Merge tag 'f2fs-for-4.19' of git://git.kernel.org/pub/scm/linux/kernel/git/jaegeuk...
[platform/kernel/linux-rpi.git] / mm / shmem.c
1 /*
2  * Resizable virtual memory filesystem for Linux.
3  *
4  * Copyright (C) 2000 Linus Torvalds.
5  *               2000 Transmeta Corp.
6  *               2000-2001 Christoph Rohland
7  *               2000-2001 SAP AG
8  *               2002 Red Hat Inc.
9  * Copyright (C) 2002-2011 Hugh Dickins.
10  * Copyright (C) 2011 Google Inc.
11  * Copyright (C) 2002-2005 VERITAS Software Corporation.
12  * Copyright (C) 2004 Andi Kleen, SuSE Labs
13  *
14  * Extended attribute support for tmpfs:
15  * Copyright (c) 2004, Luke Kenneth Casson Leighton <lkcl@lkcl.net>
16  * Copyright (c) 2004 Red Hat, Inc., James Morris <jmorris@redhat.com>
17  *
18  * tiny-shmem:
19  * Copyright (c) 2004, 2008 Matt Mackall <mpm@selenic.com>
20  *
21  * This file is released under the GPL.
22  */
23
24 #include <linux/fs.h>
25 #include <linux/init.h>
26 #include <linux/vfs.h>
27 #include <linux/mount.h>
28 #include <linux/ramfs.h>
29 #include <linux/pagemap.h>
30 #include <linux/file.h>
31 #include <linux/mm.h>
32 #include <linux/random.h>
33 #include <linux/sched/signal.h>
34 #include <linux/export.h>
35 #include <linux/swap.h>
36 #include <linux/uio.h>
37 #include <linux/khugepaged.h>
38 #include <linux/hugetlb.h>
39
40 #include <asm/tlbflush.h> /* for arch/microblaze update_mmu_cache() */
41
42 static struct vfsmount *shm_mnt;
43
44 #ifdef CONFIG_SHMEM
45 /*
46  * This virtual memory filesystem is heavily based on the ramfs. It
47  * extends ramfs by the ability to use swap and honor resource limits
48  * which makes it a completely usable filesystem.
49  */
50
51 #include <linux/xattr.h>
52 #include <linux/exportfs.h>
53 #include <linux/posix_acl.h>
54 #include <linux/posix_acl_xattr.h>
55 #include <linux/mman.h>
56 #include <linux/string.h>
57 #include <linux/slab.h>
58 #include <linux/backing-dev.h>
59 #include <linux/shmem_fs.h>
60 #include <linux/writeback.h>
61 #include <linux/blkdev.h>
62 #include <linux/pagevec.h>
63 #include <linux/percpu_counter.h>
64 #include <linux/falloc.h>
65 #include <linux/splice.h>
66 #include <linux/security.h>
67 #include <linux/swapops.h>
68 #include <linux/mempolicy.h>
69 #include <linux/namei.h>
70 #include <linux/ctype.h>
71 #include <linux/migrate.h>
72 #include <linux/highmem.h>
73 #include <linux/seq_file.h>
74 #include <linux/magic.h>
75 #include <linux/syscalls.h>
76 #include <linux/fcntl.h>
77 #include <uapi/linux/memfd.h>
78 #include <linux/userfaultfd_k.h>
79 #include <linux/rmap.h>
80 #include <linux/uuid.h>
81
82 #include <linux/uaccess.h>
83 #include <asm/pgtable.h>
84
85 #include "internal.h"
86
87 #define BLOCKS_PER_PAGE  (PAGE_SIZE/512)
88 #define VM_ACCT(size)    (PAGE_ALIGN(size) >> PAGE_SHIFT)
89
90 /* Pretend that each entry is of this size in directory's i_size */
91 #define BOGO_DIRENT_SIZE 20
92
93 /* Symlink up to this size is kmalloc'ed instead of using a swappable page */
94 #define SHORT_SYMLINK_LEN 128
95
96 /*
97  * shmem_fallocate communicates with shmem_fault or shmem_writepage via
98  * inode->i_private (with i_mutex making sure that it has only one user at
99  * a time): we would prefer not to enlarge the shmem inode just for that.
100  */
101 struct shmem_falloc {
102         wait_queue_head_t *waitq; /* faults into hole wait for punch to end */
103         pgoff_t start;          /* start of range currently being fallocated */
104         pgoff_t next;           /* the next page offset to be fallocated */
105         pgoff_t nr_falloced;    /* how many new pages have been fallocated */
106         pgoff_t nr_unswapped;   /* how often writepage refused to swap out */
107 };
108
109 #ifdef CONFIG_TMPFS
110 static unsigned long shmem_default_max_blocks(void)
111 {
112         return totalram_pages / 2;
113 }
114
115 static unsigned long shmem_default_max_inodes(void)
116 {
117         return min(totalram_pages - totalhigh_pages, totalram_pages / 2);
118 }
119 #endif
120
121 static bool shmem_should_replace_page(struct page *page, gfp_t gfp);
122 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
123                                 struct shmem_inode_info *info, pgoff_t index);
124 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
125                 struct page **pagep, enum sgp_type sgp,
126                 gfp_t gfp, struct vm_area_struct *vma,
127                 struct vm_fault *vmf, int *fault_type);
128
129 int shmem_getpage(struct inode *inode, pgoff_t index,
130                 struct page **pagep, enum sgp_type sgp)
131 {
132         return shmem_getpage_gfp(inode, index, pagep, sgp,
133                 mapping_gfp_mask(inode->i_mapping), NULL, NULL, NULL);
134 }
135
136 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
137 {
138         return sb->s_fs_info;
139 }
140
141 /*
142  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
143  * for shared memory and for shared anonymous (/dev/zero) mappings
144  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
145  * consistent with the pre-accounting of private mappings ...
146  */
147 static inline int shmem_acct_size(unsigned long flags, loff_t size)
148 {
149         return (flags & VM_NORESERVE) ?
150                 0 : security_vm_enough_memory_mm(current->mm, VM_ACCT(size));
151 }
152
153 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
154 {
155         if (!(flags & VM_NORESERVE))
156                 vm_unacct_memory(VM_ACCT(size));
157 }
158
159 static inline int shmem_reacct_size(unsigned long flags,
160                 loff_t oldsize, loff_t newsize)
161 {
162         if (!(flags & VM_NORESERVE)) {
163                 if (VM_ACCT(newsize) > VM_ACCT(oldsize))
164                         return security_vm_enough_memory_mm(current->mm,
165                                         VM_ACCT(newsize) - VM_ACCT(oldsize));
166                 else if (VM_ACCT(newsize) < VM_ACCT(oldsize))
167                         vm_unacct_memory(VM_ACCT(oldsize) - VM_ACCT(newsize));
168         }
169         return 0;
170 }
171
172 /*
173  * ... whereas tmpfs objects are accounted incrementally as
174  * pages are allocated, in order to allow large sparse files.
175  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
176  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
177  */
178 static inline int shmem_acct_block(unsigned long flags, long pages)
179 {
180         if (!(flags & VM_NORESERVE))
181                 return 0;
182
183         return security_vm_enough_memory_mm(current->mm,
184                         pages * VM_ACCT(PAGE_SIZE));
185 }
186
187 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
188 {
189         if (flags & VM_NORESERVE)
190                 vm_unacct_memory(pages * VM_ACCT(PAGE_SIZE));
191 }
192
193 static inline bool shmem_inode_acct_block(struct inode *inode, long pages)
194 {
195         struct shmem_inode_info *info = SHMEM_I(inode);
196         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
197
198         if (shmem_acct_block(info->flags, pages))
199                 return false;
200
201         if (sbinfo->max_blocks) {
202                 if (percpu_counter_compare(&sbinfo->used_blocks,
203                                            sbinfo->max_blocks - pages) > 0)
204                         goto unacct;
205                 percpu_counter_add(&sbinfo->used_blocks, pages);
206         }
207
208         return true;
209
210 unacct:
211         shmem_unacct_blocks(info->flags, pages);
212         return false;
213 }
214
215 static inline void shmem_inode_unacct_blocks(struct inode *inode, long pages)
216 {
217         struct shmem_inode_info *info = SHMEM_I(inode);
218         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
219
220         if (sbinfo->max_blocks)
221                 percpu_counter_sub(&sbinfo->used_blocks, pages);
222         shmem_unacct_blocks(info->flags, pages);
223 }
224
225 static const struct super_operations shmem_ops;
226 static const struct address_space_operations shmem_aops;
227 static const struct file_operations shmem_file_operations;
228 static const struct inode_operations shmem_inode_operations;
229 static const struct inode_operations shmem_dir_inode_operations;
230 static const struct inode_operations shmem_special_inode_operations;
231 static const struct vm_operations_struct shmem_vm_ops;
232 static struct file_system_type shmem_fs_type;
233
234 bool vma_is_shmem(struct vm_area_struct *vma)
235 {
236         return vma->vm_ops == &shmem_vm_ops;
237 }
238
239 static LIST_HEAD(shmem_swaplist);
240 static DEFINE_MUTEX(shmem_swaplist_mutex);
241
242 static int shmem_reserve_inode(struct super_block *sb)
243 {
244         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
245         if (sbinfo->max_inodes) {
246                 spin_lock(&sbinfo->stat_lock);
247                 if (!sbinfo->free_inodes) {
248                         spin_unlock(&sbinfo->stat_lock);
249                         return -ENOSPC;
250                 }
251                 sbinfo->free_inodes--;
252                 spin_unlock(&sbinfo->stat_lock);
253         }
254         return 0;
255 }
256
257 static void shmem_free_inode(struct super_block *sb)
258 {
259         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
260         if (sbinfo->max_inodes) {
261                 spin_lock(&sbinfo->stat_lock);
262                 sbinfo->free_inodes++;
263                 spin_unlock(&sbinfo->stat_lock);
264         }
265 }
266
267 /**
268  * shmem_recalc_inode - recalculate the block usage of an inode
269  * @inode: inode to recalc
270  *
271  * We have to calculate the free blocks since the mm can drop
272  * undirtied hole pages behind our back.
