]> git.karo-electronics.de Git - karo-tx-linux.git/blob - mm/mmap.c
Merge remote-tracking branch 'asoc/topic/intel' into asoc-next
[karo-tx-linux.git] / mm / mmap.c
1 /*
2  * mm/mmap.c
3  *
4  * Written by obz.
5  *
6  * Address space accounting code        <alan@lxorguk.ukuu.org.uk>
7  */
8
9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
10
11 #include <linux/kernel.h>
12 #include <linux/slab.h>
13 #include <linux/backing-dev.h>
14 #include <linux/mm.h>
15 #include <linux/vmacache.h>
16 #include <linux/shm.h>
17 #include <linux/mman.h>
18 #include <linux/pagemap.h>
19 #include <linux/swap.h>
20 #include <linux/syscalls.h>
21 #include <linux/capability.h>
22 #include <linux/init.h>
23 #include <linux/file.h>
24 #include <linux/fs.h>
25 #include <linux/personality.h>
26 #include <linux/security.h>
27 #include <linux/hugetlb.h>
28 #include <linux/profile.h>
29 #include <linux/export.h>
30 #include <linux/mount.h>
31 #include <linux/mempolicy.h>
32 #include <linux/rmap.h>
33 #include <linux/mmu_notifier.h>
34 #include <linux/mmdebug.h>
35 #include <linux/perf_event.h>
36 #include <linux/audit.h>
37 #include <linux/khugepaged.h>
38 #include <linux/uprobes.h>
39 #include <linux/rbtree_augmented.h>
40 #include <linux/sched/sysctl.h>
41 #include <linux/notifier.h>
42 #include <linux/memory.h>
43 #include <linux/printk.h>
44 #include <linux/userfaultfd_k.h>
45 #include <linux/moduleparam.h>
46
47 #include <asm/uaccess.h>
48 #include <asm/cacheflush.h>
49 #include <asm/tlb.h>
50 #include <asm/mmu_context.h>
51
52 #include "internal.h"
53
54 #ifndef arch_mmap_check
55 #define arch_mmap_check(addr, len, flags)       (0)
56 #endif
57
58 #ifndef arch_rebalance_pgtables
59 #define arch_rebalance_pgtables(addr, len)              (addr)
60 #endif
61
62 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
63 const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
64 const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
65 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
66 #endif
67 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
68 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
69 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
70 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
71 #endif
72
73 static bool ignore_rlimit_data = true;
74 core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
75
76 static void unmap_region(struct mm_struct *mm,
77                 struct vm_area_struct *vma, struct vm_area_struct *prev,
78                 unsigned long start, unsigned long end);
79
80 /* description of effects of mapping type and prot in current implementation.
81  * this is due to the limited x86 page protection hardware.  The expected
82  * behavior is in parens:
83  *
84  * map_type     prot
85  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
86  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
87  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
88  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
89  *
90  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
91  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
92  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
93  *
94  */
95 pgprot_t protection_map[16] = {
96         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
97         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
98 };
99
100 pgprot_t vm_get_page_prot(unsigned long vm_flags)
101 {
102         return __pgprot(pgprot_val(protection_map[vm_flags &
103                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
104                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
105 }
106 EXPORT_SYMBOL(vm_get_page_prot);
107
108 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
109 {
110         return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
111 }
112
113 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
114 void vma_set_page_prot(struct vm_area_struct *vma)
115 {
116         unsigned long vm_flags = vma->vm_flags;
117
118         vma->vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
119         if (vma_wants_writenotify(vma)) {
120                 vm_flags &= ~VM_SHARED;
121                 vma->vm_page_prot = vm_pgprot_modify(vma->vm_page_prot,
122                                                      vm_flags);
123         }
124 }
125
126
127 int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS;  /* heuristic overcommit */
128 int sysctl_overcommit_ratio __read_mostly = 50; /* default is 50% */
129 unsigned long sysctl_overcommit_kbytes __read_mostly;
130 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
131 unsigned long sysctl_user_reserve_kbytes __read_mostly = 1UL << 17; /* 128MB */
132 unsigned long sysctl_admin_reserve_kbytes __read_mostly = 1UL << 13; /* 8MB */
133 /*
134  * Make sure vm_committed_as in one cacheline and not cacheline shared with
135  * other variables. It can be updated by several CPUs frequently.
136  */
137 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
138
139 /*
140  * The global memory commitment made in the system can be a metric
141  * that can be used to drive ballooning decisions when Linux is hosted
142  * as a guest. On Hyper-V, the host implements a policy engine for dynamically
143  * balancing memory across competing virtual machines that are hosted.
144  * Several metrics drive this policy engine including the guest reported
145  * memory commitment.
146  */
147 unsigned long vm_memory_committed(void)
148 {
149         return percpu_counter_read_positive(&vm_committed_as);
150 }
151 EXPORT_SYMBOL_GPL(vm_memory_committed);
152
153 /*
154  * Check that a process has enough memory to allocate a new virtual
155  * mapping. 0 means there is enough memory for the allocation to
156  * succeed and -ENOMEM implies there is not.
157  *
158  * We currently support three overcommit policies, which are set via the
159  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
160  *
161  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
162  * Additional code 2002 Jul 20 by Robert Love.
163  *
164  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
165  *
166  * Note this is a helper function intended to be used by LSMs which
167  * wish to use this logic.
168  */
169 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
170 {
171         long free, allowed, reserve;
172
173         VM_WARN_ONCE(percpu_counter_read(&vm_committed_as) <
174                         -(s64)vm_committed_as_batch * num_online_cpus(),
175                         "memory commitment underflow");
176
177         vm_acct_memory(pages);
178
179         /*
180          * Sometimes we want to use more memory than we have
181          */
182         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
183                 return 0;
184
185         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
186                 free = global_page_state(NR_FREE_PAGES);
187                 free += global_page_state(NR_FILE_PAGES);
188
189                 /*
190                  * shmem pages shouldn't be counted as free in this
191                  * case, they can't be purged, only swapped out, and
192                  * that won't affect the overall amount of available
193                  * memory in the system.
194                  */
195                 free -= global_page_state(NR_SHMEM);
196
197                 free += get_nr_swap_pages();
198
199                 /*
200                  * Any slabs which are created with the
201                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
202                  * which are reclaimable, under pressure.  The dentry
203                  * cache and most inode caches should fall into this
204                  */
205                 free += global_page_state(NR_SLAB_RECLAIMABLE);
206
207                 /*
208                  * Leave reserved pages. The pages are not for anonymous pages.
209                  */
210                 if (free <= totalreserve_pages)
211                         goto error;
212                 else
213                         free -= totalreserve_pages;
214
215                 /*
216                  * Reserve some for root
217                  */
218                 if (!cap_sys_admin)
219                         free -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10);
220
221                 if (free > pages)
222                         return 0;
223
224                 goto error;
225         }
226
227         allowed = vm_commit_limit();
228         /*
229          * Reserve some for root
230          */
231         if (!cap_sys_admin)
232                 allowed -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10);
233
234         /*
235          * Don't let a single process grow so big a user can't recover
236          */
237         if (mm) {
238                 reserve = sysctl_user_reserve_kbytes >> (PAGE_SHIFT - 10);
239                 allowed -= min_t(long, mm->total_vm / 32, reserve);
240         }
241
242         if (percpu_counter_read_positive(&vm_committed_as) < allowed)
243                 return 0;
244 error:
245         vm_unacct_memory(pages);
246
247         return -ENOMEM;
248 }
249
250 /*
251  * Requires inode->i_mapping->i_mmap_rwsem
252  */
253 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
254                 struct file *file, struct address_space *mapping)
255 {
256         if (vma->vm_flags & VM_DENYWRITE)
257                 atomic_inc(&file_inode(file)->i_writecount);
258         if (vma->vm_flags & VM_SHARED)
259                 mapping_unmap_writable(mapping);
260
261         flush_dcache_mmap_lock(mapping);
262         vma_interval_tree_remove(vma, &mapping->i_mmap);
263         flush_dcache_mmap_unlock(mapping);
264 }
265
266 /*
267  * Unlink a file-based vm structure from its interval tree, to hide
268  * vma from rmap and vmtruncate before freeing its page tables.
269  */
270 void unlink_file_vma(struct vm_area_struct *vma)
271 {
272         struct file *file = vma->vm_file;
273
274         if (file) {
275                 struct address_space *mapping = file->f_mapping;
276                 i_mmap_lock_write(mapping);
277                 __remove_shared_vm_struct(vma, file, mapping);
278                 i_mmap_unlock_write(mapping);
279         }
280 }
281
282 /*
283  * Close a vm structure and free it, returning the next.
284  */
285 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
286 {
287         struct vm_area_struct *next = vma->vm_next;
288
289         might_sleep();
290         if (vma->vm_ops && vma->vm_ops->close)
291                 vma->vm_ops->close(vma);
292         if (vma->vm_file)
293                 fput(vma->vm_file);
294         mpol_put(vma_policy(vma));
295         kmem_cache_free(vm_area_cachep, vma);
296         return next;
297 }
298
299 static unsigned long do_brk(unsigned long addr, unsigned long len);
300
301 SYSCALL_DEFINE1(brk, unsigned long, brk)
302 {
303         unsigned long retval;
304         unsigned long newbrk, oldbrk;
305         struct mm_struct *mm = current->mm;
306         unsigned long min_brk;
307         bool populate;
308
309         down_write(&mm->mmap_sem);
310
311 #ifdef CONFIG_COMPAT_BRK
312         /*
313          * CONFIG_COMPAT_BRK can still be overridden by setting
314          * randomize_va_space to 2, which will still cause mm->start_brk
315          * to be arbitrarily shifted
316          */
317         if (current->brk_randomized)
318                 min_brk = mm->start_brk;
319         else
320                 min_brk = mm->end_data;
321 #else
322         min_brk = mm->start_brk;
323 #endif
324         if (brk < min_brk)
325                 goto out;
326
327         /*
328          * Check against rlimit here. If this check is done later after the test
329          * of oldbrk with newbrk then it can escape the test and let the data
330          * segment grow beyond its set limit the in case where the limit is
331          * not page aligned -Ram Gupta
332          */
333         if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
334                               mm->end_data, mm->start_data))
335                 goto out;
336
337         newbrk = PAGE_ALIGN(brk);
338         oldbrk = PAGE_ALIGN(mm->brk);
339         if (oldbrk == newbrk)
340                 goto set_brk;
341
342         /* Always allow shrinking brk. */
343         if (brk <= mm->brk) {
344                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
345                         goto set_brk;
346                 goto out;
347         }
348
349         /* Check against existing mmap mappings. */
350         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
351                 goto out;
352
353         /* Ok, looks good - let it rip. */
354         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
355                 goto out;
356
357 set_brk:
358         mm->brk = brk;
359         populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
360         up_write(&mm->mmap_sem);
361         if (populate)
362                 mm_populate(oldbrk, newbrk - oldbrk);
363         return brk;
364
365 out:
366         retval = mm->brk;
367         up_write(&mm->mmap_sem);
368         return retval;
369 }
370
371 static long vma_compute_subtree_gap(struct vm_area_struct *vma)
372 {
373         unsigned long max, subtree_gap;
374         max = vma->vm_start;
375         if (vma->vm_prev)
376                 max -= vma->vm_prev->vm_end;
377         if (vma->vm_rb.rb_left) {
378                 subtree_gap = rb_entry(vma->vm_rb.rb_left,
379                                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
380                 if (subtree_gap > max)
381                         max = subtree_gap;
382         }
383         if (vma->vm_rb.rb_right) {
384                 subtree_gap = rb_entry(vma->vm_rb.rb_right,
385                                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
386                 if (subtree_gap > max)
387                         max = subtree_gap;
388         }
389         return max;
390 }
391
392 #ifdef CONFIG_DEBUG_VM_RB
393 static int browse_rb(struct mm_struct *mm)
394 {
395         struct rb_root *root = &mm->mm_rb;
396         int i = 0, j, bug = 0;
397         struct rb_node *nd, *pn = NULL;
398         unsigned long prev = 0, pend = 0;
399
400         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
401                 struct vm_area_struct *vma;
402                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
403                 if (vma->vm_start < prev) {
404                         pr_emerg("vm_start %lx < prev %lx\n",
405                                   vma->vm_start, prev);
406                         bug = 1;
407                 }
408                 if (vma->vm_start < pend) {
409                         pr_emerg("vm_start %lx < pend %lx\n",
410                                   vma->vm_start, pend);
411                         bug = 1;
412                 }
413                 if (vma->vm_start > vma->vm_end) {
414                         pr_emerg("vm_start %lx > vm_end %lx\n",
415                                   vma->vm_start, vma->vm_end);
416                         bug = 1;
417                 }
418                 spin_lock(&mm->page_table_lock);
419                 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
420                         pr_emerg("free gap %lx, correct %lx\n",
421                                vma->rb_subtree_gap,
422                                vma_compute_subtree_gap(vma));
423                         bug = 1;
424                 }
425                 spin_unlock(&mm->page_table_lock);
426                 i++;
427                 pn = nd;
428                 prev = vma->vm_start;
429                 pend = vma->vm_end;
430         }
431         j = 0;
432         for (nd = pn; nd; nd = rb_prev(nd))
433                 j++;
434         if (i != j) {
435                 pr_emerg("backwards %d, forwards %d\n", j, i);
436                 bug = 1;
437         }
438         return bug ? -1 : i;
439 }
440
441 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
442 {
443         struct rb_node *nd;
444
445         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
446                 struct vm_area_struct *vma;
447                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
448                 VM_BUG_ON_VMA(vma != ignore &&
449                         vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
450                         vma);
451         }
452 }
453
454 static void validate_mm(struct mm_struct *mm)
455 {
456         int bug = 0;
457         int i = 0;
458         unsigned long highest_address = 0;
459         struct vm_area_struct *vma = mm->mmap;
460
461         while (vma) {
462                 struct anon_vma *anon_vma = vma->anon_vma;
463                 struct anon_vma_chain *avc;
464
465                 if (anon_vma) {
466                         anon_vma_lock_read(anon_vma);
467                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
468                                 anon_vma_interval_tree_verify(avc);
469                         anon_vma_unlock_read(anon_vma);
470                 }
471
472                 highest_address = vma->vm_end;
473                 vma = vma->vm_next;
474                 i++;
475         }
476         if (i != mm->map_count) {
477                 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
478                 bug = 1;
479         }
480         if (highest_address != mm->highest_vm_end) {
481                 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
482                           mm->highest_vm_end, highest_address);
483                 bug = 1;
484         }
485         i = browse_rb(mm);
486         if (i != mm->map_count) {
487                 if (i != -1)
488                         pr_emerg("map_count %d rb %d\n", mm->map_count, i);
489                 bug = 1;
490         }
491         VM_BUG_ON_MM(bug, mm);
492 }
493 #else
494 #define validate_mm_rb(root, ignore) do { } while (0)
495 #define validate_mm(mm) do { } while (0)
496 #endif
497
498 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
499                      unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
500
501 /*
502  * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
503  * vma->vm_prev->vm_end values changed, without modifying the vma's position
504  * in the rbtree.
