mm: ignore MAP_DENYWRITE in ksys_mmap_pgoff()
[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  * On arm64, PROT_EXEC has the following behaviour for both MAP_SHARED and
98  * MAP_PRIVATE (with Enhanced PAN supported):
99  *                                                              r: (no) no
100  *                                                              w: (no) no
101  *                                                              x: (yes) yes
102  */
103 pgprot_t protection_map[16] __ro_after_init = {
104         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
105         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
106 };
107
108 #ifndef CONFIG_ARCH_HAS_FILTER_PGPROT
109 static inline pgprot_t arch_filter_pgprot(pgprot_t prot)
110 {
111         return prot;
112 }
113 #endif
114
115 pgprot_t vm_get_page_prot(unsigned long vm_flags)
116 {
117         pgprot_t ret = __pgprot(pgprot_val(protection_map[vm_flags &
118                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
119                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
120
121         return arch_filter_pgprot(ret);
122 }
123 EXPORT_SYMBOL(vm_get_page_prot);
124
125 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
126 {
127         return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
128 }
129
130 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
131 void vma_set_page_prot(struct vm_area_struct *vma)
132 {
133         unsigned long vm_flags = vma->vm_flags;
134         pgprot_t vm_page_prot;
135
136         vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
137         if (vma_wants_writenotify(vma, vm_page_prot)) {
138                 vm_flags &= ~VM_SHARED;
139                 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
140         }
141         /* remove_protection_ptes reads vma->vm_page_prot without mmap_lock */
142         WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
143 }
144
145 /*
146  * Requires inode->i_mapping->i_mmap_rwsem
147  */
148 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
149                 struct file *file, struct address_space *mapping)
150 {
151         if (vma->vm_flags & VM_SHARED)
152                 mapping_unmap_writable(mapping);
153
154         flush_dcache_mmap_lock(mapping);
155         vma_interval_tree_remove(vma, &mapping->i_mmap);
156         flush_dcache_mmap_unlock(mapping);
157 }
158
159 /*
160  * Unlink a file-based vm structure from its interval tree, to hide
161  * vma from rmap and vmtruncate before freeing its page tables.
162  */
163 void unlink_file_vma(struct vm_area_struct *vma)
164 {
165         struct file *file = vma->vm_file;
166
167         if (file) {
168                 struct address_space *mapping = file->f_mapping;
169                 i_mmap_lock_write(mapping);
170                 __remove_shared_vm_struct(vma, file, mapping);
171                 i_mmap_unlock_write(mapping);
172         }
173 }
174
175 /*
176  * Close a vm structure and free it, returning the next.
177  */
178 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
179 {
180         struct vm_area_struct *next = vma->vm_next;
181
182         might_sleep();
183         if (vma->vm_ops && vma->vm_ops->close)
184                 vma->vm_ops->close(vma);
185         if (vma->vm_file)
186                 fput(vma->vm_file);
187         mpol_put(vma_policy(vma));
188         vm_area_free(vma);
189         return next;
190 }
191
192 static int do_brk_flags(unsigned long addr, unsigned long request, unsigned long flags,
193                 struct list_head *uf);
194 SYSCALL_DEFINE1(brk, unsigned long, brk)
195 {
196         unsigned long newbrk, oldbrk, origbrk;
197         struct mm_struct *mm = current->mm;
198         struct vm_area_struct *next;
199         unsigned long min_brk;
200         bool populate;
201         bool downgraded = false;
202         LIST_HEAD(uf);
203
204         if (mmap_write_lock_killable(mm))
205                 return -EINTR;
206
207         origbrk = mm->brk;
208
209 #ifdef CONFIG_COMPAT_BRK
210         /*
211          * CONFIG_COMPAT_BRK can still be overridden by setting
212          * randomize_va_space to 2, which will still cause mm->start_brk
213          * to be arbitrarily shifted
214          */
215         if (current->brk_randomized)
216                 min_brk = mm->start_brk;
217         else
218                 min_brk = mm->end_data;
219 #else
220         min_brk = mm->start_brk;
221 #endif
222         if (brk < min_brk)
223                 goto out;
224
225         /*
226          * Check against rlimit here. If this check is done later after the test
227          * of oldbrk with newbrk then it can escape the test and let the data
228          * segment grow beyond its set limit the in case where the limit is
229          * not page aligned -Ram Gupta
230          */
231         if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
232                               mm->end_data, mm->start_data))
233                 goto out;
234
235         newbrk = PAGE_ALIGN(brk);
236         oldbrk = PAGE_ALIGN(mm->brk);
237         if (oldbrk == newbrk) {
238                 mm->brk = brk;
239                 goto success;
240         }
241
242         /*
243          * Always allow shrinking brk.
244          * __do_munmap() may downgrade mmap_lock to read.
245          */
246         if (brk <= mm->brk) {
247                 int ret;
248
249                 /*
250                  * mm->brk must to be protected by write mmap_lock so update it
251                  * before downgrading mmap_lock. When __do_munmap() fails,
252                  * mm->brk will be restored from origbrk.
253                  */
254                 mm->brk = brk;
255                 ret = __do_munmap(mm, newbrk, oldbrk-newbrk, &uf, true);
256                 if (ret < 0) {
257                         mm->brk = origbrk;
258                         goto out;
259                 } else if (ret == 1) {
260                         downgraded = true;
261                 }
262                 goto success;
263         }
264
265         /* Check against existing mmap mappings. */
266         next = find_vma(mm, oldbrk);
267         if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
268                 goto out;
269
270         /* Ok, looks good - let it rip. */
271         if (do_brk_flags(oldbrk, newbrk-oldbrk, 0, &uf) < 0)
272                 goto out;
273         mm->brk = brk;
274
275 success:
276         populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
277         if (downgraded)
278                 mmap_read_unlock(mm);
279         else
280                 mmap_write_unlock(mm);
281         userfaultfd_unmap_complete(mm, &uf);
282         if (populate)
283                 mm_populate(oldbrk, newbrk - oldbrk);
284         return brk;
285
286 out:
287         mmap_write_unlock(mm);
288         return origbrk;
289 }
290
291 static inline unsigned long vma_compute_gap(struct vm_area_struct *vma)
292 {
293         unsigned long gap, prev_end;
294
295         /*
296          * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
297          * allow two stack_guard_gaps between them here, and when choosing
298          * an unmapped area; whereas when expanding we only require one.
299          * That's a little inconsistent, but keeps the code here simpler.
300          */
301         gap = vm_start_gap(vma);
302         if (vma->vm_prev) {
303                 prev_end = vm_end_gap(vma->vm_prev);
304                 if (gap > prev_end)
305                         gap -= prev_end;
306                 else
307                         gap = 0;
308         }
309         return gap;
310 }
311
312 #ifdef CONFIG_DEBUG_VM_RB
313 static unsigned long vma_compute_subtree_gap(struct vm_area_struct *vma)
314 {
315         unsigned long max = vma_compute_gap(vma), subtree_gap;
316         if (vma->vm_rb.rb_left) {
317                 subtree_gap = rb_entry(vma->vm_rb.rb_left,
318                                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
319                 if (subtree_gap > max)
320                         max = subtree_gap;
321         }
322         if (vma->vm_rb.rb_right) {
323                 subtree_gap = rb_entry(vma->vm_rb.rb_right,
324                                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
325                 if (subtree_gap > max)
326                         max = subtree_gap;
327         }
328         return max;
329 }
330
331 static int browse_rb(struct mm_struct *mm)
332 {
333         struct rb_root *root = &mm->mm_rb;
334         int i = 0, j, bug = 0;
335         struct rb_node *nd, *pn = NULL;
336         unsigned long prev = 0, pend = 0;
337
338         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
339                 struct vm_area_struct *vma;
340                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
341                 if (vma->vm_start < prev) {
342                         pr_emerg("vm_start %lx < prev %lx\n",
343                                   vma->vm_start, prev);
344                         bug = 1;
345                 }
346                 if (vma->vm_start < pend) {
347                         pr_emerg("vm_start %lx < pend %lx\n",
348                                   vma->vm_start, pend);
349                         bug = 1;
350                 }
351                 if (vma->vm_start > vma->vm_end) {
352                         pr_emerg("vm_start %lx > vm_end %lx\n",
353                                   vma->vm_start, vma->vm_end);
354                         bug = 1;
355                 }
356                 spin_lock(&mm->page_table_lock);
357                 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
358                         pr_emerg("free gap %lx, correct %lx\n",
359                                vma->rb_subtree_gap,
360                                vma_compute_subtree_gap(vma));
361                         bug = 1;
362                 }
363                 spin_unlock(&mm->page_table_lock);
364                 i++;
365                 pn = nd;
366                 prev = vma->vm_start;
367                 pend = vma->vm_end;
368         }
369         j = 0;
370         for (nd = pn; nd; nd = rb_prev(nd))
371                 j++;
372         if (i != j) {
373                 pr_emerg("backwards %d, forwards %d\n", j, i);
374                 bug = 1;
375         }
376         return bug ? -1 : i;
377 }
378
379 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
380 {
381         struct rb_node *nd;
382
383         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
384                 struct vm_area_struct *vma;
385                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
386                 VM_BUG_ON_VMA(vma != ignore &&
387                         vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
388                         vma);
389         }
390 }
391
392 static void validate_mm(struct mm_struct *mm)
393 {
394         int bug = 0;
395         int i = 0;
396         unsigned long highest_address = 0;
397         struct vm_area_struct *vma = mm->mmap;
398
399         while (vma) {
400                 struct anon_vma *anon_vma = vma->anon_vma;
401                 struct anon_vma_chain *avc;
402
403                 if (anon_vma) {
404                         anon_vma_lock_read(anon_vma);
405                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
406                                 anon_vma_interval_tree_verify(avc);
407                         anon_vma_unlock_read(anon_vma);
408                 }
409
410                 highest_address = vm_end_gap(vma);
411                 vma = vma->vm_next;
412                 i++;
413         }
414         if (i != mm->map_count) {
415                 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
416                 bug = 1;
417         }
418         if (highest_address != mm->highest_vm_end) {
419                 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
420                           mm->highest_vm_end, highest_address);
421                 bug = 1;
422         }
423         i = browse_rb(mm);
424         if (i != mm->map_count) {
425                 if (i != -1)
426                         pr_emerg("map_count %d rb %d\n", mm->map_count, i);
427                 bug = 1;
428         }
429         VM_BUG_ON_MM(bug, mm);
430 }
431 #else
432 #define validate_mm_rb(root, ignore) do { } while (0)
433 #define validate_mm(mm) do { } while (0)
434 #endif
435
436 RB_DECLARE_CALLBACKS_MAX(static, vma_gap_callbacks,
437                          struct vm_area_struct, vm_rb,
438                          unsigned long, rb_subtree_gap, vma_compute_gap)
439
440 /*
441  * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
442  * vma->vm_prev->vm_end values changed, without modifying the vma's position
443  * in the rbtree.
444  */
445 static void vma_gap_update(struct vm_area_struct *vma)
446 {
447         /*
448          * As it turns out, RB_DECLARE_CALLBACKS_MAX() already created
449          * a callback function that does exactly what we want.
450          */
451         vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
452 }
453
454 static inline void vma_rb_insert(struct vm_area_struct *vma,
455                                  struct rb_root *root)
456 {
457         /* All rb_subtree_gap values must be consistent prior to insertion */
458         validate_mm_rb(root, NULL);
459
460         rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
461 }
462
463 static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
464 {
465         /*
466          * Note rb_erase_augmented is a fairly large inline function,
467          * so make sure we instantiate it only once with our desired
468          * augmented rbtree callbacks.
469          */
470         rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
471 }
472
473 static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
474                                                 struct rb_root *root,
475                                                 struct vm_area_struct *ignore)
476 {
477         /*
478          * All rb_subtree_gap values must be consistent prior to erase,
479          * with the possible exception of
480          *
481          * a. the "next" vma being erased if next->vm_start was reduced in
482          *    __vma_adjust() -> __vma_unlink()
483          * b. the vma being erased in detach_vmas_to_be_unmapped() ->
484          *    vma_rb_erase()
485          */
486         validate_mm_rb(root, ignore);
487
488         __vma_rb_erase(vma, root);
489 }
490
491 static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
492                                          struct rb_root *root)
493 {
494         vma_rb_erase_ignore(vma, root, vma);
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_lock 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 /*
564  * vma_next() - Get the next VMA.
565  * @mm: The mm_struct.
566  * @vma: The current vma.
567  *
568  * If @vma is NULL, return the first vma in the mm.
569  *
570  * Returns: The next VMA after @vma.
571  */
572 static inline struct vm_area_struct *vma_next(struct mm_struct *mm,
573                                          struct vm_area_struct *vma)
574 {
575         if (!vma)
576                 return mm->mmap;
577
578         return vma->vm_next;
579 }
580
581 /*
582  * munmap_vma_range() - munmap VMAs that overlap a range.
583  * @mm: The mm struct
584  * @start: The start of the range.
585  * @len: The length of the range.
586  * @pprev: pointer to the pointer that will be set to previous vm_area_struct
587  * @rb_link: the rb_node
588  * @rb_parent: the parent rb_node
589  *
590  * Find all the vm_area_struct that overlap from @start to
591  * @end and munmap them.  Set @pprev to the previous vm_area_struct.
592  *
593  * Returns: -ENOMEM on munmap failure or 0 on success.
594  */
595 static inline int
596 munmap_vma_range(struct mm_struct *mm, unsigned long start, unsigned long len,
597                  struct vm_area_struct **pprev, struct rb_node ***link,
598                  struct rb_node **parent, struct list_head *uf)
599 {
600
601         while (find_vma_links(mm, start, start + len, pprev, link, parent))
602                 if (do_munmap(mm, start, len, uf))
603                         return -ENOMEM;
604
605         return 0;
606 }
607 static unsigned long count_vma_pages_range(struct mm_struct *mm,
608                 unsigned long addr, unsigned long end)
609 {
610         unsigned long nr_pages = 0;
611         struct vm_area_struct *vma;
612
613         /* Find first overlapping mapping */
614         vma = find_vma_intersection(mm, addr, end);
615         if (!vma)
616                 return 0;
617
618         nr_pages = (min(end, vma->vm_end) -
619                 max(addr, vma->vm_start)) >> PAGE_SHIFT;
620
621         /* Iterate over the rest of the overlaps */
622         for (vma = vma->vm_next; vma; vma = vma->vm_next) {
623                 unsigned long overlap_len;
624
625                 if (vma->vm_start > end)
626                         break;
627
628                 overlap_len = min(end, vma->vm_end) - vma->vm_start;
629                 nr_pages += overlap_len >> PAGE_SHIFT;
630         }
631
632         return nr_pages;
633 }
634
635 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
636                 struct rb_node **rb_link, struct rb_node *rb_parent)
637 {
638         /* Update tracking information for the gap following the new vma. */
639         if (vma->vm_next)
640                 vma_gap_update(vma->vm_next);
641         else
642                 mm->highest_vm_end = vm_end_gap(vma);
643
644         /*
645          * vma->vm_prev wasn't known when we followed the rbtree to find the
646          * correct insertion point for that vma. As a result, we could not
647          * update the vma vm_rb parents rb_subtree_gap values on the way down.