273  *
274  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
275  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
276  *
277  * It has to be called with the spinlock held.
278  */
279 static void shmem_recalc_inode(struct inode *inode)
280 {
281         struct shmem_inode_info *info = SHMEM_I(inode);
282         long freed;
283
284         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
285         if (freed > 0) {
286                 info->alloced -= freed;
287                 inode->i_blocks -= freed * BLOCKS_PER_PAGE;
288                 shmem_inode_unacct_blocks(inode, freed);
289         }
290 }
291
292 bool shmem_charge(struct inode *inode, long pages)
293 {
294         struct shmem_inode_info *info = SHMEM_I(inode);
295         unsigned long flags;
296
297         if (!shmem_inode_acct_block(inode, pages))
298                 return false;
299
300         spin_lock_irqsave(&info->lock, flags);
301         info->alloced += pages;
302         inode->i_blocks += pages * BLOCKS_PER_PAGE;
303         shmem_recalc_inode(inode);
304         spin_unlock_irqrestore(&info->lock, flags);
305         inode->i_mapping->nrpages += pages;
306
307         return true;
308 }
309
310 void shmem_uncharge(struct inode *inode, long pages)
311 {
312         struct shmem_inode_info *info = SHMEM_I(inode);
313         unsigned long flags;
314
315         spin_lock_irqsave(&info->lock, flags);
316         info->alloced -= pages;
317         inode->i_blocks -= pages * BLOCKS_PER_PAGE;
318         shmem_recalc_inode(inode);
319         spin_unlock_irqrestore(&info->lock, flags);
320
321         shmem_inode_unacct_blocks(inode, pages);
322 }
323
324 /*
325  * Replace item expected in radix tree by a new item, while holding tree lock.
326  */
327 static int shmem_radix_tree_replace(struct address_space *mapping,
328                         pgoff_t index, void *expected, void *replacement)
329 {
330         struct radix_tree_node *node;
331         void __rcu **pslot;
332         void *item;
333
334         VM_BUG_ON(!expected);
335         VM_BUG_ON(!replacement);
336         item = __radix_tree_lookup(&mapping->i_pages, index, &node, &pslot);
337         if (!item)
338                 return -ENOENT;
339         if (item != expected)
340                 return -ENOENT;
341         __radix_tree_replace(&mapping->i_pages, node, pslot,
342                              replacement, NULL);
343         return 0;
344 }
345
346 /*
347  * Sometimes, before we decide whether to proceed or to fail, we must check
348  * that an entry was not already brought back from swap by a racing thread.
349  *
350  * Checking page is not enough: by the time a SwapCache page is locked, it
351  * might be reused, and again be SwapCache, using the same swap as before.
352  */
353 static bool shmem_confirm_swap(struct address_space *mapping,
354                                pgoff_t index, swp_entry_t swap)
355 {
356         void *item;
357
358         rcu_read_lock();
359         item = radix_tree_lookup(&mapping->i_pages, index);
360         rcu_read_unlock();
361         return item == swp_to_radix_entry(swap);
362 }
363
364 /*
365  * Definitions for "huge tmpfs": tmpfs mounted with the huge= option
366  *
367  * SHMEM_HUGE_NEVER:
368  *      disables huge pages for the mount;
369  * SHMEM_HUGE_ALWAYS:
370  *      enables huge pages for the mount;
371  * SHMEM_HUGE_WITHIN_SIZE:
372  *      only allocate huge pages if the page will be fully within i_size,
373  *      also respect fadvise()/madvise() hints;
374  * SHMEM_HUGE_ADVISE:
375  *      only allocate huge pages if requested with fadvise()/madvise();
376  */
377
378 #define SHMEM_HUGE_NEVER        0
379 #define SHMEM_HUGE_ALWAYS       1
380 #define SHMEM_HUGE_WITHIN_SIZE  2
381 #define SHMEM_HUGE_ADVISE       3
382
383 /*
384  * Special values.
385  * Only can be set via /sys/kernel/mm/transparent_hugepage/shmem_enabled:
386  *
387  * SHMEM_HUGE_DENY:
388  *      disables huge on shm_mnt and all mounts, for emergency use;
389  * SHMEM_HUGE_FORCE:
390  *      enables huge on shm_mnt and all mounts, w/o needing option, for testing;
391  *
392  */
393 #define SHMEM_HUGE_DENY         (-1)
394 #define SHMEM_HUGE_FORCE        (-2)
395
396 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
397 /* ifdef here to avoid bloating shmem.o when not necessary */
398
399 static int shmem_huge __read_mostly;
400
401 #if defined(CONFIG_SYSFS) || defined(CONFIG_TMPFS)
402 static int shmem_parse_huge(const char *str)
403 {
404         if (!strcmp(str, "never"))
405                 return SHMEM_HUGE_NEVER;
406         if (!strcmp(str, "always"))
407                 return SHMEM_HUGE_ALWAYS;
408         if (!strcmp(str, "within_size"))
409                 return SHMEM_HUGE_WITHIN_SIZE;
410         if (!strcmp(str, "advise"))
411                 return SHMEM_HUGE_ADVISE;
412         if (!strcmp(str, "deny"))
413                 return SHMEM_HUGE_DENY;
414         if (!strcmp(str, "force"))
415                 return SHMEM_HUGE_FORCE;
416         return -EINVAL;
417 }
418
419 static const char *shmem_format_huge(int huge)
420 {
421         switch (huge) {
422         case SHMEM_HUGE_NEVER:
423                 return "never";
424         case SHMEM_HUGE_ALWAYS:
425                 return "always";
426         case SHMEM_HUGE_WITHIN_SIZE:
427                 return "within_size";
428         case SHMEM_HUGE_ADVISE:
429                 return "advise";
430         case SHMEM_HUGE_DENY:
431                 return "deny";
432         case SHMEM_HUGE_FORCE:
433                 return "force";
434         default:
435                 VM_BUG_ON(1);
436                 return "bad_val";
437         }
438 }
439 #endif
440
441 static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
442                 struct shrink_control *sc, unsigned long nr_to_split)
443 {
444         LIST_HEAD(list), *pos, *next;
445         LIST_HEAD(to_remove);
446         struct inode *inode;
447         struct shmem_inode_info *info;
448         struct page *page;
449         unsigned long batch = sc ? sc->nr_to_scan : 128;
450         int removed = 0, split = 0;
451
452         if (list_empty(&sbinfo->shrinklist))
453                 return SHRINK_STOP;
454
455         spin_lock(&sbinfo->shrinklist_lock);
456         list_for_each_safe(pos, next, &sbinfo->shrinklist) {
457                 info = list_entry(pos, struct shmem_inode_info, shrinklist);
458
459                 /* pin the inode */
460                 inode = igrab(&info->vfs_inode);
461
462                 /* inode is about to be evicted */
463                 if (!inode) {
464                         list_del_init(&info->shrinklist);
465                         removed++;
466                         goto next;
467                 }
468
469                 /* Check if there's anything to gain */
470                 if (round_up(inode->i_size, PAGE_SIZE) ==
471                                 round_up(inode->i_size, HPAGE_PMD_SIZE)) {
472                         list_move(&info->shrinklist, &to_remove);
473                         removed++;
474                         goto next;
475                 }
476
477                 list_move(&info->shrinklist, &list);
478 next:
479                 if (!--batch)
480                         break;
481         }
482         spin_unlock(&sbinfo->shrinklist_lock);
483
484         list_for_each_safe(pos, next, &to_remove) {
485                 info = list_entry(pos, struct shmem_inode_info, shrinklist);
486                 inode = &info->vfs_inode;
487                 list_del_init(&info->shrinklist);
488                 iput(inode);
489         }
490
491         list_for_each_safe(pos, next, &list) {
492                 int ret;
493
494                 info = list_entry(pos, struct shmem_inode_info, shrinklist);
495                 inode = &info->vfs_inode;
496
497                 if (nr_to_split && split >= nr_to_split)
498                         goto leave;
499
500                 page = find_get_page(inode->i_mapping,
501                                 (inode->i_size & HPAGE_PMD_MASK) >> PAGE_SHIFT);
502                 if (!page)
503                         goto drop;
504
505                 /* No huge page at the end of the file: nothing to split */
506                 if (!PageTransHuge(page)) {
507                         put_page(page);
508                         goto drop;
509                 }
510
511                 /*
512                  * Leave the inode on the list if we failed to lock
513                  * the page at this time.
514                  *
515                  * Waiting for the lock may lead to deadlock in the
516                  * reclaim path.