505  */
506 static void vma_gap_update(struct vm_area_struct *vma)
507 {
508         /*
509          * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
510          * function that does exacltly what we want.
511          */
512         vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
513 }
514
515 static inline void vma_rb_insert(struct vm_area_struct *vma,
516                                  struct rb_root *root)
517 {
518         /* All rb_subtree_gap values must be consistent prior to insertion */
519         validate_mm_rb(root, NULL);
520
521         rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
522 }
523
524 static void vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
525 {
526         /*
527          * All rb_subtree_gap values must be consistent prior to erase,
528          * with the possible exception of the vma being erased.
529          */
530         validate_mm_rb(root, vma);
531
532         /*
533          * Note rb_erase_augmented is a fairly large inline function,
534          * so make sure we instantiate it only once with our desired
535          * augmented rbtree callbacks.
536          */
537         rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
538 }
539
540 /*
541  * vma has some anon_vma assigned, and is already inserted on that
542  * anon_vma's interval trees.
543  *
544  * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
545  * vma must be removed from the anon_vma's interval trees using
546  * anon_vma_interval_tree_pre_update_vma().
547  *
548  * After the update, the vma will be reinserted using
549  * anon_vma_interval_tree_post_update_vma().
550  *
551  * The entire update must be protected by exclusive mmap_sem and by
552  * the root anon_vma's mutex.
553  */
554 static inline void
555 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
556 {
557         struct anon_vma_chain *avc;
558
559         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
560                 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
561 }
562
563 static inline void
564 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
565 {
566         struct anon_vma_chain *avc;
567
568         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
569                 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
570 }
571
572 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
573                 unsigned long end, struct vm_area_struct **pprev,
574                 struct rb_node ***rb_link, struct rb_node **rb_parent)
575 {
576         struct rb_node **__rb_link, *__rb_parent, *rb_prev;
577
578         __rb_link = &mm->mm_rb.rb_node;
579         rb_prev = __rb_parent = NULL;
580
581         while (*__rb_link) {
582                 struct vm_area_struct *vma_tmp;
583
584                 __rb_parent = *__rb_link;
585                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
586
587                 if (vma_tmp->vm_end > addr) {
588                         /* Fail if an existing vma overlaps the area */
589                         if (vma_tmp->vm_start < end)
590                                 return -ENOMEM;
591                         __rb_link = &__rb_parent->rb_left;
592                 } else {
593                         rb_prev = __rb_parent;
594                         __rb_link = &__rb_parent->rb_right;
595                 }
596         }
597
598         *pprev = NULL;
599         if (rb_prev)
600                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
601         *rb_link = __rb_link;
602         *rb_parent = __rb_parent;
603         return 0;
604 }
605
606 static unsigned long count_vma_pages_range(struct mm_struct *mm,
607                 unsigned long addr, unsigned long end)
608 {
609         unsigned long nr_pages = 0;
610         struct vm_area_struct *vma;
611
612         /* Find first overlaping mapping */
613         vma = find_vma_intersection(mm, addr, end);
614         if (!vma)
615                 return 0;
616
617         nr_pages = (min(end, vma->vm_end) -
618                 max(addr, vma->vm_start)) >> PAGE_SHIFT;
619
620         /* Iterate over the rest of the overlaps */
621         for (vma = vma->vm_next; vma; vma = vma->vm_next) {
622                 unsigned long overlap_len;
623
624                 if (vma->vm_start > end)
625                         break;
626
627                 overlap_len = min(end, vma->vm_end) - vma->vm_start;
628                 nr_pages += overlap_len >> PAGE_SHIFT;
629         }
630
631         return nr_pages;
632 }
633
634 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
635                 struct rb_node **rb_link, struct rb_node *rb_parent)
636 {
637         /* Update tracking information for the gap following the new vma. */
638         if (vma->vm_next)
639                 vma_gap_update(vma->vm_next);
640         else
641                 mm->highest_vm_end = vma->vm_end;
642
643         /*
644          * vma->vm_prev wasn't known when we followed the rbtree to find the
645          * correct insertion point for that vma. As a result, we could not
646          * update the vma vm_rb parents rb_subtree_gap values on the way down.
647          * So, we first insert the vma with a zero rb_subtree_gap value
648          * (to be consistent with what we did on the way down), and then
649          * immediately update the gap to the correct value. Finally we
650          * rebalance the rbtree after all augmented values have been set.
651          */
652         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
653         vma->rb_subtree_gap = 0;
654         vma_gap_update(vma);
655         vma_rb_insert(vma, &mm->mm_rb);
656 }
657
658 static void __vma_link_file(struct vm_area_struct *vma)
659 {
660         struct file *file;
661
662         file = vma->vm_file;
663         if (file) {
664                 struct address_space *mapping = file->f_mapping;
665
666                 if (vma->vm_flags & VM_DENYWRITE)
667                         atomic_dec(&file_inode(file)->i_writecount);
668                 if (vma->vm_flags & VM_SHARED)
669                         atomic_inc(&mapping->i_mmap_writable);
670
671                 flush_dcache_mmap_lock(mapping);
672                 vma_interval_tree_insert(vma, &mapping->i_mmap);
673                 flush_dcache_mmap_unlock(mapping);
674         }
675 }
676
677 static void
678 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
679         struct vm_area_struct *prev, struct rb_node **rb_link,
680         struct rb_node *rb_parent)
681 {
682         __vma_link_list(mm, vma, prev, rb_parent);
683         __vma_link_rb(mm, vma, rb_link, rb_parent);
684 }
685
686 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
687                         struct vm_area_struct *prev, struct rb_node **rb_link,
688                         struct rb_node *rb_parent)
689 {
690         struct address_space *mapping = NULL;
691
692         if (vma->vm_file) {
693                 mapping = vma->vm_file->f_mapping;
694                 i_mmap_lock_write(mapping);
695         }
696
697         __vma_link(mm, vma, prev, rb_link, rb_parent);
698         __vma_link_file(vma);
699
700         if (mapping)
701                 i_mmap_unlock_write(mapping);
702
703         mm->map_count++;
704         validate_mm(mm);
705 }
706
707 /*
708  * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
709  * mm's list and rbtree.  It has already been inserted into the interval tree.
710  */
711 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
712 {
713         struct vm_area_struct *prev;
714         struct rb_node **rb_link, *rb_parent;
715
716         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
717                            &prev, &rb_link, &rb_parent))
718                 BUG();
719         __vma_link(mm, vma, prev, rb_link, rb_parent);
720         mm->map_count++;
721 }
722
723 static inline void
724 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
725                 struct vm_area_struct *prev)
726 {
727         struct vm_area_struct *next;
728
729         vma_rb_erase(vma, &mm->mm_rb);
730         prev->vm_next = next = vma->vm_next;
731         if (next)
732                 next->vm_prev = prev;
733
734         /* Kill the cache */
735         vmacache_invalidate(mm);
736 }
737
738 /*
739  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
740  * is already present in an i_mmap tree without adjusting the tree.
741  * The following helper function should be used when such adjustments
742  * are necessary.  The "insert" vma (if any) is to be inserted
743  * before we drop the necessary locks.
744  */
745 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
746         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
747 {
748         struct mm_struct *mm = vma->vm_mm;
749         struct vm_area_struct *next = vma->vm_next;
750         struct vm_area_struct *importer = NULL;
751         struct address_space *mapping = NULL;
752         struct rb_root *root = NULL;
753         struct anon_vma *anon_vma = NULL;
754         struct file *file = vma->vm_file;
755         bool start_changed = false, end_changed = false;
756         long adjust_next = 0;
757         int remove_next = 0;
758
759         if (next && !insert) {
760                 struct vm_area_struct *exporter = NULL;
761
762                 if (end >= next->vm_end) {
763                         /*
764                          * vma expands, overlapping all the next, and
765                          * perhaps the one after too (mprotect case 6).
766                          */
767 again:                  remove_next = 1 + (end > next->vm_end);
768                         end = next->vm_end;
769                         exporter = next;
770                         importer = vma;
771                 } else if (end > next->vm_start) {
772                         /*
773                          * vma expands, overlapping part of the next:
774                          * mprotect case 5 shifting the boundary up.
775                          */
776                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
777                         exporter = next;
778                         importer = vma;
779                 } else if (end < vma->vm_end) {
780                         /*
781                          * vma shrinks, and !insert tells it's not
782                          * split_vma inserting another: so it must be
783                          * mprotect case 4 shifting the boundary down.
784                          */
785                         adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT);
786                         exporter = vma;
787                         importer = next;
788                 }
789
790                 /*
791                  * Easily overlooked: when mprotect shifts the boundary,
792                  * make sure the expanding vma has anon_vma set if the
793                  * shrinking vma had, to cover any anon pages imported.
794                  */
795                 if (exporter && exporter->anon_vma && !importer->anon_vma) {
796                         int error;
797
798                         importer->anon_vma = exporter->anon_vma;
799                         error = anon_vma_clone(importer, exporter);
800                         if (error)
801                                 return error;
802                 }
803         }
804
805         if (file) {
806                 mapping = file->f_mapping;
807                 root = &mapping->i_mmap;
808                 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
809
810                 if (adjust_next)
811                         uprobe_munmap(next, next->vm_start, next->vm_end);
812
813                 i_mmap_lock_write(mapping);
814                 if (insert) {
815                         /*
816                          * Put into interval tree now, so instantiated pages
817                          * are visible to arm/parisc __flush_dcache_page
818                          * throughout; but we cannot insert into address
819                          * space until vma start or end is updated.
820                          */
821                         __vma_link_file(insert);
822                 }
823         }
824
825         vma_adjust_trans_huge(vma, start, end, adjust_next);
826
827         anon_vma = vma->anon_vma;
828         if (!anon_vma && adjust_next)
829                 anon_vma = next->anon_vma;
830         if (anon_vma) {
831                 VM_BUG_ON_VMA(adjust_next && next->anon_vma &&
832                           anon_vma != next->anon_vma, next);
833                 anon_vma_lock_write(anon_vma);
834                 anon_vma_interval_tree_pre_update_vma(vma);
835                 if (adjust_next)
836                         anon_vma_interval_tree_pre_update_vma(next);
837         }
838
839         if (root) {
840                 flush_dcache_mmap_lock(mapping);
841                 vma_interval_tree_remove(vma, root);
842                 if (adjust_next)
843                         vma_interval_tree_remove(next, root);
844         }
845
846         if (start != vma->vm_start) {
847                 vma->vm_start = start;
848                 start_changed = true;
849         }
850         if (end != vma->vm_end) {
851                 vma->vm_end = end;
852                 end_changed = true;
853         }
854         vma->vm_pgoff = pgoff;
855         if (adjust_next) {
856                 next->vm_start += adjust_next << PAGE_SHIFT;
857                 next->vm_pgoff += adjust_next;
858         }
859
860         if (root) {
861                 if (adjust_next)
862                         vma_interval_tree_insert(next, root);
863                 vma_interval_tree_insert(vma, root);
864                 flush_dcache_mmap_unlock(mapping);
865         }
866
867         if (remove_next) {
868                 /*
869                  * vma_merge has merged next into vma, and needs
870                  * us to remove next before dropping the locks.
871                  */
872                 __vma_unlink(mm, next, vma);
873                 if (file)
874                         __remove_shared_vm_struct(next, file, mapping);
875         } else if (insert) {
876                 /*
877                  * split_vma has split insert from vma, and needs
878                  * us to insert it before dropping the locks
879                  * (it may either follow vma or precede it).
880                  */
881                 __insert_vm_struct(mm, insert);
882         } else {
883                 if (start_changed)
884                         vma_gap_update(vma);
885                 if (end_changed) {
886                         if (!next)
887                                 mm->highest_vm_end = end;
888                         else if (!adjust_next)
889                                 vma_gap_update(next);
890                 }
891         }
892
893         if (anon_vma) {
894                 anon_vma_interval_tree_post_update_vma(vma);
895                 if (adjust_next)
896                         anon_vma_interval_tree_post_update_vma(next);
897                 anon_vma_unlock_write(anon_vma);
898         }
899         if (mapping)
900                 i_mmap_unlock_write(mapping);
901
902         if (root) {
903                 uprobe_mmap(vma);
904
905                 if (adjust_next)
906                         uprobe_mmap(next);
907         }
908
909         if (remove_next) {
910                 if (file) {
911                         uprobe_munmap(next, next->vm_start, next->vm_end);
912                         fput(file);
913                 }
914                 if (next->anon_vma)
915                         anon_vma_merge(vma, next);
916                 mm->map_count--;
917                 mpol_put(vma_policy(next));
918                 kmem_cache_free(vm_area_cachep, next);
919                 /*
920                  * In mprotect's case 6 (see comments on vma_merge),
921                  * we must remove another next too. It would clutter
922                  * up the code too much to do both in one go.