648          * So, we first insert the vma with a zero rb_subtree_gap value
649          * (to be consistent with what we did on the way down), and then
650          * immediately update the gap to the correct value. Finally we
651          * rebalance the rbtree after all augmented values have been set.
652          */
653         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
654         vma->rb_subtree_gap = 0;
655         vma_gap_update(vma);
656         vma_rb_insert(vma, &mm->mm_rb);
657 }
658
659 static void __vma_link_file(struct vm_area_struct *vma)
660 {
661         struct file *file;
662
663         file = vma->vm_file;
664         if (file) {
665                 struct address_space *mapping = file->f_mapping;
666
667                 if (vma->vm_flags & VM_SHARED)
668                         mapping_allow_writable(mapping);
669
670                 flush_dcache_mmap_lock(mapping);
671                 vma_interval_tree_insert(vma, &mapping->i_mmap);
672                 flush_dcache_mmap_unlock(mapping);
673         }
674 }
675
676 static void
677 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
678         struct vm_area_struct *prev, struct rb_node **rb_link,
679         struct rb_node *rb_parent)
680 {
681         __vma_link_list(mm, vma, prev);
682         __vma_link_rb(mm, vma, rb_link, rb_parent);
683 }
684
685 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
686                         struct vm_area_struct *prev, struct rb_node **rb_link,
687                         struct rb_node *rb_parent)
688 {
689         struct address_space *mapping = NULL;
690
691         if (vma->vm_file) {
692                 mapping = vma->vm_file->f_mapping;
693                 i_mmap_lock_write(mapping);
694         }
695
696         __vma_link(mm, vma, prev, rb_link, rb_parent);
697         __vma_link_file(vma);
698
699         if (mapping)
700                 i_mmap_unlock_write(mapping);
701
702         mm->map_count++;
703         validate_mm(mm);
704 }
705
706 /*
707  * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
708  * mm's list and rbtree.  It has already been inserted into the interval tree.
709  */
710 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
711 {
712         struct vm_area_struct *prev;
713         struct rb_node **rb_link, *rb_parent;
714
715         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
716                            &prev, &rb_link, &rb_parent))
717                 BUG();
718         __vma_link(mm, vma, prev, rb_link, rb_parent);
719         mm->map_count++;
720 }
721
722 static __always_inline void __vma_unlink(struct mm_struct *mm,
723                                                 struct vm_area_struct *vma,
724                                                 struct vm_area_struct *ignore)
725 {
726         vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
727         __vma_unlink_list(mm, vma);
728         /* Kill the cache */
729         vmacache_invalidate(mm);
730 }
731
732 /*
733  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
734  * is already present in an i_mmap tree without adjusting the tree.
735  * The following helper function should be used when such adjustments
736  * are necessary.  The "insert" vma (if any) is to be inserted
737  * before we drop the necessary locks.
738  */
739 int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
740         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
741         struct vm_area_struct *expand)
742 {
743         struct mm_struct *mm = vma->vm_mm;
744         struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
745         struct address_space *mapping = NULL;
746         struct rb_root_cached *root = NULL;
747         struct anon_vma *anon_vma = NULL;
748         struct file *file = vma->vm_file;
749         bool start_changed = false, end_changed = false;
750         long adjust_next = 0;
751         int remove_next = 0;
752
753         if (next && !insert) {
754                 struct vm_area_struct *exporter = NULL, *importer = NULL;
755
756                 if (end >= next->vm_end) {
757                         /*
758                          * vma expands, overlapping all the next, and
759                          * perhaps the one after too (mprotect case 6).
760                          * The only other cases that gets here are
761                          * case 1, case 7 and case 8.
762                          */
763                         if (next == expand) {
764                                 /*
765                                  * The only case where we don't expand "vma"
766                                  * and we expand "next" instead is case 8.
767                                  */
768                                 VM_WARN_ON(end != next->vm_end);
769                                 /*
770                                  * remove_next == 3 means we're
771                                  * removing "vma" and that to do so we
772                                  * swapped "vma" and "next".
773                                  */
774                                 remove_next = 3;
775                                 VM_WARN_ON(file != next->vm_file);
776                                 swap(vma, next);
777                         } else {
778                                 VM_WARN_ON(expand != vma);
779                                 /*
780                                  * case 1, 6, 7, remove_next == 2 is case 6,
781                                  * remove_next == 1 is case 1 or 7.
782                                  */
783                                 remove_next = 1 + (end > next->vm_end);
784                                 VM_WARN_ON(remove_next == 2 &&
785                                            end != next->vm_next->vm_end);
786                                 /* trim end to next, for case 6 first pass */
787                                 end = next->vm_end;
788                         }
789
790                         exporter = next;
791                         importer = vma;
792
793                         /*
794                          * If next doesn't have anon_vma, import from vma after
795                          * next, if the vma overlaps with it.
796                          */
797                         if (remove_next == 2 && !next->anon_vma)
798                                 exporter = next->vm_next;
799
800                 } else if (end > next->vm_start) {
801                         /*
802                          * vma expands, overlapping part of the next:
803                          * mprotect case 5 shifting the boundary up.
804                          */
805                         adjust_next = (end - next->vm_start);
806                         exporter = next;
807                         importer = vma;
808                         VM_WARN_ON(expand != importer);
809                 } else if (end < vma->vm_end) {
810                         /*
811                          * vma shrinks, and !insert tells it's not
812                          * split_vma inserting another: so it must be
813                          * mprotect case 4 shifting the boundary down.
814                          */
815                         adjust_next = -(vma->vm_end - end);
816                         exporter = vma;
817                         importer = next;
818                         VM_WARN_ON(expand != importer);
819                 }
820
821                 /*
822                  * Easily overlooked: when mprotect shifts the boundary,
823                  * make sure the expanding vma has anon_vma set if the
824                  * shrinking vma had, to cover any anon pages imported.
825                  */
826                 if (exporter && exporter->anon_vma && !importer->anon_vma) {
827                         int error;
828
829                         importer->anon_vma = exporter->anon_vma;
830                         error = anon_vma_clone(importer, exporter);
831                         if (error)
832                                 return error;
833                 }
834         }
835 again:
836         vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
837
838         if (file) {
839                 mapping = file->f_mapping;
840                 root = &mapping->i_mmap;
841                 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
842
843                 if (adjust_next)
844                         uprobe_munmap(next, next->vm_start, next->vm_end);
845
846                 i_mmap_lock_write(mapping);
847                 if (insert) {
848                         /*
849                          * Put into interval tree now, so instantiated pages
850                          * are visible to arm/parisc __flush_dcache_page
851                          * throughout; but we cannot insert into address
852                          * space until vma start or end is updated.
853                          */
854                         __vma_link_file(insert);
855                 }
856         }
857
858         anon_vma = vma->anon_vma;
859         if (!anon_vma && adjust_next)
860                 anon_vma = next->anon_vma;
861         if (anon_vma) {
862                 VM_WARN_ON(adjust_next && next->anon_vma &&
863                            anon_vma != next->anon_vma);
864                 anon_vma_lock_write(anon_vma);
865                 anon_vma_interval_tree_pre_update_vma(vma);
866                 if (adjust_next)
867                         anon_vma_interval_tree_pre_update_vma(next);
868         }
869
870         if (file) {
871                 flush_dcache_mmap_lock(mapping);
872                 vma_interval_tree_remove(vma, root);
873                 if (adjust_next)
874                         vma_interval_tree_remove(next, root);
875         }
876
877         if (start != vma->vm_start) {
878                 vma->vm_start = start;
879                 start_changed = true;
880         }
881         if (end != vma->vm_end) {
882                 vma->vm_end = end;
883                 end_changed = true;
884         }
885         vma->vm_pgoff = pgoff;
886         if (adjust_next) {
887                 next->vm_start += adjust_next;
888                 next->vm_pgoff += adjust_next >> PAGE_SHIFT;
889         }
890
891         if (file) {
892                 if (adjust_next)
893                         vma_interval_tree_insert(next, root);
894                 vma_interval_tree_insert(vma, root);
895                 flush_dcache_mmap_unlock(mapping);
896         }
897
898         if (remove_next) {
899                 /*
900                  * vma_merge has merged next into vma, and needs
901                  * us to remove next before dropping the locks.
902                  */
903                 if (remove_next != 3)
904                         __vma_unlink(mm, next, next);
905                 else
906                         /*
907                          * vma is not before next if they've been
908                          * swapped.
909                          *
910                          * pre-swap() next->vm_start was reduced so
911                          * tell validate_mm_rb to ignore pre-swap()
912                          * "next" (which is stored in post-swap()
913                          * "vma").
914                          */
915                         __vma_unlink(mm, next, vma);
916                 if (file)
917                         __remove_shared_vm_struct(next, file, mapping);
918         } else if (insert) {
919                 /*
920                  * split_vma has split insert from vma, and needs
921                  * us to insert it before dropping the locks
922                  * (it may either follow vma or precede it).
923                  */
924                 __insert_vm_struct(mm, insert);
925         } else {
926                 if (start_changed)
927                         vma_gap_update(vma);
928                 if (end_changed) {
929                         if (!next)
930                                 mm->highest_vm_end = vm_end_gap(vma);
931                         else if (!adjust_next)
932                                 vma_gap_update(next);
933                 }
934         }
935
936         if (anon_vma) {
937                 anon_vma_interval_tree_post_update_vma(vma);
938                 if (adjust_next)
939                         anon_vma_interval_tree_post_update_vma(next);
940                 anon_vma_unlock_write(anon_vma);
941         }
942
943         if (file) {
944                 i_mmap_unlock_write(mapping);
945                 uprobe_mmap(vma);
946
947                 if (adjust_next)
948                         uprobe_mmap(next);
949         }
950
951         if (remove_next) {
952                 if (file) {
953                         uprobe_munmap(next, next->vm_start, next->vm_end);
954                         fput(file);
955                 }
956                 if (next->anon_vma)
957                         anon_vma_merge(vma, next);
958                 mm->map_count--;
959                 mpol_put(vma_policy(next));
960                 vm_area_free(next);
961                 /*
962                  * In mprotect's case 6 (see comments on vma_merge),
963                  * we must remove another next too. It would clutter
964                  * up the code too much to do both in one go.
965                  */
966                 if (remove_next != 3) {
967                         /*
968                          * If "next" was removed and vma->vm_end was
969                          * expanded (up) over it, in turn
970                          * "next->vm_prev->vm_end" changed and the
971                          * "vma->vm_next" gap must be updated.
972                          */
973                         next = vma->vm_next;
974                 } else {
975                         /*
976                          * For the scope of the comment "next" and
977                          * "vma" considered pre-swap(): if "vma" was
978                          * removed, next->vm_start was expanded (down)
979                          * over it and the "next" gap must be updated.
980                          * Because of the swap() the post-swap() "vma"
981                          * actually points to pre-swap() "next"
982                          * (post-swap() "next" as opposed is now a
983                          * dangling pointer).
984                          */
985                         next = vma;
986                 }
987                 if (remove_next == 2) {
988                         remove_next = 1;
989                         end = next->vm_end;
990                         goto again;
991                 }
992                 else if (next)
993                         vma_gap_update(next);
994                 else {
995                         /*
996                          * If remove_next == 2 we obviously can't
997                          * reach this path.
998                          *
999                          * If remove_next == 3 we can't reach this
1000                          * path because pre-swap() next is always not
1001                          * NULL. pre-swap() "next" is not being
1002                          * removed and its next->vm_end is not altered
1003                          * (and furthermore "end" already matches
1004                          * next->vm_end in remove_next == 3).
1005                          *
1006                          * We reach this only in the remove_next == 1
1007                          * case if the "next" vma that was removed was
1008                          * the highest vma of the mm. However in such
1009                          * case next->vm_end == "end" and the extended
1010                          * "vma" has vma->vm_end == next->vm_end so
1011                          * mm->highest_vm_end doesn't need any update
1012                          * in remove_next == 1 case.
1013                          */
1014                         VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
1015                 }
1016         }
1017         if (insert && file)
1018                 uprobe_mmap(insert);
1019
1020         validate_mm(mm);
1021
1022         return 0;
1023 }
1024
1025 /*
1026  * If the vma has a ->close operation then the driver probably needs to release
1027  * per-vma resources, so we don't attempt to merge those.
1028  */
1029 static inline int is_mergeable_vma(struct vm_area_struct *vma,
1030                                 struct file *file, unsigned long vm_flags,
1031                                 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1032 {
1033         /*
1034          * VM_SOFTDIRTY should not prevent from VMA merging, if we
1035          * match the flags but dirty bit -- the caller should mark
1036          * merged VMA as dirty. If dirty bit won't be excluded from
1037          * comparison, we increase pressure on the memory system forcing
1038          * the kernel to generate new VMAs when old one could be
1039          * extended instead.
1040          */
1041         if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
1042                 return 0;
1043         if (vma->vm_file != file)
1044                 return 0;
1045         if (vma->vm_ops && vma->vm_ops->close)
1046                 return 0;
1047         if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
1048                 return 0;
1049         return 1;
1050 }
1051
1052 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
1053                                         struct anon_vma *anon_vma2,
1054                                         struct vm_area_struct *vma)
1055 {
1056         /*
1057          * The list_is_singular() test is to avoid merging VMA cloned from
1058          * parents. This can improve scalability caused by anon_vma lock.
1059          */
1060         if ((!anon_vma1 || !anon_vma2) && (!vma ||
1061                 list_is_singular(&vma->anon_vma_chain)))
1062                 return 1;
1063         return anon_vma1 == anon_vma2;
1064 }
1065
1066 /*
1067  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1068  * in front of (at a lower virtual address and file offset than) the vma.
1069  *
1070  * We cannot merge two vmas if they have differently assigned (non-NULL)
1071  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1072  *
1073  * We don't check here for the merged mmap wrapping around the end of pagecache
1074  * indices (16TB on ia32) because do_mmap() does not permit mmap's which
1075  * wrap, nor mmaps which cover the final page at index -1UL.
1076  */
1077 static int
1078 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1079                      struct anon_vma *anon_vma, struct file *file,
1080                      pgoff_t vm_pgoff,
1081                      struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1082 {
1083         if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1084             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1085                 if (vma->vm_pgoff == vm_pgoff)
1086                         return 1;
1087         }
1088         return 0;
1089 }
1090
1091 /*
1092  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1093  * beyond (at a higher virtual address and file offset than) the vma.
1094  *
1095  * We cannot merge two vmas if they have differently assigned (non-NULL)
1096  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1097  */
1098 static int
1099 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1100                     struct anon_vma *anon_vma, struct file *file,
1101                     pgoff_t vm_pgoff,
1102                     struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1103 {
1104         if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1105             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1106                 pgoff_t vm_pglen;
1107                 vm_pglen = vma_pages(vma);
1108                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1109                         return 1;
1110         }
1111         return 0;
1112 }
1113
1114 /*
1115  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1116  * whether that can be merged with its predecessor or its successor.
