt to merge those.
987  */
988 static inline int is_mergeable_vma(struct vm_area_struct *vma,
989                                 struct file *file, unsigned long vm_flags,
990                                 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
991 {
992         /*
993          * VM_SOFTDIRTY should not prevent from VMA merging, if we
994          * match the flags but dirty bit -- the caller should mark
995          * merged VMA as dirty. If dirty bit won't be excluded from
996          * comparison, we increase pressure on the memory system forcing
997          * the kernel to generate new VMAs when old one could be
998          * extended instead.
999          */
1000         if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
1001                 return 0;
1002         if (vma->vm_file != file)
1003                 return 0;
1004         if (vma->vm_ops && vma->vm_ops->close)
1005                 return 0;
1006         if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
1007                 return 0;
1008         return 1;
1009 }
1010
1011 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
1012                                         struct anon_vma *anon_vma2,
1013                                         struct vm_area_struct *vma)
1014 {
1015         /*
1016          * The list_is_singular() test is to avoid merging VMA cloned from
1017          * parents. This can improve scalability caused by anon_vma lock.
1018          */
1019         if ((!anon_vma1 || !anon_vma2) && (!vma ||
1020                 list_is_singular(&vma->anon_vma_chain)))
1021                 return 1;
1022         return anon_vma1 == anon_vma2;
1023 }
1024
1025 /*
1026  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1027  * in front of (at a lower virtual address and file offset than) the vma.
1028  *
1029  * We cannot merge two vmas if they have differently assigned (non-NULL)
1030  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1031  *
1032  * We don't check here for the merged mmap wrapping around the end of pagecache
1033  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
1034  * wrap, nor mmaps which cover the final page at index -1UL.
1035  */
1036 static int
1037 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1038                      struct anon_vma *anon_vma, struct file *file,
1039                      pgoff_t vm_pgoff,
1040                      struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1041 {
1042         if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1043             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1044                 if (vma->vm_pgoff == vm_pgoff)
1045                         return 1;
1046         }
1047         return 0;
1048 }
1049
1050 /*
1051  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1052  * beyond (at a higher virtual address and file offset than) the vma.
1053  *
1054  * We cannot merge two vmas if they have differently assigned (non-NULL)
1055  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1056  */
1057 static int
1058 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1059                     struct anon_vma *anon_vma, struct file *file,
1060                     pgoff_t vm_pgoff,
1061                     struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1062 {
1063         if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1064             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1065                 pgoff_t vm_pglen;
1066                 vm_pglen = vma_pages(vma);
1067                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1068                         return 1;
1069         }
1070         return 0;
1071 }
1072
1073 /*
1074  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1075  * whether that can be merged with its predecessor or its successor.
1076  * Or both (it neatly fills a hole).
1077  *
1078  * In most cases - when called for mmap, brk or mremap - [addr,end) is
1079  * certain not to be mapped by the time vma_merge is called; but when
1080  * called for mprotect, it is certain to be already mapped (either at
1081  * an offset within prev, or at the start of next), and the flags of
1082  * this area are about to be changed to vm_flags - and the no-change
1083  * case has already been eliminated.
1084  *
1085  * The following mprotect cases have to be considered, where AAAA is
1086  * the area passed down from mprotect_fixup, never extending beyond one
1087  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1088  *
1089  *     AAAA             AAAA                   AAAA
1090  *    PPPPPPNNNNNN    PPPPPPNNNNNN       PPPPPPNNNNNN
1091  *    cannot merge    might become       might become
1092  *                    PPNNNNNNNNNN       PPPPPPPPPPNN
1093  *    mmap, brk or    case 4 below       case 5 below
1094  *    mremap move:
1095  *                        AAAA               AAAA
1096  *                    PPPP    NNNN       PPPPNNNNXXXX
1097  *                    might become       might become
1098  *                    PPPPPPPPPPPP 1 or  PPPPPPPPPPPP 6 or
1099  *                    PPPPPPPPNNNN 2 or  PPPPPPPPXXXX 7 or
1100  *                    PPPPNNNNNNNN 3     PPPPXXXXXXXX 8
1101  *
1102  * It is important for case 8 that the vma NNNN overlapping the
1103  * region AAAA is never going to extended over XXXX. Instead XXXX must
1104  * be extended in region AAAA and NNNN must be removed. This way in
1105  * all cases where vma_merge succeeds, the moment vma_adjust drops the
1106  * rmap_locks, the properties of the merged vma will be already
1107  * correct for the whole merged range. Some of those properties like
1108  * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1109  * be correct for the whole merged range immediately after the
1110  * rmap_locks are released. Otherwise if XXXX would be removed and
1111  * NNNN would be extended over the XXXX range, remove_migration_ptes
1112  * or other rmap walkers (if working on addresses beyond the "end"
1113  * parameter) may establish ptes with the wrong permissions of NNNN
1114  * instead of the right permissions of XXXX.
1115  */
1116 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1117                         struct vm_area_struct *prev, unsigned long addr,
1118                         unsigned long end, unsigned long vm_flags,
1119                         struct anon_vma *anon_vma, struct file *file,
1120                         pgoff_t pgoff, struct mempolicy *policy,
1121                         struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1122 {
1123         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1124         struct vm_area_struct *area, *next;
1125         int err;
1126
1127         /*
1128          * We later require that vma->vm_flags == vm_flags,
1129          * so this tests vma->vm_flags & VM_SPECIAL, too.
1130          */
1131         if (vm_flags & VM_SPECIAL)
1132                 return NULL;
1133
1134         if (prev)
1135                 next = prev->vm_next;
1136         else
1137                 next = mm->mmap;
1138         area = next;
1139         if (area && area->vm_end == end)                /* cases 6, 7, 8 */
1140                 next = next->vm_next;
1141
1142         /* verify some invariant that must be enforced by the caller */
1143         VM_WARN_ON(prev && addr <= prev->vm_start);
1144         VM_WARN_ON(area && end > area->vm_end);
1145         VM_WARN_ON(addr >= end);
1146
1147         /*
1148          * Can it merge with the predecessor?
1149          */
1150         if (prev && prev->vm_end == addr &&
1151                         mpol_equal(vma_policy(prev), policy) &&
1152                         can_vma_merge_after(prev, vm_flags,
1153                                             anon_vma, file, pgoff,
1154                                             vm_userfaultfd_ctx)) {
1155                 /*
1156                  * OK, it can.  Can we now merge in the successor as well?
1157                  */
1158                 if (next && end == next->vm_start &&
1159                                 mpol_equal(policy, vma_policy(next)) &&
1160                                 can_vma_merge_before(next, vm_flags,
1161                                                      anon_vma, file,
1162                                                      pgoff+pglen,
1163                                                      vm_userfaultfd_ctx) &&
1164                                 is_mergeable_anon_vma(prev->anon_vma,
1165                                                       next->anon_vma, NULL)) {
1166                                                         /* cases 1, 6 */
1167                         err = __vma_adjust(prev, prev->vm_start,
1168                                          next->vm_end, prev->vm_pgoff, NULL,
1169                                          prev);
1170                 } else                                  /* cases 2, 5, 7 */
1171                         err = __vma_adjust(prev, prev->vm_start,
1172                                          end, prev->vm_pgoff, NULL, prev);
1173                 if (err)
1174                         return NULL;
1175                 khugepaged_enter_vma_merge(prev, vm_flags);
1176                 return prev;
1177         }
1178
1179         /*
1180          * Can this new request be merged in front of next?
1181          */
1182         if (next && end == next->vm_start &&
1183                         mpol_equal(policy, vma_policy(next)) &&
1184                         can_vma_merge_before(next, vm_flags,
1185                                              anon_vma, file, pgoff+pglen,
1186                                              vm_userfaultfd_ctx)) {
1187                 if (prev && addr < prev->vm_end)        /* case 4 */
1188                         err = __vma_adjust(prev, prev->vm_start,
1189                                          addr, prev->vm_pgoff, NULL, next);
1190                 else {                                  /* cases 3, 8 */
1191                         err = __vma_adjust(area, addr, next->vm_end,
1192                                          next->vm_pgoff - pglen, NULL, next);
1193                         /*
1194                          * In case 3 area is already equal to next and
1195                          * this is a noop, but in case 8 "area" has
1196                          * been removed and next was expanded over it.
1197                          */
1198                         area = next;
1199                 }
1200                 if (err)
1201                         return NULL;
1202                 khugepaged_enter_vma_merge(area, vm_flags);
1203                 return area;
1204         }
1205
1206         return NULL;
1207 }
1208
1209 /*
1210  * Rough compatbility check to quickly see if it's even worth looking
1211  * at sharing an anon_vma.
1212  *
1213  * They need to have the same vm_file, and the flags can only differ
1214  * in things that mprotect may change.
1215  *
1216  * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1217  * we can merge the two vma's. For example, we refuse to merge a vma if
1218  * there is a vm_ops->close() function, because that indicates that the
1219  * driver is doing some kind of reference counting. But that doesn't
1220  * really matter for the anon_vma sharing case.
1221  */
1222 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1223 {
1224         return a->vm_end == b->vm_start &&
1225                 mpol_equal(vma_policy(a), vma_policy(b)) &&
1226                 a->vm_file == b->vm_file &&
1227                 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) &&
1228                 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1229 }
1230
1231 /*
1232  * Do some basic sanity checking to see if we can re-use the anon_vma
1233  * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1234  * the same as 'old', the other will be the new one that is trying
1235  * to share the anon_vma.
1236  *
1237  * NOTE! This runs with mm_sem held for reading, so it is possible that
1238  * the anon_vma of 'old' is concurrently in the process of being set up
1239  * by another page fault trying to merge _that_. But that's ok: if it
1240  * is being set up, that automatically means that it will be a singleton
1241  * acceptable for merging, so we can do all of this optimistically. But
1242  * we do that READ_ONCE() to make sure that we never re-load the pointer.
1243  *
1244  * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1245  * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1246  * is to return an anon_vma that is "complex" due to having gone through
1247  * a fork).
1248  *
1249  * We also make sure that the two vma's are compatible (adjacent,
1250  * and with the same memory policies). That's all stable, even with just
1251  * a read lock on the mm_sem.
1252  */
1253 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1254 {
1255         if (anon_vma_compatible(a, b)) {
1256                 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1257
1258                 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1259                         return anon_vma;
1260         }
1261         return NULL;
1262 }
1263
1264 /*
1265  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1266  * neighbouring vmas for a suitable anon_vma, before it goes off
1267  * to allocate a new anon_vma.  It checks because a repetitive
1268  * sequence of mprotects and faults may otherwise lead to distinct
1269  * anon_vmas being allocated, preventing vma merge in subsequent
1270  * mprotect.
1271  */
1272 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1273 {
1274         struct anon_vma *anon_vma = NULL;
1275
1276         /* Try next first. */
1277         if (vma->vm_next) {
1278                 anon_vma = reusable_anon_vma(vma->vm_next, vma, vma->vm_next);
1279                 if (anon_vma)
1280                         return anon_vma;
1281         }
1282
1283         /* Try prev next. */
1284         if (vma->vm_prev)
1285                 anon_vma = reusable_anon_vma(vma->vm_prev, vma->vm_prev, vma);
1286
1287         /*
1288          * We might reach here with anon_vma == NULL if we can't find
1289          * any reusable anon_vma.
1290          * There's no absolute need to look only at touching neighbours:
1291          * we could search further afield for "compatible" anon_vmas.
1292          * But it would probably just be a waste of time searching,
1293          * or lead to too many vmas hanging off the same anon_vma.
1294          * We're trying to allow mprotect remerging later on,
1295          * not trying to minimize memory used for anon_vmas.