517                  */
518                 if (!trylock_page(page)) {
519                         put_page(page);
520                         goto leave;
521                 }
522
523                 ret = split_huge_page(page);
524                 unlock_page(page);
525                 put_page(page);
526
527                 /* If split failed leave the inode on the list */
528                 if (ret)
529                         goto leave;
530
531                 split++;
532 drop:
533                 list_del_init(&info->shrinklist);
534                 removed++;
535 leave:
536                 iput(inode);
537         }
538
539         spin_lock(&sbinfo->shrinklist_lock);
540         list_splice_tail(&list, &sbinfo->shrinklist);
541         sbinfo->shrinklist_len -= removed;
542         spin_unlock(&sbinfo->shrinklist_lock);
543
544         return split;
545 }
546
547 static long shmem_unused_huge_scan(struct super_block *sb,
548                 struct shrink_control *sc)
549 {
550         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
551
552         if (!READ_ONCE(sbinfo->shrinklist_len))
553                 return SHRINK_STOP;
554
555         return shmem_unused_huge_shrink(sbinfo, sc, 0);
556 }
557
558 static long shmem_unused_huge_count(struct super_block *sb,
559                 struct shrink_control *sc)
560 {
561         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
562         return READ_ONCE(sbinfo->shrinklist_len);
563 }
564 #else /* !CONFIG_TRANSPARENT_HUGE_PAGECACHE */
565
566 #define shmem_huge SHMEM_HUGE_DENY
567
568 static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
569                 struct shrink_control *sc, unsigned long nr_to_split)
570 {
571         return 0;
572 }
573 #endif /* CONFIG_TRANSPARENT_HUGE_PAGECACHE */
574
575 static inline bool is_huge_enabled(struct shmem_sb_info *sbinfo)
576 {
577         if (IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE) &&
578             (shmem_huge == SHMEM_HUGE_FORCE || sbinfo->huge) &&
579             shmem_huge != SHMEM_HUGE_DENY)
580                 return true;
581         return false;
582 }
583
584 /*
585  * Like add_to_page_cache_locked, but error if expected item has gone.
586  */
587 static int shmem_add_to_page_cache(struct page *page,
588                                    struct address_space *mapping,
589                                    pgoff_t index, void *expected)
590 {
591         int error, nr = hpage_nr_pages(page);
592
593         VM_BUG_ON_PAGE(PageTail(page), page);
594         VM_BUG_ON_PAGE(index != round_down(index, nr), page);
595         VM_BUG_ON_PAGE(!PageLocked(page), page);
596         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
597         VM_BUG_ON(expected && PageTransHuge(page));
598
599         page_ref_add(page, nr);
600         page->mapping = mapping;
601         page->index = index;
602
603         xa_lock_irq(&mapping->i_pages);
604         if (PageTransHuge(page)) {
605                 void __rcu **results;
606                 pgoff_t idx;
607                 int i;
608
609                 error = 0;
610                 if (radix_tree_gang_lookup_slot(&mapping->i_pages,
611                                         &results, &idx, index, 1) &&
612                                 idx < index + HPAGE_PMD_NR) {
613                         error = -EEXIST;
614                 }
615
616                 if (!error) {
617                         for (i = 0; i < HPAGE_PMD_NR; i++) {
618                                 error = radix_tree_insert(&mapping->i_pages,
619                                                 index + i, page + i);
620                                 VM_BUG_ON(error);
621                         }
622                         count_vm_event(THP_FILE_ALLOC);
623                 }
624         } else if (!expected) {
625                 error = radix_tree_insert(&mapping->i_pages, index, page);
626         } else {
627                 error = shmem_radix_tree_replace(mapping, index, expected,
628                                                                  page);
629         }
630
631         if (!error) {
632                 mapping->nrpages += nr;
633                 if (PageTransHuge(page))
634                         __inc_node_page_state(page, NR_SHMEM_THPS);
635                 __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, nr);
636                 __mod_node_page_state(page_pgdat(page), NR_SHMEM, nr);
637                 xa_unlock_irq(&mapping->i_pages);
638         } else {
639                 page->mapping = NULL;
640                 xa_unlock_irq(&mapping->i_pages);
641                 page_ref_sub(page, nr);
642         }
643         return error;
644 }
645
646 /*
647  * Like delete_from_page_cache, but substitutes swap for page.
648  */
649 static void shmem_delete_from_page_cache(struct page *page, void *radswap)
650 {
651         struct address_space *mapping = page->mapping;
652         int error;
653
654         VM_BUG_ON_PAGE(PageCompound(page), page);
655
656         xa_lock_irq(&mapping->i_pages);
657         error = shmem_radix_tree_replace(mapping, page->index, page, radswap);
658         page->mapping = NULL;
659         mapping->nrpages--;
660         __dec_node_page_state(page, NR_FILE_PAGES);
661         __dec_node_page_state(page, NR_SHMEM);
662         xa_unlock_irq(&mapping->i_pages);
663         put_page(page);
664         BUG_ON(error);
665 }
666
667 /*
668  * Remove swap entry from radix tree, free the swap and its page cache.
669  */
670 static int shmem_free_swap(struct address_space *mapping,
671                            pgoff_t index, void *radswap)
672 {
673         void *old;
674
675         xa_lock_irq(&mapping->i_pages);
676         old = radix_tree_delete_item(&mapping->i_pages, index, radswap);
677         xa_unlock_irq(&mapping->i_pages);
678         if (old != radswap)
679                 return -ENOENT;
680         free_swap_and_cache(radix_to_swp_entry(radswap));
681         return 0;
682 }
683
684 /*
685  * Determine (in bytes) how many of the shmem object's pages mapped by the
686  * given offsets are swapped out.
687  *
688  * This is safe to call without i_mutex or the i_pages lock thanks to RCU,
689  * as long as the inode doesn't go away and racy results are not a problem.
690  */
691 unsigned long shmem_partial_swap_usage(struct address_space *mapping,
692                                                 pgoff_t start, pgoff_t end)
693 {
694         struct radix_tree_iter iter;
695         void __rcu **slot;
696         struct page *page;
697         unsigned long swapped = 0;
698
699         rcu_read_lock();
700
701         radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start) {
702                 if (iter.index >= end)
703                         break;
704
705                 page = radix_tree_deref_slot(slot);
706
707                 if (radix_tree_deref_retry(page)) {
708                         slot = radix_tree_iter_retry(&iter);
709                         continue;
710                 }
711
712                 if (radix_tree_exceptional_entry(page))
713                         swapped++;
714
715                 if (need_resched()) {
716                         slot = radix_tree_iter_resume(slot, &iter);
717                         cond_resched_rcu();
718                 }
719         }
720
721         rcu_read_unlock();
722
723         return swapped << PAGE_SHIFT;
724 }
725
726 /*
727  * Determine (in bytes) how many of the shmem object's pages mapped by the
728  * given vma is swapped out.
729  *
730  * This is safe to call without i_mutex or the i_pages lock thanks to RCU,
731  * as long as the inode doesn't go away and racy results are not a problem.
732  */
733 unsigned long shmem_swap_usage(struct vm_area_struct *vma)
734 {
735         struct inode *inode = file_inode(vma->vm_file);
736         struct shmem_inode_info *info = SHMEM_I(inode);
737         struct address_space *mapping = inode->i_mapping;
738         unsigned long swapped;
739
740         /* Be careful as we don't hold info->lock */
741         swapped = READ_ONCE(info->swapped);
742
743         /*
744          * The easier cases are when the shmem object has nothing in swap, or
745          * the vma maps it whole. Then we can simply use the stats that we
746          * already track.
747          */
748         if (!swapped)
749                 return 0;
750
751         if (!vma->vm_pgoff && vma->vm_end - vma->vm_start >= inode->i_size)
752                 return swapped << PAGE_SHIFT;
753
754         /* Here comes the more involved part */
755         return shmem_partial_swap_usage(mapping,
756                         linear_page_index(vma, vma->vm_start),
757                         linear_page_index(vma, vma->vm_end));
758 }
759
760 /*
761  * SysV IPC SHM_UNLOCK restore Unevictable pages to their evictable lists.
762  */
763 void shmem_unlock_mapping(struct address_space *mapping)
764 {
765         struct pagevec pvec;
766         pgoff_t indices[PAGEVEC_SIZE];
767         pgoff_t index = 0;
768
769         pagevec_init(&pvec);
770         /*
771          * Minor point, but we might as well stop if someone else SHM_LOCKs it.
772          */
773         while (!mapping_unevictable(mapping)) {
774                 /*
775                  * Avoid pagevec_lookup(): find_get_pages() returns 0 as if it
776                  * has finished, if it hits a row of PAGEVEC_SIZE swap entries.
777                  */
778                 pvec.nr = find_get_entries(mapping, index,
779                                            PAGEVEC_SIZE, pvec.pages, indices);
780                 if (!pvec.nr)
781                         break;
782                 index = indices[pvec.nr - 1] + 1;
783                 pagevec_remove_exceptionals(&pvec);
784                 check_move_unevictable_pages(pvec.pages, pvec.nr);
785                 pagevec_release(&pvec);
786                 cond_resched();
787         }
788 }
789
790 /*
791  * Remove range of pages and swap entries from radix tree, and free them.
792  * If !unfalloc, truncate or punch hole; if unfalloc, undo failed fallocate.