1117  * Or both (it neatly fills a hole).
1118  *
1119  * In most cases - when called for mmap, brk or mremap - [addr,end) is
1120  * certain not to be mapped by the time vma_merge is called; but when
1121  * called for mprotect, it is certain to be already mapped (either at
1122  * an offset within prev, or at the start of next), and the flags of
1123  * this area are about to be changed to vm_flags - and the no-change
1124  * case has already been eliminated.
1125  *
1126  * The following mprotect cases have to be considered, where AAAA is
1127  * the area passed down from mprotect_fixup, never extending beyond one
1128  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1129  *
1130  *     AAAA             AAAA                   AAAA
1131  *    PPPPPPNNNNNN    PPPPPPNNNNNN       PPPPPPNNNNNN
1132  *    cannot merge    might become       might become
1133  *                    PPNNNNNNNNNN       PPPPPPPPPPNN
1134  *    mmap, brk or    case 4 below       case 5 below
1135  *    mremap move:
1136  *                        AAAA               AAAA
1137  *                    PPPP    NNNN       PPPPNNNNXXXX
1138  *                    might become       might become
1139  *                    PPPPPPPPPPPP 1 or  PPPPPPPPPPPP 6 or
1140  *                    PPPPPPPPNNNN 2 or  PPPPPPPPXXXX 7 or
1141  *                    PPPPNNNNNNNN 3     PPPPXXXXXXXX 8
1142  *
1143  * It is important for case 8 that the vma NNNN overlapping the
1144  * region AAAA is never going to extended over XXXX. Instead XXXX must
1145  * be extended in region AAAA and NNNN must be removed. This way in
1146  * all cases where vma_merge succeeds, the moment vma_adjust drops the
1147  * rmap_locks, the properties of the merged vma will be already
1148  * correct for the whole merged range. Some of those properties like
1149  * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1150  * be correct for the whole merged range immediately after the
1151  * rmap_locks are released. Otherwise if XXXX would be removed and
1152  * NNNN would be extended over the XXXX range, remove_migration_ptes
1153  * or other rmap walkers (if working on addresses beyond the "end"
1154  * parameter) may establish ptes with the wrong permissions of NNNN
1155  * instead of the right permissions of XXXX.
1156  */
1157 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1158                         struct vm_area_struct *prev, unsigned long addr,
1159                         unsigned long end, unsigned long vm_flags,
1160                         struct anon_vma *anon_vma, struct file *file,
1161                         pgoff_t pgoff, struct mempolicy *policy,
1162                         struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1163 {
1164         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1165         struct vm_area_struct *area, *next;
1166         int err;
1167
1168         /*
1169          * We later require that vma->vm_flags == vm_flags,
1170          * so this tests vma->vm_flags & VM_SPECIAL, too.
1171          */
1172         if (vm_flags & VM_SPECIAL)
1173                 return NULL;
1174
1175         next = vma_next(mm, prev);
1176         area = next;
1177         if (area && area->vm_end == end)                /* cases 6, 7, 8 */
1178                 next = next->vm_next;
1179
1180         /* verify some invariant that must be enforced by the caller */
1181         VM_WARN_ON(prev && addr <= prev->vm_start);
1182         VM_WARN_ON(area && end > area->vm_end);
1183         VM_WARN_ON(addr >= end);
1184
1185         /*
1186          * Can it merge with the predecessor?
1187          */
1188         if (prev && prev->vm_end == addr &&
1189                         mpol_equal(vma_policy(prev), policy) &&
1190                         can_vma_merge_after(prev, vm_flags,
1191                                             anon_vma, file, pgoff,
1192                                             vm_userfaultfd_ctx)) {
1193                 /*
1194                  * OK, it can.  Can we now merge in the successor as well?
1195                  */
1196                 if (next && end == next->vm_start &&
1197                                 mpol_equal(policy, vma_policy(next)) &&
1198                                 can_vma_merge_before(next, vm_flags,
1199                                                      anon_vma, file,
1200                                                      pgoff+pglen,
1201                                                      vm_userfaultfd_ctx) &&
1202                                 is_mergeable_anon_vma(prev->anon_vma,
1203                                                       next->anon_vma, NULL)) {
1204                                                         /* cases 1, 6 */
1205                         err = __vma_adjust(prev, prev->vm_start,
1206                                          next->vm_end, prev->vm_pgoff, NULL,
1207                                          prev);
1208                 } else                                  /* cases 2, 5, 7 */
1209                         err = __vma_adjust(prev, prev->vm_start,
1210                                          end, prev->vm_pgoff, NULL, prev);
1211                 if (err)
1212                         return NULL;
1213                 khugepaged_enter_vma_merge(prev, vm_flags);
1214                 return prev;
1215         }
1216
1217         /*
1218          * Can this new request be merged in front of next?
1219          */
1220         if (next && end == next->vm_start &&
1221                         mpol_equal(policy, vma_policy(next)) &&
1222                         can_vma_merge_before(next, vm_flags,
1223                                              anon_vma, file, pgoff+pglen,
1224                                              vm_userfaultfd_ctx)) {
1225                 if (prev && addr < prev->vm_end)        /* case 4 */
1226                         err = __vma_adjust(prev, prev->vm_start,
1227                                          addr, prev->vm_pgoff, NULL, next);
1228                 else {                                  /* cases 3, 8 */
1229                         err = __vma_adjust(area, addr, next->vm_end,
1230                                          next->vm_pgoff - pglen, NULL, next);
1231                         /*
1232                          * In case 3 area is already equal to next and
1233                          * this is a noop, but in case 8 "area" has
1234                          * been removed and next was expanded over it.
1235                          */
1236                         area = next;
1237                 }
1238                 if (err)
1239                         return NULL;
1240                 khugepaged_enter_vma_merge(area, vm_flags);
1241                 return area;
1242         }
1243
1244         return NULL;
1245 }
1246
1247 /*
1248  * Rough compatibility check to quickly see if it's even worth looking
1249  * at sharing an anon_vma.
1250  *
1251  * They need to have the same vm_file, and the flags can only differ
1252  * in things that mprotect may change.
1253  *
1254  * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1255  * we can merge the two vma's. For example, we refuse to merge a vma if
1256  * there is a vm_ops->close() function, because that indicates that the
1257  * driver is doing some kind of reference counting. But that doesn't
1258  * really matter for the anon_vma sharing case.
1259  */
1260 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1261 {
1262         return a->vm_end == b->vm_start &&
1263                 mpol_equal(vma_policy(a), vma_policy(b)) &&
1264                 a->vm_file == b->vm_file &&
1265                 !((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) &&
1266                 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1267 }
1268
1269 /*
1270  * Do some basic sanity checking to see if we can re-use the anon_vma
1271  * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1272  * the same as 'old', the other will be the new one that is trying
1273  * to share the anon_vma.
1274  *
1275  * NOTE! This runs with mm_sem held for reading, so it is possible that
1276  * the anon_vma of 'old' is concurrently in the process of being set up
1277  * by another page fault trying to merge _that_. But that's ok: if it
1278  * is being set up, that automatically means that it will be a singleton
1279  * acceptable for merging, so we can do all of this optimistically. But
1280  * we do that READ_ONCE() to make sure that we never re-load the pointer.
1281  *
1282  * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1283  * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1284  * is to return an anon_vma that is "complex" due to having gone through
1285  * a fork).
1286  *
1287  * We also make sure that the two vma's are compatible (adjacent,
1288  * and with the same memory policies). That's all stable, even with just
1289  * a read lock on the mm_sem.
1290  */
1291 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1292 {
1293         if (anon_vma_compatible(a, b)) {
1294                 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1295
1296                 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1297                         return anon_vma;
1298         }
1299         return NULL;
1300 }
1301
1302 /*
1303  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1304  * neighbouring vmas for a suitable anon_vma, before it goes off
1305  * to allocate a new anon_vma.  It checks because a repetitive
1306  * sequence of mprotects and faults may otherwise lead to distinct
1307  * anon_vmas being allocated, preventing vma merge in subsequent
1308  * mprotect.
1309  */
1310 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1311 {
1312         struct anon_vma *anon_vma = NULL;
1313
1314         /* Try next first. */
1315         if (vma->vm_next) {
1316                 anon_vma = reusable_anon_vma(vma->vm_next, vma, vma->vm_next);
1317                 if (anon_vma)
1318                         return anon_vma;
1319         }
1320
1321         /* Try prev next. */
1322         if (vma->vm_prev)
1323                 anon_vma = reusable_anon_vma(vma->vm_prev, vma->vm_prev, vma);
1324
1325         /*
1326          * We might reach here with anon_vma == NULL if we can't find
1327          * any reusable anon_vma.
1328          * There's no absolute need to look only at touching neighbours:
1329          * we could search further afield for "compatible" anon_vmas.
1330          * But it would probably just be a waste of time searching,
1331          * or lead to too many vmas hanging off the same anon_vma.
1332          * We're trying to allow mprotect remerging later on,
1333          * not trying to minimize memory used for anon_vmas.
1334          */
1335         return anon_vma;
1336 }
1337
1338 /*
1339  * If a hint addr is less than mmap_min_addr change hint to be as
1340  * low as possible but still greater than mmap_min_addr
1341  */
1342 static inline unsigned long round_hint_to_min(unsigned long hint)
1343 {
1344         hint &= PAGE_MASK;
1345         if (((void *)hint != NULL) &&
1346             (hint < mmap_min_addr))
1347                 return PAGE_ALIGN(mmap_min_addr);
1348         return hint;
1349 }
1350
1351 int mlock_future_check(struct mm_struct *mm, unsigned long flags,
1352                        unsigned long len)
1353 {
1354         unsigned long locked, lock_limit;
1355
1356         /*  mlock MCL_FUTURE? */
1357         if (flags & VM_LOCKED) {
1358                 locked = len >> PAGE_SHIFT;
1359                 locked += mm->locked_vm;
1360                 lock_limit = rlimit(RLIMIT_MEMLOCK);
1361                 lock_limit >>= PAGE_SHIFT;
1362                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1363                         return -EAGAIN;
1364         }
1365         return 0;
1366 }
1367
1368 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1369 {
1370         if (S_ISREG(inode->i_mode))
1371                 return MAX_LFS_FILESIZE;
1372
1373         if (S_ISBLK(inode->i_mode))
1374                 return MAX_LFS_FILESIZE;
1375
1376         if (S_ISSOCK(inode->i_mode))
1377                 return MAX_LFS_FILESIZE;
1378
1379         /* Special "we do even unsigned file positions" case */
1380         if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1381                 return 0;
1382
1383         /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1384         return ULONG_MAX;
1385 }
1386
1387 static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1388                                 unsigned long pgoff, unsigned long len)
1389 {
1390         u64 maxsize = file_mmap_size_max(file, inode);
1391
1392         if (maxsize && len > maxsize)
1393                 return false;
1394         maxsize -= len;
1395         if (pgoff > maxsize >> PAGE_SHIFT)
1396                 return false;
1397         return true;
1398 }
1399
1400 /*
1401  * The caller must write-lock current->mm->mmap_lock.
1402  */
1403 unsigned long do_mmap(struct file *file, unsigned long addr,
1404                         unsigned long len, unsigned long prot,
1405                         unsigned long flags, unsigned long pgoff,
1406                         unsigned long *populate, struct list_head *uf)
1407 {
1408         struct mm_struct *mm = current->mm;
1409         vm_flags_t vm_flags;
1410         int pkey = 0;
1411
1412         *populate = 0;
1413
1414         if (!len)
1415                 return -EINVAL;
1416
1417         /*
1418          * Does the application expect PROT_READ to imply PROT_EXEC?
1419          *
1420          * (the exception is when the underlying filesystem is noexec
1421          *  mounted, in which case we dont add PROT_EXEC.)
1422          */
1423         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1424                 if (!(file && path_noexec(&file->f_path)))
1425                         prot |= PROT_EXEC;
1426
1427         /* force arch specific MAP_FIXED handling in get_unmapped_area */
1428         if (flags & MAP_FIXED_NOREPLACE)
1429                 flags |= MAP_FIXED;
1430
1431         if (!(flags & MAP_FIXED))
1432                 addr = round_hint_to_min(addr);
1433
1434         /* Careful about overflows.. */
1435         len = PAGE_ALIGN(len);
1436         if (!len)
1437                 return -ENOMEM;
1438
1439         /* offset overflow? */
1440         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1441                 return -EOVERFLOW;
1442
1443         /* Too many mappings? */
1444         if (mm->map_count > sysctl_max_map_count)
1445                 return -ENOMEM;
1446
1447         /* Obtain the address to map to. we verify (or select) it and ensure
1448          * that it represents a valid section of the address space.
1449          */
1450         addr = get_unmapped_area(file, addr, len, pgoff, flags);
1451         if (IS_ERR_VALUE(addr))
1452                 return addr;
1453
1454         if (flags & MAP_FIXED_NOREPLACE) {
1455                 if (find_vma_intersection(mm, addr, addr + len))
1456                         return -EEXIST;
1457         }
1458
1459         if (prot == PROT_EXEC) {
1460                 pkey = execute_only_pkey(mm);
1461                 if (pkey < 0)
1462                         pkey = 0;
1463         }
1464
1465         /* Do simple checking here so the lower-level routines won't have
1466          * to. we assume access permissions have been handled by the open
1467          * of the memory object, so we don't do any here.
1468          */
1469         vm_flags = calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1470                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1471
1472         if (flags & MAP_LOCKED)
1473                 if (!can_do_mlock())
1474                         return -EPERM;
1475
1476         if (mlock_future_check(mm, vm_flags, len))
1477                 return -EAGAIN;
1478
1479         if (file) {
1480                 struct inode *inode = file_inode(file);
1481                 unsigned long flags_mask;
1482
1483                 if (!file_mmap_ok(file, inode, pgoff, len))
1484                         return -EOVERFLOW;
1485
1486                 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1487
1488                 switch (flags & MAP_TYPE) {
1489                 case MAP_SHARED:
1490                         /*
1491                          * Force use of MAP_SHARED_VALIDATE with non-legacy
1492                          * flags. E.g. MAP_SYNC is dangerous to use with
1493                          * MAP_SHARED as you don't know which consistency model
1494                          * you will get. We silently ignore unsupported flags
1495                          * with MAP_SHARED to preserve backward compatibility.
1496                          */
1497                         flags &= LEGACY_MAP_MASK;
1498                         fallthrough;
1499                 case MAP_SHARED_VALIDATE:
1500                         if (flags & ~flags_mask)
1501                                 return -EOPNOTSUPP;
1502                         if (prot & PROT_WRITE) {
1503                                 if (!(file->f_mode & FMODE_WRITE))
1504                                         return -EACCES;
1505                                 if (IS_SWAPFILE(file->f_mapping->host))
1506                                         return -ETXTBSY;
1507                         }
1508
1509                         /*
1510                          * Make sure we don't allow writing to an append-only
1511                          * file..