1296          */
1297         return anon_vma;
1298 }
1299
1300 /*
1301  * If a hint addr is less than mmap_min_addr change hint to be as
1302  * low as possible but still greater than mmap_min_addr
1303  */
1304 static inline unsigned long round_hint_to_min(unsigned long hint)
1305 {
1306         hint &= PAGE_MASK;
1307         if (((void *)hint != NULL) &&
1308             (hint < mmap_min_addr))
1309                 return PAGE_ALIGN(mmap_min_addr);
1310         return hint;
1311 }
1312
1313 static inline int mlock_future_check(struct mm_struct *mm,
1314                                      unsigned long flags,
1315                                      unsigned long len)
1316 {
1317         unsigned long locked, lock_limit;
1318
1319         /*  mlock MCL_FUTURE? */
1320         if (flags & VM_LOCKED) {
1321                 locked = len >> PAGE_SHIFT;
1322                 locked += mm->locked_vm;
1323                 lock_limit = rlimit(RLIMIT_MEMLOCK);
1324                 lock_limit >>= PAGE_SHIFT;
1325                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1326                         return -EAGAIN;
1327         }
1328         return 0;
1329 }
1330
1331 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1332 {
1333         if (S_ISREG(inode->i_mode))
1334                 return MAX_LFS_FILESIZE;
1335
1336         if (S_ISBLK(inode->i_mode))
1337                 return MAX_LFS_FILESIZE;
1338
1339         if (S_ISSOCK(inode->i_mode))
1340                 return MAX_LFS_FILESIZE;
1341
1342         /* Special "we do even unsigned file positions" case */
1343         if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1344                 return 0;
1345
1346         /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1347         return ULONG_MAX;
1348 }
1349
1350 static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1351                                 unsigned long pgoff, unsigned long len)
1352 {
1353         u64 maxsize = file_mmap_size_max(file, inode);
1354
1355         if (maxsize && len > maxsize)
1356                 return false;
1357         maxsize -= len;
1358         if (pgoff > maxsize >> PAGE_SHIFT)
1359                 return false;
1360         return true;
1361 }
1362
1363 /*
1364  * The caller must hold down_write(&current->mm->mmap_sem).
1365  */
1366 unsigned long do_mmap(struct file *file, unsigned long addr,
1367                         unsigned long len, unsigned long prot,
1368                         unsigned long flags, vm_flags_t vm_flags,
1369                         unsigned long pgoff, unsigned long *populate,
1370                         struct list_head *uf)
1371 {
1372         struct mm_struct *mm = current->mm;
1373         int pkey = 0;
1374
1375         *populate = 0;
1376
1377         if (!len)
1378                 return -EINVAL;
1379
1380         /*
1381          * Does the application expect PROT_READ to imply PROT_EXEC?
1382          *
1383          * (the exception is when the underlying filesystem is noexec
1384          *  mounted, in which case we dont add PROT_EXEC.)
1385          */
1386         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1387                 if (!(file && path_noexec(&file->f_path)))
1388                         prot |= PROT_EXEC;
1389
1390         /* force arch specific MAP_FIXED handling in get_unmapped_area */
1391         if (flags & MAP_FIXED_NOREPLACE)
1392                 flags |= MAP_FIXED;
1393
1394         if (!(flags & MAP_FIXED))
1395                 addr = round_hint_to_min(addr);
1396
1397         /* Careful about overflows.. */
1398         len = PAGE_ALIGN(len);
1399         if (!len)
1400                 return -ENOMEM;
1401
1402         /* offset overflow? */
1403         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1404                 return -EOVERFLOW;
1405
1406         /* Too many mappings? */
1407         if (mm->map_count > sysctl_max_map_count)
1408                 return -ENOMEM;
1409
1410         /* Obtain the address to map to. we verify (or select) it and ensure
1411          * that it represents a valid section of the address space.
1412          */
1413         addr = get_unmapped_area(file, addr, len, pgoff, flags);
1414         if (IS_ERR_VALUE(addr))
1415                 return addr;
1416
1417         if (flags & MAP_FIXED_NOREPLACE) {
1418                 struct vm_area_struct *vma = find_vma(mm, addr);
1419
1420                 if (vma && vma->vm_start < addr + len)
1421                         return -EEXIST;
1422         }
1423
1424         if (prot == PROT_EXEC) {
1425                 pkey = execute_only_pkey(mm);
1426                 if (pkey < 0)
1427                         pkey = 0;
1428         }
1429
1430         /* Do simple checking here so the lower-level routines won't have
1431          * to. we assume access permissions have been handled by the open
1432          * of the memory object, so we don't do any here.
1433          */
1434         vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1435                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1436
1437         if (flags & MAP_LOCKED)
1438                 if (!can_do_mlock())
1439                         return -EPERM;
1440
1441         if (mlock_future_check(mm, vm_flags, len))
1442                 return -EAGAIN;
1443
1444         if (file) {
1445                 struct inode *inode = file_inode(file);
1446                 unsigned long flags_mask;
1447
1448                 if (!file_mmap_ok(file, inode, pgoff, len))
1449                         return -EOVERFLOW;
1450
1451                 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1452
1453                 switch (flags & MAP_TYPE) {
1454                 case MAP_SHARED:
1455                         /*
1456                          * Force use of MAP_SHARED_VALIDATE with non-legacy
1457                          * flags. E.g. MAP_SYNC is dangerous to use with
1458                          * MAP_SHARED as you don't know which consistency model
1459                          * you will get. We silently ignore unsupported flags
1460                          * with MAP_SHARED to preserve backward compatibility.
1461                          */
1462                         flags &= LEGACY_MAP_MASK;
1463                         fallthrough;
1464                 case MAP_SHARED_VALIDATE:
1465                         if (flags & ~flags_mask)
1466                                 return -EOPNOTSUPP;
1467                         if (prot & PROT_WRITE) {
1468                                 if (!(file->f_mode & FMODE_WRITE))
1469                                         return -EACCES;
1470                                 if (IS_SWAPFILE(file->f_mapping->host))
1471                                         return -ETXTBSY;
1472                         }
1473
1474                         /*
1475                          * Make sure we don't allow writing to an append-only
1476                          * file..
1477                          */
1478                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1479                                 return -EACCES;
1480
1481                         /*
1482                          * Make sure there are no mandatory locks on the file.
1483                          */
1484                         if (locks_verify_locked(file))
1485                                 return -EAGAIN;
1486
1487                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1488                         if (!(file->f_mode & FMODE_WRITE))
1489                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1490                         fallthrough;
1491                 case MAP_PRIVATE:
1492                         if (!(file->f_mode & FMODE_READ))
1493                                 return -EACCES;
1494                         if (path_noexec(&file->f_path)) {
1495                                 if (vm_flags & VM_EXEC)
1496                                         return -EPERM;
1497                                 vm_flags &= ~VM_MAYEXEC;
1498                         }
1499
1500                         if (!file->f_op->mmap)
1501                                 return -ENODEV;
1502                         if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1503                                 return -EINVAL;
1504                         break;
1505
1506                 default:
1507                         return -EINVAL;
1508                 }
1509         } else {
1510                 switch (flags & MAP_TYPE) {
1511                 case MAP_SHARED:
1512                         if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1513                                 return -EINVAL;
1514                         /*
1515                          * Ignore pgoff.
1516                          */
1517                         pgoff = 0;
1518                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1519                         break;
1520                 case MAP_PRIVATE:
1521                         /*
1522                          * Set pgoff according to addr for anon_vma.
1523                          */
1524                         pgoff = addr >> PAGE_SHIFT;
1525                         break;
1526                 default:
1527                         return -EINVAL;
1528                 }
1529         }
1530
1531         /*
1532          * Set 'VM_NORESERVE' if we should not account for the
1533          * memory use of this mapping.
1534          */
1535         if (flags & MAP_NORESERVE) {
1536                 /* We honor MAP_NORESERVE if allowed to overcommit */
1537                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1538                         vm_flags |= VM_NORESERVE;
1539
1540                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1541                 if (file && is_file_hugepages(file))
1542                         vm_flags |= VM_NORESERVE;
1543         }
1544
1545         addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1546         if (!IS_ERR_VALUE(addr) &&
1547             ((vm_flags & VM_LOCKED) ||
1548              (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1549                 *populate = len;
1550         return addr;
1551 }
1552
1553 unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1554                               unsigned long prot, unsigned long flags,
1555                               unsigned long fd, unsigned long pgoff)
1556 {
1557         struct file *file = NULL;
1558         unsigned long retval;
1559
1560         if (!(flags & MAP_ANONYMOUS)) {
1561                 audit_mmap_fd(fd, flags);
1562                 file = fget(fd);
1563                 if (!file)
1564                         return -EBADF;
1565                 if (is_file_hugepages(file))
1566                         len = ALIGN(len, huge_page_size(hstate_file(file)));
1567                 retval = -EINVAL;
1568                 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1569                         goto out_fput;
1570         } else if (flags & MAP_HUGETLB) {
1571                 struct user_struct *user = NULL;
1572                 struct hstate *hs;
1573
1574                 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1575                 if (!hs)
1576                         return -EINVAL;
1577
1578                 len = ALIGN(len, huge_page_size(hs));
1579                 /*
1580                  * VM_NORESERVE is used because the reservations will be
1581                  * taken when vm_ops->mmap() is called
1582                  * A dummy user value is used because we are not locking
1583                  * memory so no accounting is necessary
1584                  */
1585                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1586                                 VM_NORESERVE,
1587                                 &user, HUGETLB_ANONHUGE_INODE,
1588                                 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1589                 if (IS_ERR(file))
1590                         return PTR_ERR(file);
1591         }
1592
1593         flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1594
1595         retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1596 out_fput:
1597         if (file)
1598                 fput(file);
1599         return retval;
1600 }
1601
1602 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1603                 unsigned long, prot, unsigned long, flags,
1604                 unsigned long, fd, unsigned long, pgoff)
1605 {
1606         return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1607 }
1608
1609 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1610 struct mmap_arg_struct {
1611         unsigned long addr;
1612         unsigned long len;
1613         unsigned long prot;
1614         unsigned long flags;
1615         unsigned long fd;
1616         unsigned long offset;
1617 };
1618
1619 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1620 {
1621         struct mmap_arg_struct a;
1622
1623         if (copy_from_user(&a, arg, sizeof(a)))
1624                 return -EFAULT;
1625         if (offset_in_page(a.offset))
1626                 return -EINVAL;
1627
1628         return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1629                                a.offset >> PAGE_SHIFT);
1630 }
1631 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1632
1633 /*
1634  * Some shared mappings will want the pages marked read-only
1635  * to track write events. If so, we'll downgrade vm_page_prot
1636  * to the private version (using protection_map[] without the
1637  * VM_SHARED bit).
1638  */
1639 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1640 {
1641         vm_flags_t vm_flags = vma->vm_flags;
1642         const struct vm_operations_struct *vm_ops = vma->vm_ops;
1643
1644         /* If it was private or non-writable, the write bit is already clear */
1645         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1646                 return 0;
1647
1648         /* The backer wishes to know when pages are first written to? */
1649         if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1650                 return 1;
1651
1652         /* The open routine did something to the protections that pgprot_modify
1653          * won't preserve? */
1654         if (pgprot_val(vm_page_prot) !=
1655             pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1656                 return 0;
1657
1658         /* Do we need to track softdirty? */
1659         if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1660                 return 1;
1661
1662         /* Specialty mapping? */
1663         if (vm_flags & VM_PFNMAP)
1664                 return 0;
1665
1666         /* Can the mapping track the dirty pages? */
1667         return vma->vm_file && vma->vm_file->f_mapping &&
1668                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1669 }
1670
1671 /*
1672  * We account for memory if it's a private writeable mapping,
1673  * not hugepages and VM_NORESERVE wasn't set.