793  */
794 static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
795                                                                  bool unfalloc)
796 {
797         struct address_space *mapping = inode->i_mapping;
798         struct shmem_inode_info *info = SHMEM_I(inode);
799         pgoff_t start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
800         pgoff_t end = (lend + 1) >> PAGE_SHIFT;
801         unsigned int partial_start = lstart & (PAGE_SIZE - 1);
802         unsigned int partial_end = (lend + 1) & (PAGE_SIZE - 1);
803         struct pagevec pvec;
804         pgoff_t indices[PAGEVEC_SIZE];
805         long nr_swaps_freed = 0;
806         pgoff_t index;
807         int i;
808
809         if (lend == -1)
810                 end = -1;       /* unsigned, so actually very big */
811
812         pagevec_init(&pvec);
813         index = start;
814         while (index < end) {
815                 pvec.nr = find_get_entries(mapping, index,
816                         min(end - index, (pgoff_t)PAGEVEC_SIZE),
817                         pvec.pages, indices);
818                 if (!pvec.nr)
819                         break;
820                 for (i = 0; i < pagevec_count(&pvec); i++) {
821                         struct page *page = pvec.pages[i];
822
823                         index = indices[i];
824                         if (index >= end)
825                                 break;
826
827                         if (radix_tree_exceptional_entry(page)) {
828                                 if (unfalloc)
829                                         continue;
830                                 nr_swaps_freed += !shmem_free_swap(mapping,
831                                                                 index, page);
832                                 continue;
833                         }
834
835                         VM_BUG_ON_PAGE(page_to_pgoff(page) != index, page);
836
837                         if (!trylock_page(page))
838                                 continue;
839
840                         if (PageTransTail(page)) {
841                                 /* Middle of THP: zero out the page */
842                                 clear_highpage(page);
843                                 unlock_page(page);
844                                 continue;
845                         } else if (PageTransHuge(page)) {
846                                 if (index == round_down(end, HPAGE_PMD_NR)) {
847                                         /*
848                                          * Range ends in the middle of THP:
849                                          * zero out the page
850                                          */
851                                         clear_highpage(page);
852                                         unlock_page(page);
853                                         continue;
854                                 }
855                                 index += HPAGE_PMD_NR - 1;
856                                 i += HPAGE_PMD_NR - 1;
857                         }
858
859                         if (!unfalloc || !PageUptodate(page)) {
860                                 VM_BUG_ON_PAGE(PageTail(page), page);
861                                 if (page_mapping(page) == mapping) {
862                                         VM_BUG_ON_PAGE(PageWriteback(page), page);
863                                         truncate_inode_page(mapping, page);
864                                 }
865                         }
866                         unlock_page(page);
867                 }
868                 pagevec_remove_exceptionals(&pvec);
869                 pagevec_release(&pvec);
870                 cond_resched();
871                 index++;
872         }
873
874         if (partial_start) {
875                 struct page *page = NULL;
876                 shmem_getpage(inode, start - 1, &page, SGP_READ);
877                 if (page) {
878                         unsigned int top = PAGE_SIZE;
879                         if (start > end) {
880                                 top = partial_end;
881                                 partial_end = 0;
882                         }
883                         zero_user_segment(page, partial_start, top);
884                         set_page_dirty(page);
885                         unlock_page(page);
886                         put_page(page);
887                 }
888         }
889         if (partial_end) {
890                 struct page *page = NULL;
891                 shmem_getpage(inode, end, &page, SGP_READ);
892                 if (page) {
893                         zero_user_segment(page, 0, partial_end);
894                         set_page_dirty(page);
895                         unlock_page(page);
896                         put_page(page);
897                 }
898         }
899         if (start >= end)
900                 return;
901
902         index = start;
903         while (index < end) {
904                 cond_resched();
905
906                 pvec.nr = find_get_entries(mapping, index,
907                                 min(end - index, (pgoff_t)PAGEVEC_SIZE),
908                                 pvec.pages, indices);
909                 if (!pvec.nr) {
910                         /* If all gone or hole-punch or unfalloc, we're done */
911                         if (index == start || end != -1)
912                                 break;
913                         /* But if truncating, restart to make sure all gone */
914                         index = start;
915                         continue;
916                 }
917                 for (i = 0; i < pagevec_count(&pvec); i++) {
918                         struct page *page = pvec.pages[i];
919
920                         index = indices[i];
921                         if (index >= end)
922                                 break;
923
924                         if (radix_tree_exceptional_entry(page)) {
925                                 if (unfalloc)
926                                         continue;
927                                 if (shmem_free_swap(mapping, index, page)) {
928                                         /* Swap was replaced by page: retry */
929                                         index--;
930                                         break;
931                                 }
932                                 nr_swaps_freed++;
933                                 continue;
934                         }
935
936                         lock_page(page);
937
938                         if (PageTransTail(page)) {
939                                 /* Middle of THP: zero out the page */
940                                 clear_highpage(page);
941                                 unlock_page(page);
942                                 /*
943                                  * Partial thp truncate due 'start' in middle
944                                  * of THP: don't need to look on these pages
945                                  * again on !pvec.nr restart.
946                                  */
947                                 if (index != round_down(end, HPAGE_PMD_NR))
948                                         start++;
949                                 continue;
950                         } else if (PageTransHuge(page)) {
951                                 if (index == round_down(end, HPAGE_PMD_NR)) {
952                                         /*
953                                          * Range ends in the middle of THP:
954                                          * zero out the page
955                                          */
956                                         clear_highpage(page);
957                                         unlock_page(page);
958                                         continue;
959                                 }
960                                 index += HPAGE_PMD_NR - 1;
961                                 i += HPAGE_PMD_NR - 1;
962                         }
963
964                         if (!unfalloc || !PageUptodate(page)) {
965                                 VM_BUG_ON_PAGE(PageTail(page), page);
966                                 if (page_mapping(page) == mapping) {
967                                         VM_BUG_ON_PAGE(PageWriteback(page), page);
968                                         truncate_inode_page(mapping, page);
969                                 } else {
970                                         /* Page was replaced by swap: retry */
971                                         unlock_page(page);
972                                         index--;
973                                         break;
974                                 }
975                         }
976                         unlock_page(page);
977                 }
978                 pagevec_remove_exceptionals(&pvec);
979                 pagevec_release(&pvec);
980                 index++;
981         }
982
983         spin_lock_irq(&info->lock);
984         info->swapped -= nr_swaps_freed;
985         shmem_recalc_inode(inode);
986         spin_unlock_irq(&info->lock);
987 }
988
989 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
990 {
991         shmem_undo_range(inode, lstart, lend, false);
992         inode->i_ctime = inode->i_mtime = current_time(inode);
993 }
994 EXPORT_SYMBOL_GPL(shmem_truncate_range);
995
996 static int shmem_getattr(const struct path *path, struct kstat *stat,
997                          u32 request_mask, unsigned int query_flags)
998 {
999         struct inode *inode = path->dentry->d_inode;
1000         struct shmem_inode_info *info = SHMEM_I(inode);
1001         struct shmem_sb_info *sb_info = SHMEM_SB(inode->i_sb);
1002
1003         if (info->alloced - info->swapped != inode->i_mapping->nrpages) {
1004                 spin_lock_irq(&info->lock);
1005                 shmem_recalc_inode(inode);
1006                 spin_unlock_irq(&info->lock);
1007         }
1008         generic_fillattr(inode, stat);
1009
1010         if (is_huge_enabled(sb_info))
1011                 stat->blksize = HPAGE_PMD_SIZE;
1012
1013         return 0;
1014 }
1015
1016 static int shmem_setattr(struct dentry *dentry, struct iattr *attr)
1017 {
1018         struct inode *inode = d_inode(dentry);
1019         struct shmem_inode_info *info = SHMEM_I(inode);
1020         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1021         int error;
1022
1023         error = setattr_prepare(dentry, attr);
1024         if (error)
1025                 return error;
1026
1027         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
1028                 loff_t oldsize = inode->i_size;
1029                 loff_t newsize = attr->ia_size;
1030
1031                 /* protected by i_mutex */
1032                 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
1033                     (newsize > oldsize && (info->seals & F_SEAL_GROW)))
1034                         return -EPERM;
1035
1036                 if (newsize != oldsize) {
1037                         error = shmem_reacct_size(SHMEM_I(inode)->flags,
1038                                         oldsize, newsize);
1039                         if (error)
1040                                 return error;
1041                         i_size_write(inode, newsize);
1042                         inode->i_ctime = inode->i_mtime = current_time(inode);
1043                 }
1044                 if (newsize <= oldsize) {
1045                         loff_t holebegin = round_up(newsize, PAGE_SIZE);
1046                         if (oldsize > holebegin)
1047                                 unmap_mapping_range(inode->i_mapping,
1048                                                         holebegin, 0, 1);
1049                         if (info->alloced)
1050                                 shmem_truncate_range(inode,
1051                                                         newsize, (loff_t)-1);
1052                         /* unmap again to remove racily COWed private pages */
1053                         if (oldsize > holebegin)
1054                                 unmap_mapping_range(inode->i_mapping,
1055                                                         holebegin, 0, 1);
1056
1057                         /*
1058                          * Part of the huge page can be beyond i_size: subject
1059                          * to shrink under memory pressure.
1060                          */
1061                         if (IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
1062                                 spin_lock(&sbinfo->shrinklist_lock);
1063                                 /*
1064                                  * _careful to defend against unlocked access to
1065                                  * ->shrink_list in shmem_unused_huge_shrink()
1066                                  */
1067                                 if (list_empty_careful(&info->shrinklist)) {
1068                                         list_add_tail(&info->shrinklist,
1069                                                         &sbinfo->shrinklist);
1070                                         sbinfo->shrinklist_len++;
1071                                 }
1072                                 spin_unlock(&sbinfo->shrinklist_lock);
1073                         }
1074                 }
1075         }
1076
1077         setattr_copy(inode, attr);
1078         if (attr->ia_valid & ATTR_MODE)
1079                 error = posix_acl_chmod(inode, inode->i_mode);
1080         return error;
1081 }
1082
1083 static void shmem_evict_inode(struct inode *inode)
1084 {
1085         struct shmem_inode_info *info = SHMEM_I(inode);
1086         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1087
1088         if (inode->i_mapping->a_ops == &shmem_aops) {
1089                 shmem_unacct_size(info->flags, inode->i_size);
1090                 inode->i_size = 0;
1091                 shmem_truncate_range(inode, 0, (loff_t)-1);
1092                 if (!list_empty(&info->shrinklist)) {
1093                         spin_lock(&sbinfo->shrinklist_lock);
1094                         if (!list_empty(&info->shrinklist)) {
1095                                 list_del_init(&info->shrinklist);
1096                                 sbinfo->shrinklist_len--;
1097                         }
1098                         spin_unlock(&sbinfo->shrinklist_lock);
1099                 }
1100                 if (!list_empty(&info->swaplist)) {
1101                         mutex_lock(&shmem_swaplist_mutex);
1102                         list_del_init(&info->swaplist);
1103                         mutex_unlock(&shmem_swaplist_mutex);
1104                 }
1105         }
1106
1107         simple_xattrs_free(&info->xattrs);
1108         WARN_ON(inode->i_blocks);
1109         shmem_free_inode(inode->i_sb);
1110         clear_inode(inode);
1111 }
1112
1113 static unsigned long find_swap_entry(struct radix_tree_root *root, void *item)
1114 {
1115         struct radix_tree_iter iter;
1116         void __rcu **slot;
1117         unsigned long found = -1;
1118         unsigned int checked = 0;
1119
1120         rcu_read_lock();
1121         radix_tree_for_each_slot(slot, root, &iter, 0) {
1122                 void *entry = radix_tree_deref_slot(slot);
1123
1124                 if (radix_tree_deref_retry(entry)) {
1125                         slot = radix_tree_iter_retry(&iter);
1126                         continue;
1127                 }
1128                 if (entry == item) {
1129                         found = iter.index;
1130                         break;
1131                 }
1132                 checked++;
1133                 if ((checked % 4096) != 0)
1134                         continue;
1135                 slot = radix_tree_iter_resume(slot, &iter);
1136                 cond_resched_rcu();
1137         }
1138
1139         rcu_read_unlock();
1140         return found;
1141 }
1142
1143 /*
1144  * If swap found in inode, free it and move page from swapcache to filecache.