1512                          */
1513                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1514                                 return -EACCES;
1515
1516                         /*
1517                          * Make sure there are no mandatory locks on the file.
1518                          */
1519                         if (locks_verify_locked(file))
1520                                 return -EAGAIN;
1521
1522                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1523                         if (!(file->f_mode & FMODE_WRITE))
1524                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1525                         fallthrough;
1526                 case MAP_PRIVATE:
1527                         if (!(file->f_mode & FMODE_READ))
1528                                 return -EACCES;
1529                         if (path_noexec(&file->f_path)) {
1530                                 if (vm_flags & VM_EXEC)
1531                                         return -EPERM;
1532                                 vm_flags &= ~VM_MAYEXEC;
1533                         }
1534
1535                         if (!file->f_op->mmap)
1536                                 return -ENODEV;
1537                         if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1538                                 return -EINVAL;
1539                         break;
1540
1541                 default:
1542                         return -EINVAL;
1543                 }
1544         } else {
1545                 switch (flags & MAP_TYPE) {
1546                 case MAP_SHARED:
1547                         if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1548                                 return -EINVAL;
1549                         /*
1550                          * Ignore pgoff.
1551                          */
1552                         pgoff = 0;
1553                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1554                         break;
1555                 case MAP_PRIVATE:
1556                         /*
1557                          * Set pgoff according to addr for anon_vma.
1558                          */
1559                         pgoff = addr >> PAGE_SHIFT;
1560                         break;
1561                 default:
1562                         return -EINVAL;
1563                 }
1564         }
1565
1566         /*
1567          * Set 'VM_NORESERVE' if we should not account for the
1568          * memory use of this mapping.
1569          */
1570         if (flags & MAP_NORESERVE) {
1571                 /* We honor MAP_NORESERVE if allowed to overcommit */
1572                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1573                         vm_flags |= VM_NORESERVE;
1574
1575                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1576                 if (file && is_file_hugepages(file))
1577                         vm_flags |= VM_NORESERVE;
1578         }
1579
1580         addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1581         if (!IS_ERR_VALUE(addr) &&
1582             ((vm_flags & VM_LOCKED) ||
1583              (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1584                 *populate = len;
1585         return addr;
1586 }
1587
1588 unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1589                               unsigned long prot, unsigned long flags,
1590                               unsigned long fd, unsigned long pgoff)
1591 {
1592         struct file *file = NULL;
1593         unsigned long retval;
1594
1595         if (!(flags & MAP_ANONYMOUS)) {
1596                 audit_mmap_fd(fd, flags);
1597                 file = fget(fd);
1598                 if (!file)
1599                         return -EBADF;
1600                 if (is_file_hugepages(file)) {
1601                         len = ALIGN(len, huge_page_size(hstate_file(file)));
1602                 } else if (unlikely(flags & MAP_HUGETLB)) {
1603                         retval = -EINVAL;
1604                         goto out_fput;
1605                 }
1606         } else if (flags & MAP_HUGETLB) {
1607                 struct ucounts *ucounts = NULL;
1608                 struct hstate *hs;
1609
1610                 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1611                 if (!hs)
1612                         return -EINVAL;
1613
1614                 len = ALIGN(len, huge_page_size(hs));
1615                 /*
1616                  * VM_NORESERVE is used because the reservations will be
1617                  * taken when vm_ops->mmap() is called
1618                  * A dummy user value is used because we are not locking
1619                  * memory so no accounting is necessary
1620                  */
1621                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1622                                 VM_NORESERVE,
1623                                 &ucounts, HUGETLB_ANONHUGE_INODE,
1624                                 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1625                 if (IS_ERR(file))
1626                         return PTR_ERR(file);
1627         }
1628
1629         retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1630 out_fput:
1631         if (file)
1632                 fput(file);
1633         return retval;
1634 }
1635
1636 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1637                 unsigned long, prot, unsigned long, flags,
1638                 unsigned long, fd, unsigned long, pgoff)
1639 {
1640         return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1641 }
1642
1643 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1644 struct mmap_arg_struct {
1645         unsigned long addr;
1646         unsigned long len;
1647         unsigned long prot;
1648         unsigned long flags;
1649         unsigned long fd;
1650         unsigned long offset;
1651 };
1652
1653 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1654 {
1655         struct mmap_arg_struct a;
1656
1657         if (copy_from_user(&a, arg, sizeof(a)))
1658                 return -EFAULT;
1659         if (offset_in_page(a.offset))
1660                 return -EINVAL;
1661
1662         return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1663                                a.offset >> PAGE_SHIFT);
1664 }
1665 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1666
1667 /*
1668  * Some shared mappings will want the pages marked read-only
1669  * to track write events. If so, we'll downgrade vm_page_prot
1670  * to the private version (using protection_map[] without the
1671  * VM_SHARED bit).
1672  */
1673 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1674 {
1675         vm_flags_t vm_flags = vma->vm_flags;
1676         const struct vm_operations_struct *vm_ops = vma->vm_ops;
1677
1678         /* If it was private or non-writable, the write bit is already clear */
1679         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1680                 return 0;
1681
1682         /* The backer wishes to know when pages are first written to? */
1683         if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1684                 return 1;
1685
1686         /* The open routine did something to the protections that pgprot_modify
1687          * won't preserve? */
1688         if (pgprot_val(vm_page_prot) !=
1689             pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1690                 return 0;
1691
1692         /* Do we need to track softdirty? */
1693         if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1694                 return 1;
1695
1696         /* Specialty mapping? */
1697         if (vm_flags & VM_PFNMAP)
1698                 return 0;
1699
1700         /* Can the mapping track the dirty pages? */
1701         return vma->vm_file && vma->vm_file->f_mapping &&
1702                 mapping_can_writeback(vma->vm_file->f_mapping);
1703 }
1704
1705 /*
1706  * We account for memory if it's a private writeable mapping,
1707  * not hugepages and VM_NORESERVE wasn't set.
1708  */
1709 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1710 {
1711         /*
1712          * hugetlb has its own accounting separate from the core VM
1713          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1714          */
1715         if (file && is_file_hugepages(file))
1716                 return 0;
1717
1718         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1719 }
1720
1721 unsigned long mmap_region(struct file *file, unsigned long addr,
1722                 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
1723                 struct list_head *uf)
1724 {
1725         struct mm_struct *mm = current->mm;
1726         struct vm_area_struct *vma, *prev, *merge;
1727         int error;
1728         struct rb_node **rb_link, *rb_parent;
1729         unsigned long charged = 0;
1730
1731         /* Check against address space limit. */
1732         if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1733                 unsigned long nr_pages;
1734
1735                 /*
1736                  * MAP_FIXED may remove pages of mappings that intersects with
1737                  * requested mapping. Account for the pages it would unmap.
1738                  */
1739                 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1740
1741                 if (!may_expand_vm(mm, vm_flags,
1742                                         (len >> PAGE_SHIFT) - nr_pages))
1743                         return -ENOMEM;
1744         }
1745
1746         /* Clear old maps, set up prev, rb_link, rb_parent, and uf */
1747         if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf))
1748                 return -ENOMEM;
1749         /*
1750          * Private writable mapping: check memory availability
1751          */
1752         if (accountable_mapping(file, vm_flags)) {
1753                 charged = len >> PAGE_SHIFT;
1754                 if (security_vm_enough_memory_mm(mm, charged))
1755                         return -ENOMEM;
1756                 vm_flags |= VM_ACCOUNT;
1757         }
1758
1759         /*
1760          * Can we just expand an old mapping?
1761          */
1762         vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1763                         NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1764         if (vma)
1765                 goto out;
1766
1767         /*
1768          * Determine the object being mapped and call the appropriate
1769          * specific mapper. the address has already been validated, but
1770          * not unmapped, but the maps are removed from the list.
1771          */
1772         vma = vm_area_alloc(mm);
1773         if (!vma) {
1774                 error = -ENOMEM;
1775                 goto unacct_error;
1776         }
1777
1778         vma->vm_start = addr;
1779         vma->vm_end = addr + len;
1780         vma->vm_flags = vm_flags;
1781         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1782         vma->vm_pgoff = pgoff;
1783
1784         if (file) {
1785                 if (vm_flags & VM_SHARED) {
1786                         error = mapping_map_writable(file->f_mapping);
1787                         if (error)
1788                                 goto free_vma;
1789                 }
1790
1791                 vma->vm_file = get_file(file);
1792                 error = call_mmap(file, vma);
1793                 if (error)
1794                         goto unmap_and_free_vma;
1795
1796                 /* Can addr have changed??
1797                  *
1798                  * Answer: Yes, several device drivers can do it in their
1799                  *         f_op->mmap method. -DaveM
1800                  * Bug: If addr is changed, prev, rb_link, rb_parent should
1801                  *      be updated for vma_link()
1802                  */
1803                 WARN_ON_ONCE(addr != vma->vm_start);
1804
1805                 addr = vma->vm_start;
1806
1807                 /* If vm_flags changed after call_mmap(), we should try merge vma again
1808                  * as we may succeed this time.
1809                  */
1810                 if (unlikely(vm_flags != vma->vm_flags && prev)) {
1811                         merge = vma_merge(mm, prev, vma->vm_start, vma->vm_end, vma->vm_flags,
1812                                 NULL, vma->vm_file, vma->vm_pgoff, NULL, NULL_VM_UFFD_CTX);
1813                         if (merge) {
1814                                 /* ->mmap() can change vma->vm_file and fput the original file. So
1815                                  * fput the vma->vm_file here or we would add an extra fput for file
1816                                  * and cause general protection fault ultimately.
1817                                  */
1818                                 fput(vma->vm_file);
1819                                 vm_area_free(vma);
1820                                 vma = merge;
1821                                 /* Update vm_flags to pick up the change. */
1822                                 vm_flags = vma->vm_flags;
1823                                 goto unmap_writable;
1824                         }
1825                 }
1826
1827                 vm_flags = vma->vm_flags;
1828         } else if (vm_flags & VM_SHARED) {
1829                 error = shmem_zero_setup(vma);
1830                 if (error)
1831                         goto free_vma;
1832         } else {
1833                 vma_set_anonymous(vma);
1834         }
1835
1836         /* Allow architectures to sanity-check the vm_flags */
1837         if (!arch_validate_flags(vma->vm_flags)) {
1838                 error = -EINVAL;
1839                 if (file)
1840                         goto unmap_and_free_vma;
1841                 else
1842                         goto free_vma;
1843         }
1844
1845         vma_link(mm, vma, prev, rb_link, rb_parent);
1846         /* Once vma denies write, undo our temporary denial count */
1847 unmap_writable:
1848         if (file && vm_flags & VM_SHARED)
1849                 mapping_unmap_writable(file->f_mapping);
1850         file = vma->vm_file;
1851 out:
1852         perf_event_mmap(vma);
1853
1854         vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1855         if (vm_flags & VM_LOCKED) {
1856                 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
1857                                         is_vm_hugetlb_page(vma) ||
1858                                         vma == get_gate_vma(current->mm))
1859                         vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1860                 else
1861                         mm->locked_vm += (len >> PAGE_SHIFT);
1862         }
1863
1864         if (file)
1865                 uprobe_mmap(vma);
1866
1867         /*
1868          * New (or expanded) vma always get soft dirty status.
1869          * Otherwise user-space soft-dirty page tracker won't
1870          * be able to distinguish situation when vma area unmapped,
1871          * then new mapped in-place (which must be aimed as
1872          * a completely new data area).
1873          */
1874         vma->vm_flags |= VM_SOFTDIRTY;
1875
1876         vma_set_page_prot(vma);
1877
1878         return addr;
1879
1880 unmap_and_free_vma:
1881         fput(vma->vm_file);
1882         vma->vm_file = NULL;
1883
1884         /* Undo any partial mapping done by a device driver. */
1885         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1886         charged = 0;
1887         if (vm_flags & VM_SHARED)
1888                 mapping_unmap_writable(file->f_mapping);
1889 free_vma:
1890         vm_area_free(vma);
1891 unacct_error:
1892         if (charged)
1893                 vm_unacct_memory(charged);
1894         return error;
1895 }
1896
1897 static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1898 {
1899         /*
1900          * We implement the search by looking for an rbtree node that
1901          * immediately follows a suitable gap. That is,
1902          * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1903          * - gap_end   = vma->vm_start        >= info->low_limit  + length;
1904          * - gap_end - gap_start >= length
1905          */
1906
1907         struct mm_struct *mm = current->mm;
1908         struct vm_area_struct *vma;
1909         unsigned long length, low_limit, high_limit, gap_start, gap_end;
1910
1911         /* Adjust search length to account for worst case alignment overhead */
1912         length = info->length + info->align_mask;
1913         if (length < info->length)
1914                 return -ENOMEM;
1915
1916         /* Adjust search limits by the desired length */
1917         if (info->high_limit < length)
1918                 return -ENOMEM;
1919         high_limit = info->high_limit - length;
1920
1921         if (info->low_limit > high_limit)
1922                 return -ENOMEM;
1923         low_limit = info->low_limit + length;
1924
1925         /* Check if rbtree root looks promising */
1926         if (RB_EMPTY_ROOT(&mm->mm_rb))
1927                 goto check_highest;
1928         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1929         if (vma->rb_subtree_gap < length)
1930                 goto check_highest;
1931
1932         while (true) {
1933                 /* Visit left subtree if it looks promising */
1934                 gap_end = vm_start_gap(vma);
1935                 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1936                         struct vm_area_struct *left =
1937                                 rb_entry(vma->vm_rb.rb_left,
1938                                          struct vm_area_struct, vm_rb);
1939                         if (left->rb_subtree_gap >= length) {
1940                                 vma = left;
1941                                 continue;
1942                         }
1943                 }
1944
1945                 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1946 check_current:
1947                 /* Check if current node has a suitable gap */
1948                 if (gap_start > high_limit)
1949                         return -ENOMEM;
1950                 if (gap_end >= low_limit &&
1951                     gap_end > gap_start && gap_end - gap_start >= length)
1952                         goto found;
1953
1954                 /* Visit right subtree if it looks promising */
1955                 if (vma->vm_rb.rb_right) {
1956                         struct vm_area_struct *right =
1957                                 rb_entry(vma->vm_rb.rb_right,
1958                                          struct vm_area_struct, vm_rb);
1959                         if (right->rb_subtree_gap >= length) {
1960                                 vma = right;
1961                                 continue;
1962                         }
1963                 }
1964
1965                 /* Go back up the rbtree to find next candidate node */
1966                 while (true) {
1967                         struct rb_node *prev = &vma->vm_rb;
1968                         if (!rb_parent(prev))
1969                                 goto check_highest;
1970                         vma = rb_entry(rb_parent(prev),
1971                                        struct vm_area_struct, vm_rb);
1972                         if (prev == vma->vm_rb.rb_left) {
1973                                 gap_start = vm_end_gap(vma->vm_prev);
1974                                 gap_end = vm_start_gap(vma);
1975                                 goto check_current;
1976                         }
1977                 }
1978         }
1979
1980 check_highest:
1981         /* Check highest gap, which does not precede any rbtree node */
1982         gap_start = mm->highest_vm_end;
1983         gap_end = ULONG_MAX;  /* Only for VM_BUG_ON below */
1984         if (gap_start > high_limit)
1985                 return -ENOMEM;
1986
1987 found:
1988         /* We found a suitable gap. Clip it with the original low_limit. */
1989         if (gap_start < info->low_limit)
1990                 gap_start = info->low_limit;
1991
1992         /* Adjust gap address to the desired alignment */
1993         gap_start += (info->align_offset - gap_start) & info->align_mask;
1994
1995         VM_BUG_ON(gap_start + info->length > info->high_limit);
1996         VM_BUG_ON(gap_start + info->length > gap_end);
1997         return gap_start;
1998 }
1999
2000 static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
2001 {
2002         struct mm_struct *mm = current->mm;
2003         struct vm_area_struct *vma;
2004         unsigned long length, low_limit, high_limit, gap_start, gap_end;
2005
2006         /* Adjust search length to account for worst case alignment overhead */
2007         length = info->length + info->align_mask;
2008         if (length < info->length)
2009                 return -ENOMEM;
2010
2011         /*
2012          * Adjust search limits by the desired length.