923                  */
924                 next = vma->vm_next;
925                 if (remove_next == 2)
926                         goto again;
927                 else if (next)
928                         vma_gap_update(next);
929                 else
930                         mm->highest_vm_end = end;
931         }
932         if (insert && file)
933                 uprobe_mmap(insert);
934
935         validate_mm(mm);
936
937         return 0;
938 }
939
940 /*
941  * If the vma has a ->close operation then the driver probably needs to release
942  * per-vma resources, so we don't attempt to merge those.
943  */
944 static inline int is_mergeable_vma(struct vm_area_struct *vma,
945                                 struct file *file, unsigned long vm_flags,
946                                 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
947 {
948         /*
949          * VM_SOFTDIRTY should not prevent from VMA merging, if we
950          * match the flags but dirty bit -- the caller should mark
951          * merged VMA as dirty. If dirty bit won't be excluded from
952          * comparison, we increase pressue on the memory system forcing
953          * the kernel to generate new VMAs when old one could be
954          * extended instead.
955          */
956         if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
957                 return 0;
958         if (vma->vm_file != file)
959                 return 0;
960         if (vma->vm_ops && vma->vm_ops->close)
961                 return 0;
962         if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
963                 return 0;
964         return 1;
965 }
966
967 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
968                                         struct anon_vma *anon_vma2,
969                                         struct vm_area_struct *vma)
970 {
971         /*
972          * The list_is_singular() test is to avoid merging VMA cloned from
973          * parents. This can improve scalability caused by anon_vma lock.
974          */
975         if ((!anon_vma1 || !anon_vma2) && (!vma ||
976                 list_is_singular(&vma->anon_vma_chain)))
977                 return 1;
978         return anon_vma1 == anon_vma2;
979 }
980
981 /*
982  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
983  * in front of (at a lower virtual address and file offset than) the vma.
984  *
985  * We cannot merge two vmas if they have differently assigned (non-NULL)
986  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
987  *
988  * We don't check here for the merged mmap wrapping around the end of pagecache
989  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
990  * wrap, nor mmaps which cover the final page at index -1UL.
991  */
992 static int
993 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
994                      struct anon_vma *anon_vma, struct file *file,
995                      pgoff_t vm_pgoff,
996                      struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
997 {
998         if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
999             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1000                 if (vma->vm_pgoff == vm_pgoff)
1001                         return 1;
1002         }
1003         return 0;
1004 }
1005
1006 /*
1007  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1008  * beyond (at a higher virtual address and file offset than) the vma.
1009  *
1010  * We cannot merge two vmas if they have differently assigned (non-NULL)
1011  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1012  */
1013 static int
1014 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1015                     struct anon_vma *anon_vma, struct file *file,
1016                     pgoff_t vm_pgoff,
1017                     struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1018 {
1019         if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1020             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1021                 pgoff_t vm_pglen;
1022                 vm_pglen = vma_pages(vma);
1023                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1024                         return 1;
1025         }
1026         return 0;
1027 }
1028
1029 /*
1030  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1031  * whether that can be merged with its predecessor or its successor.
1032  * Or both (it neatly fills a hole).
1033  *
1034  * In most cases - when called for mmap, brk or mremap - [addr,end) is
1035  * certain not to be mapped by the time vma_merge is called; but when
1036  * called for mprotect, it is certain to be already mapped (either at
1037  * an offset within prev, or at the start of next), and the flags of
1038  * this area are about to be changed to vm_flags - and the no-change
1039  * case has already been eliminated.
1040  *
1041  * The following mprotect cases have to be considered, where AAAA is
1042  * the area passed down from mprotect_fixup, never extending beyond one
1043  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1044  *
1045  *     AAAA             AAAA                AAAA          AAAA
1046  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
1047  *    cannot merge    might become    might become    might become
1048  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
1049  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
1050  *    mremap move:                                    PPPPNNNNNNNN 8
1051  *        AAAA
1052  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
1053  *    might become    case 1 below    case 2 below    case 3 below
1054  *
1055  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
1056  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
1057  */
1058 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1059                         struct vm_area_struct *prev, unsigned long addr,
1060                         unsigned long end, unsigned long vm_flags,
1061                         struct anon_vma *anon_vma, struct file *file,
1062                         pgoff_t pgoff, struct mempolicy *policy,
1063                         struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1064 {
1065         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1066         struct vm_area_struct *area, *next;
1067         int err;
1068
1069         /*
1070          * We later require that vma->vm_flags == vm_flags,
1071          * so this tests vma->vm_flags & VM_SPECIAL, too.
1072          */
1073         if (vm_flags & VM_SPECIAL)
1074                 return NULL;
1075
1076         if (prev)
1077                 next = prev->vm_next;
1078         else
1079                 next = mm->mmap;
1080         area = next;
1081         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
1082                 next = next->vm_next;
1083
1084         /*
1085          * Can it merge with the predecessor?
1086          */
1087         if (prev && prev->vm_end == addr &&
1088                         mpol_equal(vma_policy(prev), policy) &&
1089                         can_vma_merge_after(prev, vm_flags,
1090                                             anon_vma, file, pgoff,
1091                                             vm_userfaultfd_ctx)) {
1092                 /*
1093                  * OK, it can.  Can we now merge in the successor as well?
1094                  */
1095                 if (next && end == next->vm_start &&
1096                                 mpol_equal(policy, vma_policy(next)) &&
1097                                 can_vma_merge_before(next, vm_flags,
1098                                                      anon_vma, file,
1099                                                      pgoff+pglen,
1100                                                      vm_userfaultfd_ctx) &&
1101                                 is_mergeable_anon_vma(prev->anon_vma,
1102                                                       next->anon_vma, NULL)) {
1103                                                         /* cases 1, 6 */
1104                         err = vma_adjust(prev, prev->vm_start,
1105                                 next->vm_end, prev->vm_pgoff, NULL);
1106                 } else                                  /* cases 2, 5, 7 */
1107                         err = vma_adjust(prev, prev->vm_start,
1108                                 end, prev->vm_pgoff, NULL);
1109                 if (err)
1110                         return NULL;
1111                 khugepaged_enter_vma_merge(prev, vm_flags);
1112                 return prev;
1113         }
1114
1115         /*
1116          * Can this new request be merged in front of next?
1117          */
1118         if (next && end == next->vm_start &&
1119                         mpol_equal(policy, vma_policy(next)) &&
1120                         can_vma_merge_before(next, vm_flags,
1121                                              anon_vma, file, pgoff+pglen,
1122                                              vm_userfaultfd_ctx)) {
1123                 if (prev && addr < prev->vm_end)        /* case 4 */
1124                         err = vma_adjust(prev, prev->vm_start,
1125                                 addr, prev->vm_pgoff, NULL);
1126                 else                                    /* cases 3, 8 */
1127                         err = vma_adjust(area, addr, next->vm_end,
1128                                 next->vm_pgoff - pglen, NULL);
1129                 if (err)
1130                         return NULL;
1131                 khugepaged_enter_vma_merge(area, vm_flags);
1132                 return area;
1133         }
1134
1135         return NULL;
1136 }
1137
1138 /*
1139  * Rough compatbility check to quickly see if it's even worth looking
1140  * at sharing an anon_vma.
1141  *
1142  * They need to have the same vm_file, and the flags can only differ
1143  * in things that mprotect may change.
1144  *
1145  * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1146  * we can merge the two vma's. For example, we refuse to merge a vma if
1147  * there is a vm_ops->close() function, because that indicates that the
1148  * driver is doing some kind of reference counting. But that doesn't
1149  * really matter for the anon_vma sharing case.
1150  */
1151 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1152 {
1153         return a->vm_end == b->vm_start &&
1154                 mpol_equal(vma_policy(a), vma_policy(b)) &&
1155                 a->vm_file == b->vm_file &&
1156                 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) &&
1157                 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1158 }
1159
1160 /*
1161  * Do some basic sanity checking to see if we can re-use the anon_vma
1162  * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1163  * the same as 'old', the other will be the new one that is trying
1164  * to share the anon_vma.
1165  *
1166  * NOTE! This runs with mm_sem held for reading, so it is possible that
1167  * the anon_vma of 'old' is concurrently in the process of being set up
1168  * by another page fault trying to merge _that_. But that's ok: if it
1169  * is being set up, that automatically means that it will be a singleton
1170  * acceptable for merging, so we can do all of this optimistically. But
1171  * we do that READ_ONCE() to make sure that we never re-load the pointer.
1172  *
1173  * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1174  * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1175  * is to return an anon_vma that is "complex" due to having gone through
1176  * a fork).
1177  *
1178  * We also make sure that the two vma's are compatible (adjacent,
1179  * and with the same memory policies). That's all stable, even with just
1180  * a read lock on the mm_sem.
1181  */
1182 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1183 {
1184         if (anon_vma_compatible(a, b)) {
1185                 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1186
1187                 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1188                         return anon_vma;
1189         }
1190         return NULL;
1191 }
1192
1193 /*
1194  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1195  * neighbouring vmas for a suitable anon_vma, before it goes off
1196  * to allocate a new anon_vma.  It checks because a repetitive
1197  * sequence of mprotects and faults may otherwise lead to distinct
1198  * anon_vmas being allocated, preventing vma merge in subsequent
1199  * mprotect.
1200  */
1201 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1202 {
1203         struct anon_vma *anon_vma;
1204         struct vm_area_struct *near;
1205
1206         near = vma->vm_next;
1207         if (!near)
1208                 goto try_prev;
1209
1210         anon_vma = reusable_anon_vma(near, vma, near);
1211         if (anon_vma)
1212                 return anon_vma;
1213 try_prev:
1214         near = vma->vm_prev;
1215         if (!near)
1216                 goto none;
1217
1218         anon_vma = reusable_anon_vma(near, near, vma);
1219         if (anon_vma)
1220                 return anon_vma;
1221 none:
1222         /*
1223          * There's no absolute need to look only at touching neighbours:
1224          * we could search further afield for "compatible" anon_vmas.
1225          * But it would probably just be a waste of time searching,
1226          * or lead to too many vmas hanging off the same anon_vma.
1227          * We're trying to allow mprotect remerging later on,
1228          * not trying to minimize memory used for anon_vmas.
1229          */
1230         return NULL;
1231 }
1232
1233 /*
1234  * If a hint addr is less than mmap_min_addr change hint to be as
1235  * low as possible but still greater than mmap_min_addr
1236  */
1237 static inline unsigned long round_hint_to_min(unsigned long hint)
1238 {
1239         hint &= PAGE_MASK;
1240         if (((void *)hint != NULL) &&
1241             (hint < mmap_min_addr))
1242                 return PAGE_ALIGN(mmap_min_addr);
1243         return hint;
1244 }
1245
1246 static inline int mlock_future_check(struct mm_struct *mm,
1247                                      unsigned long flags,
1248                                      unsigned long len)
1249 {
1250         unsigned long locked, lock_limit;
1251
1252         /*  mlock MCL_FUTURE? */
1253         if (flags & VM_LOCKED) {
1254                 locked = len >> PAGE_SHIFT;
1255                 locked += mm->locked_vm;
1256                 lock_limit = rlimit(RLIMIT_MEMLOCK);
1257                 lock_limit >>= PAGE_SHIFT;
1258                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1259                         return -EAGAIN;
1260         }
1261         return 0;
1262 }
1263
1264 /*
1265  * The caller must hold down_write(&current->mm->mmap_sem).
1266  */
1267 unsigned long do_mmap(struct file *file, unsigned long addr,
1268                         unsigned long len, unsigned long prot,
1269                         unsigned long flags, vm_flags_t vm_flags,
1270                         unsigned long pgoff, unsigned long *populate)
1271 {
1272         struct mm_struct *mm = current->mm;
1273
1274         *populate = 0;
1275
1276         if (!len)
1277                 return -EINVAL;
1278
1279         /*
1280          * Does the application expect PROT_READ to imply PROT_EXEC?
1281          *
1282          * (the exception is when the underlying filesystem is noexec
1283          *  mounted, in which case we dont add PROT_EXEC.)
1284          */
1285         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1286                 if (!(file && path_noexec(&file->f_path)))
1287                         prot |= PROT_EXEC;
1288
1289         if (!(flags & MAP_FIXED))
1290                 addr = round_hint_to_min(addr);
1291
1292         /* Careful about overflows.. */
1293         len = PAGE_ALIGN(len);
1294         if (!len)
1295                 return -ENOMEM;
1296
1297         /* offset overflow? */
1298         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1299                 return -EOVERFLOW;
1300
1301         /* Too many mappings? */
1302         if (mm->map_count > sysctl_max_map_count)
1303                 return -ENOMEM;
1304
1305         /* Obtain the address to map to. we verify (or select) it and ensure
1306          * that it represents a valid section of the address space.
1307          */
1308         addr = get_unmapped_area(file, addr, len, pgoff, flags);
1309         if (offset_in_page(addr))
1310                 return addr;
1311
1312         /* Do simple checking here so the lower-level routines won't have
1313          * to. we assume access permissions have been handled by the open
1314          * of the memory object, so we don't do any here.
1315          */
1316         vm_flags |= calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1317                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1318
1319         if (flags & MAP_LOCKED)
1320                 if (!can_do_mlock())
1321                         return -EPERM;
1322
1323         if (mlock_future_check(mm, vm_flags, len))
1324                 return -EAGAIN;
1325
1326         if (file) {
1327                 struct inode *inode = file_inode(file);
1328
1329                 switch (flags & MAP_TYPE) {
1330                 case MAP_SHARED:
1331                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1332                                 return -EACCES;
1333
1334                         /*
1335                          * Make sure we don't allow writing to an append-only
1336                          * file..