1674  */
1675 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1676 {
1677         /*
1678          * hugetlb has its own accounting separate from the core VM
1679          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1680          */
1681         if (file && is_file_hugepages(file))
1682                 return 0;
1683
1684         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1685 }
1686
1687 unsigned long mmap_region(struct file *file, unsigned long addr,
1688                 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
1689                 struct list_head *uf)
1690 {
1691         struct mm_struct *mm = current->mm;
1692         struct vm_area_struct *vma, *prev;
1693         int error;
1694         struct rb_node **rb_link, *rb_parent;
1695         unsigned long charged = 0;
1696
1697         /* Check against address space limit. */
1698         if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1699                 unsigned long nr_pages;
1700
1701                 /*
1702                  * MAP_FIXED may remove pages of mappings that intersects with
1703                  * requested mapping. Account for the pages it would unmap.
1704                  */
1705                 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1706
1707                 if (!may_expand_vm(mm, vm_flags,
1708                                         (len >> PAGE_SHIFT) - nr_pages))
1709                         return -ENOMEM;
1710         }
1711
1712         /* Clear old maps */
1713         while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
1714                               &rb_parent)) {
1715                 if (do_munmap(mm, addr, len, uf))
1716                         return -ENOMEM;
1717         }
1718
1719         /*
1720          * Private writable mapping: check memory availability
1721          */
1722         if (accountable_mapping(file, vm_flags)) {
1723                 charged = len >> PAGE_SHIFT;
1724                 if (security_vm_enough_memory_mm(mm, charged))
1725                         return -ENOMEM;
1726                 vm_flags |= VM_ACCOUNT;
1727         }
1728
1729         /*
1730          * Can we just expand an old mapping?
1731          */
1732         vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1733                         NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1734         if (vma)
1735                 goto out;
1736
1737         /*
1738          * Determine the object being mapped and call the appropriate
1739          * specific mapper. the address has already been validated, but
1740          * not unmapped, but the maps are removed from the list.
1741          */
1742         vma = vm_area_alloc(mm);
1743         if (!vma) {
1744                 error = -ENOMEM;
1745                 goto unacct_error;
1746         }
1747
1748         vma->vm_start = addr;
1749         vma->vm_end = addr + len;
1750         vma->vm_flags = vm_flags;
1751         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1752         vma->vm_pgoff = pgoff;
1753
1754         if (file) {
1755                 if (vm_flags & VM_DENYWRITE) {
1756                         error = deny_write_access(file);
1757                         if (error)
1758                                 goto free_vma;
1759                 }
1760                 if (vm_flags & VM_SHARED) {
1761                         error = mapping_map_writable(file->f_mapping);
1762                         if (error)
1763                                 goto allow_write_and_free_vma;
1764                 }
1765
1766                 /* ->mmap() can change vma->vm_file, but must guarantee that
1767                  * vma_link() below can deny write-access if VM_DENYWRITE is set
1768                  * and map writably if VM_SHARED is set. This usually means the
1769                  * new file must not have been exposed to user-space, yet.
1770                  */
1771                 vma->vm_file = get_file(file);
1772                 error = call_mmap(file, vma);
1773                 if (error)
1774                         goto unmap_and_free_vma;
1775
1776                 /* Can addr have changed??
1777                  *
1778                  * Answer: Yes, several device drivers can do it in their
1779                  *         f_op->mmap method. -DaveM
1780                  * Bug: If addr is changed, prev, rb_link, rb_parent should
1781                  *      be updated for vma_link()
1782                  */
1783                 WARN_ON_ONCE(addr != vma->vm_start);
1784
1785                 addr = vma->vm_start;
1786                 vm_flags = vma->vm_flags;
1787         } else if (vm_flags & VM_SHARED) {
1788                 error = shmem_zero_setup(vma);
1789                 if (error)
1790                         goto free_vma;
1791         } else {
1792                 vma_set_anonymous(vma);
1793         }
1794
1795         vma_link(mm, vma, prev, rb_link, rb_parent);
1796         /* Once vma denies write, undo our temporary denial count */
1797         if (file) {
1798                 if (vm_flags & VM_SHARED)
1799                         mapping_unmap_writable(file->f_mapping);
1800                 if (vm_flags & VM_DENYWRITE)
1801                         allow_write_access(file);
1802         }
1803         file = vma->vm_file;
1804 out:
1805         perf_event_mmap(vma);
1806
1807         vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1808         if (vm_flags & VM_LOCKED) {
1809                 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
1810                                         is_vm_hugetlb_page(vma) ||
1811                                         vma == get_gate_vma(current->mm))
1812                         vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1813                 else
1814                         mm->locked_vm += (len >> PAGE_SHIFT);
1815         }
1816
1817         if (file)
1818                 uprobe_mmap(vma);
1819
1820         /*
1821          * New (or expanded) vma always get soft dirty status.
1822          * Otherwise user-space soft-dirty page tracker won't
1823          * be able to distinguish situation when vma area unmapped,
1824          * then new mapped in-place (which must be aimed as
1825          * a completely new data area).
1826          */
1827         vma->vm_flags |= VM_SOFTDIRTY;
1828
1829         vma_set_page_prot(vma);
1830
1831         return addr;
1832
1833 unmap_and_free_vma:
1834         vma->vm_file = NULL;
1835         fput(file);
1836
1837         /* Undo any partial mapping done by a device driver. */
1838         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1839         charged = 0;
1840         if (vm_flags & VM_SHARED)
1841                 mapping_unmap_writable(file->f_mapping);
1842 allow_write_and_free_vma:
1843         if (vm_flags & VM_DENYWRITE)
1844                 allow_write_access(file);
1845 free_vma:
1846         vm_area_free(vma);
1847 unacct_error:
1848         if (charged)
1849                 vm_unacct_memory(charged);
1850         return error;
1851 }
1852
1853 static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1854 {
1855         /*
1856          * We implement the search by looking for an rbtree node that
1857          * immediately follows a suitable gap. That is,
1858          * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1859          * - gap_end   = vma->vm_start        >= info->low_limit  + length;
1860          * - gap_end - gap_start >= length
1861          */
1862
1863         struct mm_struct *mm = current->mm;
1864         struct vm_area_struct *vma;
1865         unsigned long length, low_limit, high_limit, gap_start, gap_end;
1866
1867         /* Adjust search length to account for worst case alignment overhead */
1868         length = info->length + info->align_mask;
1869         if (length < info->length)
1870                 return -ENOMEM;
1871
1872         /* Adjust search limits by the desired length */
1873         if (info->high_limit < length)
1874                 return -ENOMEM;
1875         high_limit = info->high_limit - length;
1876
1877         if (info->low_limit > high_limit)
1878                 return -ENOMEM;
1879         low_limit = info->low_limit + length;
1880
1881         /* Check if rbtree root looks promising */
1882         if (RB_EMPTY_ROOT(&mm->mm_rb))
1883                 goto check_highest;
1884         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1885         if (vma->rb_subtree_gap < length)
1886                 goto check_highest;
1887
1888         while (true) {
1889                 /* Visit left subtree if it looks promising */
1890                 gap_end = vm_start_gap(vma);
1891                 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1892                         struct vm_area_struct *left =
1893                                 rb_entry(vma->vm_rb.rb_left,
1894                                          struct vm_area_struct, vm_rb);
1895                         if (left->rb_subtree_gap >= length) {
1896                                 vma = left;
1897                                 continue;
1898                         }
1899                 }
1900
1901                 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1902 check_current:
1903                 /* Check if current node has a suitable gap */
1904                 if (gap_start > high_limit)
1905                         return -ENOMEM;
1906                 if (gap_end >= low_limit &&
1907                     gap_end > gap_start && gap_end - gap_start >= length)
1908                         goto found;
1909
1910                 /* Visit right subtree if it looks promising */
1911                 if (vma->vm_rb.rb_right) {
1912                         struct vm_area_struct *right =
1913                                 rb_entry(vma->vm_rb.rb_right,
1914                                          struct vm_area_struct, vm_rb);
1915                         if (right->rb_subtree_gap >= length) {
1916                                 vma = right;
1917                                 continue;
1918                         }
1919                 }
1920
1921                 /* Go back up the rbtree to find next candidate node */
1922                 while (true) {
1923                         struct rb_node *prev = &vma->vm_rb;
1924                         if (!rb_parent(prev))
1925                                 goto check_highest;
1926                         vma = rb_entry(rb_parent(prev),
1927                                        struct vm_area_struct, vm_rb);
1928                         if (prev == vma->vm_rb.rb_left) {
1929                                 gap_start = vm_end_gap(vma->vm_prev);
1930                                 gap_end = vm_start_gap(vma);
1931                                 goto check_current;
1932                         }
1933                 }
1934         }
1935
1936 check_highest:
1937         /* Check highest gap, which does not precede any rbtree node */
1938         gap_start = mm->highest_vm_end;
1939         gap_end = ULONG_MAX;  /* Only for VM_BUG_ON below */
1940         if (gap_start > high_limit)
1941                 return -ENOMEM;
1942
1943 found:
1944         /* We found a suitable gap. Clip it with the original low_limit. */
1945         if (gap_start < info->low_limit)
1946                 gap_start = info->low_limit;
1947
1948         /* Adjust gap address to the desired alignment */
1949         gap_start += (info->align_offset - gap_start) & info->align_mask;
1950
1951         VM_BUG_ON(gap_start + info->length > info->high_limit);
1952         VM_BUG_ON(gap_start + info->length > gap_end);
1953         return gap_start;
1954 }
1955
1956 static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1957 {
1958         struct mm_struct *mm = current->mm;
1959         struct vm_area_struct *vma;
1960         unsigned long length, low_limit, high_limit, gap_start, gap_end;
1961
1962         /* Adjust search length to account for worst case alignment overhead */
1963         length = info->length + info->align_mask;
1964         if (length < info->length)
1965                 return -ENOMEM;
1966
1967         /*
1968          * Adjust search limits by the desired length.
1969          * See implementation comment at top of unmapped_area().
1970          */
1971         gap_end = info->high_limit;
1972         if (gap_end < length)
1973                 return -ENOMEM;
1974         high_limit = gap_end - length;
1975
1976         if (info->low_limit > high_limit)
1977                 return -ENOMEM;
1978         low_limit = info->low_limit + length;
1979
1980         /* Check highest gap, which does not precede any rbtree node */
1981         gap_start = mm->highest_vm_end;
1982         if (gap_start <= high_limit)
1983                 goto found_highest;
1984
1985         /* Check if rbtree root looks promising */
1986         if (RB_EMPTY_ROOT(&mm->mm_rb))
1987                 return -ENOMEM;
1988         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1989         if (vma->rb_subtree_gap < length)
1990                 return -ENOMEM;
1991
1992         while (true) {
1993                 /* Visit right subtree if it looks promising */
1994                 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1995                 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1996                         struct vm_area_struct *right =
1997                                 rb_entry(vma->vm_rb.rb_right,
1998                                          struct vm_area_struct, vm_rb);
1999                         if (right->rb_subtree_gap >= length) {
2000                                 vma = right;
2001                                 continue;
2002                         }
2003                 }
2004
2005 check_current:
2006                 /* Check if current node has a suitable gap */
2007                 gap_end = vm_start_gap(vma);
2008                 if (gap_end < low_limit)
2009                         return -ENOMEM;
2010                 if (gap_start <= high_limit &&
2011                     gap_end > gap_start && gap_end - gap_start >= length)
2012                         goto found;
2013
2014                 /* Visit left subtree if it looks promising */
2015                 if (vma->vm_rb.rb_left) {
2016                         struct vm_area_struct *left =
2017                                 rb_entry(vma->vm_rb.rb_left,
2018                                          struct vm_area_struct, vm_rb);
2019                         if (left->rb_subtree_gap >= length) {
2020                                 vma = left;
2021                                 continue;
2022                         }
2023                 }
2024
2025                 /* Go back up the rbtree to find next candidate node */
2026                 while (true) {
2027                         struct rb_node *prev = &vma->vm_rb;
2028                         if (!rb_parent(prev))
2029                                 return -ENOMEM;
2030                         vma = rb_entry(rb_parent(prev),
2031                                        struct vm_area_struct, vm_rb);
2032                         if (prev == vma->vm_rb.rb_right) {
2033                                 gap_start = vma->vm_prev ?