2013          * See implementation comment at top of unmapped_area().
2014          */
2015         gap_end = info->high_limit;
2016         if (gap_end < length)
2017                 return -ENOMEM;
2018         high_limit = gap_end - length;
2019
2020         if (info->low_limit > high_limit)
2021                 return -ENOMEM;
2022         low_limit = info->low_limit + length;
2023
2024         /* Check highest gap, which does not precede any rbtree node */
2025         gap_start = mm->highest_vm_end;
2026         if (gap_start <= high_limit)
2027                 goto found_highest;
2028
2029         /* Check if rbtree root looks promising */
2030         if (RB_EMPTY_ROOT(&mm->mm_rb))
2031                 return -ENOMEM;
2032         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
2033         if (vma->rb_subtree_gap < length)
2034                 return -ENOMEM;
2035
2036         while (true) {
2037                 /* Visit right subtree if it looks promising */
2038                 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
2039                 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
2040                         struct vm_area_struct *right =
2041                                 rb_entry(vma->vm_rb.rb_right,
2042                                          struct vm_area_struct, vm_rb);
2043                         if (right->rb_subtree_gap >= length) {
2044                                 vma = right;
2045                                 continue;
2046                         }
2047                 }
2048
2049 check_current:
2050                 /* Check if current node has a suitable gap */
2051                 gap_end = vm_start_gap(vma);
2052                 if (gap_end < low_limit)
2053                         return -ENOMEM;
2054                 if (gap_start <= high_limit &&
2055                     gap_end > gap_start && gap_end - gap_start >= length)
2056                         goto found;
2057
2058                 /* Visit left subtree if it looks promising */
2059                 if (vma->vm_rb.rb_left) {
2060                         struct vm_area_struct *left =
2061                                 rb_entry(vma->vm_rb.rb_left,
2062                                          struct vm_area_struct, vm_rb);
2063                         if (left->rb_subtree_gap >= length) {
2064                                 vma = left;
2065                                 continue;
2066                         }
2067                 }
2068
2069                 /* Go back up the rbtree to find next candidate node */
2070                 while (true) {
2071                         struct rb_node *prev = &vma->vm_rb;
2072                         if (!rb_parent(prev))
2073                                 return -ENOMEM;
2074                         vma = rb_entry(rb_parent(prev),
2075                                        struct vm_area_struct, vm_rb);
2076                         if (prev == vma->vm_rb.rb_right) {
2077                                 gap_start = vma->vm_prev ?
2078                                         vm_end_gap(vma->vm_prev) : 0;
2079                                 goto check_current;
2080                         }
2081                 }
2082         }
2083
2084 found:
2085         /* We found a suitable gap. Clip it with the original high_limit. */
2086         if (gap_end > info->high_limit)
2087                 gap_end = info->high_limit;
2088
2089 found_highest:
2090         /* Compute highest gap address at the desired alignment */
2091         gap_end -= info->length;
2092         gap_end -= (gap_end - info->align_offset) & info->align_mask;
2093
2094         VM_BUG_ON(gap_end < info->low_limit);
2095         VM_BUG_ON(gap_end < gap_start);
2096         return gap_end;
2097 }
2098
2099 /*
2100  * Search for an unmapped address range.
2101  *
2102  * We are looking for a range that:
2103  * - does not intersect with any VMA;
2104  * - is contained within the [low_limit, high_limit) interval;
2105  * - is at least the desired size.
2106  * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
2107  */
2108 unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
2109 {
2110         unsigned long addr;
2111
2112         if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
2113                 addr = unmapped_area_topdown(info);
2114         else
2115                 addr = unmapped_area(info);
2116
2117         trace_vm_unmapped_area(addr, info);
2118         return addr;
2119 }
2120
2121 #ifndef arch_get_mmap_end
2122 #define arch_get_mmap_end(addr) (TASK_SIZE)
2123 #endif
2124
2125 #ifndef arch_get_mmap_base
2126 #define arch_get_mmap_base(addr, base) (base)
2127 #endif
2128
2129 /* Get an address range which is currently unmapped.
2130  * For shmat() with addr=0.
2131  *
2132  * Ugly calling convention alert:
2133  * Return value with the low bits set means error value,
2134  * ie
2135  *      if (ret & ~PAGE_MASK)
2136  *              error = ret;
2137  *
2138  * This function "knows" that -ENOMEM has the bits set.
2139  */
2140 #ifndef HAVE_ARCH_UNMAPPED_AREA
2141 unsigned long
2142 arch_get_unmapped_area(struct file *filp, unsigned long addr,
2143                 unsigned long len, unsigned long pgoff, unsigned long flags)
2144 {
2145         struct mm_struct *mm = current->mm;
2146         struct vm_area_struct *vma, *prev;
2147         struct vm_unmapped_area_info info;
2148         const unsigned long mmap_end = arch_get_mmap_end(addr);
2149
2150         if (len > mmap_end - mmap_min_addr)
2151                 return -ENOMEM;
2152
2153         if (flags & MAP_FIXED)
2154                 return addr;
2155
2156         if (addr) {
2157                 addr = PAGE_ALIGN(addr);
2158                 vma = find_vma_prev(mm, addr, &prev);
2159                 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2160                     (!vma || addr + len <= vm_start_gap(vma)) &&
2161                     (!prev || addr >= vm_end_gap(prev)))
2162                         return addr;
2163         }
2164
2165         info.flags = 0;
2166         info.length = len;
2167         info.low_limit = mm->mmap_base;
2168         info.high_limit = mmap_end;
2169         info.align_mask = 0;
2170         info.align_offset = 0;
2171         return vm_unmapped_area(&info);
2172 }
2173 #endif
2174
2175 /*
2176  * This mmap-allocator allocates new areas top-down from below the
2177  * stack's low limit (the base):
2178  */
2179 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2180 unsigned long
2181 arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
2182                           unsigned long len, unsigned long pgoff,
2183                           unsigned long flags)
2184 {
2185         struct vm_area_struct *vma, *prev;
2186         struct mm_struct *mm = current->mm;
2187         struct vm_unmapped_area_info info;
2188         const unsigned long mmap_end = arch_get_mmap_end(addr);
2189
2190         /* requested length too big for entire address space */
2191         if (len > mmap_end - mmap_min_addr)
2192                 return -ENOMEM;
2193
2194         if (flags & MAP_FIXED)
2195                 return addr;
2196
2197         /* requesting a specific address */
2198         if (addr) {
2199                 addr = PAGE_ALIGN(addr);
2200                 vma = find_vma_prev(mm, addr, &prev);
2201                 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2202                                 (!vma || addr + len <= vm_start_gap(vma)) &&
2203                                 (!prev || addr >= vm_end_gap(prev)))
2204                         return addr;
2205         }
2206
2207         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2208         info.length = len;
2209         info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2210         info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
2211         info.align_mask = 0;
2212         info.align_offset = 0;
2213         addr = vm_unmapped_area(&info);
2214
2215         /*
2216          * A failed mmap() very likely causes application failure,
2217          * so fall back to the bottom-up function here. This scenario
2218          * can happen with large stack limits and large mmap()
2219          * allocations.
2220          */
2221         if (offset_in_page(addr)) {
2222                 VM_BUG_ON(addr != -ENOMEM);
2223                 info.flags = 0;
2224                 info.low_limit = TASK_UNMAPPED_BASE;
2225                 info.high_limit = mmap_end;
2226                 addr = vm_unmapped_area(&info);
2227         }
2228
2229         return addr;
2230 }
2231 #endif
2232
2233 unsigned long
2234 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2235                 unsigned long pgoff, unsigned long flags)
2236 {
2237         unsigned long (*get_area)(struct file *, unsigned long,
2238                                   unsigned long, unsigned long, unsigned long);
2239
2240         unsigned long error = arch_mmap_check(addr, len, flags);
2241         if (error)
2242                 return error;
2243
2244         /* Careful about overflows.. */
2245         if (len > TASK_SIZE)
2246                 return -ENOMEM;
2247
2248         get_area = current->mm->get_unmapped_area;
2249         if (file) {
2250                 if (file->f_op->get_unmapped_area)
2251                         get_area = file->f_op->get_unmapped_area;
2252         } else if (flags & MAP_SHARED) {
2253                 /*
2254                  * mmap_region() will call shmem_zero_setup() to create a file,
2255                  * so use shmem's get_unmapped_area in case it can be huge.
2256                  * do_mmap() will clear pgoff, so match alignment.
2257                  */
2258                 pgoff = 0;
2259                 get_area = shmem_get_unmapped_area;
2260         }
2261
2262         addr = get_area(file, addr, len, pgoff, flags);
2263         if (IS_ERR_VALUE(addr))
2264                 return addr;
2265
2266         if (addr > TASK_SIZE - len)
2267                 return -ENOMEM;
2268         if (offset_in_page(addr))
2269                 return -EINVAL;
2270
2271         error = security_mmap_addr(addr);
2272         return error ? error : addr;
2273 }
2274
2275 EXPORT_SYMBOL(get_unmapped_area);
2276
2277 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
2278 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2279 {
2280         struct rb_node *rb_node;
2281         struct vm_area_struct *vma;
2282
2283         /* Check the cache first. */
2284         vma = vmacache_find(mm, addr);
2285         if (likely(vma))
2286                 return vma;
2287
2288         rb_node = mm->mm_rb.rb_node;
2289
2290         while (rb_node) {
2291                 struct vm_area_struct *tmp;
2292
2293                 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2294
2295                 if (tmp->vm_end > addr) {
2296                         vma = tmp;
2297                         if (tmp->vm_start <= addr)
2298                                 break;
2299                         rb_node = rb_node->rb_left;
2300                 } else
2301                         rb_node = rb_node->rb_right;
2302         }
2303
2304         if (vma)
2305                 vmacache_update(addr, vma);
2306         return vma;
2307 }
2308
2309 EXPORT_SYMBOL(find_vma);
2310
2311 /*
2312  * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2313  */
2314 struct vm_area_struct *
2315 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2316                         struct vm_area_struct **pprev)
2317 {
2318         struct vm_area_struct *vma;
2319
2320         vma = find_vma(mm, addr);
2321         if (vma) {
2322                 *pprev = vma->vm_prev;
2323         } else {
2324                 struct rb_node *rb_node = rb_last(&mm->mm_rb);
2325
2326                 *pprev = rb_node ? rb_entry(rb_node, struct vm_area_struct, vm_rb) : NULL;
2327         }
2328         return vma;
2329 }
2330
2331 /*
2332  * Verify that the stack growth is acceptable and
2333  * update accounting. This is shared with both the
2334  * grow-up and grow-down cases.
2335  */
2336 static int acct_stack_growth(struct vm_area_struct *vma,
2337                              unsigned long size, unsigned long grow)
2338 {
2339         struct mm_struct *mm = vma->vm_mm;
2340         unsigned long new_start;
2341
2342         /* address space limit tests */
2343         if (!may_expand_vm(mm, vma->vm_flags, grow))
2344                 return -ENOMEM;
2345
2346         /* Stack limit test */
2347         if (size > rlimit(RLIMIT_STACK))
2348                 return -ENOMEM;
2349
2350         /* mlock limit tests */
2351         if (vma->vm_flags & VM_LOCKED) {
2352                 unsigned long locked;
2353                 unsigned long limit;
2354                 locked = mm->locked_vm + grow;
2355                 limit = rlimit(RLIMIT_MEMLOCK);
2356                 limit >>= PAGE_SHIFT;
2357                 if (locked > limit && !capable(CAP_IPC_LOCK))
2358                         return -ENOMEM;
2359         }
2360
2361         /* Check to ensure the stack will not grow into a hugetlb-only region */
2362         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2363                         vma->vm_end - size;
2364         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2365                 return -EFAULT;
2366
2367         /*
2368          * Overcommit..  This must be the final test, as it will
2369          * update security statistics.
2370          */
2371         if (security_vm_enough_memory_mm(mm, grow))
2372                 return -ENOMEM;
2373
2374         return 0;
2375 }
2376
2377 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2378 /*
2379  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2380  * vma is the last one with address > vma->vm_end.  Have to extend vma.
2381  */
2382 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2383 {
2384         struct mm_struct *mm = vma->vm_mm;
2385         struct vm_area_struct *next;
2386         unsigned long gap_addr;
2387         int error = 0;
2388
2389         if (!(vma->vm_flags & VM_GROWSUP))
2390                 return -EFAULT;
2391
2392         /* Guard against exceeding limits of the address space. */
2393         address &= PAGE_MASK;
2394         if (address >= (TASK_SIZE & PAGE_MASK))
2395                 return -ENOMEM;
2396         address += PAGE_SIZE;
2397
2398         /* Enforce stack_guard_gap */
2399         gap_addr = address + stack_guard_gap;
2400
2401         /* Guard against overflow */
2402         if (gap_addr < address || gap_addr > TASK_SIZE)
2403                 gap_addr = TASK_SIZE;
2404
2405         next = vma->vm_next;
2406         if (next && next->vm_start < gap_addr && vma_is_accessible(next)) {
2407                 if (!(next->vm_flags & VM_GROWSUP))
2408                         return -ENOMEM;
2409                 /* Check that both stack segments have the same anon_vma? */
2410         }
2411
2412         /* We must make sure the anon_vma is allocated. */
2413         if (unlikely(anon_vma_prepare(vma)))
2414                 return -ENOMEM;
2415
2416         /*
2417          * vma->vm_start/vm_end cannot change under us because the caller
2418          * is required to hold the mmap_lock in read mode.  We need the
2419          * anon_vma lock to serialize against concurrent expand_stacks.
2420          */
2421         anon_vma_lock_write(vma->anon_vma);
2422
2423         /* Somebody else might have raced and expanded it already */
2424         if (address > vma->vm_end) {
2425                 unsigned long size, grow;
2426
2427                 size = address - vma->vm_start;
2428                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2429
2430                 error = -ENOMEM;
2431                 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2432                         error = acct_stack_growth(vma, size, grow);
2433                         if (!error) {
2434                                 /*
2435                                  * vma_gap_update() doesn't support concurrent
2436                                  * updates, but we only hold a shared mmap_lock
2437                                  * lock here, so we need to protect against
2438                                  * concurrent vma expansions.