1337                          */
1338                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1339                                 return -EACCES;
1340
1341                         /*
1342                          * Make sure there are no mandatory locks on the file.
1343                          */
1344                         if (locks_verify_locked(file))
1345                                 return -EAGAIN;
1346
1347                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1348                         if (!(file->f_mode & FMODE_WRITE))
1349                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1350
1351                         /* fall through */
1352                 case MAP_PRIVATE:
1353                         if (!(file->f_mode & FMODE_READ))
1354                                 return -EACCES;
1355                         if (path_noexec(&file->f_path)) {
1356                                 if (vm_flags & VM_EXEC)
1357                                         return -EPERM;
1358                                 vm_flags &= ~VM_MAYEXEC;
1359                         }
1360
1361                         if (!file->f_op->mmap)
1362                                 return -ENODEV;
1363                         if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1364                                 return -EINVAL;
1365                         break;
1366
1367                 default:
1368                         return -EINVAL;
1369                 }
1370         } else {
1371                 switch (flags & MAP_TYPE) {
1372                 case MAP_SHARED:
1373                         if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1374                                 return -EINVAL;
1375                         /*
1376                          * Ignore pgoff.
1377                          */
1378                         pgoff = 0;
1379                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1380                         break;
1381                 case MAP_PRIVATE:
1382                         /*
1383                          * Set pgoff according to addr for anon_vma.
1384                          */
1385                         pgoff = addr >> PAGE_SHIFT;
1386                         break;
1387                 default:
1388                         return -EINVAL;
1389                 }
1390         }
1391
1392         /*
1393          * Set 'VM_NORESERVE' if we should not account for the
1394          * memory use of this mapping.
1395          */
1396         if (flags & MAP_NORESERVE) {
1397                 /* We honor MAP_NORESERVE if allowed to overcommit */
1398                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1399                         vm_flags |= VM_NORESERVE;
1400
1401                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1402                 if (file && is_file_hugepages(file))
1403                         vm_flags |= VM_NORESERVE;
1404         }
1405
1406         addr = mmap_region(file, addr, len, vm_flags, pgoff);
1407         if (!IS_ERR_VALUE(addr) &&
1408             ((vm_flags & VM_LOCKED) ||
1409              (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1410                 *populate = len;
1411         return addr;
1412 }
1413
1414 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1415                 unsigned long, prot, unsigned long, flags,
1416                 unsigned long, fd, unsigned long, pgoff)
1417 {
1418         struct file *file = NULL;
1419         unsigned long retval;
1420
1421         if (!(flags & MAP_ANONYMOUS)) {
1422                 audit_mmap_fd(fd, flags);
1423                 file = fget(fd);
1424                 if (!file)
1425                         return -EBADF;
1426                 if (is_file_hugepages(file))
1427                         len = ALIGN(len, huge_page_size(hstate_file(file)));
1428                 retval = -EINVAL;
1429                 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1430                         goto out_fput;
1431         } else if (flags & MAP_HUGETLB) {
1432                 struct user_struct *user = NULL;
1433                 struct hstate *hs;
1434
1435                 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & SHM_HUGE_MASK);
1436                 if (!hs)
1437                         return -EINVAL;
1438
1439                 len = ALIGN(len, huge_page_size(hs));
1440                 /*
1441                  * VM_NORESERVE is used because the reservations will be
1442                  * taken when vm_ops->mmap() is called
1443                  * A dummy user value is used because we are not locking
1444                  * memory so no accounting is necessary
1445                  */
1446                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1447                                 VM_NORESERVE,
1448                                 &user, HUGETLB_ANONHUGE_INODE,
1449                                 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1450                 if (IS_ERR(file))
1451                         return PTR_ERR(file);
1452         }
1453
1454         flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1455
1456         retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1457 out_fput:
1458         if (file)
1459                 fput(file);
1460         return retval;
1461 }
1462
1463 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1464 struct mmap_arg_struct {
1465         unsigned long addr;
1466         unsigned long len;
1467         unsigned long prot;
1468         unsigned long flags;
1469         unsigned long fd;
1470         unsigned long offset;
1471 };
1472
1473 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1474 {
1475         struct mmap_arg_struct a;
1476
1477         if (copy_from_user(&a, arg, sizeof(a)))
1478                 return -EFAULT;
1479         if (offset_in_page(a.offset))
1480                 return -EINVAL;
1481
1482         return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1483                               a.offset >> PAGE_SHIFT);
1484 }
1485 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1486
1487 /*
1488  * Some shared mappigns will want the pages marked read-only
1489  * to track write events. If so, we'll downgrade vm_page_prot
1490  * to the private version (using protection_map[] without the
1491  * VM_SHARED bit).
1492  */
1493 int vma_wants_writenotify(struct vm_area_struct *vma)
1494 {
1495         vm_flags_t vm_flags = vma->vm_flags;
1496         const struct vm_operations_struct *vm_ops = vma->vm_ops;
1497
1498         /* If it was private or non-writable, the write bit is already clear */
1499         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1500                 return 0;
1501
1502         /* The backer wishes to know when pages are first written to? */
1503         if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1504                 return 1;
1505
1506         /* The open routine did something to the protections that pgprot_modify
1507          * won't preserve? */
1508         if (pgprot_val(vma->vm_page_prot) !=
1509             pgprot_val(vm_pgprot_modify(vma->vm_page_prot, vm_flags)))
1510                 return 0;
1511
1512         /* Do we need to track softdirty? */
1513         if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1514                 return 1;
1515
1516         /* Specialty mapping? */
1517         if (vm_flags & VM_PFNMAP)
1518                 return 0;
1519
1520         /* Can the mapping track the dirty pages? */
1521         return vma->vm_file && vma->vm_file->f_mapping &&
1522                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1523 }
1524
1525 /*
1526  * We account for memory if it's a private writeable mapping,
1527  * not hugepages and VM_NORESERVE wasn't set.
1528  */
1529 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1530 {
1531         /*
1532          * hugetlb has its own accounting separate from the core VM
1533          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1534          */
1535         if (file && is_file_hugepages(file))
1536                 return 0;
1537
1538         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1539 }
1540
1541 unsigned long mmap_region(struct file *file, unsigned long addr,
1542                 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff)
1543 {
1544         struct mm_struct *mm = current->mm;
1545         struct vm_area_struct *vma, *prev;
1546         int error;
1547         struct rb_node **rb_link, *rb_parent;
1548         unsigned long charged = 0;
1549
1550         /* Check against address space limit. */
1551         if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1552                 unsigned long nr_pages;
1553
1554                 /*
1555                  * MAP_FIXED may remove pages of mappings that intersects with
1556                  * requested mapping. Account for the pages it would unmap.
1557                  */
1558                 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1559
1560                 if (!may_expand_vm(mm, vm_flags,
1561                                         (len >> PAGE_SHIFT) - nr_pages))
1562                         return -ENOMEM;
1563         }
1564
1565         /* Clear old maps */
1566         while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
1567                               &rb_parent)) {
1568                 if (do_munmap(mm, addr, len))
1569                         return -ENOMEM;
1570         }
1571
1572         /*
1573          * Private writable mapping: check memory availability
1574          */
1575         if (accountable_mapping(file, vm_flags)) {
1576                 charged = len >> PAGE_SHIFT;
1577                 if (security_vm_enough_memory_mm(mm, charged))
1578                         return -ENOMEM;
1579                 vm_flags |= VM_ACCOUNT;
1580         }
1581
1582         /*
1583          * Can we just expand an old mapping?
1584          */
1585         vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1586                         NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1587         if (vma)
1588                 goto out;
1589
1590         /*
1591          * Determine the object being mapped and call the appropriate
1592          * specific mapper. the address has already been validated, but
1593          * not unmapped, but the maps are removed from the list.
1594          */
1595         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1596         if (!vma) {
1597                 error = -ENOMEM;
1598                 goto unacct_error;
1599         }
1600
1601         vma->vm_mm = mm;
1602         vma->vm_start = addr;
1603         vma->vm_end = addr + len;
1604         vma->vm_flags = vm_flags;
1605         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1606         vma->vm_pgoff = pgoff;
1607         INIT_LIST_HEAD(&vma->anon_vma_chain);
1608
1609         if (file) {
1610                 if (vm_flags & VM_DENYWRITE) {
1611                         error = deny_write_access(file);
1612                         if (error)
1613                                 goto free_vma;
1614                 }
1615                 if (vm_flags & VM_SHARED) {
1616                         error = mapping_map_writable(file->f_mapping);
1617                         if (error)
1618                                 goto allow_write_and_free_vma;
1619                 }
1620
1621                 /* ->mmap() can change vma->vm_file, but must guarantee that
1622                  * vma_link() below can deny write-access if VM_DENYWRITE is set
1623                  * and map writably if VM_SHARED is set. This usually means the
1624                  * new file must not have been exposed to user-space, yet.
1625                  */
1626                 vma->vm_file = get_file(file);
1627                 error = file->f_op->mmap(file, vma);
1628                 if (error)
1629                         goto unmap_and_free_vma;
1630
1631                 /* Can addr have changed??
1632                  *
1633                  * Answer: Yes, several device drivers can do it in their
1634                  *         f_op->mmap method. -DaveM
1635                  * Bug: If addr is changed, prev, rb_link, rb_parent should
1636                  *      be updated for vma_link()
1637                  */
1638                 WARN_ON_ONCE(addr != vma->vm_start);
1639
1640                 addr = vma->vm_start;
1641                 vm_flags = vma->vm_flags;
1642         } else if (vm_flags & VM_SHARED) {
1643                 error = shmem_zero_setup(vma);
1644                 if (error)
1645                         goto free_vma;
1646         }
1647
1648         vma_link(mm, vma, prev, rb_link, rb_parent);
1649         /* Once vma denies write, undo our temporary denial count */
1650         if (file) {
1651                 if (vm_flags & VM_SHARED)
1652                         mapping_unmap_writable(file->f_mapping);
1653                 if (vm_flags & VM_DENYWRITE)
1654                         allow_write_access(file);
1655         }
1656         file = vma->vm_file;
1657 out:
1658         perf_event_mmap(vma);
1659
1660         vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1661         if (vm_flags & VM_LOCKED) {
1662                 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1663                                         vma == get_gate_vma(current->mm)))
1664                         mm->locked_vm += (len >> PAGE_SHIFT);
1665                 else
1666                         vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1667         }
1668
1669         if (file)
1670                 uprobe_mmap(vma);
1671
1672         /*
1673          * New (or expanded) vma always get soft dirty status.
1674          * Otherwise user-space soft-dirty page tracker won't
1675          * be able to distinguish situation when vma area unmapped,
1676          * then new mapped in-place (which must be aimed as
1677          * a completely new data area).
1678          */
1679         vma->vm_flags |= VM_SOFTDIRTY;
1680
1681         vma_set_page_prot(vma);
1682
1683         return addr;
1684
1685 unmap_and_free_vma:
1686         vma->vm_file = NULL;
1687         fput(file);
1688
1689         /* Undo any partial mapping done by a device driver. */
1690         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1691         charged = 0;
1692         if (vm_flags & VM_SHARED)
1693                 mapping_unmap_writable(file->f_mapping);
1694 allow_write_and_free_vma:
1695         if (vm_flags & VM_DENYWRITE)
1696                 allow_write_access(file);
1697 free_vma:
1698         kmem_cache_free(vm_area_cachep, vma);
1699 unacct_error:
1700         if (charged)
1701                 vm_unacct_memory(charged);
1702         return error;
1703 }
1704
1705 unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1706 {
1707         /*
1708          * We implement the search by looking for an rbtree node that
1709          * immediately follows a suitable gap. That is,
1710          * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1711          * - gap_end   = vma->vm_start        >= info->low_limit  + length;
1712          * - gap_end - gap_start >= length
1713          */
1714
1715         struct mm_struct *mm = current->mm;
1716         struct vm_area_struct *vma;
1717         unsigned long length, low_limit, high_limit, gap_start, gap_end;
1718
1719         /* Adjust search length to account for worst case alignment overhead */
1720         length = info->length + info->align_mask;
1721         if (length < info->length)
1722                 return -ENOMEM;
1723
1724         /* Adjust search limits by the desired length */
1725         if (info->high_limit < length)
1726                 return -ENOMEM;
1727         high_limit = info->high_limit - length;
1728
1729         if (info->low_limit > high_limit)
1730                 return -ENOMEM;
1731         low_limit = info->low_limit + length;
1732
1733         /* Check if rbtree root looks promising */
1734         if (RB_EMPTY_ROOT(&mm->mm_rb))
1735                 goto check_highest;
1736         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1737         if (vma->rb_subtree_gap < length)
1738                 goto check_highest;
1739
1740         while (true) {
1741                 /* Visit left subtree if it looks promising */
1742                 gap_end = vma->vm_start;
1743                 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1744                         struct vm_area_struct *left =
1745                                 rb_entry(vma->vm_rb.rb_left,
1746                                          struct vm_area_struct, vm_rb);
1747                         if (left->rb_subtree_gap >= length) {
1748                                 vma = left;
1749                                 continue;
1750                         }
1751                 }
1752
1753                 gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0;
1754 check_current:
1755                 /* Check if current node has a suitable gap */
1756                 if (gap_start > high_limit)
1757                         return -ENOMEM;
1758                 if (gap_end >= low_limit && gap_end - gap_start >= length)
1759                         goto found;
1760
1761                 /* Visit right subtree if it looks promising */
1762                 if (vma->vm_rb.rb_right) {
1763                         struct vm_area_struct *right =
1764                                 rb_entry(vma->vm_rb.rb_right,
1765                                          struct vm_area_struct, vm_rb);
1766                         if (right->rb_subtree_gap >= length) {
1767                                 vma = right;
1768                                 continue;
1769                         }
1770                 }
1771
1772                 /* Go back up the rbtree to find next candidate node */
1773                 while (true) {
1774                         struct rb_node *prev = &vma->vm_rb;
1775                         if (!rb_parent(prev))
1776                                 goto check_highest;
1777                         vma = rb_entry(rb_parent(prev),
1778                                        struct vm_area_struct, vm_rb);
1779                         if (prev == vma->vm_rb.rb_left) {
1780                                 gap_start = vma->vm_prev->vm_end;
1781                                 gap_end = vma->vm_start;
1782                                 goto check_current;
1783                         }
1784                 }
1785         }
1786
1787 check_highest:
1788         /* Check highest gap, which does not precede any rbtree node */
1789         gap_start = mm->highest_vm_end;
1790         gap_end = ULONG_MAX;  /* Only for VM_BUG_ON below */
1791         if (gap_start > high_limit)
1792                 return -ENOMEM;
1793
1794 found:
1795         /* We found a suitable gap. Clip it with the original low_limit. */
1796         if (gap_start < info->low_limit)
1797                 gap_start = info->low_limit;
1798
1799         /* Adjust gap address to the desired alignment */
1800         gap_start += (info->align_offset - gap_start) & info->align_mask;
1801
1802         VM_BUG_ON(gap_start + info->length > info->high_limit);
1803         VM_BUG_ON(gap_start + info->length > gap_end);
1804         return gap_start;
1805 }
1806
1807 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1808 {
1809         struct mm_struct *mm = current->mm;
1810         struct vm_area_struct *vma;
1811         unsigned long length, low_limit, high_limit, gap_start, gap_end;
1812
1813         /* Adjust search length to account for worst case alignment overhead */
1814         length = info->length + info->align_mask;
1815         if (length < info->length)
1816                 return -ENOMEM;
1817
1818         /*
1819          * Adjust search limits by the desired length.