1145  */
1146 static int shmem_unuse_inode(struct shmem_inode_info *info,
1147                              swp_entry_t swap, struct page **pagep)
1148 {
1149         struct address_space *mapping = info->vfs_inode.i_mapping;
1150         void *radswap;
1151         pgoff_t index;
1152         gfp_t gfp;
1153         int error = 0;
1154
1155         radswap = swp_to_radix_entry(swap);
1156         index = find_swap_entry(&mapping->i_pages, radswap);
1157         if (index == -1)
1158                 return -EAGAIN; /* tell shmem_unuse we found nothing */
1159
1160         /*
1161          * Move _head_ to start search for next from here.
1162          * But be careful: shmem_evict_inode checks list_empty without taking
1163          * mutex, and there's an instant in list_move_tail when info->swaplist
1164          * would appear empty, if it were the only one on shmem_swaplist.
1165          */
1166         if (shmem_swaplist.next != &info->swaplist)
1167                 list_move_tail(&shmem_swaplist, &info->swaplist);
1168
1169         gfp = mapping_gfp_mask(mapping);
1170         if (shmem_should_replace_page(*pagep, gfp)) {
1171                 mutex_unlock(&shmem_swaplist_mutex);
1172                 error = shmem_replace_page(pagep, gfp, info, index);
1173                 mutex_lock(&shmem_swaplist_mutex);
1174                 /*
1175                  * We needed to drop mutex to make that restrictive page
1176                  * allocation, but the inode might have been freed while we
1177                  * dropped it: although a racing shmem_evict_inode() cannot
1178                  * complete without emptying the radix_tree, our page lock
1179                  * on this swapcache page is not enough to prevent that -
1180                  * free_swap_and_cache() of our swap entry will only
1181                  * trylock_page(), removing swap from radix_tree whatever.
1182                  *
1183                  * We must not proceed to shmem_add_to_page_cache() if the
1184                  * inode has been freed, but of course we cannot rely on
1185                  * inode or mapping or info to check that.  However, we can
1186                  * safely check if our swap entry is still in use (and here
1187                  * it can't have got reused for another page): if it's still
1188                  * in use, then the inode cannot have been freed yet, and we
1189                  * can safely proceed (if it's no longer in use, that tells
1190                  * nothing about the inode, but we don't need to unuse swap).
1191                  */
1192                 if (!page_swapcount(*pagep))
1193                         error = -ENOENT;
1194         }
1195
1196         /*
1197          * We rely on shmem_swaplist_mutex, not only to protect the swaplist,
1198          * but also to hold up shmem_evict_inode(): so inode cannot be freed
1199          * beneath us (pagelock doesn't help until the page is in pagecache).
1200          */
1201         if (!error)
1202                 error = shmem_add_to_page_cache(*pagep, mapping, index,
1203                                                 radswap);
1204         if (error != -ENOMEM) {
1205                 /*
1206                  * Truncation and eviction use free_swap_and_cache(), which
1207                  * only does trylock page: if we raced, best clean up here.
1208                  */
1209                 delete_from_swap_cache(*pagep);
1210                 set_page_dirty(*pagep);
1211                 if (!error) {
1212                         spin_lock_irq(&info->lock);
1213                         info->swapped--;
1214                         spin_unlock_irq(&info->lock);
1215                         swap_free(swap);
1216                 }
1217         }
1218         return error;
1219 }
1220
1221 /*
1222  * Search through swapped inodes to find and replace swap by page.
1223  */
1224 int shmem_unuse(swp_entry_t swap, struct page *page)
1225 {
1226         struct list_head *this, *next;
1227         struct shmem_inode_info *info;
1228         struct mem_cgroup *memcg;
1229         int error = 0;
1230
1231         /*
1232          * There's a faint possibility that swap page was replaced before
1233          * caller locked it: caller will come back later with the right page.
1234          */
1235         if (unlikely(!PageSwapCache(page) || page_private(page) != swap.val))
1236                 goto out;
1237
1238         /*
1239          * Charge page using GFP_KERNEL while we can wait, before taking
1240          * the shmem_swaplist_mutex which might hold up shmem_writepage().
1241          * Charged back to the user (not to caller) when swap account is used.
1242          */
1243         error = mem_cgroup_try_charge_delay(page, current->mm, GFP_KERNEL,
1244                                             &memcg, false);
1245         if (error)
1246                 goto out;
1247         /* No radix_tree_preload: swap entry keeps a place for page in tree */
1248         error = -EAGAIN;
1249
1250         mutex_lock(&shmem_swaplist_mutex);
1251         list_for_each_safe(this, next, &shmem_swaplist) {
1252                 info = list_entry(this, struct shmem_inode_info, swaplist);
1253                 if (info->swapped)
1254                         error = shmem_unuse_inode(info, swap, &page);
1255                 else
1256                         list_del_init(&info->swaplist);
1257                 cond_resched();
1258                 if (error != -EAGAIN)
1259                         break;
1260                 /* found nothing in this: move on to search the next */
1261         }
1262         mutex_unlock(&shmem_swaplist_mutex);
1263
1264         if (error) {
1265                 if (error != -ENOMEM)
1266                         error = 0;
1267                 mem_cgroup_cancel_charge(page, memcg, false);
1268         } else
1269                 mem_cgroup_commit_charge(page, memcg, true, false);
1270 out:
1271         unlock_page(page);
1272         put_page(page);
1273         return error;
1274 }
1275
1276 /*
1277  * Move the page from the page cache to the swap cache.
1278  */
1279 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1280 {
1281         struct shmem_inode_info *info;
1282         struct address_space *mapping;
1283         struct inode *inode;
1284         swp_entry_t swap;
1285         pgoff_t index;
1286
1287         VM_BUG_ON_PAGE(PageCompound(page), page);
1288         BUG_ON(!PageLocked(page));
1289         mapping = page->mapping;
1290         index = page->index;
1291         inode = mapping->host;
1292         info = SHMEM_I(inode);
1293         if (info->flags & VM_LOCKED)
1294                 goto redirty;
1295         if (!total_swap_pages)
1296                 goto redirty;
1297
1298         /*
1299          * Our capabilities prevent regular writeback or sync from ever calling
1300          * shmem_writepage; but a stacking filesystem might use ->writepage of
1301          * its underlying filesystem, in which case tmpfs should write out to
1302          * swap only in response to memory pressure, and not for the writeback
1303          * threads or sync.
1304          */
1305         if (!wbc->for_reclaim) {
1306                 WARN_ON_ONCE(1);        /* Still happens? Tell us about it! */
1307                 goto redirty;
1308         }
1309
1310         /*
1311          * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
1312          * value into swapfile.c, the only way we can correctly account for a
1313          * fallocated page arriving here is now to initialize it and write it.
1314          *
1315          * That's okay for a page already fallocated earlier, but if we have
1316          * not yet completed the fallocation, then (a) we want to keep track
1317          * of this page in case we have to undo it, and (b) it may not be a
1318          * good idea to continue anyway, once we're pushing into swap.  So
1319          * reactivate the page, and let shmem_fallocate() quit when too many.
1320          */
1321         if (!PageUptodate(page)) {
1322                 if (inode->i_private) {
1323                         struct shmem_falloc *shmem_falloc;
1324                         spin_lock(&inode->i_lock);
1325                         shmem_falloc = inode->i_private;
1326                         if (shmem_falloc &&
1327                             !shmem_falloc->waitq &&
1328                             index >= shmem_falloc->start &&
1329                             index < shmem_falloc->next)
1330                                 shmem_falloc->nr_unswapped++;
1331                         else
1332                                 shmem_falloc = NULL;
1333                         spin_unlock(&inode->i_lock);
1334                         if (shmem_falloc)
1335                                 goto redirty;
1336                 }
1337                 clear_highpage(page);
1338                 flush_dcache_page(page);
1339                 SetPageUptodate(page);
1340         }
1341
1342         swap = get_swap_page(page);
1343         if (!swap.val)
1344                 goto redirty;
1345
1346         /*
1347          * Add inode to shmem_unuse()'s list of swapped-out inodes,
1348          * if it's not already there.  Do it now before the page is
1349          * moved to swap cache, when its pagelock no longer protects
1350          * the inode from eviction.  But don't unlock the mutex until
1351          * we've incremented swapped, because shmem_unuse_inode() will
1352          * prune a !swapped inode from the swaplist under this mutex.