2439                                  * anon_vma_lock_write() doesn't help here, as
2440                                  * we don't guarantee that all growable vmas
2441                                  * in a mm share the same root anon vma.
2442                                  * So, we reuse mm->page_table_lock to guard
2443                                  * against concurrent vma expansions.
2444                                  */
2445                                 spin_lock(&mm->page_table_lock);
2446                                 if (vma->vm_flags & VM_LOCKED)
2447                                         mm->locked_vm += grow;
2448                                 vm_stat_account(mm, vma->vm_flags, grow);
2449                                 anon_vma_interval_tree_pre_update_vma(vma);
2450                                 vma->vm_end = address;
2451                                 anon_vma_interval_tree_post_update_vma(vma);
2452                                 if (vma->vm_next)
2453                                         vma_gap_update(vma->vm_next);
2454                                 else
2455                                         mm->highest_vm_end = vm_end_gap(vma);
2456                                 spin_unlock(&mm->page_table_lock);
2457
2458                                 perf_event_mmap(vma);
2459                         }
2460                 }
2461         }
2462         anon_vma_unlock_write(vma->anon_vma);
2463         khugepaged_enter_vma_merge(vma, vma->vm_flags);
2464         validate_mm(mm);
2465         return error;
2466 }
2467 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2468
2469 /*
2470  * vma is the first one with address < vma->vm_start.  Have to extend vma.
2471  */
2472 int expand_downwards(struct vm_area_struct *vma,
2473                                    unsigned long address)
2474 {
2475         struct mm_struct *mm = vma->vm_mm;
2476         struct vm_area_struct *prev;
2477         int error = 0;
2478
2479         address &= PAGE_MASK;
2480         if (address < mmap_min_addr)
2481                 return -EPERM;
2482
2483         /* Enforce stack_guard_gap */
2484         prev = vma->vm_prev;
2485         /* Check that both stack segments have the same anon_vma? */
2486         if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2487                         vma_is_accessible(prev)) {
2488                 if (address - prev->vm_end < stack_guard_gap)
2489                         return -ENOMEM;
2490         }
2491
2492         /* We must make sure the anon_vma is allocated. */
2493         if (unlikely(anon_vma_prepare(vma)))
2494                 return -ENOMEM;
2495
2496         /*
2497          * vma->vm_start/vm_end cannot change under us because the caller
2498          * is required to hold the mmap_lock in read mode.  We need the
2499          * anon_vma lock to serialize against concurrent expand_stacks.
2500          */
2501         anon_vma_lock_write(vma->anon_vma);
2502
2503         /* Somebody else might have raced and expanded it already */
2504         if (address < vma->vm_start) {
2505                 unsigned long size, grow;
2506
2507                 size = vma->vm_end - address;
2508                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2509
2510                 error = -ENOMEM;
2511                 if (grow <= vma->vm_pgoff) {
2512                         error = acct_stack_growth(vma, size, grow);
2513                         if (!error) {
2514                                 /*
2515                                  * vma_gap_update() doesn't support concurrent
2516                                  * updates, but we only hold a shared mmap_lock
2517                                  * lock here, so we need to protect against
2518                                  * concurrent vma expansions.
2519                                  * anon_vma_lock_write() doesn't help here, as
2520                                  * we don't guarantee that all growable vmas
2521                                  * in a mm share the same root anon vma.
2522                                  * So, we reuse mm->page_table_lock to guard
2523                                  * against concurrent vma expansions.
2524                                  */
2525                                 spin_lock(&mm->page_table_lock);
2526                                 if (vma->vm_flags & VM_LOCKED)
2527                                         mm->locked_vm += grow;
2528                                 vm_stat_account(mm, vma->vm_flags, grow);
2529                                 anon_vma_interval_tree_pre_update_vma(vma);
2530                                 vma->vm_start = address;
2531                                 vma->vm_pgoff -= grow;
2532                                 anon_vma_interval_tree_post_update_vma(vma);
2533                                 vma_gap_update(vma);
2534                                 spin_unlock(&mm->page_table_lock);
2535
2536                                 perf_event_mmap(vma);
2537                         }
2538                 }
2539         }
2540         anon_vma_unlock_write(vma->anon_vma);
2541         khugepaged_enter_vma_merge(vma, vma->vm_flags);
2542         validate_mm(mm);
2543         return error;
2544 }
2545
2546 /* enforced gap between the expanding stack and other mappings. */
2547 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2548
2549 static int __init cmdline_parse_stack_guard_gap(char *p)
2550 {
2551         unsigned long val;
2552         char *endptr;
2553
2554         val = simple_strtoul(p, &endptr, 10);
2555         if (!*endptr)
2556                 stack_guard_gap = val << PAGE_SHIFT;
2557
2558         return 0;
2559 }
2560 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2561
2562 #ifdef CONFIG_STACK_GROWSUP
2563 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2564 {
2565         return expand_upwards(vma, address);
2566 }
2567
2568 struct vm_area_struct *
2569 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2570 {
2571         struct vm_area_struct *vma, *prev;
2572
2573         addr &= PAGE_MASK;
2574         vma = find_vma_prev(mm, addr, &prev);
2575         if (vma && (vma->vm_start <= addr))
2576                 return vma;
2577         /* don't alter vm_end if the coredump is running */
2578         if (!prev || expand_stack(prev, addr))
2579                 return NULL;
2580         if (prev->vm_flags & VM_LOCKED)
2581                 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2582         return prev;
2583 }
2584 #else
2585 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2586 {
2587         return expand_downwards(vma, address);
2588 }
2589
2590 struct vm_area_struct *
2591 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2592 {
2593         struct vm_area_struct *vma;
2594         unsigned long start;
2595
2596         addr &= PAGE_MASK;
2597         vma = find_vma(mm, addr);
2598         if (!vma)
2599                 return NULL;
2600         if (vma->vm_start <= addr)
2601                 return vma;
2602         if (!(vma->vm_flags & VM_GROWSDOWN))
2603                 return NULL;
2604         start = vma->vm_start;
2605         if (expand_stack(vma, addr))
2606                 return NULL;
2607         if (vma->vm_flags & VM_LOCKED)
2608                 populate_vma_page_range(vma, addr, start, NULL);
2609         return vma;
2610 }
2611 #endif
2612
2613 EXPORT_SYMBOL_GPL(find_extend_vma);
2614
2615 /*
2616  * Ok - we have the memory areas we should free on the vma list,
2617  * so release them, and do the vma updates.
2618  *
2619  * Called with the mm semaphore held.
2620  */
2621 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2622 {
2623         unsigned long nr_accounted = 0;
2624
2625         /* Update high watermark before we lower total_vm */
2626         update_hiwater_vm(mm);
2627         do {
2628                 long nrpages = vma_pages(vma);
2629
2630                 if (vma->vm_flags & VM_ACCOUNT)
2631                         nr_accounted += nrpages;
2632                 vm_stat_account(mm, vma->vm_flags, -nrpages);
2633                 vma = remove_vma(vma);
2634         } while (vma);
2635         vm_unacct_memory(nr_accounted);
2636         validate_mm(mm);
2637 }
2638
2639 /*
2640  * Get rid of page table information in the indicated region.
2641  *
2642  * Called with the mm semaphore held.
2643  */
2644 static void unmap_region(struct mm_struct *mm,
2645                 struct vm_area_struct *vma, struct vm_area_struct *prev,
2646                 unsigned long start, unsigned long end)
2647 {
2648         struct vm_area_struct *next = vma_next(mm, prev);
2649         struct mmu_gather tlb;
2650
2651         lru_add_drain();
2652         tlb_gather_mmu(&tlb, mm);
2653         update_hiwater_rss(mm);
2654         unmap_vmas(&tlb, vma, start, end);
2655         free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2656                                  next ? next->vm_start : USER_PGTABLES_CEILING);
2657         tlb_finish_mmu(&tlb);
2658 }
2659
2660 /*
2661  * Create a list of vma's touched by the unmap, removing them from the mm's
2662  * vma list as we go..
2663  */
2664 static bool
2665 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2666         struct vm_area_struct *prev, unsigned long end)
2667 {
2668         struct vm_area_struct **insertion_point;
2669         struct vm_area_struct *tail_vma = NULL;
2670
2671         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2672         vma->vm_prev = NULL;
2673         do {
2674                 vma_rb_erase(vma, &mm->mm_rb);
2675                 mm->map_count--;
2676                 tail_vma = vma;
2677                 vma = vma->vm_next;
2678         } while (vma && vma->vm_start < end);
2679         *insertion_point = vma;
2680         if (vma) {
2681                 vma->vm_prev = prev;
2682                 vma_gap_update(vma);
2683         } else
2684                 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2685         tail_vma->vm_next = NULL;
2686
2687         /* Kill the cache */
2688         vmacache_invalidate(mm);
2689
2690         /*
2691          * Do not downgrade mmap_lock if we are next to VM_GROWSDOWN or
2692          * VM_GROWSUP VMA. Such VMAs can change their size under
2693          * down_read(mmap_lock) and collide with the VMA we are about to unmap.
2694          */
2695         if (vma && (vma->vm_flags & VM_GROWSDOWN))
2696                 return false;
2697         if (prev && (prev->vm_flags & VM_GROWSUP))
2698                 return false;
2699         return true;
2700 }
2701
2702 /*
2703  * __split_vma() bypasses sysctl_max_map_count checking.  We use this where it
2704  * has already been checked or doesn't make sense to fail.
2705  */
2706 int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2707                 unsigned long addr, int new_below)
2708 {
2709         struct vm_area_struct *new;
2710         int err;
2711
2712         if (vma->vm_ops && vma->vm_ops->may_split) {
2713                 err = vma->vm_ops->may_split(vma, addr);
2714                 if (err)
2715                         return err;
2716         }
2717
2718         new = vm_area_dup(vma);
2719         if (!new)
2720                 return -ENOMEM;
2721
2722         if (new_below)
2723                 new->vm_end = addr;
2724         else {
2725                 new->vm_start = addr;
2726                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2727         }
2728
2729         err = vma_dup_policy(vma, new);
2730         if (err)
2731                 goto out_free_vma;
2732
2733         err = anon_vma_clone(new, vma);
2734         if (err)
2735                 goto out_free_mpol;
2736
2737         if (new->vm_file)
2738                 get_file(new->vm_file);
2739
2740         if (new->vm_ops && new->vm_ops->open)
2741                 new->vm_ops->open(new);
2742
2743         if (new_below)
2744                 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2745                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
2746         else
2747                 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2748
2749         /* Success. */
2750         if (!err)
2751                 return 0;
2752
2753         /* Clean everything up if vma_adjust failed. */
2754         if (new->vm_ops && new->vm_ops->close)
2755                 new->vm_ops->close(new);
2756         if (new->vm_file)
2757                 fput(new->vm_file);
2758         unlink_anon_vmas(new);
2759  out_free_mpol:
2760         mpol_put(vma_policy(new));
2761  out_free_vma:
2762         vm_area_free(new);
2763         return err;
2764 }
2765
2766 /*
2767  * Split a vma into two pieces at address 'addr', a new vma is allocated
2768  * either for the first part or the tail.
2769  */
2770 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2771               unsigned long addr, int new_below)
2772 {
2773         if (mm->map_count >= sysctl_max_map_count)
2774                 return -ENOMEM;
2775
2776         return __split_vma(mm, vma, addr, new_below);
2777 }
2778
2779 static inline void
2780 unlock_range(struct vm_area_struct *start, unsigned long limit)
2781 {
2782         struct mm_struct *mm = start->vm_mm;
2783         struct vm_area_struct *tmp = start;
2784
2785         while (tmp && tmp->vm_start < limit) {
2786                 if (tmp->vm_flags & VM_LOCKED) {
2787                         mm->locked_vm -= vma_pages(tmp);
2788                         munlock_vma_pages_all(tmp);
2789                 }
2790
2791                 tmp = tmp->vm_next;
2792         }
2793 }
2794
2795 /* Munmap is split into 2 main parts -- this part which finds
2796  * what needs doing, and the areas themselves, which do the
2797  * work.  This now handles partial unmappings.
2798  * Jeremy Fitzhardinge <jeremy@goop.org>
2799  */
2800 int __do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2801                 struct list_head *uf, bool downgrade)
2802 {
2803         unsigned long end;
2804         struct vm_area_struct *vma, *prev, *last;
2805
2806         if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2807                 return -EINVAL;
2808
2809         len = PAGE_ALIGN(len);
2810         end = start + len;
2811         if (len == 0)
2812                 return -EINVAL;
2813
2814         /*
2815          * arch_unmap() might do unmaps itself.  It must be called
2816          * and finish any rbtree manipulation before this code
2817          * runs and also starts to manipulate the rbtree.
2818          */
2819         arch_unmap(mm, start, end);
2820
2821         /* Find the first overlapping VMA where start < vma->vm_end */
2822         vma = find_vma_intersection(mm, start, end);
2823         if (!vma)
2824                 return 0;
2825         prev = vma->vm_prev;
2826
2827         /*
2828          * If we need to split any vma, do it now to save pain later.
2829          *
2830          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2831          * unmapped vm_area_struct will remain in use: so lower split_vma
2832          * places tmp vma above, and higher split_vma places tmp vma below.
2833          */
2834         if (start > vma->vm_start) {
2835                 int error;
2836
2837                 /*
2838                  * Make sure that map_count on return from munmap() will
2839                  * not exceed its limit; but let map_count go just above
2840                  * its limit temporarily, to help free resources as expected.
2841                  */
2842                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2843                         return -ENOMEM;
2844
2845                 error = __split_vma(mm, vma, start, 0);
2846                 if (error)
2847                         return error;
2848                 prev = vma;
2849         }
2850
2851         /* Does it split the last one? */
2852         last = find_vma(mm, end);
2853         if (last && end > last->vm_start) {
2854                 int error = __split_vma(mm, last, end, 1);
2855                 if (error)
2856                         return error;
2857         }
2858         vma = vma_next(mm, prev);
2859
2860         if (unlikely(uf)) {
2861                 /*
2862                  * If userfaultfd_unmap_prep returns an error the vmas
2863                  * will remain split, but userland will get a
2864                  * highly unexpected error anyway. This is no
2865                  * different than the case where the first of the two
2866                  * __split_vma fails, but we don't undo the first
2867                  * split, despite we could. This is unlikely enough
2868                  * failure that it's not worth optimizing it for.