1820          * See implementation comment at top of unmapped_area().
1821          */
1822         gap_end = info->high_limit;
1823         if (gap_end < length)
1824                 return -ENOMEM;
1825         high_limit = gap_end - length;
1826
1827         if (info->low_limit > high_limit)
1828                 return -ENOMEM;
1829         low_limit = info->low_limit + length;
1830
1831         /* Check highest gap, which does not precede any rbtree node */
1832         gap_start = mm->highest_vm_end;
1833         if (gap_start <= high_limit)
1834                 goto found_highest;
1835
1836         /* Check if rbtree root looks promising */
1837         if (RB_EMPTY_ROOT(&mm->mm_rb))
1838                 return -ENOMEM;
1839         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1840         if (vma->rb_subtree_gap < length)
1841                 return -ENOMEM;
1842
1843         while (true) {
1844                 /* Visit right subtree if it looks promising */
1845                 gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0;
1846                 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1847                         struct vm_area_struct *right =
1848                                 rb_entry(vma->vm_rb.rb_right,
1849                                          struct vm_area_struct, vm_rb);
1850                         if (right->rb_subtree_gap >= length) {
1851                                 vma = right;
1852                                 continue;
1853                         }
1854                 }
1855
1856 check_current:
1857                 /* Check if current node has a suitable gap */
1858                 gap_end = vma->vm_start;
1859                 if (gap_end < low_limit)
1860                         return -ENOMEM;
1861                 if (gap_start <= high_limit && gap_end - gap_start >= length)
1862                         goto found;
1863
1864                 /* Visit left subtree if it looks promising */
1865                 if (vma->vm_rb.rb_left) {
1866                         struct vm_area_struct *left =
1867                                 rb_entry(vma->vm_rb.rb_left,
1868                                          struct vm_area_struct, vm_rb);
1869                         if (left->rb_subtree_gap >= length) {
1870                                 vma = left;
1871                                 continue;
1872                         }
1873                 }
1874
1875                 /* Go back up the rbtree to find next candidate node */
1876                 while (true) {
1877                         struct rb_node *prev = &vma->vm_rb;
1878                         if (!rb_parent(prev))
1879                                 return -ENOMEM;
1880                         vma = rb_entry(rb_parent(prev),
1881                                        struct vm_area_struct, vm_rb);
1882                         if (prev == vma->vm_rb.rb_right) {
1883                                 gap_start = vma->vm_prev ?
1884                                         vma->vm_prev->vm_end : 0;
1885                                 goto check_current;
1886                         }
1887                 }
1888         }
1889
1890 found:
1891         /* We found a suitable gap. Clip it with the original high_limit. */
1892         if (gap_end > info->high_limit)
1893                 gap_end = info->high_limit;
1894
1895 found_highest:
1896         /* Compute highest gap address at the desired alignment */
1897         gap_end -= info->length;
1898         gap_end -= (gap_end - info->align_offset) & info->align_mask;
1899
1900         VM_BUG_ON(gap_end < info->low_limit);
1901         VM_BUG_ON(gap_end < gap_start);
1902         return gap_end;
1903 }
1904
1905 /* Get an address range which is currently unmapped.
1906  * For shmat() with addr=0.
1907  *
1908  * Ugly calling convention alert:
1909  * Return value with the low bits set means error value,
1910  * ie
1911  *      if (ret & ~PAGE_MASK)
1912  *              error = ret;
1913  *
1914  * This function "knows" that -ENOMEM has the bits set.
1915  */
1916 #ifndef HAVE_ARCH_UNMAPPED_AREA
1917 unsigned long
1918 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1919                 unsigned long len, unsigned long pgoff, unsigned long flags)
1920 {
1921         struct mm_struct *mm = current->mm;
1922         struct vm_area_struct *vma;
1923         struct vm_unmapped_area_info info;
1924
1925         if (len > TASK_SIZE - mmap_min_addr)
1926                 return -ENOMEM;
1927
1928         if (flags & MAP_FIXED)
1929                 return addr;
1930
1931         if (addr) {
1932                 addr = PAGE_ALIGN(addr);
1933                 vma = find_vma(mm, addr);
1934                 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
1935                     (!vma || addr + len <= vma->vm_start))
1936                         return addr;
1937         }
1938
1939         info.flags = 0;
1940         info.length = len;
1941         info.low_limit = mm->mmap_base;
1942         info.high_limit = TASK_SIZE;
1943         info.align_mask = 0;
1944         return vm_unmapped_area(&info);
1945 }
1946 #endif
1947
1948 /*
1949  * This mmap-allocator allocates new areas top-down from below the
1950  * stack's low limit (the base):
1951  */
1952 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1953 unsigned long
1954 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1955                           const unsigned long len, const unsigned long pgoff,
1956                           const unsigned long flags)
1957 {
1958         struct vm_area_struct *vma;
1959         struct mm_struct *mm = current->mm;
1960         unsigned long addr = addr0;
1961         struct vm_unmapped_area_info info;
1962
1963         /* requested length too big for entire address space */
1964         if (len > TASK_SIZE - mmap_min_addr)
1965                 return -ENOMEM;
1966
1967         if (flags & MAP_FIXED)
1968                 return addr;
1969
1970         /* requesting a specific address */
1971         if (addr) {
1972                 addr = PAGE_ALIGN(addr);
1973                 vma = find_vma(mm, addr);
1974                 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
1975                                 (!vma || addr + len <= vma->vm_start))
1976                         return addr;
1977         }
1978
1979         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
1980         info.length = len;
1981         info.low_limit = max(PAGE_SIZE, mmap_min_addr);
1982         info.high_limit = mm->mmap_base;
1983         info.align_mask = 0;
1984         addr = vm_unmapped_area(&info);
1985
1986         /*
1987          * A failed mmap() very likely causes application failure,
1988          * so fall back to the bottom-up function here. This scenario
1989          * can happen with large stack limits and large mmap()
1990          * allocations.
1991          */
1992         if (offset_in_page(addr)) {
1993                 VM_BUG_ON(addr != -ENOMEM);
1994                 info.flags = 0;
1995                 info.low_limit = TASK_UNMAPPED_BASE;
1996                 info.high_limit = TASK_SIZE;
1997                 addr = vm_unmapped_area(&info);
1998         }
1999
2000         return addr;
2001 }
2002 #endif
2003
2004 unsigned long
2005 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2006                 unsigned long pgoff, unsigned long flags)
2007 {
2008         unsigned long (*get_area)(struct file *, unsigned long,
2009                                   unsigned long, unsigned long, unsigned long);
2010
2011         unsigned long error = arch_mmap_check(addr, len, flags);
2012         if (error)
2013                 return error;
2014
2015         /* Careful about overflows.. */
2016         if (len > TASK_SIZE)
2017                 return -ENOMEM;
2018
2019         get_area = current->mm->get_unmapped_area;
2020         if (file && file->f_op->get_unmapped_area)
2021                 get_area = file->f_op->get_unmapped_area;
2022         addr = get_area(file, addr, len, pgoff, flags);
2023         if (IS_ERR_VALUE(addr))
2024                 return addr;
2025
2026         if (addr > TASK_SIZE - len)
2027                 return -ENOMEM;
2028         if (offset_in_page(addr))
2029                 return -EINVAL;
2030
2031         addr = arch_rebalance_pgtables(addr, len);
2032         error = security_mmap_addr(addr);
2033         return error ? error : addr;
2034 }
2035
2036 EXPORT_SYMBOL(get_unmapped_area);
2037
2038 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
2039 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2040 {
2041         struct rb_node *rb_node;
2042         struct vm_area_struct *vma;
2043
2044         /* Check the cache first. */
2045         vma = vmacache_find(mm, addr);
2046         if (likely(vma))
2047                 return vma;
2048
2049         rb_node = mm->mm_rb.rb_node;
2050
2051         while (rb_node) {
2052                 struct vm_area_struct *tmp;
2053
2054                 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2055
2056                 if (tmp->vm_end > addr) {
2057                         vma = tmp;
2058                         if (tmp->vm_start <= addr)
2059                                 break;
2060                         rb_node = rb_node->rb_left;
2061                 } else
2062                         rb_node = rb_node->rb_right;
2063         }
2064
2065         if (vma)
2066                 vmacache_update(addr, vma);
2067         return vma;
2068 }
2069
2070 EXPORT_SYMBOL(find_vma);
2071
2072 /*
2073  * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2074  */
2075 struct vm_area_struct *
2076 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2077                         struct vm_area_struct **pprev)
2078 {
2079         struct vm_area_struct *vma;
2080
2081         vma = find_vma(mm, addr);
2082         if (vma) {
2083                 *pprev = vma->vm_prev;
2084         } else {
2085                 struct rb_node *rb_node = mm->mm_rb.rb_node;
2086                 *pprev = NULL;
2087                 while (rb_node) {
2088                         *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2089                         rb_node = rb_node->rb_right;
2090                 }
2091         }
2092         return vma;
2093 }
2094
2095 /*
2096  * Verify that the stack growth is acceptable and
2097  * update accounting. This is shared with both the
2098  * grow-up and grow-down cases.
2099  */
2100 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
2101 {
2102         struct mm_struct *mm = vma->vm_mm;
2103         struct rlimit *rlim = current->signal->rlim;
2104         unsigned long new_start, actual_size;
2105
2106         /* address space limit tests */
2107         if (!may_expand_vm(mm, vma->vm_flags, grow))
2108                 return -ENOMEM;
2109
2110         /* Stack limit test */
2111         actual_size = size;
2112         if (size && (vma->vm_flags & (VM_GROWSUP | VM_GROWSDOWN)))
2113                 actual_size -= PAGE_SIZE;
2114         if (actual_size > READ_ONCE(rlim[RLIMIT_STACK].rlim_cur))
2115                 return -ENOMEM;
2116
2117         /* mlock limit tests */
2118         if (vma->vm_flags & VM_LOCKED) {
2119                 unsigned long locked;
2120                 unsigned long limit;
2121                 locked = mm->locked_vm + grow;
2122                 limit = READ_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
2123                 limit >>= PAGE_SHIFT;
2124                 if (locked > limit && !capable(CAP_IPC_LOCK))
2125                         return -ENOMEM;
2126         }
2127
2128         /* Check to ensure the stack will not grow into a hugetlb-only region */
2129         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2130                         vma->vm_end - size;
2131         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2132                 return -EFAULT;
2133
2134         /*
2135          * Overcommit..  This must be the final test, as it will
2136          * update security statistics.
2137          */
2138         if (security_vm_enough_memory_mm(mm, grow))
2139                 return -ENOMEM;
2140
2141         return 0;
2142 }
2143
2144 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2145 /*
2146  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2147  * vma is the last one with address > vma->vm_end.  Have to extend vma.
2148  */
2149 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2150 {
2151         struct mm_struct *mm = vma->vm_mm;
2152         int error = 0;
2153
2154         if (!(vma->vm_flags & VM_GROWSUP))
2155                 return -EFAULT;
2156
2157         /* Guard against wrapping around to address 0. */
2158         if (address < PAGE_ALIGN(address+4))
2159                 address = PAGE_ALIGN(address+4);
2160         else
2161                 return -ENOMEM;
2162
2163         /* We must make sure the anon_vma is allocated. */
2164         if (unlikely(anon_vma_prepare(vma)))
2165                 return -ENOMEM;
2166
2167         /*
2168          * vma->vm_start/vm_end cannot change under us because the caller
2169          * is required to hold the mmap_sem in read mode.  We need the
2170          * anon_vma lock to serialize against concurrent expand_stacks.
2171          */
2172         anon_vma_lock_write(vma->anon_vma);
2173
2174         /* Somebody else might have raced and expanded it already */
2175         if (address > vma->vm_end) {
2176                 unsigned long size, grow;
2177
2178                 size = address - vma->vm_start;
2179                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2180
2181                 error = -ENOMEM;
2182                 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2183                         error = acct_stack_growth(vma, size, grow);
2184                         if (!error) {
2185                                 /*
2186                                  * vma_gap_update() doesn't support concurrent
2187                                  * updates, but we only hold a shared mmap_sem
2188                                  * lock here, so we need to protect against
2189                                  * concurrent vma expansions.