1353          */
1354         mutex_lock(&shmem_swaplist_mutex);
1355         if (list_empty(&info->swaplist))
1356                 list_add_tail(&info->swaplist, &shmem_swaplist);
1357
1358         if (add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
1359                 spin_lock_irq(&info->lock);
1360                 shmem_recalc_inode(inode);
1361                 info->swapped++;
1362                 spin_unlock_irq(&info->lock);
1363
1364                 swap_shmem_alloc(swap);
1365                 shmem_delete_from_page_cache(page, swp_to_radix_entry(swap));
1366
1367                 mutex_unlock(&shmem_swaplist_mutex);
1368                 BUG_ON(page_mapped(page));
1369                 swap_writepage(page, wbc);
1370                 return 0;
1371         }
1372
1373         mutex_unlock(&shmem_swaplist_mutex);
1374         put_swap_page(page, swap);
1375 redirty:
1376         set_page_dirty(page);
1377         if (wbc->for_reclaim)
1378                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
1379         unlock_page(page);
1380         return 0;
1381 }
1382
1383 #if defined(CONFIG_NUMA) && defined(CONFIG_TMPFS)
1384 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1385 {
1386         char buffer[64];
1387
1388         if (!mpol || mpol->mode == MPOL_DEFAULT)
1389                 return;         /* show nothing */
1390
1391         mpol_to_str(buffer, sizeof(buffer), mpol);
1392
1393         seq_printf(seq, ",mpol=%s", buffer);
1394 }
1395
1396 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1397 {
1398         struct mempolicy *mpol = NULL;
1399         if (sbinfo->mpol) {
1400                 spin_lock(&sbinfo->stat_lock);  /* prevent replace/use races */
1401                 mpol = sbinfo->mpol;
1402                 mpol_get(mpol);
1403                 spin_unlock(&sbinfo->stat_lock);
1404         }
1405         return mpol;
1406 }
1407 #else /* !CONFIG_NUMA || !CONFIG_TMPFS */
1408 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1409 {
1410 }
1411 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1412 {
1413         return NULL;
1414 }
1415 #endif /* CONFIG_NUMA && CONFIG_TMPFS */
1416 #ifndef CONFIG_NUMA
1417 #define vm_policy vm_private_data
1418 #endif
1419
1420 static void shmem_pseudo_vma_init(struct vm_area_struct *vma,
1421                 struct shmem_inode_info *info, pgoff_t index)
1422 {
1423         /* Create a pseudo vma that just contains the policy */
1424         vma_init(vma, NULL);
1425         /* Bias interleave by inode number to distribute better across nodes */
1426         vma->vm_pgoff = index + info->vfs_inode.i_ino;
1427         vma->vm_policy = mpol_shared_policy_lookup(&info->policy, index);
1428 }
1429
1430 static void shmem_pseudo_vma_destroy(struct vm_area_struct *vma)
1431 {
1432         /* Drop reference taken by mpol_shared_policy_lookup() */
1433         mpol_cond_put(vma->vm_policy);
1434 }
1435
1436 static struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
1437                         struct shmem_inode_info *info, pgoff_t index)
1438 {
1439         struct vm_area_struct pvma;
1440         struct page *page;
1441         struct vm_fault vmf;
1442
1443         shmem_pseudo_vma_init(&pvma, info, index);
1444         vmf.vma = &pvma;
1445         vmf.address = 0;
1446         page = swap_cluster_readahead(swap, gfp, &vmf);
1447         shmem_pseudo_vma_destroy(&pvma);
1448
1449         return page;
1450 }
1451
1452 static struct page *shmem_alloc_hugepage(gfp_t gfp,
1453                 struct shmem_inode_info *info, pgoff_t index)
1454 {
1455         struct vm_area_struct pvma;
1456         struct inode *inode = &info->vfs_inode;
1457         struct address_space *mapping = inode->i_mapping;
1458         pgoff_t idx, hindex;
1459         void __rcu **results;
1460         struct page *page;
1461
1462         if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
1463                 return NULL;
1464
1465         hindex = round_down(index, HPAGE_PMD_NR);
1466         rcu_read_lock();
1467         if (radix_tree_gang_lookup_slot(&mapping->i_pages, &results, &idx,
1468                                 hindex, 1) && idx < hindex + HPAGE_PMD_NR) {
1469                 rcu_read_unlock();
1470                 return NULL;
1471         }
1472         rcu_read_unlock();
1473
1474         shmem_pseudo_vma_init(&pvma, info, hindex);
1475         page = alloc_pages_vma(gfp | __GFP_COMP | __GFP_NORETRY | __GFP_NOWARN,
1476                         HPAGE_PMD_ORDER, &pvma, 0, numa_node_id(), true);
1477         shmem_pseudo_vma_destroy(&pvma);
1478         if (page)
1479                 prep_transhuge_page(page);
1480         return page;
1481 }
1482
1483 static struct page *shmem_alloc_page(gfp_t gfp,
1484                         struct shmem_inode_info *info, pgoff_t index)
1485 {
1486         struct vm_area_struct pvma;
1487         struct page *page;
1488
1489         shmem_pseudo_vma_init(&pvma, info, index);
1490         page = alloc_page_vma(gfp, &pvma, 0);
1491         shmem_pseudo_vma_destroy(&pvma);
1492
1493         return page;
1494 }
1495
1496 static struct page *shmem_alloc_and_acct_page(gfp_t gfp,
1497                 struct inode *inode,
1498                 pgoff_t index, bool huge)
1499 {
1500         struct shmem_inode_info *info = SHMEM_I(inode);
1501         struct page *page;
1502         int nr;
1503         int err = -ENOSPC;
1504
1505         if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
1506                 huge = false;
1507         nr = huge ? HPAGE_PMD_NR : 1;
1508
1509         if (!shmem_inode_acct_block(inode, nr))
1510                 goto failed;
1511
1512         if (huge)
1513                 page = shmem_alloc_hugepage(gfp, info, index);
1514         else
1515                 page = shmem_alloc_page(gfp, info, index);
1516         if (page) {
1517                 __SetPageLocked(page);
1518                 __SetPageSwapBacked(page);
1519                 return page;
1520         }
1521
1522         err = -ENOMEM;
1523         shmem_inode_unacct_blocks(inode, nr);
1524 failed:
1525         return ERR_PTR(err);
1526 }
1527
1528 /*
1529  * When a page is moved from swapcache to shmem filecache (either by the
1530  * usual swapin of shmem_getpage_gfp(), or by the less common swapoff of
1531  * shmem_unuse_inode()), it may have been read in earlier from swap, in
1532  * ignorance of the mapping it belongs to.  If that mapping has special
1533  * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
1534  * we may need to copy to a suitable page before moving to filecache.
1535  *
1536  * In a future release, this may well be extended to respect cpuset and
1537  * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
1538  * but for now it is a simple matter of zone.
1539  */
1540 static bool shmem_should_replace_page(struct page *page, gfp_t gfp)
1541 {
1542         return page_zonenum(page) > gfp_zone(gfp);
1543 }
1544
1545 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
1546                                 struct shmem_inode_info *info, pgoff_t index)
1547 {
1548         struct page *oldpage, *newpage;
1549         struct address_space *swap_mapping;
1550         pgoff_t swap_index;
1551         int error;
1552
1553         oldpage = *pagep;
1554         swap_index = page_private(oldpage);
1555         swap_mapping = page_mapping(oldpage);
1556
1557         /*
1558          * We have arrived here because our zones are constrained, so don't
1559          * limit chance of success by further cpuset and node constraints.
1560          */
1561         gfp &= ~GFP_CONSTRAINT_MASK;
1562         newpage = shmem_alloc_page(gfp, info, index);
1563         if (!newpage)
1564                 return -ENOMEM;
1565
1566         get_page(newpage);
1567         copy_highpage(newpage, oldpage);
1568         flush_dcache_page(newpage);
1569
1570         __SetPageLocked(newpage);
1571         __SetPageSwapBacked(newpage);
1572         SetPageUptodate(newpage);
1573         set_page_private(newpage, swap_index);
1574         SetPageSwapCache(newpage);
1575
1576         /*
1577          * Our caller will very soon move newpage out of swapcache, but it's
1578          * a nice clean interface for us to replace oldpage by newpage there.
1579          */
1580         xa_lock_irq(&swap_mapping->i_pages);
1581         error = shmem_radix_tree_replace(swap_mapping, swap_index, oldpage,
1582                                                                    newpage);
1583         if (!error) {
1584                 __inc_node_page_state(newpage, NR_FILE_PAGES);
1585                 __dec_node_page_state(oldpage, NR_FILE_PAGES);
1586         }
1587         xa_unlock_irq(&swap_mapping->i_pages);
1588
1589         if (unlikely(error)) {
1590                 /*
1591                  * Is this possible?  I think not, now that our callers check
1592                  * both PageSwapCache and page_private after getting page lock;
1593                  * but be defensive.  Reverse old to newpage for clear and free.
1594                  */
1595                 oldpage = newpage;
1596         } else {
1597                 mem_cgroup_migrate(oldpage, newpage);
1598                 lru_cache_add_anon(newpage);
1599                 *pagep = newpage;
1600         }
1601
1602         ClearPageSwapCache(oldpage);
1603         set_page_private(oldpage, 0);
1604
1605         unlock_page(oldpage);
1606         put_page(oldpage);
1607         put_page(oldpage);
1608         return error;
1609 }
1610
1611 /*
1612  * shmem_getpage_gfp - find page in cache, or get from swap, or allocate
1613  *
1614  * If we allocate a new one we do not mark it dirty. That's up to the
1615  * vm. If we swap it in we mark it dirty since we also free the swap
1616  * entry since a page cannot live in both the swap and page cache.