2869                  */
2870                 int error = userfaultfd_unmap_prep(vma, start, end, uf);
2871                 if (error)
2872                         return error;
2873         }
2874
2875         /*
2876          * unlock any mlock()ed ranges before detaching vmas
2877          */
2878         if (mm->locked_vm)
2879                 unlock_range(vma, end);
2880
2881         /* Detach vmas from rbtree */
2882         if (!detach_vmas_to_be_unmapped(mm, vma, prev, end))
2883                 downgrade = false;
2884
2885         if (downgrade)
2886                 mmap_write_downgrade(mm);
2887
2888         unmap_region(mm, vma, prev, start, end);
2889
2890         /* Fix up all other VM information */
2891         remove_vma_list(mm, vma);
2892
2893         return downgrade ? 1 : 0;
2894 }
2895
2896 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2897               struct list_head *uf)
2898 {
2899         return __do_munmap(mm, start, len, uf, false);
2900 }
2901
2902 static int __vm_munmap(unsigned long start, size_t len, bool downgrade)
2903 {
2904         int ret;
2905         struct mm_struct *mm = current->mm;
2906         LIST_HEAD(uf);
2907
2908         if (mmap_write_lock_killable(mm))
2909                 return -EINTR;
2910
2911         ret = __do_munmap(mm, start, len, &uf, downgrade);
2912         /*
2913          * Returning 1 indicates mmap_lock is downgraded.
2914          * But 1 is not legal return value of vm_munmap() and munmap(), reset
2915          * it to 0 before return.
2916          */
2917         if (ret == 1) {
2918                 mmap_read_unlock(mm);
2919                 ret = 0;
2920         } else
2921                 mmap_write_unlock(mm);
2922
2923         userfaultfd_unmap_complete(mm, &uf);
2924         return ret;
2925 }
2926
2927 int vm_munmap(unsigned long start, size_t len)
2928 {
2929         return __vm_munmap(start, len, false);
2930 }
2931 EXPORT_SYMBOL(vm_munmap);
2932
2933 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2934 {
2935         addr = untagged_addr(addr);
2936         profile_munmap(addr);
2937         return __vm_munmap(addr, len, true);
2938 }
2939
2940
2941 /*
2942  * Emulation of deprecated remap_file_pages() syscall.
2943  */
2944 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2945                 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2946 {
2947
2948         struct mm_struct *mm = current->mm;
2949         struct vm_area_struct *vma;
2950         unsigned long populate = 0;
2951         unsigned long ret = -EINVAL;
2952         struct file *file;
2953
2954         pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.rst.\n",
2955                      current->comm, current->pid);
2956
2957         if (prot)
2958                 return ret;
2959         start = start & PAGE_MASK;
2960         size = size & PAGE_MASK;
2961
2962         if (start + size <= start)
2963                 return ret;
2964
2965         /* Does pgoff wrap? */
2966         if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2967                 return ret;
2968
2969         if (mmap_write_lock_killable(mm))
2970                 return -EINTR;
2971
2972         vma = find_vma(mm, start);
2973
2974         if (!vma || !(vma->vm_flags & VM_SHARED))
2975                 goto out;
2976
2977         if (start < vma->vm_start)
2978                 goto out;
2979
2980         if (start + size > vma->vm_end) {
2981                 struct vm_area_struct *next;
2982
2983                 for (next = vma->vm_next; next; next = next->vm_next) {
2984                         /* hole between vmas ? */
2985                         if (next->vm_start != next->vm_prev->vm_end)
2986                                 goto out;
2987
2988                         if (next->vm_file != vma->vm_file)
2989                                 goto out;
2990
2991                         if (next->vm_flags != vma->vm_flags)
2992                                 goto out;
2993
2994                         if (start + size <= next->vm_end)
2995                                 break;
2996                 }
2997
2998                 if (!next)
2999                         goto out;
3000         }
3001
3002         prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
3003         prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
3004         prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
3005
3006         flags &= MAP_NONBLOCK;
3007         flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
3008         if (vma->vm_flags & VM_LOCKED)
3009                 flags |= MAP_LOCKED;
3010
3011         file = get_file(vma->vm_file);
3012         ret = do_mmap(vma->vm_file, start, size,
3013                         prot, flags, pgoff, &populate, NULL);
3014         fput(file);
3015 out:
3016         mmap_write_unlock(mm);
3017         if (populate)
3018                 mm_populate(ret, populate);
3019         if (!IS_ERR_VALUE(ret))
3020                 ret = 0;
3021         return ret;
3022 }
3023
3024 /*
3025  *  this is really a simplified "do_mmap".  it only handles
3026  *  anonymous maps.  eventually we may be able to do some
3027  *  brk-specific accounting here.
3028  */
3029 static int do_brk_flags(unsigned long addr, unsigned long len, unsigned long flags, struct list_head *uf)
3030 {
3031         struct mm_struct *mm = current->mm;
3032         struct vm_area_struct *vma, *prev;
3033         struct rb_node **rb_link, *rb_parent;
3034         pgoff_t pgoff = addr >> PAGE_SHIFT;
3035         int error;
3036         unsigned long mapped_addr;
3037
3038         /* Until we need other flags, refuse anything except VM_EXEC. */
3039         if ((flags & (~VM_EXEC)) != 0)
3040                 return -EINVAL;
3041         flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
3042
3043         mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
3044         if (IS_ERR_VALUE(mapped_addr))
3045                 return mapped_addr;
3046
3047         error = mlock_future_check(mm, mm->def_flags, len);
3048         if (error)
3049                 return error;
3050
3051         /* Clear old maps, set up prev, rb_link, rb_parent, and uf */
3052         if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf))
3053                 return -ENOMEM;
3054
3055         /* Check against address space limits *after* clearing old maps... */
3056         if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
3057                 return -ENOMEM;
3058
3059         if (mm->map_count > sysctl_max_map_count)
3060                 return -ENOMEM;
3061
3062         if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
3063                 return -ENOMEM;
3064
3065         /* Can we just expand an old private anonymous mapping? */
3066         vma = vma_merge(mm, prev, addr, addr + len, flags,
3067                         NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
3068         if (vma)
3069                 goto out;
3070
3071         /*
3072          * create a vma struct for an anonymous mapping
3073          */
3074         vma = vm_area_alloc(mm);
3075         if (!vma) {
3076                 vm_unacct_memory(len >> PAGE_SHIFT);
3077                 return -ENOMEM;
3078         }
3079
3080         vma_set_anonymous(vma);
3081         vma->vm_start = addr;
3082         vma->vm_end = addr + len;
3083         vma->vm_pgoff = pgoff;
3084         vma->vm_flags = flags;
3085         vma->vm_page_prot = vm_get_page_prot(flags);
3086         vma_link(mm, vma, prev, rb_link, rb_parent);
3087 out:
3088         perf_event_mmap(vma);
3089         mm->total_vm += len >> PAGE_SHIFT;
3090         mm->data_vm += len >> PAGE_SHIFT;
3091         if (flags & VM_LOCKED)
3092                 mm->locked_vm += (len >> PAGE_SHIFT);
3093         vma->vm_flags |= VM_SOFTDIRTY;
3094         return 0;
3095 }
3096
3097 int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3098 {
3099         struct mm_struct *mm = current->mm;
3100         unsigned long len;
3101         int ret;
3102         bool populate;
3103         LIST_HEAD(uf);
3104
3105         len = PAGE_ALIGN(request);
3106         if (len < request)
3107                 return -ENOMEM;
3108         if (!len)
3109                 return 0;
3110
3111         if (mmap_write_lock_killable(mm))
3112                 return -EINTR;
3113
3114         ret = do_brk_flags(addr, len, flags, &uf);
3115         populate = ((mm->def_flags & VM_LOCKED) != 0);
3116         mmap_write_unlock(mm);
3117         userfaultfd_unmap_complete(mm, &uf);
3118         if (populate && !ret)
3119                 mm_populate(addr, len);
3120         return ret;
3121 }
3122 EXPORT_SYMBOL(vm_brk_flags);
3123
3124 int vm_brk(unsigned long addr, unsigned long len)
3125 {
3126         return vm_brk_flags(addr, len, 0);
3127 }
3128 EXPORT_SYMBOL(vm_brk);
3129
3130 /* Release all mmaps. */
3131 void exit_mmap(struct mm_struct *mm)
3132 {
3133         struct mmu_gather tlb;
3134         struct vm_area_struct *vma;
3135         unsigned long nr_accounted = 0;
3136
3137         /* mm's last user has gone, and its about to be pulled down */
3138         mmu_notifier_release(mm);
3139
3140         if (unlikely(mm_is_oom_victim(mm))) {
3141                 /*
3142                  * Manually reap the mm to free as much memory as possible.
3143                  * Then, as the oom reaper does, set MMF_OOM_SKIP to disregard
3144                  * this mm from further consideration.  Taking mm->mmap_lock for
3145                  * write after setting MMF_OOM_SKIP will guarantee that the oom
3146                  * reaper will not run on this mm again after mmap_lock is
3147                  * dropped.
3148                  *
3149                  * Nothing can be holding mm->mmap_lock here and the above call
3150                  * to mmu_notifier_release(mm) ensures mmu notifier callbacks in
3151                  * __oom_reap_task_mm() will not block.
3152                  *
3153                  * This needs to be done before calling munlock_vma_pages_all(),
3154                  * which clears VM_LOCKED, otherwise the oom reaper cannot
3155                  * reliably test it.
3156                  */
3157                 (void)__oom_reap_task_mm(mm);
3158
3159                 set_bit(MMF_OOM_SKIP, &mm->flags);
3160                 mmap_write_lock(mm);
3161                 mmap_write_unlock(mm);
3162         }
3163
3164         if (mm->locked_vm)
3165                 unlock_range(mm->mmap, ULONG_MAX);
3166
3167         arch_exit_mmap(mm);
3168
3169         vma = mm->mmap;
3170         if (!vma)       /* Can happen if dup_mmap() received an OOM */
3171                 return;
3172
3173         lru_add_drain();
3174         flush_cache_mm(mm);
3175         tlb_gather_mmu_fullmm(&tlb, mm);
3176         /* update_hiwater_rss(mm) here? but nobody should be looking */
3177         /* Use -1 here to ensure all VMAs in the mm are unmapped */
3178         unmap_vmas(&tlb, vma, 0, -1);
3179         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3180         tlb_finish_mmu(&tlb);
3181
3182         /*
3183          * Walk the list again, actually closing and freeing it,
3184          * with preemption enabled, without holding any MM locks.
3185          */
3186         while (vma) {
3187                 if (vma->vm_flags & VM_ACCOUNT)
3188                         nr_accounted += vma_pages(vma);
3189                 vma = remove_vma(vma);
3190                 cond_resched();
3191         }
3192         vm_unacct_memory(nr_accounted);
3193 }
3194
3195 /* Insert vm structure into process list sorted by address
3196  * and into the inode's i_mmap tree.  If vm_file is non-NULL
3197  * then i_mmap_rwsem is taken here.
3198  */
3199 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3200 {
3201         struct vm_area_struct *prev;
3202         struct rb_node **rb_link, *rb_parent;
3203
3204         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3205                            &prev, &rb_link, &rb_parent))
3206                 return -ENOMEM;
3207         if ((vma->vm_flags & VM_ACCOUNT) &&
3208              security_vm_enough_memory_mm(mm, vma_pages(vma)))
3209                 return -ENOMEM;
3210
3211         /*
3212          * The vm_pgoff of a purely anonymous vma should be irrelevant
3213          * until its first write fault, when page's anon_vma and index
3214          * are set.  But now set the vm_pgoff it will almost certainly
3215          * end up with (unless mremap moves it elsewhere before that
3216          * first wfault), so /proc/pid/maps tells a consistent story.
3217          *
3218          * By setting it to reflect the virtual start address of the
3219          * vma, merges and splits can happen in a seamless way, just
3220          * using the existing file pgoff checks and manipulations.
3221          * Similarly in do_mmap and in do_brk_flags.
3222          */
3223         if (vma_is_anonymous(vma)) {
3224                 BUG_ON(vma->anon_vma);
3225                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3226         }
3227
3228         vma_link(mm, vma, prev, rb_link, rb_parent);
3229         return 0;
3230 }
3231
3232 /*
3233  * Copy the vma structure to a new location in the same mm,
3234  * prior to moving page table entries, to effect an mremap move.
3235  */
3236 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3237         unsigned long addr, unsigned long len, pgoff_t pgoff,
3238         bool *need_rmap_locks)
3239 {
3240         struct vm_area_struct *vma = *vmap;
3241         unsigned long vma_start = vma->vm_start;
3242         struct mm_struct *mm = vma->vm_mm;
3243         struct vm_area_struct *new_vma, *prev;
3244         struct rb_node **rb_link, *rb_parent;
3245         bool faulted_in_anon_vma = true;
3246
3247         /*
3248          * If anonymous vma has not yet been faulted, update new pgoff
3249          * to match new location, to increase its chance of merging.
3250          */
3251         if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3252                 pgoff = addr >> PAGE_SHIFT;
3253                 faulted_in_anon_vma = false;
3254         }
3255
3256         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3257                 return NULL;    /* should never get here */
3258         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3259                             vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3260                             vma->vm_userfaultfd_ctx);
3261         if (new_vma) {
3262                 /*
3263                  * Source vma may have been merged into new_vma
3264                  */
3265                 if (unlikely(vma_start >= new_vma->vm_start &&
3266                              vma_start < new_vma->vm_end)) {
3267                         /*
3268                          * The only way we can get a vma_merge with
3269                          * self during an mremap is if the vma hasn't
3270                          * been faulted in yet and we were allowed to
3271                          * reset the dst vma->vm_pgoff to the
3272                          * destination address of the mremap to allow
3273                          * the merge to happen. mremap must change the
3274                          * vm_pgoff linearity between src and dst vmas
3275                          * (in turn preventing a vma_merge) to be
3276                          * safe. It is only safe to keep the vm_pgoff
3277                          * linear if there are no pages mapped yet.
3278                          */
3279                         VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3280                         *vmap = vma = new_vma;
3281                 }
3282                 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3283         } else {
3284                 new_vma = vm_area_dup(vma);
3285                 if (!new_vma)
3286                         goto out;
3287                 new_vma->vm_start = addr;
3288                 new_vma->vm_end = addr + len;
3289                 new_vma->vm_pgoff = pgoff;
3290                 if (vma_dup_policy(vma, new_vma))
3291                         goto out_free_vma;
3292                 if (anon_vma_clone(new_vma, vma))
3293                         goto out_free_mempol;
3294                 if (new_vma->vm_file)
3295                         get_file(new_vma->vm_file);
3296                 if (new_vma->vm_ops && new_vma->vm_ops->open)
3297                         new_vma->vm_ops->open(new_vma);
3298                 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3299                 *need_rmap_locks = false;
3300         }
3301         return new_vma;
3302
3303 out_free_mempol:
3304         mpol_put(vma_policy(new_vma));
3305 out_free_vma:
3306         vm_area_free(new_vma);
3307 out:
3308         return NULL;
3309 }
3310
3311 /*
3312  * Return true if the calling process may expand its vm space by the passed
3313  * number of pages
3314  */
3315 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3316 {
3317         if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3318                 return false;
3319
3320         if (is_data_mapping(flags) &&
3321             mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3322                 /* Workaround for Valgrind */
3323                 if (rlimit(RLIMIT_DATA) == 0 &&
3324                     mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3325                         return true;
3326
3327                 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3328                              current->comm, current->pid,
3329                              (mm->data_vm + npages) << PAGE_SHIFT,
3330                              rlimit(RLIMIT_DATA),
3331                              ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3332
3333                 if (!ignore_rlimit_data)
3334                         return false;
3335         }
3336
3337         return true;
3338 }
3339
3340 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3341 {
3342         mm->total_vm += npages;
3343
3344         if (is_exec_mapping(flags))
3345                 mm->exec_vm += npages;
3346         else if (is_stack_mapping(flags))
3347                 mm->stack_vm += npages;
3348         else if (is_data_mapping(flags))
3349                 mm->data_vm += npages;
3350 }
3351
3352 static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3353
3354 /*
3355  * Having a close hook prevents vma merging regardless of flags.