2190                                  * anon_vma_lock_write() doesn't help here, as
2191                                  * we don't guarantee that all growable vmas
2192                                  * in a mm share the same root anon vma.
2193                                  * So, we reuse mm->page_table_lock to guard
2194                                  * against concurrent vma expansions.
2195                                  */
2196                                 spin_lock(&mm->page_table_lock);
2197                                 if (vma->vm_flags & VM_LOCKED)
2198                                         mm->locked_vm += grow;
2199                                 vm_stat_account(mm, vma->vm_flags, grow);
2200                                 anon_vma_interval_tree_pre_update_vma(vma);
2201                                 vma->vm_end = address;
2202                                 anon_vma_interval_tree_post_update_vma(vma);
2203                                 if (vma->vm_next)
2204                                         vma_gap_update(vma->vm_next);
2205                                 else
2206                                         mm->highest_vm_end = address;
2207                                 spin_unlock(&mm->page_table_lock);
2208
2209                                 perf_event_mmap(vma);
2210                         }
2211                 }
2212         }
2213         anon_vma_unlock_write(vma->anon_vma);
2214         khugepaged_enter_vma_merge(vma, vma->vm_flags);
2215         validate_mm(mm);
2216         return error;
2217 }
2218 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2219
2220 /*
2221  * vma is the first one with address < vma->vm_start.  Have to extend vma.
2222  */
2223 int expand_downwards(struct vm_area_struct *vma,
2224                                    unsigned long address)
2225 {
2226         struct mm_struct *mm = vma->vm_mm;
2227         int error;
2228
2229         address &= PAGE_MASK;
2230         error = security_mmap_addr(address);
2231         if (error)
2232                 return error;
2233
2234         /* We must make sure the anon_vma is allocated. */
2235         if (unlikely(anon_vma_prepare(vma)))
2236                 return -ENOMEM;
2237
2238         /*
2239          * vma->vm_start/vm_end cannot change under us because the caller
2240          * is required to hold the mmap_sem in read mode.  We need the
2241          * anon_vma lock to serialize against concurrent expand_stacks.
2242          */
2243         anon_vma_lock_write(vma->anon_vma);
2244
2245         /* Somebody else might have raced and expanded it already */
2246         if (address < vma->vm_start) {
2247                 unsigned long size, grow;
2248
2249                 size = vma->vm_end - address;
2250                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2251
2252                 error = -ENOMEM;
2253                 if (grow <= vma->vm_pgoff) {
2254                         error = acct_stack_growth(vma, size, grow);
2255                         if (!error) {
2256                                 /*
2257                                  * vma_gap_update() doesn't support concurrent
2258                                  * updates, but we only hold a shared mmap_sem
2259                                  * lock here, so we need to protect against
2260                                  * concurrent vma expansions.
2261                                  * anon_vma_lock_write() doesn't help here, as
2262                                  * we don't guarantee that all growable vmas
2263                                  * in a mm share the same root anon vma.
2264                                  * So, we reuse mm->page_table_lock to guard
2265                                  * against concurrent vma expansions.
2266                                  */
2267                                 spin_lock(&mm->page_table_lock);
2268                                 if (vma->vm_flags & VM_LOCKED)
2269                                         mm->locked_vm += grow;
2270                                 vm_stat_account(mm, vma->vm_flags, grow);
2271                                 anon_vma_interval_tree_pre_update_vma(vma);
2272                                 vma->vm_start = address;
2273                                 vma->vm_pgoff -= grow;
2274                                 anon_vma_interval_tree_post_update_vma(vma);
2275                                 vma_gap_update(vma);
2276                                 spin_unlock(&mm->page_table_lock);
2277
2278                                 perf_event_mmap(vma);
2279                         }
2280                 }
2281         }
2282         anon_vma_unlock_write(vma->anon_vma);
2283         khugepaged_enter_vma_merge(vma, vma->vm_flags);
2284         validate_mm(mm);
2285         return error;
2286 }
2287
2288 /*
2289  * Note how expand_stack() refuses to expand the stack all the way to
2290  * abut the next virtual mapping, *unless* that mapping itself is also
2291  * a stack mapping. We want to leave room for a guard page, after all
2292  * (the guard page itself is not added here, that is done by the
2293  * actual page faulting logic)
2294  *
2295  * This matches the behavior of the guard page logic (see mm/memory.c:
2296  * check_stack_guard_page()), which only allows the guard page to be
2297  * removed under these circumstances.
2298  */
2299 #ifdef CONFIG_STACK_GROWSUP
2300 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2301 {
2302         struct vm_area_struct *next;
2303
2304         address &= PAGE_MASK;
2305         next = vma->vm_next;
2306         if (next && next->vm_start == address + PAGE_SIZE) {
2307                 if (!(next->vm_flags & VM_GROWSUP))
2308                         return -ENOMEM;
2309         }
2310         return expand_upwards(vma, address);
2311 }
2312
2313 struct vm_area_struct *
2314 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2315 {
2316         struct vm_area_struct *vma, *prev;
2317
2318         addr &= PAGE_MASK;
2319         vma = find_vma_prev(mm, addr, &prev);
2320         if (vma && (vma->vm_start <= addr))
2321                 return vma;
2322         if (!prev || expand_stack(prev, addr))
2323                 return NULL;
2324         if (prev->vm_flags & VM_LOCKED)
2325                 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2326         return prev;
2327 }
2328 #else
2329 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2330 {
2331         struct vm_area_struct *prev;
2332
2333         address &= PAGE_MASK;
2334         prev = vma->vm_prev;
2335         if (prev && prev->vm_end == address) {
2336                 if (!(prev->vm_flags & VM_GROWSDOWN))
2337                         return -ENOMEM;
2338         }
2339         return expand_downwards(vma, address);
2340 }
2341
2342 struct vm_area_struct *
2343 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2344 {
2345         struct vm_area_struct *vma;
2346         unsigned long start;
2347
2348         addr &= PAGE_MASK;
2349         vma = find_vma(mm, addr);
2350         if (!vma)
2351                 return NULL;
2352         if (vma->vm_start <= addr)
2353                 return vma;
2354         if (!(vma->vm_flags & VM_GROWSDOWN))
2355                 return NULL;
2356         start = vma->vm_start;
2357         if (expand_stack(vma, addr))
2358                 return NULL;
2359         if (vma->vm_flags & VM_LOCKED)
2360                 populate_vma_page_range(vma, addr, start, NULL);
2361         return vma;
2362 }
2363 #endif
2364
2365 EXPORT_SYMBOL_GPL(find_extend_vma);
2366
2367 /*
2368  * Ok - we have the memory areas we should free on the vma list,
2369  * so release them, and do the vma updates.
2370  *
2371  * Called with the mm semaphore held.
2372  */
2373 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2374 {
2375         unsigned long nr_accounted = 0;
2376
2377         /* Update high watermark before we lower total_vm */
2378         update_hiwater_vm(mm);
2379         do {
2380                 long nrpages = vma_pages(vma);
2381
2382                 if (vma->vm_flags & VM_ACCOUNT)
2383                         nr_accounted += nrpages;
2384                 vm_stat_account(mm, vma->vm_flags, -nrpages);
2385                 vma = remove_vma(vma);
2386         } while (vma);
2387         vm_unacct_memory(nr_accounted);
2388         validate_mm(mm);
2389 }
2390
2391 /*
2392  * Get rid of page table information in the indicated region.
2393  *
2394  * Called with the mm semaphore held.
2395  */
2396 static void unmap_region(struct mm_struct *mm,
2397                 struct vm_area_struct *vma, struct vm_area_struct *prev,
2398                 unsigned long start, unsigned long end)
2399 {
2400         struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap;
2401         struct mmu_gather tlb;
2402
2403         lru_add_drain();
2404         tlb_gather_mmu(&tlb, mm, start, end);
2405         update_hiwater_rss(mm);
2406         unmap_vmas(&tlb, vma, start, end);
2407         free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2408                                  next ? next->vm_start : USER_PGTABLES_CEILING);
2409         tlb_finish_mmu(&tlb, start, end);
2410 }
2411
2412 /*
2413  * Create a list of vma's touched by the unmap, removing them from the mm's
2414  * vma list as we go..
2415  */
2416 static void
2417 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2418         struct vm_area_struct *prev, unsigned long end)
2419 {
2420         struct vm_area_struct **insertion_point;
2421         struct vm_area_struct *tail_vma = NULL;
2422
2423         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2424         vma->vm_prev = NULL;
2425         do {
2426                 vma_rb_erase(vma, &mm->mm_rb);
2427                 mm->map_count--;
2428                 tail_vma = vma;
2429                 vma = vma->vm_next;
2430         } while (vma && vma->vm_start < end);
2431         *insertion_point = vma;
2432         if (vma) {
2433                 vma->vm_prev = prev;
2434                 vma_gap_update(vma);
2435         } else
2436                 mm->highest_vm_end = prev ? prev->vm_end : 0;
2437         tail_vma->vm_next = NULL;
2438
2439         /* Kill the cache */
2440         vmacache_invalidate(mm);
2441 }
2442
2443 /*
2444  * __split_vma() bypasses sysctl_max_map_count checking.  We use this on the
2445  * munmap path where it doesn't make sense to fail.
2446  */
2447 static int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2448               unsigned long addr, int new_below)
2449 {
2450         struct vm_area_struct *new;
2451         int err;
2452
2453         if (is_vm_hugetlb_page(vma) && (addr &
2454                                         ~(huge_page_mask(hstate_vma(vma)))))
2455                 return -EINVAL;
2456
2457         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2458         if (!new)
2459                 return -ENOMEM;
2460
2461         /* most fields are the same, copy all, and then fixup */
2462         *new = *vma;
2463
2464         INIT_LIST_HEAD(&new->anon_vma_chain);
2465
2466         if (new_below)
2467                 new->vm_end = addr;
2468         else {
2469                 new->vm_start = addr;
2470                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2471         }
2472
2473         err = vma_dup_policy(vma, new);
2474         if (err)
2475                 goto out_free_vma;
2476
2477         err = anon_vma_clone(new, vma);
2478         if (err)
2479                 goto out_free_mpol;
2480
2481         if (new->vm_file)
2482                 get_file(new->vm_file);
2483
2484         if (new->vm_ops && new->vm_ops->open)
2485                 new->vm_ops->open(new);
2486
2487         if (new_below)
2488                 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2489                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
2490         else
2491                 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2492
2493         /* Success. */
2494         if (!err)
2495                 return 0;
2496
2497         /* Clean everything up if vma_adjust failed. */
2498         if (new->vm_ops && new->vm_ops->close)
2499                 new->vm_ops->close(new);
2500         if (new->vm_file)
2501                 fput(new->vm_file);
2502         unlink_anon_vmas(new);
2503  out_free_mpol:
2504         mpol_put(vma_policy(new));
2505  out_free_vma:
2506         kmem_cache_free(vm_area_cachep, new);
2507         return err;
2508 }
2509
2510 /*
2511  * Split a vma into two pieces at address 'addr', a new vma is allocated
2512  * either for the first part or the tail.
2513  */
2514 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2515               unsigned long addr, int new_below)
2516 {
2517         if (mm->map_count >= sysctl_max_map_count)
2518                 return -ENOMEM;
2519
2520         return __split_vma(mm, vma, addr, new_below);
2521 }
2522
2523 /* Munmap is split into 2 main parts -- this part which finds
2524  * what needs doing, and the areas themselves, which do the
2525  * work.  This now handles partial unmappings.
2526  * Jeremy Fitzhardinge <jeremy@goop.org>
2527  */
2528 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2529 {
2530         unsigned long end;
2531         struct vm_area_struct *vma, *prev, *last;
2532
2533         if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2534                 return -EINVAL;
2535
2536         len = PAGE_ALIGN(len);
2537         if (len == 0)
2538                 return -EINVAL;
2539
2540         /* Find the first overlapping VMA */
2541         vma = find_vma(mm, start);
2542         if (!vma)
2543                 return 0;
2544         prev = vma->vm_prev;
2545         /* we have  start < vma->vm_end  */
2546
2547         /* if it doesn't overlap, we have nothing.. */
2548         end = start + len;
2549         if (vma->vm_start >= end)
2550                 return 0;
2551
2552         /*
2553          * If we need to split any vma, do it now to save pain later.
2554          *
2555          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2556          * unmapped vm_area_struct will remain in use: so lower split_vma
2557          * places tmp vma above, and higher split_vma places tmp vma below.
2558          */
2559         if (start > vma->vm_start) {
2560                 int error;
2561
2562                 /*
2563                  * Make sure that map_count on return from munmap() will
2564                  * not exceed its limit; but let map_count go just above
2565                  * its limit temporarily, to help free resources as expected.