2034                                         vm_end_gap(vma->vm_prev) : 0;
2035                                 goto check_current;
2036                         }
2037                 }
2038         }
2039
2040 found:
2041         /* We found a suitable gap. Clip it with the original high_limit. */
2042         if (gap_end > info->high_limit)
2043                 gap_end = info->high_limit;
2044
2045 found_highest:
2046         /* Compute highest gap address at the desired alignment */
2047         gap_end -= info->length;
2048         gap_end -= (gap_end - info->align_offset) & info->align_mask;
2049
2050         VM_BUG_ON(gap_end < info->low_limit);
2051         VM_BUG_ON(gap_end < gap_start);
2052         return gap_end;
2053 }
2054
2055 /*
2056  * Search for an unmapped address range.
2057  *
2058  * We are looking for a range that:
2059  * - does not intersect with any VMA;
2060  * - is contained within the [low_limit, high_limit) interval;
2061  * - is at least the desired size.
2062  * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
2063  */
2064 unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
2065 {
2066         unsigned long addr;
2067
2068         if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
2069                 addr = unmapped_area_topdown(info);
2070         else
2071                 addr = unmapped_area(info);
2072
2073         trace_vm_unmapped_area(addr, info);
2074         return addr;
2075 }
2076
2077 #ifndef arch_get_mmap_end
2078 #define arch_get_mmap_end(addr) (TASK_SIZE)
2079 #endif
2080
2081 #ifndef arch_get_mmap_base
2082 #define arch_get_mmap_base(addr, base) (base)
2083 #endif
2084
2085 /* Get an address range which is currently unmapped.
2086  * For shmat() with addr=0.
2087  *
2088  * Ugly calling convention alert:
2089  * Return value with the low bits set means error value,
2090  * ie
2091  *      if (ret & ~PAGE_MASK)
2092  *              error = ret;
2093  *
2094  * This function "knows" that -ENOMEM has the bits set.
2095  */
2096 #ifndef HAVE_ARCH_UNMAPPED_AREA
2097 unsigned long
2098 arch_get_unmapped_area(struct file *filp, unsigned long addr,
2099                 unsigned long len, unsigned long pgoff, unsigned long flags)
2100 {
2101         struct mm_struct *mm = current->mm;
2102         struct vm_area_struct *vma, *prev;
2103         struct vm_unmapped_area_info info;
2104         const unsigned long mmap_end = arch_get_mmap_end(addr);
2105
2106         if (len > mmap_end - mmap_min_addr)
2107                 return -ENOMEM;
2108
2109         if (flags & MAP_FIXED)
2110                 return addr;
2111
2112         if (addr) {
2113                 addr = PAGE_ALIGN(addr);
2114                 vma = find_vma_prev(mm, addr, &prev);
2115                 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2116                     (!vma || addr + len <= vm_start_gap(vma)) &&
2117                     (!prev || addr >= vm_end_gap(prev)))
2118                         return addr;
2119         }
2120
2121         info.flags = 0;
2122         info.length = len;
2123         info.low_limit = mm->mmap_base;
2124         info.high_limit = mmap_end;
2125         info.align_mask = 0;
2126         return vm_unmapped_area(&info);
2127 }
2128 #endif
2129
2130 /*
2131  * This mmap-allocator allocates new areas top-down from below the
2132  * stack's low limit (the base):
2133  */
2134 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2135 unsigned long
2136 arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
2137                           unsigned long len, unsigned long pgoff,
2138                           unsigned long flags)
2139 {
2140         struct vm_area_struct *vma, *prev;
2141         struct mm_struct *mm = current->mm;
2142         struct vm_unmapped_area_info info;
2143         const unsigned long mmap_end = arch_get_mmap_end(addr);
2144
2145         /* requested length too big for entire address space */
2146         if (len > mmap_end - mmap_min_addr)
2147                 return -ENOMEM;
2148
2149         if (flags & MAP_FIXED)
2150                 return addr;
2151
2152         /* requesting a specific address */
2153         if (addr) {
2154                 addr = PAGE_ALIGN(addr);
2155                 vma = find_vma_prev(mm, addr, &prev);
2156                 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2157                                 (!vma || addr + len <= vm_start_gap(vma)) &&
2158                                 (!prev || addr >= vm_end_gap(prev)))
2159                         return addr;
2160         }
2161
2162         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2163         info.length = len;
2164         info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2165         info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
2166         info.align_mask = 0;
2167         addr = vm_unmapped_area(&info);
2168
2169         /*
2170          * A failed mmap() very likely causes application failure,
2171          * so fall back to the bottom-up function here. This scenario
2172          * can happen with large stack limits and large mmap()
2173          * allocations.
2174          */
2175         if (offset_in_page(addr)) {
2176                 VM_BUG_ON(addr != -ENOMEM);
2177                 info.flags = 0;
2178                 info.low_limit = TASK_UNMAPPED_BASE;
2179                 info.high_limit = mmap_end;
2180                 addr = vm_unmapped_area(&info);
2181         }
2182
2183         return addr;
2184 }
2185 #endif
2186
2187 unsigned long
2188 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2189                 unsigned long pgoff, unsigned long flags)
2190 {
2191         unsigned long (*get_area)(struct file *, unsigned long,
2192                                   unsigned long, unsigned long, unsigned long);
2193
2194         unsigned long error = arch_mmap_check(addr, len, flags);
2195         if (error)
2196                 return error;
2197
2198         /* Careful about overflows.. */
2199         if (len > TASK_SIZE)
2200                 return -ENOMEM;
2201
2202         get_area = current->mm->get_unmapped_area;
2203         if (file) {
2204                 if (file->f_op->get_unmapped_area)
2205                         get_area = file->f_op->get_unmapped_area;
2206         } else if (flags & MAP_SHARED) {
2207                 /*
2208                  * mmap_region() will call shmem_zero_setup() to create a file,
2209                  * so use shmem's get_unmapped_area in case it can be huge.
2210                  * do_mmap_pgoff() will clear pgoff, so match alignment.
2211                  */
2212                 pgoff = 0;
2213                 get_area = shmem_get_unmapped_area;
2214         }
2215
2216         addr = get_area(file, addr, len, pgoff, flags);
2217         if (IS_ERR_VALUE(addr))
2218                 return addr;
2219
2220         if (addr > TASK_SIZE - len)
2221                 return -ENOMEM;
2222         if (offset_in_page(addr))
2223                 return -EINVAL;
2224
2225         error = security_mmap_addr(addr);
2226         return error ? error : addr;
2227 }
2228
2229 EXPORT_SYMBOL(get_unmapped_area);
2230
2231 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
2232 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2233 {
2234         struct rb_node *rb_node;
2235         struct vm_area_struct *vma;
2236
2237         /* Check the cache first. */
2238         vma = vmacache_find(mm, addr);
2239         if (likely(vma))
2240                 return vma;
2241
2242         rb_node = mm->mm_rb.rb_node;
2243
2244         while (rb_node) {
2245                 struct vm_area_struct *tmp;
2246
2247                 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2248
2249                 if (tmp->vm_end > addr) {
2250                         vma = tmp;
2251                         if (tmp->vm_start <= addr)
2252                                 break;
2253                         rb_node = rb_node->rb_left;
2254                 } else
2255                         rb_node = rb_node->rb_right;
2256         }
2257
2258         if (vma)
2259                 vmacache_update(addr, vma);
2260         return vma;
2261 }
2262
2263 EXPORT_SYMBOL(find_vma);
2264
2265 /*
2266  * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2267  */
2268 struct vm_area_struct *
2269 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2270                         struct vm_area_struct **pprev)
2271 {
2272         struct vm_area_struct *vma;
2273
2274         vma = find_vma(mm, addr);
2275         if (vma) {
2276                 *pprev = vma->vm_prev;
2277         } else {
2278                 struct rb_node *rb_node = rb_last(&mm->mm_rb);
2279
2280                 *pprev = rb_node ? rb_entry(rb_node, struct vm_area_struct, vm_rb) : NULL;
2281         }
2282         return vma;
2283 }
2284
2285 /*
2286  * Verify that the stack growth is acceptable and
2287  * update accounting. This is shared with both the
2288  * grow-up and grow-down cases.
2289  */
2290 static int acct_stack_growth(struct vm_area_struct *vma,
2291                              unsigned long size, unsigned long grow)
2292 {
2293         struct mm_struct *mm = vma->vm_mm;
2294         unsigned long new_start;
2295
2296         /* address space limit tests */
2297         if (!may_expand_vm(mm, vma->vm_flags, grow))
2298                 return -ENOMEM;
2299
2300         /* Stack limit test */
2301         if (size > rlimit(RLIMIT_STACK))
2302                 return -ENOMEM;
2303
2304         /* mlock limit tests */
2305         if (vma->vm_flags & VM_LOCKED) {
2306                 unsigned long locked;
2307                 unsigned long limit;
2308                 locked = mm->locked_vm + grow;
2309                 limit = rlimit(RLIMIT_MEMLOCK);
2310                 limit >>= PAGE_SHIFT;
2311                 if (locked > limit && !capable(CAP_IPC_LOCK))
2312                         return -ENOMEM;
2313         }
2314
2315         /* Check to ensure the stack will not grow into a hugetlb-only region */
2316         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2317                         vma->vm_end - size;
2318         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2319                 return -EFAULT;
2320
2321         /*
2322          * Overcommit..  This must be the final test, as it will
2323          * update security statistics.
2324          */
2325         if (security_vm_enough_memory_mm(mm, grow))
2326                 return -ENOMEM;
2327
2328         return 0;
2329 }
2330
2331 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2332 /*
2333  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2334  * vma is the last one with address > vma->vm_end.  Have to extend vma.
2335  */
2336 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2337 {
2338         struct mm_struct *mm = vma->vm_mm;
2339         struct vm_area_struct *next;
2340         unsigned long gap_addr;
2341         int error = 0;
2342
2343         if (!(vma->vm_flags & VM_GROWSUP))
2344                 return -EFAULT;
2345
2346         /* Guard against exceeding limits of the address space. */
2347         address &= PAGE_MASK;
2348         if (address >= (TASK_SIZE & PAGE_MASK))
2349                 return -ENOMEM;
2350         address += PAGE_SIZE;
2351
2352         /* Enforce stack_guard_gap */
2353         gap_addr = address + stack_guard_gap;
2354
2355         /* Guard against overflow */
2356         if (gap_addr < address || gap_addr > TASK_SIZE)
2357                 gap_addr = TASK_SIZE;
2358
2359         next = vma->vm_next;
2360         if (next && next->vm_start < gap_addr && vma_is_accessible(next)) {
2361                 if (!(next->vm_flags & VM_GROWSUP))
2362                         return -ENOMEM;
2363                 /* Check that both stack segments have the same anon_vma? */
2364         }
2365
2366         /* We must make sure the anon_vma is allocated. */
2367         if (unlikely(anon_vma_prepare(vma)))
2368                 return -ENOMEM;
2369
2370         /*
2371          * vma->vm_start/vm_end cannot change under us because the caller
2372          * is required to hold the mmap_sem in read mode.  We need the
2373          * anon_vma lock to serialize against concurrent expand_stacks.