3356  */
3357 static void special_mapping_close(struct vm_area_struct *vma)
3358 {
3359 }
3360
3361 static const char *special_mapping_name(struct vm_area_struct *vma)
3362 {
3363         return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3364 }
3365
3366 static int special_mapping_mremap(struct vm_area_struct *new_vma)
3367 {
3368         struct vm_special_mapping *sm = new_vma->vm_private_data;
3369
3370         if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3371                 return -EFAULT;
3372
3373         if (sm->mremap)
3374                 return sm->mremap(sm, new_vma);
3375
3376         return 0;
3377 }
3378
3379 static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr)
3380 {
3381         /*
3382          * Forbid splitting special mappings - kernel has expectations over
3383          * the number of pages in mapping. Together with VM_DONTEXPAND
3384          * the size of vma should stay the same over the special mapping's
3385          * lifetime.
3386          */
3387         return -EINVAL;
3388 }
3389
3390 static const struct vm_operations_struct special_mapping_vmops = {
3391         .close = special_mapping_close,
3392         .fault = special_mapping_fault,
3393         .mremap = special_mapping_mremap,
3394         .name = special_mapping_name,
3395         /* vDSO code relies that VVAR can't be accessed remotely */
3396         .access = NULL,
3397         .may_split = special_mapping_split,
3398 };
3399
3400 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3401         .close = special_mapping_close,
3402         .fault = special_mapping_fault,
3403 };
3404
3405 static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3406 {
3407         struct vm_area_struct *vma = vmf->vma;
3408         pgoff_t pgoff;
3409         struct page **pages;
3410
3411         if (vma->vm_ops == &legacy_special_mapping_vmops) {
3412                 pages = vma->vm_private_data;
3413         } else {
3414                 struct vm_special_mapping *sm = vma->vm_private_data;
3415
3416                 if (sm->fault)
3417                         return sm->fault(sm, vmf->vma, vmf);
3418
3419                 pages = sm->pages;
3420         }
3421
3422         for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3423                 pgoff--;
3424
3425         if (*pages) {
3426                 struct page *page = *pages;
3427                 get_page(page);
3428                 vmf->page = page;
3429                 return 0;
3430         }
3431
3432         return VM_FAULT_SIGBUS;
3433 }
3434
3435 static struct vm_area_struct *__install_special_mapping(
3436         struct mm_struct *mm,
3437         unsigned long addr, unsigned long len,
3438         unsigned long vm_flags, void *priv,
3439         const struct vm_operations_struct *ops)
3440 {
3441         int ret;
3442         struct vm_area_struct *vma;
3443
3444         vma = vm_area_alloc(mm);
3445         if (unlikely(vma == NULL))
3446                 return ERR_PTR(-ENOMEM);
3447
3448         vma->vm_start = addr;
3449         vma->vm_end = addr + len;
3450
3451         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3452         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3453
3454         vma->vm_ops = ops;
3455         vma->vm_private_data = priv;
3456
3457         ret = insert_vm_struct(mm, vma);
3458         if (ret)
3459                 goto out;
3460
3461         vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3462
3463         perf_event_mmap(vma);
3464
3465         return vma;
3466
3467 out:
3468         vm_area_free(vma);
3469         return ERR_PTR(ret);
3470 }
3471
3472 bool vma_is_special_mapping(const struct vm_area_struct *vma,
3473         const struct vm_special_mapping *sm)
3474 {
3475         return vma->vm_private_data == sm &&
3476                 (vma->vm_ops == &special_mapping_vmops ||
3477                  vma->vm_ops == &legacy_special_mapping_vmops);
3478 }
3479
3480 /*
3481  * Called with mm->mmap_lock held for writing.
3482  * Insert a new vma covering the given region, with the given flags.
3483  * Its pages are supplied by the given array of struct page *.
3484  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3485  * The region past the last page supplied will always produce SIGBUS.
3486  * The array pointer and the pages it points to are assumed to stay alive
3487  * for as long as this mapping might exist.
3488  */
3489 struct vm_area_struct *_install_special_mapping(
3490         struct mm_struct *mm,
3491         unsigned long addr, unsigned long len,
3492         unsigned long vm_flags, const struct vm_special_mapping *spec)
3493 {
3494         return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3495                                         &special_mapping_vmops);
3496 }
3497
3498 int install_special_mapping(struct mm_struct *mm,
3499                             unsigned long addr, unsigned long len,
3500                             unsigned long vm_flags, struct page **pages)
3501 {
3502         struct vm_area_struct *vma = __install_special_mapping(
3503                 mm, addr, len, vm_flags, (void *)pages,
3504                 &legacy_special_mapping_vmops);
3505
3506         return PTR_ERR_OR_ZERO(vma);
3507 }
3508
3509 static DEFINE_MUTEX(mm_all_locks_mutex);
3510
3511 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3512 {
3513         if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3514                 /*
3515                  * The LSB of head.next can't change from under us
3516                  * because we hold the mm_all_locks_mutex.
3517                  */
3518                 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
3519                 /*
3520                  * We can safely modify head.next after taking the
3521                  * anon_vma->root->rwsem. If some other vma in this mm shares
3522                  * the same anon_vma we won't take it again.
3523                  *
3524                  * No need of atomic instructions here, head.next
3525                  * can't change from under us thanks to the
3526                  * anon_vma->root->rwsem.
3527                  */
3528                 if (__test_and_set_bit(0, (unsigned long *)
3529                                        &anon_vma->root->rb_root.rb_root.rb_node))
3530                         BUG();
3531         }
3532 }
3533
3534 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3535 {
3536         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3537                 /*
3538                  * AS_MM_ALL_LOCKS can't change from under us because
3539                  * we hold the mm_all_locks_mutex.
3540                  *
3541                  * Operations on ->flags have to be atomic because
3542                  * even if AS_MM_ALL_LOCKS is stable thanks to the
3543                  * mm_all_locks_mutex, there may be other cpus
3544                  * changing other bitflags in parallel to us.
3545                  */
3546                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3547                         BUG();
3548                 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
3549         }
3550 }
3551
3552 /*
3553  * This operation locks against the VM for all pte/vma/mm related
3554  * operations that could ever happen on a certain mm. This includes
3555  * vmtruncate, try_to_unmap, and all page faults.
3556  *
3557  * The caller must take the mmap_lock in write mode before calling
3558  * mm_take_all_locks(). The caller isn't allowed to release the
3559  * mmap_lock until mm_drop_all_locks() returns.
3560  *
3561  * mmap_lock in write mode is required in order to block all operations
3562  * that could modify pagetables and free pages without need of
3563  * altering the vma layout. It's also needed in write mode to avoid new
3564  * anon_vmas to be associated with existing vmas.
3565  *
3566  * A single task can't take more than one mm_take_all_locks() in a row
3567  * or it would deadlock.
3568  *
3569  * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3570  * mapping->flags avoid to take the same lock twice, if more than one
3571  * vma in this mm is backed by the same anon_vma or address_space.
3572  *
3573  * We take locks in following order, accordingly to comment at beginning
3574  * of mm/rmap.c:
3575  *   - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3576  *     hugetlb mapping);
3577  *   - all i_mmap_rwsem locks;
3578  *   - all anon_vma->rwseml
3579  *
3580  * We can take all locks within these types randomly because the VM code
3581  * doesn't nest them and we protected from parallel mm_take_all_locks() by
3582  * mm_all_locks_mutex.
3583  *
3584  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3585  * that may have to take thousand of locks.
3586  *
3587  * mm_take_all_locks() can fail if it's interrupted by signals.
3588  */
3589 int mm_take_all_locks(struct mm_struct *mm)
3590 {
3591         struct vm_area_struct *vma;
3592         struct anon_vma_chain *avc;
3593
3594         BUG_ON(mmap_read_trylock(mm));
3595
3596         mutex_lock(&mm_all_locks_mutex);
3597
3598         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3599                 if (signal_pending(current))
3600                         goto out_unlock;
3601                 if (vma->vm_file && vma->vm_file->f_mapping &&
3602                                 is_vm_hugetlb_page(vma))
3603                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
3604         }
3605
3606         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3607                 if (signal_pending(current))
3608                         goto out_unlock;
3609                 if (vma->vm_file && vma->vm_file->f_mapping &&
3610                                 !is_vm_hugetlb_page(vma))
3611                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
3612         }
3613
3614         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3615                 if (signal_pending(current))
3616                         goto out_unlock;
3617                 if (vma->anon_vma)
3618                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3619                                 vm_lock_anon_vma(mm, avc->anon_vma);
3620         }
3621
3622         return 0;
3623
3624 out_unlock:
3625         mm_drop_all_locks(mm);
3626         return -EINTR;
3627 }
3628
3629 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3630 {
3631         if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3632                 /*
3633                  * The LSB of head.next can't change to 0 from under
3634                  * us because we hold the mm_all_locks_mutex.
3635                  *
3636                  * We must however clear the bitflag before unlocking
3637                  * the vma so the users using the anon_vma->rb_root will
3638                  * never see our bitflag.
3639                  *
3640                  * No need of atomic instructions here, head.next
3641                  * can't change from under us until we release the
3642                  * anon_vma->root->rwsem.
3643                  */
3644                 if (!__test_and_clear_bit(0, (unsigned long *)
3645                                           &anon_vma->root->rb_root.rb_root.rb_node))
3646                         BUG();
3647                 anon_vma_unlock_write(anon_vma);
3648         }
3649 }
3650
3651 static void vm_unlock_mapping(struct address_space *mapping)
3652 {
3653         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3654                 /*
3655                  * AS_MM_ALL_LOCKS can't change to 0 from under us
3656                  * because we hold the mm_all_locks_mutex.
3657                  */
3658                 i_mmap_unlock_write(mapping);
3659                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3660                                         &mapping->flags))
3661                         BUG();
3662         }
3663 }
3664
3665 /*
3666  * The mmap_lock cannot be released by the caller until
3667  * mm_drop_all_locks() returns.
3668  */
3669 void mm_drop_all_locks(struct mm_struct *mm)
3670 {
3671         struct vm_area_struct *vma;
3672         struct anon_vma_chain *avc;
3673
3674         BUG_ON(mmap_read_trylock(mm));
3675         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3676
3677         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3678                 if (vma->anon_vma)
3679                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3680                                 vm_unlock_anon_vma(avc->anon_vma);
3681                 if (vma->vm_file && vma->vm_file->f_mapping)
3682                         vm_unlock_mapping(vma->vm_file->f_mapping);
3683         }
3684
3685         mutex_unlock(&mm_all_locks_mutex);
3686 }
3687
3688 /*
3689  * initialise the percpu counter for VM
3690  */
3691 void __init mmap_init(void)
3692 {
3693         int ret;
3694
3695         ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3696         VM_BUG_ON(ret);
3697 }
3698
3699 /*
3700  * Initialise sysctl_user_reserve_kbytes.
3701  *
3702  * This is intended to prevent a user from starting a single memory hogging
3703  * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3704  * mode.
3705  *
3706  * The default value is min(3% of free memory, 128MB)
3707  * 128MB is enough to recover with sshd/login, bash, and top/kill.
3708  */
3709 static int init_user_reserve(void)
3710 {
3711         unsigned long free_kbytes;
3712
3713         free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3714
3715         sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3716         return 0;
3717 }
3718 subsys_initcall(init_user_reserve);
3719
3720 /*
3721  * Initialise sysctl_admin_reserve_kbytes.
3722  *
3723  * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3724  * to log in and kill a memory hogging process.
3725  *
3726  * Systems with more than 256MB will reserve 8MB, enough to recover
3727  * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3728  * only reserve 3% of free pages by default.
3729  */
3730 static int init_admin_reserve(void)
3731 {
3732         unsigned long free_kbytes;
3733
3734         free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3735
3736         sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3737         return 0;
3738 }
3739 subsys_initcall(init_admin_reserve);
3740
3741 /*
3742  * Reinititalise user and admin reserves if memory is added or removed.
3743  *
3744  * The default user reserve max is 128MB, and the default max for the
3745  * admin reserve is 8MB. These are usually, but not always, enough to
3746  * enable recovery from a memory hogging process using login/sshd, a shell,
3747  * and tools like top. It may make sense to increase or even disable the
3748  * reserve depending on the existence of swap or variations in the recovery
3749  * tools. So, the admin may have changed them.
3750  *
3751  * If memory is added and the reserves have been eliminated or increased above
3752  * the default max, then we'll trust the admin.
3753  *
3754  * If memory is removed and there isn't enough free memory, then we
3755  * need to reset the reserves.
3756  *
3757  * Otherwise keep the reserve set by the admin.
3758  */
3759 static int reserve_mem_notifier(struct notifier_block *nb,
3760                              unsigned long action, void *data)
3761 {
3762         unsigned long tmp, free_kbytes;
3763
3764         switch (action) {
3765         case MEM_ONLINE:
3766                 /* Default max is 128MB. Leave alone if modified by operator. */
3767                 tmp = sysctl_user_reserve_kbytes;
3768                 if (0 < tmp && tmp < (1UL << 17))
3769                         init_user_reserve();
3770
3771                 /* Default max is 8MB.  Leave alone if modified by operator. */
3772                 tmp = sysctl_admin_reserve_kbytes;
3773                 if (0 < tmp && tmp < (1UL << 13))
3774                         init_admin_reserve();
3775
3776                 break;
3777         case MEM_OFFLINE:
3778                 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3779
3780                 if (sysctl_user_reserve_kbytes > free_kbytes) {
3781                         init_user_reserve();
3782                         pr_info("vm.user_reserve_kbytes reset to %lu\n",
3783                                 sysctl_user_reserve_kbytes);
3784                 }
3785
3786                 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3787                         init_admin_reserve();
3788                         pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3789                                 sysctl_admin_reserve_kbytes);
3790                 }
3791                 break;
3792         default:
3793                 break;
3794         }
3795         return NOTIFY_OK;
3796 }
3797
3798 static struct notifier_block reserve_mem_nb = {
3799         .notifier_call = reserve_mem_notifier,
3800 };
3801
3802 static int __meminit init_reserve_notifier(void)
3803 {
3804         if (register_hotmemory_notifier(&reserve_mem_nb))
3805                 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3806
3807         return 0;
3808 }
3809 subsys_initcall(init_reserve_notifier);