2566                  */
2567                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2568                         return -ENOMEM;
2569
2570                 error = __split_vma(mm, vma, start, 0);
2571                 if (error)
2572                         return error;
2573                 prev = vma;
2574         }
2575
2576         /* Does it split the last one? */
2577         last = find_vma(mm, end);
2578         if (last && end > last->vm_start) {
2579                 int error = __split_vma(mm, last, end, 1);
2580                 if (error)
2581                         return error;
2582         }
2583         vma = prev ? prev->vm_next : mm->mmap;
2584
2585         /*
2586          * unlock any mlock()ed ranges before detaching vmas
2587          */
2588         if (mm->locked_vm) {
2589                 struct vm_area_struct *tmp = vma;
2590                 while (tmp && tmp->vm_start < end) {
2591                         if (tmp->vm_flags & VM_LOCKED) {
2592                                 mm->locked_vm -= vma_pages(tmp);
2593                                 munlock_vma_pages_all(tmp);
2594                         }
2595                         tmp = tmp->vm_next;
2596                 }
2597         }
2598
2599         /*
2600          * Remove the vma's, and unmap the actual pages
2601          */
2602         detach_vmas_to_be_unmapped(mm, vma, prev, end);
2603         unmap_region(mm, vma, prev, start, end);
2604
2605         arch_unmap(mm, vma, start, end);
2606
2607         /* Fix up all other VM information */
2608         remove_vma_list(mm, vma);
2609
2610         return 0;
2611 }
2612
2613 int vm_munmap(unsigned long start, size_t len)
2614 {
2615         int ret;
2616         struct mm_struct *mm = current->mm;
2617
2618         down_write(&mm->mmap_sem);
2619         ret = do_munmap(mm, start, len);
2620         up_write(&mm->mmap_sem);
2621         return ret;
2622 }
2623 EXPORT_SYMBOL(vm_munmap);
2624
2625 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2626 {
2627         profile_munmap(addr);
2628         return vm_munmap(addr, len);
2629 }
2630
2631
2632 /*
2633  * Emulation of deprecated remap_file_pages() syscall.
2634  */
2635 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2636                 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2637 {
2638
2639         struct mm_struct *mm = current->mm;
2640         struct vm_area_struct *vma;
2641         unsigned long populate = 0;
2642         unsigned long ret = -EINVAL;
2643         struct file *file;
2644
2645         pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. "
2646                         "See Documentation/vm/remap_file_pages.txt.\n",
2647                         current->comm, current->pid);
2648
2649         if (prot)
2650                 return ret;
2651         start = start & PAGE_MASK;
2652         size = size & PAGE_MASK;
2653
2654         if (start + size <= start)
2655                 return ret;
2656
2657         /* Does pgoff wrap? */
2658         if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2659                 return ret;
2660
2661         down_write(&mm->mmap_sem);
2662         vma = find_vma(mm, start);
2663
2664         if (!vma || !(vma->vm_flags & VM_SHARED))
2665                 goto out;
2666
2667         if (start < vma->vm_start)
2668                 goto out;
2669
2670         if (start + size > vma->vm_end) {
2671                 struct vm_area_struct *next;
2672
2673                 for (next = vma->vm_next; next; next = next->vm_next) {
2674                         /* hole between vmas ? */
2675                         if (next->vm_start != next->vm_prev->vm_end)
2676                                 goto out;
2677
2678                         if (next->vm_file != vma->vm_file)
2679                                 goto out;
2680
2681                         if (next->vm_flags != vma->vm_flags)
2682                                 goto out;
2683
2684                         if (start + size <= next->vm_end)
2685                                 break;
2686                 }
2687
2688                 if (!next)
2689                         goto out;
2690         }
2691
2692         prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2693         prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2694         prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2695
2696         flags &= MAP_NONBLOCK;
2697         flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2698         if (vma->vm_flags & VM_LOCKED) {
2699                 struct vm_area_struct *tmp;
2700                 flags |= MAP_LOCKED;
2701
2702                 /* drop PG_Mlocked flag for over-mapped range */
2703                 for (tmp = vma; tmp->vm_start >= start + size;
2704                                 tmp = tmp->vm_next) {
2705                         munlock_vma_pages_range(tmp,
2706                                         max(tmp->vm_start, start),
2707                                         min(tmp->vm_end, start + size));
2708                 }
2709         }
2710
2711         file = get_file(vma->vm_file);
2712         ret = do_mmap_pgoff(vma->vm_file, start, size,
2713                         prot, flags, pgoff, &populate);
2714         fput(file);
2715 out:
2716         up_write(&mm->mmap_sem);
2717         if (populate)
2718                 mm_populate(ret, populate);
2719         if (!IS_ERR_VALUE(ret))
2720                 ret = 0;
2721         return ret;
2722 }
2723
2724 static inline void verify_mm_writelocked(struct mm_struct *mm)
2725 {
2726 #ifdef CONFIG_DEBUG_VM
2727         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2728                 WARN_ON(1);
2729                 up_read(&mm->mmap_sem);
2730         }
2731 #endif
2732 }
2733
2734 /*
2735  *  this is really a simplified "do_mmap".  it only handles
2736  *  anonymous maps.  eventually we may be able to do some
2737  *  brk-specific accounting here.
2738  */
2739 static unsigned long do_brk(unsigned long addr, unsigned long len)
2740 {
2741         struct mm_struct *mm = current->mm;
2742         struct vm_area_struct *vma, *prev;
2743         unsigned long flags;
2744         struct rb_node **rb_link, *rb_parent;
2745         pgoff_t pgoff = addr >> PAGE_SHIFT;
2746         int error;
2747
2748         len = PAGE_ALIGN(len);
2749         if (!len)
2750                 return addr;
2751
2752         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2753
2754         error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2755         if (offset_in_page(error))
2756                 return error;
2757
2758         error = mlock_future_check(mm, mm->def_flags, len);
2759         if (error)
2760                 return error;
2761
2762         /*
2763          * mm->mmap_sem is required to protect against another thread
2764          * changing the mappings in case we sleep.
2765          */
2766         verify_mm_writelocked(mm);
2767
2768         /*
2769          * Clear old maps.  this also does some error checking for us
2770          */
2771         while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
2772                               &rb_parent)) {
2773                 if (do_munmap(mm, addr, len))
2774                         return -ENOMEM;
2775         }
2776
2777         /* Check against address space limits *after* clearing old maps... */
2778         if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
2779                 return -ENOMEM;
2780
2781         if (mm->map_count > sysctl_max_map_count)
2782                 return -ENOMEM;
2783
2784         if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2785                 return -ENOMEM;
2786
2787         /* Can we just expand an old private anonymous mapping? */
2788         vma = vma_merge(mm, prev, addr, addr + len, flags,
2789                         NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
2790         if (vma)
2791                 goto out;
2792
2793         /*
2794          * create a vma struct for an anonymous mapping
2795          */
2796         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2797         if (!vma) {
2798                 vm_unacct_memory(len >> PAGE_SHIFT);
2799                 return -ENOMEM;
2800         }
2801
2802         INIT_LIST_HEAD(&vma->anon_vma_chain);
2803         vma->vm_mm = mm;
2804         vma->vm_start = addr;
2805         vma->vm_end = addr + len;
2806         vma->vm_pgoff = pgoff;
2807         vma->vm_flags = flags;
2808         vma->vm_page_prot = vm_get_page_prot(flags);
2809         vma_link(mm, vma, prev, rb_link, rb_parent);
2810 out:
2811         perf_event_mmap(vma);
2812         mm->total_vm += len >> PAGE_SHIFT;
2813         mm->data_vm += len >> PAGE_SHIFT;
2814         if (flags & VM_LOCKED)
2815                 mm->locked_vm += (len >> PAGE_SHIFT);
2816         vma->vm_flags |= VM_SOFTDIRTY;
2817         return addr;
2818 }
2819
2820 unsigned long vm_brk(unsigned long addr, unsigned long len)
2821 {
2822         struct mm_struct *mm = current->mm;
2823         unsigned long ret;
2824         bool populate;
2825
2826         down_write(&mm->mmap_sem);
2827         ret = do_brk(addr, len);
2828         populate = ((mm->def_flags & VM_LOCKED) != 0);
2829         up_write(&mm->mmap_sem);
2830         if (populate)
2831                 mm_populate(addr, len);
2832         return ret;
2833 }
2834 EXPORT_SYMBOL(vm_brk);
2835
2836 /* Release all mmaps. */
2837 void exit_mmap(struct mm_struct *mm)
2838 {
2839         struct mmu_gather tlb;
2840         struct vm_area_struct *vma;
2841         unsigned long nr_accounted = 0;
2842
2843         /* mm's last user has gone, and its about to be pulled down */
2844         mmu_notifier_release(mm);
2845
2846         if (mm->locked_vm) {
2847                 vma = mm->mmap;
2848                 while (vma) {
2849                         if (vma->vm_flags & VM_LOCKED)
2850                                 munlock_vma_pages_all(vma);
2851                         vma = vma->vm_next;
2852                 }
2853         }
2854
2855         arch_exit_mmap(mm);
2856
2857         vma = mm->mmap;
2858         if (!vma)       /* Can happen if dup_mmap() received an OOM */
2859                 return;
2860
2861         lru_add_drain();
2862         flush_cache_mm(mm);
2863         tlb_gather_mmu(&tlb, mm, 0, -1);
2864         /* update_hiwater_rss(mm) here? but nobody should be looking */
2865         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2866         unmap_vmas(&tlb, vma, 0, -1);
2867
2868         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
2869         tlb_finish_mmu(&tlb, 0, -1);
2870
2871         /*
2872          * Walk the list again, actually closing and freeing it,
2873          * with preemption enabled, without holding any MM locks.
2874          */
2875         while (vma) {
2876                 if (vma->vm_flags & VM_ACCOUNT)
2877                         nr_accounted += vma_pages(vma);
2878                 vma = remove_vma(vma);
2879         }
2880         vm_unacct_memory(nr_accounted);
2881 }
2882
2883 /* Insert vm structure into process list sorted by address
2884  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2885  * then i_mmap_rwsem is taken here.
2886  */
2887 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
2888 {
2889         struct vm_area_struct *prev;
2890         struct rb_node **rb_link, *rb_parent;
2891
2892         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
2893                            &prev, &rb_link, &rb_parent))
2894                 return -ENOMEM;
2895         if ((vma->vm_flags & VM_ACCOUNT) &&
2896              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2897                 return -ENOMEM;
2898
2899         /*
2900          * The vm_pgoff of a purely anonymous vma should be irrelevant
2901          * until its first write fault, when page's anon_vma and index
2902          * are set.  But now set the vm_pgoff it will almost certainly
2903          * end up with (unless mremap moves it elsewhere before that
2904          * first wfault), so /proc/pid/maps tells a consistent story.
2905          *
2906          * By setting it to reflect the virtual start address of the
2907          * vma, merges and splits can happen in a seamless way, just
2908          * using the existing file pgoff checks and manipulations.
2909          * Similarly in do_mmap_pgoff and in do_brk.
2910          */
2911         if (vma_is_anonymous(vma)) {
2912                 BUG_ON(vma->anon_vma);
2913                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2914         }
2915
2916         vma_link(mm, vma, prev, rb_link, rb_parent);
2917         return 0;
2918 }
2919
2920 /*
2921  * Copy the vma structure to a new location in the same mm,
2922  * prior to moving page table entries, to effect an mremap move.
2923  */
2924 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2925         unsigned long addr, unsigned long len, pgoff_t pgoff,
2926         bool *need_rmap_locks)
2927 {
2928         struct vm_area_struct *vma = *vmap;
2929         unsigned long vma_start = vma->vm_start;
2930         struct mm_struct *mm = vma->vm_mm;
2931         struct vm_area_struct *new_vma, *prev;
2932         struct rb_node **rb_link, *rb_parent;
2933         bool faulted_in_anon_vma = true;
2934
2935         /*
2936          * If anonymous vma has not yet been faulted, update new pgoff
2937          * to match new location, to increase its chance of merging.
2938          */
2939         if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
2940                 pgoff = addr >> PAGE_SHIFT;
2941                 faulted_in_anon_vma = false;
2942         }
2943
2944         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
2945                 return NULL;    /* should never get here */
2946         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2947                             vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
2948                             vma->vm_userfaultfd_ctx);
2949         if (new_vma) {
2950                 /*
2951                  * Source vma may have been merged into new_vma
2952                  */
2953                 if (unlikely(vma_start >= new_vma->vm_start &&
2954                              vma_start < new_vma->vm_end)) {
2955                         /*
2956                          * The only way we can get a vma_merge with
2957                          * self during an mremap is if the vma hasn't
2958                          * been faulted in yet and we were allowed to
2959                          * reset the dst vma->vm_pgoff to the
2960                          * destination address of the mremap to allow
2961                          * the merge to happen. mremap must change the
2962                          * vm_pgoff linearity between src and dst vmas
2963                          * (in turn preventing a vma_merge) to be
2964                          * safe. It is only safe to keep the vm_pgoff
2965                          * linear if there are no pages mapped yet.
2966                          */
2967                         VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
2968                         *vmap = vma = new_vma;
2969                 }
2970                 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
2971         } else {
2972                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2973                 if (!new_vma)
2974                         goto out;
2975                 *new_vma = *vma;
2976                 new_vma->vm_start = addr;
2977                 new_vma->vm_end = addr + len;
2978                 new_vma->vm_pgoff = pgoff;
2979                 if (vma_dup_policy(vma, new_vma))
2980                         goto out_free_vma;
2981                 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2982                 if (anon_vma_clone(new_vma, vma))
2983                         goto out_free_mempol;
2984                 if (new_vma->vm_file)
2985                         get_file(new_vma->vm_file);
2986                 if (new_vma->vm_ops && new_vma->vm_ops->open)
2987                         new_vma->vm_ops->open(new_vma);
2988                 vma_link(mm, new_vma, prev, rb_link, rb_parent);
2989                 *need_rmap_locks = false;
2990         }
2991         return new_vma;
2992
2993 out_free_mempol:
2994         mpol_put(vma_policy(new_vma));
2995 out_free_vma:
2996         kmem_cache_free(vm_area_cachep, new_vma);
2997 out:
2998         return NULL;
2999 }
3000
3001 /*
3002  * Return true if the calling process may expand its vm space by the passed
3003  * number of pages
3004  */
3005 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3006 {
3007         if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3008                 return false;
3009
3010         if (is_data_mapping(flags) &&
3011             mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3012                 if (ignore_rlimit_data)
3013                         pr_warn_once("%s (%d): VmData %lu exceed data ulimit "
3014                                      "%lu. Will be forbidden soon.\n",
3015                                      current->comm, current->pid,
3016                                      (mm->data_vm + npages) << PAGE_SHIFT,
3017                                      rlimit(RLIMIT_DATA));
3018                 else
3019                         return false;
3020         }
3021
3022         return true;
3023 }
3024
3025 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3026 {
3027         mm->total_vm += npages;
3028
3029         if (is_exec_mapping(flags))
3030                 mm->exec_vm += npages;
3031         else if (is_stack_mapping(flags))
3032                 mm->stack_vm += npages;
3033         else if (is_data_mapping(flags))
3034                 mm->data_vm += npages;
3035 }
3036
3037 static int special_mapping_fault(struct vm_area_struct *vma,
3038                                  struct vm_fault *vmf);
3039
3040 /*
3041  * Having a close hook prevents vma merging regardless of flags.