2374          */
2375         anon_vma_lock_write(vma->anon_vma);
2376
2377         /* Somebody else might have raced and expanded it already */
2378         if (address > vma->vm_end) {
2379                 unsigned long size, grow;
2380
2381                 size = address - vma->vm_start;
2382                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2383
2384                 error = -ENOMEM;
2385                 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2386                         error = acct_stack_growth(vma, size, grow);
2387                         if (!error) {
2388                                 /*
2389                                  * vma_gap_update() doesn't support concurrent
2390                                  * updates, but we only hold a shared mmap_sem
2391                                  * lock here, so we need to protect against
2392                                  * concurrent vma expansions.
2393                                  * anon_vma_lock_write() doesn't help here, as
2394                                  * we don't guarantee that all growable vmas
2395                                  * in a mm share the same root anon vma.
2396                                  * So, we reuse mm->page_table_lock to guard
2397                                  * against concurrent vma expansions.
2398                                  */
2399                                 spin_lock(&mm->page_table_lock);
2400                                 if (vma->vm_flags & VM_LOCKED)
2401                                         mm->locked_vm += grow;
2402                                 vm_stat_account(mm, vma->vm_flags, grow);
2403                                 anon_vma_interval_tree_pre_update_vma(vma);
2404                                 vma->vm_end = address;
2405                                 anon_vma_interval_tree_post_update_vma(vma);
2406                                 if (vma->vm_next)
2407                                         vma_gap_update(vma->vm_next);
2408                                 else
2409                                         mm->highest_vm_end = vm_end_gap(vma);
2410                                 spin_unlock(&mm->page_table_lock);
2411
2412                                 perf_event_mmap(vma);
2413                         }
2414                 }
2415         }
2416         anon_vma_unlock_write(vma->anon_vma);
2417         khugepaged_enter_vma_merge(vma, vma->vm_flags);
2418         validate_mm(mm);
2419         return error;
2420 }
2421 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2422
2423 /*
2424  * vma is the first one with address < vma->vm_start.  Have to extend vma.
2425  */
2426 int expand_downwards(struct vm_area_struct *vma,
2427                                    unsigned long address)
2428 {
2429         struct mm_struct *mm = vma->vm_mm;
2430         struct vm_area_struct *prev;
2431         int error = 0;
2432
2433         address &= PAGE_MASK;
2434         if (address < mmap_min_addr)
2435                 return -EPERM;
2436
2437         /* Enforce stack_guard_gap */
2438         prev = vma->vm_prev;
2439         /* Check that both stack segments have the same anon_vma? */
2440         if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2441                         vma_is_accessible(prev)) {
2442                 if (address - prev->vm_end < stack_guard_gap)
2443                         return -ENOMEM;
2444         }
2445
2446         /* We must make sure the anon_vma is allocated. */
2447         if (unlikely(anon_vma_prepare(vma)))
2448                 return -ENOMEM;
2449
2450         /*
2451          * vma->vm_start/vm_end cannot change under us because the caller
2452          * is required to hold the mmap_sem in read mode.  We need the
2453          * anon_vma lock to serialize against concurrent expand_stacks.
2454          */
2455         anon_vma_lock_write(vma->anon_vma);
2456
2457         /* Somebody else might have raced and expanded it already */
2458         if (address < vma->vm_start) {
2459                 unsigned long size, grow;
2460
2461                 size = vma->vm_end - address;
2462                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2463
2464                 error = -ENOMEM;
2465                 if (grow <= vma->vm_pgoff) {
2466                         error = acct_stack_growth(vma, size, grow);
2467                         if (!error) {
2468                                 /*
2469                                  * vma_gap_update() doesn't support concurrent
2470                                  * updates, but we only hold a shared mmap_sem
2471                                  * lock here, so we need to protect against
2472                                  * concurrent vma expansions.
2473                                  * anon_vma_lock_write() doesn't help here, as
2474                                  * we don't guarantee that all growable vmas
2475                                  * in a mm share the same root anon vma.
2476                                  * So, we reuse mm->page_table_lock to guard
2477                                  * against concurrent vma expansions.
2478                                  */
2479                                 spin_lock(&mm->page_table_lock);
2480                                 if (vma->vm_flags & VM_LOCKED)
2481                                         mm->locked_vm += grow;
2482                                 vm_stat_account(mm, vma->vm_flags, grow);
2483                                 anon_vma_interval_tree_pre_update_vma(vma);
2484                                 vma->vm_start = address;
2485                                 vma->vm_pgoff -= grow;
2486                                 anon_vma_interval_tree_post_update_vma(vma);
2487                                 vma_gap_update(vma);
2488                                 spin_unlock(&mm->page_table_lock);
2489
2490                                 perf_event_mmap(vma);
2491                         }
2492                 }
2493         }
2494         anon_vma_unlock_write(vma->anon_vma);
2495         khugepaged_enter_vma_merge(vma, vma->vm_flags);
2496         validate_mm(mm);
2497         return error;
2498 }
2499
2500 /* enforced gap between the expanding stack and other mappings. */
2501 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2502
2503 static int __init cmdline_parse_stack_guard_gap(char *p)
2504 {
2505         unsigned long val;
2506         char *endptr;
2507
2508         val = simple_strtoul(p, &endptr, 10);
2509         if (!*endptr)
2510                 stack_guard_gap = val << PAGE_SHIFT;
2511
2512         return 0;
2513 }
2514 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2515
2516 #ifdef CONFIG_STACK_GROWSUP
2517 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2518 {
2519         return expand_upwards(vma, address);
2520 }
2521
2522 struct vm_area_struct *
2523 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2524 {
2525         struct vm_area_struct *vma, *prev;
2526
2527         addr &= PAGE_MASK;
2528         vma = find_vma_prev(mm, addr, &prev);
2529         if (vma && (vma->vm_start <= addr))
2530                 return vma;
2531         /* don't alter vm_end if the coredump is running */
2532         if (!prev || !mmget_still_valid(mm) || expand_stack(prev, addr))
2533                 return NULL;
2534         if (prev->vm_flags & VM_LOCKED)
2535                 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2536         return prev;
2537 }
2538 #else
2539 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2540 {
2541         return expand_downwards(vma, address);
2542 }
2543
2544 struct vm_area_struct *
2545 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2546 {
2547         struct vm_area_struct *vma;
2548         unsigned long start;
2549
2550         addr &= PAGE_MASK;
2551         vma = find_vma(mm, addr);
2552         if (!vma)
2553                 return NULL;
2554         if (vma->vm_start <= addr)
2555                 return vma;
2556         if (!(vma->vm_flags & VM_GROWSDOWN))
2557                 return NULL;
2558         /* don't alter vm_start if the coredump is running */
2559         if (!mmget_still_valid(mm))
2560                 return NULL;
2561         start = vma->vm_start;
2562         if (expand_stack(vma, addr))
2563                 return NULL;
2564         if (vma->vm_flags & VM_LOCKED)
2565                 populate_vma_page_range(vma, addr, start, NULL);
2566         return vma;
2567 }
2568 #endif
2569
2570 EXPORT_SYMBOL_GPL(find_extend_vma);
2571
2572 /*
2573  * Ok - we have the memory areas we should free on the vma list,
2574  * so release them, and do the vma updates.
2575  *
2576  * Called with the mm semaphore held.
2577  */
2578 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2579 {
2580         unsigned long nr_accounted = 0;
2581
2582         /* Update high watermark before we lower total_vm */
2583         update_hiwater_vm(mm);
2584         do {
2585                 long nrpages = vma_pages(vma);
2586
2587                 if (vma->vm_flags & VM_ACCOUNT)
2588                         nr_accounted += nrpages;
2589                 vm_stat_account(mm, vma->vm_flags, -nrpages);
2590                 vma = remove_vma(vma);
2591         } while (vma);
2592         vm_unacct_memory(nr_accounted);
2593         validate_mm(mm);
2594 }
2595
2596 /*
2597  * Get rid of page table information in the indicated region.
2598  *
2599  * Called with the mm semaphore held.
2600  */
2601 static void unmap_region(struct mm_struct *mm,
2602                 struct vm_area_struct *vma, struct vm_area_struct *prev,
2603                 unsigned long start, unsigned long end)
2604 {
2605         struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap;
2606         struct mmu_gather tlb;
2607
2608         lru_add_drain();
2609         tlb_gather_mmu(&tlb, mm, start, end);
2610         update_hiwater_rss(mm);
2611         unmap_vmas(&tlb, vma, start, end);
2612         free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2613                                  next ? next->vm_start : USER_PGTABLES_CEILING);
2614         tlb_finish_mmu(&tlb, start, end);
2615 }
2616
2617 /*
2618  * Create a list of vma's touched by the unmap, removing them from the mm's
2619  * vma list as we go..
2620  */
2621 static void
2622 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2623         struct vm_area_struct *prev, unsigned long end)
2624 {
2625         struct vm_area_struct **insertion_point;
2626         struct vm_area_struct *tail_vma = NULL;
2627
2628         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2629         vma->vm_prev = NULL;
2630         do {
2631                 vma_rb_erase(vma, &mm->mm_rb);
2632                 mm->map_count--;
2633                 tail_vma = vma;
2634                 vma = vma->vm_next;
2635         } while (vma && vma->vm_start < end);
2636         *insertion_point = vma;
2637         if (vma) {
2638                 vma->vm_prev = prev;
2639                 vma_gap_update(vma);
2640         } else
2641                 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2642         tail_vma->vm_next = NULL;
2643
2644         /* Kill the cache */
2645         vmacache_invalidate(mm);
2646 }
2647
2648 /*
2649  * __split_vma() bypasses sysctl_max_map_count checking.  We use this where it
2650  * has already been checked or doesn't make sense to fail.
2651  */
2652 int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2653                 unsigned long addr, int new_below)
2654 {
2655         struct vm_area_struct *new;
2656         int err;
2657
2658         if (vma->vm_ops && vma->vm_ops->split) {
2659                 err = vma->vm_ops->split(vma, addr);
2660                 if (err)
2661                         return err;
2662         }
2663
2664         new = vm_area_dup(vma);
2665         if (!new)
2666                 return -ENOMEM;
2667
2668         if (new_below)
2669                 new->vm_end = addr;
2670         else {
2671                 new->vm_start = addr;
2672                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2673         }
2674
2675         err = vma_dup_policy(vma, new);
2676         if (err)
2677                 goto out_free_vma;
2678
2679         err = anon_vma_clone(new, vma);
2680         if (err)
2681                 goto out_free_mpol;
2682
2683         if (new->vm_file)
2684                 get_file(new->vm_file);
2685
2686         if (new->vm_ops && new->vm_ops->open)
2687                 new->vm_ops->open(new);
2688
2689         if (new_below)
2690                 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2691                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
2692         else
2693                 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2694
2695         /* Success. */
2696         if (!err)
2697                 return 0;
2698
2699         /* Clean everything up if vma_adjust failed. */
2700         if (new->vm_ops && new->vm_ops->close)
2701                 new->vm_ops->close(new);
2702         if (new->vm_file)
2703                 fput(new->vm_file);
2704         unlink_anon_vmas(new);
2705  out_free_mpol:
2706         mpol_put(vma_policy(new));
2707  out_free_vma:
2708         vm_area_free(new);
2709         return err;
2710 }
2711
2712 /*
2713  * Split a vma into two pieces at address 'addr', a new vma is allocated
2714  * either for the first part or the tail.