1617  *
1618  * fault_mm and fault_type are only supplied by shmem_fault:
1619  * otherwise they are NULL.
1620  */
1621 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
1622         struct page **pagep, enum sgp_type sgp, gfp_t gfp,
1623         struct vm_area_struct *vma, struct vm_fault *vmf, int *fault_type)
1624 {
1625         struct address_space *mapping = inode->i_mapping;
1626         struct shmem_inode_info *info = SHMEM_I(inode);
1627         struct shmem_sb_info *sbinfo;
1628         struct mm_struct *charge_mm;
1629         struct mem_cgroup *memcg;
1630         struct page *page;
1631         swp_entry_t swap;
1632         enum sgp_type sgp_huge = sgp;
1633         pgoff_t hindex = index;
1634         int error;
1635         int once = 0;
1636         int alloced = 0;
1637
1638         if (index > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
1639                 return -EFBIG;
1640         if (sgp == SGP_NOHUGE || sgp == SGP_HUGE)
1641                 sgp = SGP_CACHE;
1642 repeat:
1643         swap.val = 0;
1644         page = find_lock_entry(mapping, index);
1645         if (radix_tree_exceptional_entry(page)) {
1646                 swap = radix_to_swp_entry(page);
1647                 page = NULL;
1648         }
1649
1650         if (sgp <= SGP_CACHE &&
1651             ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1652                 error = -EINVAL;
1653                 goto unlock;
1654         }
1655
1656         if (page && sgp == SGP_WRITE)
1657                 mark_page_accessed(page);
1658
1659         /* fallocated page? */
1660         if (page && !PageUptodate(page)) {
1661                 if (sgp != SGP_READ)
1662                         goto clear;
1663                 unlock_page(page);
1664                 put_page(page);
1665                 page = NULL;
1666         }
1667         if (page || (sgp == SGP_READ && !swap.val)) {
1668                 *pagep = page;
1669                 return 0;
1670         }
1671
1672         /*
1673          * Fast cache lookup did not find it:
1674          * bring it back from swap or allocate.
1675          */
1676         sbinfo = SHMEM_SB(inode->i_sb);
1677         charge_mm = vma ? vma->vm_mm : current->mm;
1678
1679         if (swap.val) {
1680                 /* Look it up and read it in.. */
1681                 page = lookup_swap_cache(swap, NULL, 0);
1682                 if (!page) {
1683                         /* Or update major stats only when swapin succeeds?? */
1684                         if (fault_type) {
1685                                 *fault_type |= VM_FAULT_MAJOR;
1686                                 count_vm_event(PGMAJFAULT);
1687                                 count_memcg_event_mm(charge_mm, PGMAJFAULT);
1688                         }
1689                         /* Here we actually start the io */
1690                         page = shmem_swapin(swap, gfp, info, index);
1691                         if (!page) {
1692                                 error = -ENOMEM;
1693                                 goto failed;
1694                         }
1695                 }
1696
1697                 /* We have to do this with page locked to prevent races */
1698                 lock_page(page);
1699                 if (!PageSwapCache(page) || page_private(page) != swap.val ||
1700                     !shmem_confirm_swap(mapping, index, swap)) {
1701                         error = -EEXIST;        /* try again */
1702                         goto unlock;
1703                 }
1704                 if (!PageUptodate(page)) {
1705                         error = -EIO;
1706                         goto failed;
1707                 }
1708                 wait_on_page_writeback(page);
1709
1710                 if (shmem_should_replace_page(page, gfp)) {
1711                         error = shmem_replace_page(&page, gfp, info, index);
1712                         if (error)
1713                                 goto failed;
1714                 }
1715
1716                 error = mem_cgroup_try_charge_delay(page, charge_mm, gfp, &memcg,
1717                                 false);
1718                 if (!error) {
1719                         error = shmem_add_to_page_cache(page, mapping, index,
1720                                                 swp_to_radix_entry(swap));
1721                         /*
1722                          * We already confirmed swap under page lock, and make
1723                          * no memory allocation here, so usually no possibility
1724                          * of error; but free_swap_and_cache() only trylocks a
1725                          * page, so it is just possible that the entry has been
1726                          * truncated or holepunched since swap was confirmed.
1727                          * shmem_undo_range() will have done some of the
1728                          * unaccounting, now delete_from_swap_cache() will do
1729                          * the rest.
1730                          * Reset swap.val? No, leave it so "failed" goes back to
1731                          * "repeat": reading a hole and writing should succeed.
1732                          */
1733                         if (error) {
1734                                 mem_cgroup_cancel_charge(page, memcg, false);
1735                                 delete_from_swap_cache(page);
1736                         }
1737                 }
1738                 if (error)
1739                         goto failed;
1740
1741                 mem_cgroup_commit_charge(page, memcg, true, false);
1742
1743                 spin_lock_irq(&info->lock);
1744                 info->swapped--;
1745                 shmem_recalc_inode(inode);
1746                 spin_unlock_irq(&info->lock);
1747
1748                 if (sgp == SGP_WRITE)
1749                         mark_page_accessed(page);
1750
1751                 delete_from_swap_cache(page);
1752                 set_page_dirty(page);
1753                 swap_free(swap);
1754
1755         } else {
1756                 if (vma && userfaultfd_missing(vma)) {
1757                         *fault_type = handle_userfault(vmf, VM_UFFD_MISSING);
1758                         return 0;
1759                 }
1760
1761                 /* shmem_symlink() */
1762                 if (mapping->a_ops != &shmem_aops)
1763                         goto alloc_nohuge;
1764                 if (shmem_huge == SHMEM_HUGE_DENY || sgp_huge == SGP_NOHUGE)
1765                         goto alloc_nohuge;
1766                 if (shmem_huge == SHMEM_HUGE_FORCE)
1767                         goto alloc_huge;
1768                 switch (sbinfo->huge) {
1769                         loff_t i_size;
1770                         pgoff_t off;
1771                 case SHMEM_HUGE_NEVER:
1772                         goto alloc_nohuge;
1773                 case SHMEM_HUGE_WITHIN_SIZE:
1774                         off = round_up(index, HPAGE_PMD_NR);
1775                         i_size = round_up(i_size_read(inode), PAGE_SIZE);
1776                         if (i_size >= HPAGE_PMD_SIZE &&
1777                                         i_size >> PAGE_SHIFT >= off)
1778                                 goto alloc_huge;
1779                         /* fallthrough */
1780                 case SHMEM_HUGE_ADVISE:
1781                         if (sgp_huge == SGP_HUGE)
1782                                 goto alloc_huge;
1783                         /* TODO: implement fadvise() hints */
1784                         goto alloc_nohuge;
1785                 }
1786
1787 alloc_huge:
1788                 page = shmem_alloc_and_acct_page(gfp, inode, index, true);
1789                 if (IS_ERR(page)) {
1790 alloc_nohuge:           page = shmem_alloc_and_acct_page(gfp, inode,
1791                                         index, false);
1792                 }
1793                 if (IS_ERR(page)) {
1794                         int retry = 5;
1795                         error = PTR_ERR(page);
1796                         page = NULL;
1797                         if (error != -ENOSPC)
1798                                 goto failed;
1799                         /*
1800                          * Try to reclaim some spece by splitting a huge page
1801                          * beyond i_size on the filesystem.
1802                          */
1803                         while (retry--) {
1804                                 int ret;
1805                                 ret = shmem_unused_huge_shrink(sbinfo, NULL, 1);
1806                                 if (ret == SHRINK_STOP)
1807                                         break;
1808                                 if (ret)
1809                                         goto alloc_nohuge;
1810                         }
1811                         goto failed;
1812                 }
1813
1814                 if (PageTransHuge(page))
1815                         hindex = round_down(index, HPAGE_PMD_NR);
1816                 else
1817                         hindex = index;
1818
1819                 if (sgp == SGP_WRITE)
1820                         __SetPageReferenced(page);
1821
1822                 error = mem_cgroup_try_charge_delay(page, charge_mm, gfp, &memcg,
1823                                 PageTransHuge(page));
1824                 if (error)
1825                         goto unacct;
1826                 error = radix_tree_maybe_preload_order(gfp & GFP_RECLAIM_MASK,
1827                                 compound_order(page));
1828                 if (!error) {
1829                         error = shmem_add_to_page_cache(page, mapping, hindex,
1830                                                         NULL);
1831                         radix_tree_preload_end();
1832                 }
1833                 if (error) {
1834                         mem_cgroup_cancel_charge(page, memcg,
1835                                         PageTransHuge(page));
1836                         goto unacct;
1837                 }
1838                 mem_cgroup_commit_charge(page, memcg, false,
1839                                 PageTransHuge(page));
1840                 lru_cache_add_anon(page);
1841
1842                 spin_lock_irq(&info->lock);
1843                 info->alloced += 1 << compound_order(page);
1844                 inode->i_blocks += BLOCKS_PER_PAGE << compound_order(page);
1845                 shmem_recalc_inode(inode);
1846                 spin_unlock_irq(&info->lock);
1847                 alloced = true;
1848
1849                 if (PageTransHuge(page) &&
1850                                 DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE) <
1851                                 hindex + HPAGE_PMD_NR - 1) {
1852                         /*
1853                          * Part of the huge page is beyond i_size: subject
1854                          * to shrink under memory pressure.