3042  */
3043 static void special_mapping_close(struct vm_area_struct *vma)
3044 {
3045 }
3046
3047 static const char *special_mapping_name(struct vm_area_struct *vma)
3048 {
3049         return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3050 }
3051
3052 static const struct vm_operations_struct special_mapping_vmops = {
3053         .close = special_mapping_close,
3054         .fault = special_mapping_fault,
3055         .name = special_mapping_name,
3056 };
3057
3058 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3059         .close = special_mapping_close,
3060         .fault = special_mapping_fault,
3061 };
3062
3063 static int special_mapping_fault(struct vm_area_struct *vma,
3064                                 struct vm_fault *vmf)
3065 {
3066         pgoff_t pgoff;
3067         struct page **pages;
3068
3069         if (vma->vm_ops == &legacy_special_mapping_vmops)
3070                 pages = vma->vm_private_data;
3071         else
3072                 pages = ((struct vm_special_mapping *)vma->vm_private_data)->
3073                         pages;
3074
3075         for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3076                 pgoff--;
3077
3078         if (*pages) {
3079                 struct page *page = *pages;
3080                 get_page(page);
3081                 vmf->page = page;
3082                 return 0;
3083         }
3084
3085         return VM_FAULT_SIGBUS;
3086 }
3087
3088 static struct vm_area_struct *__install_special_mapping(
3089         struct mm_struct *mm,
3090         unsigned long addr, unsigned long len,
3091         unsigned long vm_flags, void *priv,
3092         const struct vm_operations_struct *ops)
3093 {
3094         int ret;
3095         struct vm_area_struct *vma;
3096
3097         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
3098         if (unlikely(vma == NULL))
3099                 return ERR_PTR(-ENOMEM);
3100
3101         INIT_LIST_HEAD(&vma->anon_vma_chain);
3102         vma->vm_mm = mm;
3103         vma->vm_start = addr;
3104         vma->vm_end = addr + len;
3105
3106         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3107         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3108
3109         vma->vm_ops = ops;
3110         vma->vm_private_data = priv;
3111
3112         ret = insert_vm_struct(mm, vma);
3113         if (ret)
3114                 goto out;
3115
3116         vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3117
3118         perf_event_mmap(vma);
3119
3120         return vma;
3121
3122 out:
3123         kmem_cache_free(vm_area_cachep, vma);
3124         return ERR_PTR(ret);
3125 }
3126
3127 /*
3128  * Called with mm->mmap_sem held for writing.
3129  * Insert a new vma covering the given region, with the given flags.
3130  * Its pages are supplied by the given array of struct page *.
3131  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3132  * The region past the last page supplied will always produce SIGBUS.
3133  * The array pointer and the pages it points to are assumed to stay alive
3134  * for as long as this mapping might exist.
3135  */
3136 struct vm_area_struct *_install_special_mapping(
3137         struct mm_struct *mm,
3138         unsigned long addr, unsigned long len,
3139         unsigned long vm_flags, const struct vm_special_mapping *spec)
3140 {
3141         return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3142                                         &special_mapping_vmops);
3143 }
3144
3145 int install_special_mapping(struct mm_struct *mm,
3146                             unsigned long addr, unsigned long len,
3147                             unsigned long vm_flags, struct page **pages)
3148 {
3149         struct vm_area_struct *vma = __install_special_mapping(
3150                 mm, addr, len, vm_flags, (void *)pages,
3151                 &legacy_special_mapping_vmops);
3152
3153         return PTR_ERR_OR_ZERO(vma);
3154 }
3155
3156 static DEFINE_MUTEX(mm_all_locks_mutex);
3157
3158 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3159 {
3160         if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3161                 /*
3162                  * The LSB of head.next can't change from under us
3163                  * because we hold the mm_all_locks_mutex.
3164                  */
3165                 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
3166                 /*
3167                  * We can safely modify head.next after taking the
3168                  * anon_vma->root->rwsem. If some other vma in this mm shares
3169                  * the same anon_vma we won't take it again.
3170                  *
3171                  * No need of atomic instructions here, head.next
3172                  * can't change from under us thanks to the
3173                  * anon_vma->root->rwsem.
3174                  */
3175                 if (__test_and_set_bit(0, (unsigned long *)
3176                                        &anon_vma->root->rb_root.rb_node))
3177                         BUG();
3178         }
3179 }
3180
3181 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3182 {
3183         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3184                 /*
3185                  * AS_MM_ALL_LOCKS can't change from under us because
3186                  * we hold the mm_all_locks_mutex.
3187                  *
3188                  * Operations on ->flags have to be atomic because
3189                  * even if AS_MM_ALL_LOCKS is stable thanks to the
3190                  * mm_all_locks_mutex, there may be other cpus
3191                  * changing other bitflags in parallel to us.
3192                  */
3193                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3194                         BUG();
3195                 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem);
3196         }
3197 }
3198
3199 /*
3200  * This operation locks against the VM for all pte/vma/mm related
3201  * operations that could ever happen on a certain mm. This includes
3202  * vmtruncate, try_to_unmap, and all page faults.
3203  *
3204  * The caller must take the mmap_sem in write mode before calling
3205  * mm_take_all_locks(). The caller isn't allowed to release the
3206  * mmap_sem until mm_drop_all_locks() returns.
3207  *
3208  * mmap_sem in write mode is required in order to block all operations
3209  * that could modify pagetables and free pages without need of
3210  * altering the vma layout. It's also needed in write mode to avoid new
3211  * anon_vmas to be associated with existing vmas.
3212  *
3213  * A single task can't take more than one mm_take_all_locks() in a row
3214  * or it would deadlock.
3215  *
3216  * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3217  * mapping->flags avoid to take the same lock twice, if more than one
3218  * vma in this mm is backed by the same anon_vma or address_space.
3219  *
3220  * We take locks in following order, accordingly to comment at beginning
3221  * of mm/rmap.c:
3222  *   - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3223  *     hugetlb mapping);
3224  *   - all i_mmap_rwsem locks;
3225  *   - all anon_vma->rwseml
3226  *
3227  * We can take all locks within these types randomly because the VM code
3228  * doesn't nest them and we protected from parallel mm_take_all_locks() by
3229  * mm_all_locks_mutex.
3230  *
3231  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3232  * that may have to take thousand of locks.
3233  *
3234  * mm_take_all_locks() can fail if it's interrupted by signals.
3235  */
3236 int mm_take_all_locks(struct mm_struct *mm)
3237 {
3238         struct vm_area_struct *vma;
3239         struct anon_vma_chain *avc;
3240
3241         BUG_ON(down_read_trylock(&mm->mmap_sem));
3242
3243         mutex_lock(&mm_all_locks_mutex);
3244
3245         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3246                 if (signal_pending(current))
3247                         goto out_unlock;
3248                 if (vma->vm_file && vma->vm_file->f_mapping &&
3249                                 is_vm_hugetlb_page(vma))
3250                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
3251         }
3252
3253         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3254                 if (signal_pending(current))
3255                         goto out_unlock;
3256                 if (vma->vm_file && vma->vm_file->f_mapping &&
3257                                 !is_vm_hugetlb_page(vma))
3258                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
3259         }
3260
3261         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3262                 if (signal_pending(current))
3263                         goto out_unlock;
3264                 if (vma->anon_vma)
3265                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3266                                 vm_lock_anon_vma(mm, avc->anon_vma);
3267         }
3268
3269         return 0;
3270
3271 out_unlock:
3272         mm_drop_all_locks(mm);
3273         return -EINTR;
3274 }
3275
3276 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3277 {
3278         if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3279                 /*
3280                  * The LSB of head.next can't change to 0 from under
3281                  * us because we hold the mm_all_locks_mutex.
3282                  *
3283                  * We must however clear the bitflag before unlocking
3284                  * the vma so the users using the anon_vma->rb_root will
3285                  * never see our bitflag.
3286                  *
3287                  * No need of atomic instructions here, head.next
3288                  * can't change from under us until we release the
3289                  * anon_vma->root->rwsem.
3290                  */
3291                 if (!__test_and_clear_bit(0, (unsigned long *)
3292                                           &anon_vma->root->rb_root.rb_node))
3293                         BUG();
3294                 anon_vma_unlock_write(anon_vma);
3295         }
3296 }
3297
3298 static void vm_unlock_mapping(struct address_space *mapping)
3299 {
3300         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3301                 /*
3302                  * AS_MM_ALL_LOCKS can't change to 0 from under us
3303                  * because we hold the mm_all_locks_mutex.
3304                  */
3305                 i_mmap_unlock_write(mapping);
3306                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3307                                         &mapping->flags))
3308                         BUG();
3309         }
3310 }
3311
3312 /*
3313  * The mmap_sem cannot be released by the caller until
3314  * mm_drop_all_locks() returns.
3315  */
3316 void mm_drop_all_locks(struct mm_struct *mm)
3317 {
3318         struct vm_area_struct *vma;
3319         struct anon_vma_chain *avc;
3320
3321         BUG_ON(down_read_trylock(&mm->mmap_sem));
3322         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3323
3324         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3325                 if (vma->anon_vma)
3326                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3327                                 vm_unlock_anon_vma(avc->anon_vma);
3328                 if (vma->vm_file && vma->vm_file->f_mapping)
3329                         vm_unlock_mapping(vma->vm_file->f_mapping);
3330         }
3331
3332         mutex_unlock(&mm_all_locks_mutex);
3333 }
3334
3335 /*
3336  * initialise the VMA slab
3337  */
3338 void __init mmap_init(void)
3339 {
3340         int ret;
3341
3342         ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3343         VM_BUG_ON(ret);
3344 }
3345
3346 /*
3347  * Initialise sysctl_user_reserve_kbytes.
3348  *
3349  * This is intended to prevent a user from starting a single memory hogging
3350  * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3351  * mode.
3352  *
3353  * The default value is min(3% of free memory, 128MB)
3354  * 128MB is enough to recover with sshd/login, bash, and top/kill.
3355  */
3356 static int init_user_reserve(void)
3357 {
3358         unsigned long free_kbytes;
3359
3360         free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3361
3362         sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3363         return 0;
3364 }
3365 subsys_initcall(init_user_reserve);
3366
3367 /*
3368  * Initialise sysctl_admin_reserve_kbytes.
3369  *
3370  * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3371  * to log in and kill a memory hogging process.
3372  *
3373  * Systems with more than 256MB will reserve 8MB, enough to recover
3374  * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3375  * only reserve 3% of free pages by default.
3376  */
3377 static int init_admin_reserve(void)
3378 {
3379         unsigned long free_kbytes;
3380
3381         free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3382
3383         sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3384         return 0;
3385 }
3386 subsys_initcall(init_admin_reserve);
3387
3388 /*
3389  * Reinititalise user and admin reserves if memory is added or removed.
3390  *
3391  * The default user reserve max is 128MB, and the default max for the
3392  * admin reserve is 8MB. These are usually, but not always, enough to
3393  * enable recovery from a memory hogging process using login/sshd, a shell,
3394  * and tools like top. It may make sense to increase or even disable the
3395  * reserve depending on the existence of swap or variations in the recovery
3396  * tools. So, the admin may have changed them.
3397  *
3398  * If memory is added and the reserves have been eliminated or increased above
3399  * the default max, then we'll trust the admin.
3400  *
3401  * If memory is removed and there isn't enough free memory, then we
3402  * need to reset the reserves.
3403  *
3404  * Otherwise keep the reserve set by the admin.
3405  */
3406 static int reserve_mem_notifier(struct notifier_block *nb,
3407                              unsigned long action, void *data)
3408 {
3409         unsigned long tmp, free_kbytes;
3410
3411         switch (action) {
3412         case MEM_ONLINE:
3413                 /* Default max is 128MB. Leave alone if modified by operator. */
3414                 tmp = sysctl_user_reserve_kbytes;
3415                 if (0 < tmp && tmp < (1UL << 17))
3416                         init_user_reserve();
3417
3418                 /* Default max is 8MB.  Leave alone if modified by operator. */
3419                 tmp = sysctl_admin_reserve_kbytes;
3420                 if (0 < tmp && tmp < (1UL << 13))
3421                         init_admin_reserve();
3422
3423                 break;
3424         case MEM_OFFLINE:
3425                 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3426
3427                 if (sysctl_user_reserve_kbytes > free_kbytes) {
3428                         init_user_reserve();
3429                         pr_info("vm.user_reserve_kbytes reset to %lu\n",
3430                                 sysctl_user_reserve_kbytes);
3431                 }
3432
3433                 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3434                         init_admin_reserve();
3435                         pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3436                                 sysctl_admin_reserve_kbytes);
3437                 }
3438                 break;
3439         default:
3440                 break;
3441         }
3442         return NOTIFY_OK;
3443 }
3444
3445 static struct notifier_block reserve_mem_nb = {
3446         .notifier_call = reserve_mem_notifier,
3447 };
3448
3449 static int __meminit init_reserve_notifier(void)
3450 {
3451         if (register_hotmemory_notifier(&reserve_mem_nb))
3452                 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3453
3454         return 0;
3455 }
3456 subsys_initcall(init_reserve_notifier);