2715  */
2716 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2717               unsigned long addr, int new_below)
2718 {
2719         if (mm->map_count >= sysctl_max_map_count)
2720                 return -ENOMEM;
2721
2722         return __split_vma(mm, vma, addr, new_below);
2723 }
2724
2725 /* Munmap is split into 2 main parts -- this part which finds
2726  * what needs doing, and the areas themselves, which do the
2727  * work.  This now handles partial unmappings.
2728  * Jeremy Fitzhardinge <jeremy@goop.org>
2729  */
2730 int __do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2731                 struct list_head *uf, bool downgrade)
2732 {
2733         unsigned long end;
2734         struct vm_area_struct *vma, *prev, *last;
2735
2736         if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2737                 return -EINVAL;
2738
2739         len = PAGE_ALIGN(len);
2740         end = start + len;
2741         if (len == 0)
2742                 return -EINVAL;
2743
2744         /*
2745          * arch_unmap() might do unmaps itself.  It must be called
2746          * and finish any rbtree manipulation before this code
2747          * runs and also starts to manipulate the rbtree.
2748          */
2749         arch_unmap(mm, start, end);
2750
2751         /* Find the first overlapping VMA */
2752         vma = find_vma(mm, start);
2753         if (!vma)
2754                 return 0;
2755         prev = vma->vm_prev;
2756         /* we have  start < vma->vm_end  */
2757
2758         /* if it doesn't overlap, we have nothing.. */
2759         if (vma->vm_start >= end)
2760                 return 0;
2761
2762         /*
2763          * If we need to split any vma, do it now to save pain later.
2764          *
2765          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2766          * unmapped vm_area_struct will remain in use: so lower split_vma
2767          * places tmp vma above, and higher split_vma places tmp vma below.
2768          */
2769         if (start > vma->vm_start) {
2770                 int error;
2771
2772                 /*
2773                  * Make sure that map_count on return from munmap() will
2774                  * not exceed its limit; but let map_count go just above
2775                  * its limit temporarily, to help free resources as expected.
2776                  */
2777                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2778                         return -ENOMEM;
2779
2780                 error = __split_vma(mm, vma, start, 0);
2781                 if (error)
2782                         return error;
2783                 prev = vma;
2784         }
2785
2786         /* Does it split the last one? */
2787         last = find_vma(mm, end);
2788         if (last && end > last->vm_start) {
2789                 int error = __split_vma(mm, last, end, 1);
2790                 if (error)
2791                         return error;
2792         }
2793         vma = prev ? prev->vm_next : mm->mmap;
2794
2795         if (unlikely(uf)) {
2796                 /*
2797                  * If userfaultfd_unmap_prep returns an error the vmas
2798                  * will remain splitted, but userland will get a
2799                  * highly unexpected error anyway. This is no
2800                  * different than the case where the first of the two
2801                  * __split_vma fails, but we don't undo the first
2802                  * split, despite we could. This is unlikely enough
2803                  * failure that it's not worth optimizing it for.
2804                  */
2805                 int error = userfaultfd_unmap_prep(vma, start, end, uf);
2806                 if (error)
2807                         return error;
2808         }
2809
2810         /*
2811          * unlock any mlock()ed ranges before detaching vmas
2812          */
2813         if (mm->locked_vm) {
2814                 struct vm_area_struct *tmp = vma;
2815                 while (tmp && tmp->vm_start < end) {
2816                         if (tmp->vm_flags & VM_LOCKED) {
2817                                 mm->locked_vm -= vma_pages(tmp);
2818                                 munlock_vma_pages_all(tmp);
2819                         }
2820
2821                         tmp = tmp->vm_next;
2822                 }
2823         }
2824
2825         /* Detach vmas from rbtree */
2826         detach_vmas_to_be_unmapped(mm, vma, prev, end);
2827
2828         if (downgrade)
2829                 downgrade_write(&mm->mmap_sem);
2830
2831         unmap_region(mm, vma, prev, start, end);
2832
2833         /* Fix up all other VM information */
2834         remove_vma_list(mm, vma);
2835
2836         return downgrade ? 1 : 0;
2837 }
2838
2839 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2840               struct list_head *uf)
2841 {
2842         return __do_munmap(mm, start, len, uf, false);
2843 }
2844
2845 static int __vm_munmap(unsigned long start, size_t len, bool downgrade)
2846 {
2847         int ret;
2848         struct mm_struct *mm = current->mm;
2849         LIST_HEAD(uf);
2850
2851         if (down_write_killable(&mm->mmap_sem))
2852                 return -EINTR;
2853
2854         ret = __do_munmap(mm, start, len, &uf, downgrade);
2855         /*
2856          * Returning 1 indicates mmap_sem is downgraded.
2857          * But 1 is not legal return value of vm_munmap() and munmap(), reset
2858          * it to 0 before return.
2859          */
2860         if (ret == 1) {
2861                 up_read(&mm->mmap_sem);
2862                 ret = 0;
2863         } else
2864                 up_write(&mm->mmap_sem);
2865
2866         userfaultfd_unmap_complete(mm, &uf);
2867         return ret;
2868 }
2869
2870 int vm_munmap(unsigned long start, size_t len)
2871 {
2872         return __vm_munmap(start, len, false);
2873 }
2874 EXPORT_SYMBOL(vm_munmap);
2875
2876 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2877 {
2878         addr = untagged_addr(addr);
2879         profile_munmap(addr);
2880         return __vm_munmap(addr, len, true);
2881 }
2882
2883
2884 /*
2885  * Emulation of deprecated remap_file_pages() syscall.
2886  */
2887 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2888                 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2889 {
2890
2891         struct mm_struct *mm = current->mm;
2892         struct vm_area_struct *vma;
2893         unsigned long populate = 0;
2894         unsigned long ret = -EINVAL;
2895         struct file *file;
2896
2897         pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.rst.\n",
2898                      current->comm, current->pid);
2899
2900         if (prot)
2901                 return ret;
2902         start = start & PAGE_MASK;
2903         size = size & PAGE_MASK;
2904
2905         if (start + size <= start)
2906                 return ret;
2907
2908         /* Does pgoff wrap? */
2909         if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2910                 return ret;
2911
2912         if (down_write_killable(&mm->mmap_sem))
2913                 return -EINTR;
2914
2915         vma = find_vma(mm, start);
2916
2917         if (!vma || !(vma->vm_flags & VM_SHARED))
2918                 goto out;
2919
2920         if (start < vma->vm_start)
2921                 goto out;
2922
2923         if (start + size > vma->vm_end) {
2924                 struct vm_area_struct *next;
2925
2926                 for (next = vma->vm_next; next; next = next->vm_next) {
2927                         /* hole between vmas ? */
2928                         if (next->vm_start != next->vm_prev->vm_end)
2929                                 goto out;
2930
2931                         if (next->vm_file != vma->vm_file)
2932                                 goto out;
2933
2934                         if (next->vm_flags != vma->vm_flags)
2935                                 goto out;
2936
2937                         if (start + size <= next->vm_end)
2938                                 break;
2939                 }
2940
2941                 if (!next)
2942                         goto out;
2943         }
2944
2945         prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2946         prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2947         prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2948
2949         flags &= MAP_NONBLOCK;
2950         flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2951         if (vma->vm_flags & VM_LOCKED) {
2952                 struct vm_area_struct *tmp;
2953                 flags |= MAP_LOCKED;
2954
2955                 /* drop PG_Mlocked flag for over-mapped range */
2956                 for (tmp = vma; tmp->vm_start >= start + size;
2957                                 tmp = tmp->vm_next) {
2958                         /*
2959                          * Split pmd and munlock page on the border
2960                          * of the range.
2961                          */
2962                         vma_adjust_trans_huge(tmp, start, start + size, 0);
2963
2964                         munlock_vma_pages_range(tmp,
2965                                         max(tmp->vm_start, start),
2966                                         min(tmp->vm_end, start + size));
2967                 }
2968         }
2969
2970         file = get_file(vma->vm_file);
2971         ret = do_mmap_pgoff(vma->vm_file, start, size,
2972                         prot, flags, pgoff, &populate, NULL);
2973         fput(file);
2974 out:
2975         up_write(&mm->mmap_sem);
2976         if (populate)
2977                 mm_populate(ret, populate);
2978         if (!IS_ERR_VALUE(ret))
2979                 ret = 0;
2980         return ret;
2981 }
2982
2983 /*
2984  *  this is really a simplified "do_mmap".  it only handles
2985  *  anonymous maps.  eventually we may be able to do some
2986  *  brk-specific accounting here.
2987  */
2988 static int do_brk_flags(unsigned long addr, unsigned long len, unsigned long flags, struct list_head *uf)
2989 {
2990         struct mm_struct *mm = current->mm;
2991         struct vm_area_struct *vma, *prev;
2992         struct rb_node **rb_link, *rb_parent;
2993         pgoff_t pgoff = addr >> PAGE_SHIFT;
2994         int error;
2995         unsigned long mapped_addr;
2996
2997         /* Until we need other flags, refuse anything except VM_EXEC. */
2998         if ((flags & (~VM_EXEC)) != 0)
2999                 return -EINVAL;
3000         flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
3001
3002         mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
3003         if (IS_ERR_VALUE(mapped_addr))
3004                 return mapped_addr;
3005
3006         error = mlock_future_check(mm, mm->def_flags, len);
3007         if (error)
3008                 return error;
3009
3010         /*
3011          * Clear old maps.  this also does some error checking for us
3012          */
3013         while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
3014                               &rb_parent)) {
3015                 if (do_munmap(mm, addr, len, uf))
3016                         return -ENOMEM;
3017         }
3018
3019         /* Check against address space limits *after* clearing old maps... */
3020         if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
3021                 return -ENOMEM;
3022
3023         if (mm->map_count > sysctl_max_map_count)
3024                 return -ENOMEM;
3025
3026         if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
3027                 return -ENOMEM;
3028
3029         /* Can we just expand an old private anonymous mapping? */
3030         vma = vma_merge(mm, prev, addr, addr + len, flags,
3031                         NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
3032         if (vma)
3033                 goto out;
3034
3035         /*
3036          * create a vma struct for an anonymous mapping
3037          */
3038         vma = vm_area_alloc(mm);
3039         if (!vma) {
3040                 vm_unacct_memory(len >> PAGE_SHIFT);
3041                 return -ENOMEM;
3042         }
3043
3044         vma_set_anonymous(vma);
3045         vma->vm_start = addr;
3046         vma->vm_end = addr + len;
3047         vma->vm_pgoff = pgoff;
3048         vma->vm_flags = flags;
3049         vma->vm_page_prot = vm_get_page_prot(flags);
3050         vma_link(mm, vma, prev, rb_link, rb_parent);
3051 out:
3052         perf_event_mmap(vma);
3053         mm->total_vm += len >> PAGE_SHIFT;
3054         mm->data_vm += len >> PAGE_SHIFT;
3055         if (flags & VM_LOCKED)
3056                 mm->locked_vm += (len >> PAGE_SHIFT);
3057         vma->vm_flags |= VM_SOFTDIRTY;
3058         return 0;
3059 }
3060
3061 int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3062 {
3063         struct mm_struct *mm = current->mm;
3064         unsigned long len;
3065         int ret;
3066         bool populate;
3067         LIST_HEAD(uf);
3068
3069         len = PAGE_ALIGN(request);
3070         if (len < request)
3071                 return -ENOMEM;
3072         if (!len)
3073                 return 0;
3074
3075         if (down_write_killable(&mm->mmap_sem))
3076                 return -EINTR;
3077
3078         ret = do_brk_flags(addr, len, flags, &uf);
3079         populate = ((mm->def_flags & VM_LOCKED) != 0);
3080         up_write(&mm->mmap_sem);
3081         userfaultfd_unmap_complete(mm, &uf);
3082         if (populate && !ret)
3083                 mm_populate(addr, len);
3084         return ret;
3085 }
3086 EXPORT_SYMBOL(vm_brk_flags);
3087
3088 int vm_brk(unsigned long addr, unsigned long len)
3089 {
3090         return vm_brk_flags(addr, len, 0);
3091 }
3092 EXPORT_SYMBOL(vm_brk);
3093
3094 /* Release all mmaps. */
3095 void exit_mmap(struct mm_struct *mm)
3096 {
3097         struct mmu_gather tlb;
3098         struct vm_area_struct *vma;
3099         unsigned long nr_accounted = 0;
3100
3101         /* mm's last user has gone, and its about to be pulled down */
3102         mmu_notifier_release(mm);
3103
3104         if (unlikely(mm_is_oom_victim(mm))) {
3105                 /*
3106                  * Manually reap the mm to free as much memory as possible.
3107                  * Then, as the oom reaper does, set MMF_OOM_SKIP to disregard
3108                  * this mm from further consideration.  Taking mm->mmap_sem for
3109                  * write after setting MMF_OOM_SKIP will guarantee that the oom
3110                  * reaper will not run on this mm again after mmap_sem is
3111                  * dropped.
3112                  *
3113                  * Nothing can be holding mm->mmap_sem here and the above call
3114                  * to mmu_notifier_release(mm) ensures mmu notifier callbacks in
3115                  * __oom_reap_task_mm() will not block.
3116                  *
3117                  * This needs to be done before calling munlock_vma_pages_all(),
3118                  * which clears VM_LOCKED, otherwise the oom reaper cannot
3119                  * reliably test it.
3120                  */
3121                 (void)__oom_reap_task_mm(mm);
3122
3123                 set_bit(MMF_OOM_SKIP, &mm->flags);
3124                 down_write(&mm->mmap_sem);
3125                 up_write(&mm->mmap_sem);
3126         }
3127
3128         if (mm->locked_vm) {
3129                 vma = mm->mmap;
3130                 while (vma) {
3131                         if (vma->vm_flags & VM_LOCKED)
3132                                 munlock_vma_pages_all(vma);
3133                         vma = vma->vm_next;
3134                 }
3135         }
3136
3137         arch_exit_mmap(mm);
3138
3139         vma = mm->mmap;
3140         if (!vma)       /* Can happen if dup_mmap() received an OOM */
3141                 return;
3142
3143         lru_add_drain();
3144         flush_cache_mm(mm);
3145         tlb_gather_mmu(&tlb, mm, 0, -1);
3146         /* update_hiwater_rss(mm) here? but nobody should be looking */
3147         /* Use -1 here to ensure all VMAs in the mm are unmapped */
3148         unmap_vmas(&tlb, vma, 0, -1);
3149         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3150         tlb_finish_mmu(&tlb, 0, -1);
3151
3152         /*
3153          * Walk the list again, actually closing and freeing it,
3154          * with preemption enabled, without holding any MM locks.
3155          */
3156         while (vma) {
3157                 if (vma->vm_flags & VM_ACCOUNT)
3158                         nr_accounted += vma_pages(vma);
3159                 vma = remove_vma(vma);
3160         }
3161         vm_unacct_memory(nr_accounted);
3162 }
3163
3164 /* Insert vm structure into process list sorted by address
3165  * and into the inode's i_mmap tree.  If vm_file is non-NULL
3166  * then i_mmap_rwsem is taken here.
3167  */
3168 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3169 {
3170         struct vm_area_struct *prev;
3171         struct rb_node **rb_link, *rb_parent;
3172
3173         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3174                            &prev, &rb_link, &rb_parent))
3175                 return -ENOMEM;
3176         if ((vma->vm_flags & VM_ACCOUNT) &&
3177              security_vm_enough_memory_mm(mm, vma_pages(vma)))
3178                 return -ENOMEM;
3179
3180         /*
3181          * The vm_pgoff of a purely anonymous vma should be irrelevant
3182          * until its first write fault, when page's anon_vma and index
3183          * are set.  But now set the vm_pgoff it will almost certainly
3184          * end up with (unless mremap moves it elsewhere before that
3185          * first wfault), so /proc/pid/maps tells a consistent story.
3186          *
3187          * By setting it to reflect the virtual start address of the
3188          * vma, merges and splits can happen in a seamless way, just
3189          * using the existing file pgoff checks and manipulations.
3190          * Similarly in do_mmap_pgoff and in do_brk.
3191          */
3192         if (vma_is_anonymous(vma)) {
3193                 BUG_ON(vma->anon_vma);
3194                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3195         }
3196
3197         vma_link(mm, vma, prev, rb_link, rb_parent);
3198         return 0;
3199 }
3200
3201 /*
3202  * Copy the vma structure to a new location in the same mm,
3203  * prior to moving page table entries, to effect an mremap move.
3204  */
3205 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3206         unsigned long addr, unsigned long len, pgoff_t pgoff,
3207         bool *need_rmap_locks)
3208 {
3209         struct vm_area_struct *vma = *vmap;
3210         unsigned long vma_start = vma->vm_start;
3211         struct mm_struct *mm = vma->vm_mm;
3212         struct vm_area_struct *new_vma, *prev;
3213         struct rb_node **rb_link, *rb_parent;
3214         bool faulted_in_anon_vma = true;
3215
3216         /*
3217          * If anonymous vma has not yet been faulted, update new pgoff
3218          * to match new location, to increase its chance of merging.
3219          */
3220         if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3221                 pgoff = addr >> PAGE_SHIFT;
3222                 faulted_in_anon_vma = false;
3223         }
3224
3225         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3226                 return NULL;    /* should never get here */
3227         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3228                             vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3229                             vma->vm_userfaultfd_ctx);
3230         if (new_vma) {
3231                 /*
3232                  * Source vma may have been merged into new_vma
3233                  */
3234                 if (unlikely(vma_start >= new_vma->vm_start &&
3235                              vma_start < new_vma->vm_end)) {
3236                         /*
3237                          * The only way we can get a vma_merge with
3238                          * self during an mremap is if the vma hasn't
3239                          * been faulted in yet and we were allowed to
3240                          * reset the dst vma->vm_pgoff to the
3241                          * destination address of the mremap to allow
3242                          * the merge to happen. mremap must change the
3243                          * vm_pgoff linearity between src and dst vmas
3244                          * (in turn preventing a vma_merge) to be
3245                          * safe. It is only safe to keep the vm_pgoff
3246                          * linear if there are no pages mapped yet.
3247                          */
3248                         VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3249                         *vmap = vma = new_vma;
3250                 }
3251                 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3252         } else {
3253                 new_vma = vm_area_dup(vma);
3254                 if (!new_vma)
3255                         goto out;
3256                 new_vma->vm_start = addr;
3257                 new_vma->vm_end = addr + len;
3258                 new_vma->vm_pgoff = pgoff;
3259                 if (vma_dup_policy(vma, new_vma))
3260                         goto out_free_vma;
3261                 if (anon_vma_clone(new_vma, vma))
3262                         goto out_free_mempol;
3263                 if (new_vma->vm_file)
3264                         get_file(new_vma->vm_file);
3265                 if (new_vma->vm_ops && new_vma->vm_ops->open)
3266                         new_vma->vm_ops->open(new_vma);
3267                 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3268                 *need_rmap_locks = false;
3269         }
3270         return new_vma;
3271
3272 out_free_mempol:
3273         mpol_put(vma_policy(new_vma));
3274 out_free_vma:
3275         vm_area_free(new_vma);
3276 out:
3277         return NULL;
3278 }
3279
3280 /*
3281  * Return true if the calling process may expand its vm space by the passed
3282  * number of pages
3283  */
3284 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3285 {
3286         if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3287                 return false;
3288
3289         if (is_data_mapping(flags) &&
3290             mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3291                 /* Workaround for Valgrind */
3292                 if (rlimit(RLIMIT_DATA) == 0 &&
3293                     mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3294                         return true;
3295
3296                 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3297                              current->comm, current->pid,
3298                              (mm->data_vm + npages) << PAGE_SHIFT,
3299                              rlimit(RLIMIT_DATA),
3300                              ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3301
3302                 if (!ignore_rlimit_data)
3303                         return false;
3304         }
3305
3306         return true;
3307 }
3308
3309 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3310 {
3311         mm->total_vm += npages;
3312
3313         if (is_exec_mapping(flags))
3314                 mm->exec_vm += npages;
3315         else if (is_stack_mapping(flags))
3316                 mm->stack_vm += npages;
3317         else if (is_data_mapping(flags))
3318                 mm->data_vm += npages;
3319 }
3320
3321 static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3322
3323 /*
3324  * Having a close hook prevents vma merging regardless of flags.
3325  */
3326 static void special_mapping_close(struct vm_area_struct *vma)
3327 {
3328 }
3329
3330 static const char *special_mapping_name(struct vm_area_struct *vma)
3331 {
3332         return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3333 }
3334
3335 static int special_mapping_mremap(struct vm_area_struct *new_vma)
3336 {
3337         struct vm_special_mapping *sm = new_vma->vm_private_data;
3338
3339         if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3340                 return -EFAULT;
3341
3342         if (sm->mremap)
3343                 return sm->mremap(sm, new_vma);
3344
3345         return 0;
3346 }
3347
3348 static const struct vm_operations_struct special_mapping_vmops = {
3349         .close = special_mapping_close,
3350         .fault = special_mapping_fault,
3351         .mremap = special_mapping_mremap,
3352         .name = special_mapping_name,
3353         /* vDSO code relies that VVAR can't be accessed remotely */
3354         .access = NULL,
3355 };
3356
3357 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3358         .close = special_mapping_close,
3359         .fault = special_mapping_fault,
3360 };
3361
3362 static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3363 {
3364         struct vm_area_struct *vma = vmf->vma;
3365         pgoff_t pgoff;
3366         struct page **pages;
3367
3368         if (vma->vm_ops == &legacy_special_mapping_vmops) {
3369                 pages = vma->vm_private_data;
3370         } else {
3371                 struct vm_special_mapping *sm = vma->vm_private_data;
3372
3373                 if (sm->fault)
3374                         return sm->fault(sm, vmf->vma, vmf);
3375
3376                 pages = sm->pages;
3377         }
3378
3379         for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3380                 pgoff--;
3381
3382         if (*pages) {
3383                 struct page *page = *pages;
3384                 get_page(page);
3385                 vmf->page = page;
3386                 return 0;
3387         }
3388
3389         return VM_FAULT_SIGBUS;
3390 }
3391
3392 static struct vm_area_struct *__install_special_mapping(
3393         struct mm_struct *mm,
3394         unsigned long addr, unsigned long len,
3395         unsigned long vm_flags, void *priv,
3396         const struct vm_operations_struct *ops)
3397 {
3398         int ret;
3399         struct vm_area_struct *vma;
3400
3401         vma = vm_area_alloc(mm);
3402         if (unlikely(vma == NULL))
3403                 return ERR_PTR(-ENOMEM);
3404
3405         vma->vm_start = addr;
3406         vma->vm_end = addr + len;
3407
3408         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3409         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3410
3411         vma->vm_ops = ops;
3412         vma->vm_private_data = priv;
3413
3414         ret = insert_vm_struct(mm, vma);
3415         if (ret)
3416                 goto out;
3417
3418         vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3419
3420         perf_event_mmap(vma);
3421
3422         return vma;
3423
3424 out:
3425         vm_area_free(vma);
3426         return ERR_PTR(ret);
3427 }
3428
3429 bool vma_is_special_mapping(const struct vm_area_struct *vma,
3430         const struct vm_special_mapping *sm)
3431 {
3432         return vma->vm_private_data == sm &&
3433                 (vma->vm_ops == &special_mapping_vmops ||
3434                  vma->vm_ops == &legacy_special_mapping_vmops);
3435 }
3436
3437 /*
3438  * Called with mm->mmap_sem held for writing.
3439  * Insert a new vma covering the given region, with the given flags.
3440  * Its pages are supplied by the given array of struct page *.
3441  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3442  * The region past the last page supplied will always produce SIGBUS.
3443  * The array pointer and the pages it points to are assumed to stay alive
3444  * for as long as this mapping might exist.
3445  */
3446 struct vm_area_struct *_install_special_mapping(
3447         struct mm_struct *mm,
3448         unsigned long addr, unsigned long len,
3449         unsigned long vm_flags, const struct vm_special_mapping *spec)
3450 {
3451         return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3452                                         &special_mapping_vmops);
3453 }
3454
3455 int install_special_mapping(struct mm_struct *mm,
3456                             unsigned long addr, unsigned long len,
3457                             unsigned long vm_flags, struct page **pages)
3458 {
3459         struct vm_area_struct *vma = __install_special_mapping(
3460                 mm, addr, len, vm_flags, (void *)pages,
3461                 &legacy_special_mapping_vmops);
3462
3463         return PTR_ERR_OR_ZERO(vma);
3464 }
3465
3466 static DEFINE_MUTEX(mm_all_locks_mutex);
3467
3468 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3469 {
3470         if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3471                 /*
3472                  * The LSB of head.next can't change from under us
3473                  * because we hold the mm_all_locks_mutex.
3474                  */
3475                 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
3476                 /*
3477                  * We can safely modify head.next after taking the
3478                  * anon_vma->root->rwsem. If some other vma in this mm shares
3479                  * the same anon_vma we won't take it again.
3480                  *
3481                  * No need of atomic instructions here, head.next
3482                  * can't change from under us thanks to the
3483                  * anon_vma->root->rwsem.
3484                  */
3485                 if (__test_and_set_bit(0, (unsigned long *)
3486                                        &anon_vma->root->rb_root.rb_root.rb_node))
3487                         BUG();
3488         }
3489 }
3490
3491 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3492 {
3493         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3494                 /*
3495                  * AS_MM_ALL_LOCKS can't change from under us because
3496                  * we hold the mm_all_locks_mutex.
3497                  *
3498                  * Operations on ->flags have to be atomic because
3499                  * even if AS_MM_ALL_LOCKS is stable thanks to the
3500                  * mm_all_locks_mutex, there may be other cpus
3501                  * changing other bitflags in parallel to us.
3502                  */
3503                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3504                         BUG();
3505                 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem);
3506         }
3507 }
3508
3509 /*
3510  * This operation locks against the VM for all pte/vma/mm related
3511  * operations that could ever happen on a certain mm. This includes
3512  * vmtruncate, try_to_unmap, and all page faults.
3513  *
3514  * The caller must take the mmap_sem in write mode before calling
3515  * mm_take_all_locks(). The caller isn't allowed to release the
3516  * mmap_sem until mm_drop_all_locks() returns.
3517  *
3518  * mmap_sem in write mode is required in order to block all operations
3519  * that could modify pagetables and free pages without need of
3520  * altering the vma layout. It's also needed in write mode to avoid new
3521  * anon_vmas to be associated with existing vmas.
3522  *
3523  * A single task can't take more than one mm_take_all_locks() in a row
3524  * or it would deadlock.
3525  *
3526  * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3527  * mapping->flags avoid to take the same lock twice, if more than one
3528  * vma in this mm is backed by the same anon_vma or address_space.
3529  *
3530  * We take locks in following order, accordingly to comment at beginning
3531  * of mm/rmap.c:
3532  *   - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3533  *     hugetlb mapping);
3534  *   - all i_mmap_rwsem locks;
3535  *   - all anon_vma->rwseml
3536  *
3537  * We can take all locks within these types randomly because the VM code
3538  * doesn't nest them and we protected from parallel mm_take_all_locks() by
3539  * mm_all_locks_mutex.
3540  *
3541  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3542  * that may have to take thousand of locks.
3543  *
3544  * mm_take_all_locks() can fail if it's interrupted by signals.
3545  */
3546 int mm_take_all_locks(struct mm_struct *mm)
3547 {
3548         struct vm_area_struct *vma;
3549         struct anon_vma_chain *avc;
3550
3551         BUG_ON(down_read_trylock(&mm->mmap_sem));
3552
3553         mutex_lock(&mm_all_locks_mutex);
3554
3555         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3556                 if (signal_pending(current))
3557                         goto out_unlock;
3558                 if (vma->vm_file && vma->vm_file->f_mapping &&
3559                                 is_vm_hugetlb_page(vma))
3560                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
3561         }
3562
3563         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3564                 if (signal_pending(current))
3565                         goto out_unlock;
3566                 if (vma->vm_file && vma->vm_file->f_mapping &&
3567                                 !is_vm_hugetlb_page(vma))
3568                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
3569         }
3570
3571         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3572                 if (signal_pending(current))
3573                         goto out_unlock;
3574                 if (vma->anon_vma)
3575                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3576                                 vm_lock_anon_vma(mm, avc->anon_vma);
3577         }
3578
3579         return 0;
3580
3581 out_unlock:
3582         mm_drop_all_locks(mm);
3583         return -EINTR;
3584 }
3585
3586 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3587 {
3588         if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3589                 /*
3590                  * The LSB of head.next can't change to 0 from under
3591                  * us because we hold the mm_all_locks_mutex.
3592                  *
3593                  * We must however clear the bitflag before unlocking
3594                  * the vma so the users using the anon_vma->rb_root will
3595                  * never see our bitflag.
3596                  *
3597                  * No need of atomic instructions here, head.next
3598                  * can't change from under us until we release the
3599                  * anon_vma->root->rwsem.
3600                  */
3601                 if (!__test_and_clear_bit(0, (unsigned long *)
3602                                           &anon_vma->root->rb_root.rb_root.rb_node))
3603                         BUG();
3604                 anon_vma_unlock_write(anon_vma);
3605         }
3606 }
3607
3608 static void vm_unlock_mapping(struct address_space *mapping)
3609 {
3610         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3611                 /*
3612                  * AS_MM_ALL_LOCKS can't change to 0 from under us
3613                  * because we hold the mm_all_locks_mutex.
3614                  */
3615                 i_mmap_unlock_write(mapping);
3616                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3617                                         &mapping->flags))
3618                         BUG();
3619         }
3620 }
3621
3622 /*
3623  * The mmap_sem cannot be released by the caller until
3624  * mm_drop_all_locks() returns.
3625  */
3626 void mm_drop_all_locks(struct mm_struct *mm)
3627 {
3628         struct vm_area_struct *vma;
3629         struct anon_vma_chain *avc;
3630
3631         BUG_ON(down_read_trylock(&mm->mmap_sem));
3632         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3633
3634         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3635                 if (vma->anon_vma)
3636                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3637                                 vm_unlock_anon_vma(avc->anon_vma);
3638                 if (vma->vm_file && vma->vm_file->f_mapping)
3639                         vm_unlock_mapping(vma->vm_file->f_mapping);
3640         }
3641
3642         mutex_unlock(&mm_all_locks_mutex);
3643 }
3644
3645 /*
3646  * initialise the percpu counter for VM
3647  */
3648 void __init mmap_init(void)
3649 {
3650         int ret;
3651
3652         ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3653         VM_BUG_ON(ret);
3654 }
3655
3656 /*
3657  * Initialise sysctl_user_reserve_kbytes.
3658  *
3659  * This is intended to prevent a user from starting a single memory hogging
3660  * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3661  * mode.
3662  *
3663  * The default value is min(3% of free memory, 128MB)
3664  * 128MB is enough to recover with sshd/login, bash, and top/kill.
3665  */
3666 static int init_user_reserve(void)
3667 {
3668         unsigned long free_kbytes;
3669
3670         free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3671
3672         sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3673         return 0;
3674 }
3675 subsys_initcall(init_user_reserve);
3676
3677 /*
3678  * Initialise sysctl_admin_reserve_kbytes.
3679  *
3680  * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3681  * to log in and kill a memory hogging process.
3682  *
3683  * Systems with more than 256MB will reserve 8MB, enough to recover
3684  * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3685  * only reserve 3% of free pages by default.
3686  */
3687 static int init_admin_reserve(void)
3688 {
3689         unsigned long free_kbytes;
3690
3691         free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3692
3693         sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3694         return 0;
3695 }
3696 subsys_initcall(init_admin_reserve);
3697
3698 /*
3699  * Reinititalise user and admin reserves if memory is added or removed.
3700  *
3701  * The default user reserve max is 128MB, and the default max for the
3702  * admin reserve is 8MB. These are usually, but not always, enough to
3703  * enable recovery from a memory hogging process using login/sshd, a shell,
3704  * and tools like top. It may make sense to increase or even disable the
3705  * reserve depending on the existence of swap or variations in the recovery
3706  * tools. So, the admin may have changed them.
3707  *
3708  * If memory is added and the reserves have been eliminated or increased above
3709  * the default max, then we'll trust the admin.
3710  *
3711  * If memory is removed and there isn't enough free memory, then we
3712  * need to reset the reserves.
3713  *
3714  * Otherwise keep the reserve set by the admin.
3715  */
3716 static int reserve_mem_notifier(struct notifier_block *nb,
3717                              unsigned long action, void *data)
3718 {
3719         unsigned long tmp, free_kbytes;
3720
3721         switch (action) {
3722         case MEM_ONLINE:
3723                 /* Default max is 128MB. Leave alone if modified by operator. */
3724                 tmp = sysctl_user_reserve_kbytes;
3725                 if (0 < tmp && tmp < (1UL << 17))
3726                         init_user_reserve();
3727
3728                 /* Default max is 8MB.  Leave alone if modified by operator. */
3729                 tmp = sysctl_admin_reserve_kbytes;
3730                 if (0 < tmp && tmp < (1UL << 13))
3731                         init_admin_reserve();
3732
3733                 break;
3734         case MEM_OFFLINE:
3735                 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3736
3737                 if (sysctl_user_reserve_kbytes > free_kbytes) {
3738                         init_user_reserve();
3739                         pr_info("vm.user_reserve_kbytes reset to %lu\n",
3740                                 sysctl_user_reserve_kbytes);
3741                 }
3742
3743                 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3744                         init_admin_reserve();
3745                         pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3746                                 sysctl_admin_reserve_kbytes);
3747                 }
3748                 break;
3749         default:
3750                 break;
3751         }
3752         return NOTIFY_OK;
3753 }
3754
3755 static struct notifier_block reserve_mem_nb = {
3756         .notifier_call = reserve_mem_notifier,
3757 };
3758
3759 static int __meminit init_reserve_notifier(void)
3760 {
3761         if (register_hotmemory_notifier(&reserve_mem_nb))
3762                 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3763
3764         return 0;
3765 }
3766 subsys_initcall(init_reserve_notifier);