1855                          */
1856                         spin_lock(&sbinfo->shrinklist_lock);
1857                         /*
1858                          * _careful to defend against unlocked access to
1859                          * ->shrink_list in shmem_unused_huge_shrink()
1860                          */
1861                         if (list_empty_careful(&info->shrinklist)) {
1862                                 list_add_tail(&info->shrinklist,
1863                                                 &sbinfo->shrinklist);
1864                                 sbinfo->shrinklist_len++;
1865                         }
1866                         spin_unlock(&sbinfo->shrinklist_lock);
1867                 }
1868
1869                 /*
1870                  * Let SGP_FALLOC use the SGP_WRITE optimization on a new page.
1871                  */
1872                 if (sgp == SGP_FALLOC)
1873                         sgp = SGP_WRITE;
1874 clear:
1875                 /*
1876                  * Let SGP_WRITE caller clear ends if write does not fill page;
1877                  * but SGP_FALLOC on a page fallocated earlier must initialize
1878                  * it now, lest undo on failure cancel our earlier guarantee.
1879                  */
1880                 if (sgp != SGP_WRITE && !PageUptodate(page)) {
1881                         struct page *head = compound_head(page);
1882                         int i;
1883
1884                         for (i = 0; i < (1 << compound_order(head)); i++) {
1885                                 clear_highpage(head + i);
1886                                 flush_dcache_page(head + i);
1887                         }
1888                         SetPageUptodate(head);
1889                 }
1890         }
1891
1892         /* Perhaps the file has been truncated since we checked */
1893         if (sgp <= SGP_CACHE &&
1894             ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1895                 if (alloced) {
1896                         ClearPageDirty(page);
1897                         delete_from_page_cache(page);
1898                         spin_lock_irq(&info->lock);
1899                         shmem_recalc_inode(inode);
1900                         spin_unlock_irq(&info->lock);
1901                 }
1902                 error = -EINVAL;
1903                 goto unlock;
1904         }
1905         *pagep = page + index - hindex;
1906         return 0;
1907
1908         /*
1909          * Error recovery.
1910          */
1911 unacct:
1912         shmem_inode_unacct_blocks(inode, 1 << compound_order(page));
1913
1914         if (PageTransHuge(page)) {
1915                 unlock_page(page);
1916                 put_page(page);
1917                 goto alloc_nohuge;
1918         }
1919 failed:
1920         if (swap.val && !shmem_confirm_swap(mapping, index, swap))
1921                 error = -EEXIST;
1922 unlock:
1923         if (page) {
1924                 unlock_page(page);
1925                 put_page(page);
1926         }
1927         if (error == -ENOSPC && !once++) {
1928                 spin_lock_irq(&info->lock);
1929                 shmem_recalc_inode(inode);
1930                 spin_unlock_irq(&info->lock);
1931                 goto repeat;
1932         }
1933         if (error == -EEXIST)   /* from above or from radix_tree_insert */
1934                 goto repeat;
1935         return error;
1936 }
1937
1938 /*
1939  * This is like autoremove_wake_function, but it removes the wait queue
1940  * entry unconditionally - even if something else had already woken the
1941  * target.
1942  */
1943 static int synchronous_wake_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
1944 {
1945         int ret = default_wake_function(wait, mode, sync, key);
1946         list_del_init(&wait->entry);
1947         return ret;
1948 }
1949
1950 static vm_fault_t shmem_fault(struct vm_fault *vmf)
1951 {
1952         struct vm_area_struct *vma = vmf->vma;
1953         struct inode *inode = file_inode(vma->vm_file);
1954         gfp_t gfp = mapping_gfp_mask(inode->i_mapping);
1955         enum sgp_type sgp;
1956         int err;
1957         vm_fault_t ret = VM_FAULT_LOCKED;
1958
1959         /*
1960          * Trinity finds that probing a hole which tmpfs is punching can
1961          * prevent the hole-punch from ever completing: which in turn
1962          * locks writers out with its hold on i_mutex.  So refrain from
1963          * faulting pages into the hole while it's being punched.  Although
1964          * shmem_undo_range() does remove the additions, it may be unable to
1965          * keep up, as each new page needs its own unmap_mapping_range() call,
1966          * and the i_mmap tree grows ever slower to scan if new vmas are added.
1967          *
1968          * It does not matter if we sometimes reach this check just before the
1969          * hole-punch begins, so that one fault then races with the punch:
1970          * we just need to make racing faults a rare case.
1971          *
1972          * The implementation below would be much simpler if we just used a
1973          * standard mutex or completion: but we cannot take i_mutex in fault,
1974          * and bloating every shmem inode for this unlikely case would be sad.
1975          */
1976         if (unlikely(inode->i_private)) {
1977                 struct shmem_falloc *shmem_falloc;
1978
1979                 spin_lock(&inode->i_lock);
1980                 shmem_falloc = inode->i_private;
1981                 if (shmem_falloc &&
1982                     shmem_falloc->waitq &&
1983                     vmf->pgoff >= shmem_falloc->start &&
1984                     vmf->pgoff < shmem_falloc->next) {
1985                         wait_queue_head_t *shmem_falloc_waitq;
1986                         DEFINE_WAIT_FUNC(shmem_fault_wait, synchronous_wake_function);
1987
1988                         ret = VM_FAULT_NOPAGE;
1989                         if ((vmf->flags & FAULT_FLAG_ALLOW_RETRY) &&
1990                            !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
1991                                 /* It's polite to up mmap_sem if we can */
1992                                 up_read(&vma->vm_mm->mmap_sem);
1993                                 ret = VM_FAULT_RETRY;
1994                         }
1995
1996                         shmem_falloc_waitq = shmem_falloc->waitq;
1997                         prepare_to_wait(shmem_falloc_waitq, &shmem_fault_wait,
1998                                         TASK_UNINTERRUPTIBLE);
1999                         spin_unlock(&inode->i_lock);
2000                         schedule();
2001
2002                         /*
2003                          * shmem_falloc_waitq points into the shmem_fallocate()
2004                          * stack of the hole-punching task: shmem_falloc_waitq
2005                          * is usually invalid by the time we reach here, but
2006                          * finish_wait() does not dereference it in that case;
2007                          * though i_lock needed lest racing with wake_up_all().
2008                          */
2009                         spin_lock(&inode->i_lock);
2010                         finish_wait(shmem_falloc_waitq, &shmem_fault_wait);
2011                         spin_unlock(&inode->i_lock);
2012                         return ret;
2013                 }
2014                 spin_unlock(&inode->i_lock);
2015         }
2016
2017         sgp = SGP_CACHE;
2018
2019         if ((vma->vm_flags & VM_NOHUGEPAGE) ||
2020             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
2021                 sgp = SGP_NOHUGE;
2022         else if (vma->vm_flags & VM_HUGEPAGE)
2023                 sgp = SGP_HUGE;
2024
2025         err = shmem_getpage_gfp(inode, vmf->pgoff, &vmf->page, sgp,
2026                                   gfp, vma, vmf, &ret);
2027         if (err)
2028                 return vmf_error(err);
2029         return ret;
2030 }
2031
2032 unsigned long shmem_get_unmapped_area(struct file *file,
2033                                       unsigned long uaddr, unsigned long len,
2034                                       unsigned long pgoff, unsigned long flags)
2035 {
2036         unsigned long (*get_area)(struct file *,
2037                 unsigned long, unsigned long, unsigned long, unsigned long);
2038         unsigned long addr;
2039         unsigned long offset;
2040         unsigned long inflated_len;
2041         unsigned long inflated_addr;
2042         unsigned long inflated_offset;
2043
2044         if (len > TASK_SIZE)
2045                 return -ENOMEM;
2046
2047         get_area = current->mm->get_unmapped_area;
2048         addr = get_area(file, uaddr, len, pgoff, flags);
2049
2050         if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
2051                 return addr;
2052         if (IS_ERR_VALUE(addr))
2053                 return addr;
2054         if (addr & ~PAGE_MASK)
2055                 return addr;
2056         if (addr > TASK_SIZE - len)
2057                 return addr;
2058
2059         if (shmem_huge == SHMEM_HUGE_DENY)
2060                 return addr;
2061         if (len < HPAGE_PMD_SIZE)
2062                 return addr;
2063         if (flags & MAP_FIXED)
2064                 return addr;
2065         /*
2066          * Our priority is to support MAP_SHARED mapped hugely;
2067          * and support MAP_PRIVATE mapped hugely too, until it is COWed.
2068          * But if caller specified an address hint, respect that as before.
2069          */
2070         if (uaddr)
2071                 return addr;
2072
2073         if (shmem_huge != SHMEM_HUGE_FORCE) {
2074                 struct super_block *sb;
2075
2076                 if (file) {
2077                         VM_BUG_ON(file->f_op != &shmem_file_operations);
2078                         sb = file_inode(file)->i_sb;
2079                 } else {
2080                         /*
2081                          * Called directly from mm/mmap.c, or drivers/char/mem.c
2082                          * for "/dev/zero", to create a shared anonymous object.
2083                          */
2084                         if (IS_ERR(shm_mnt))
2085                                 return addr;
2086                         sb = shm_mnt->mnt_sb;
2087                 }
2088                 if (SHMEM_SB(sb)->huge == SHMEM_HUGE_NEVER)
2089                         return addr;
2090         }
2091
2092         offset = (pgoff << PAGE_SHIFT) & (HPAGE_PMD_SIZE-1);
2093         if (offset && offset + len < 2 * HPAGE_PMD_SIZE)
2094                 return addr;
2095         if ((addr & (HPAGE_PMD_SIZE-1)) == offset)
2096                 return addr;
2097
2098         inflated_len = len + HPAGE_PMD_SIZE - PAGE_SIZE;
2099         if (inflated_len > TASK_SIZE)
2100                 return addr;
2101         if (inflated_len < len)
2102                 return addr;
2103
2104         inflated_addr = get_area(NULL, 0, inflated_len, 0, flags);
2105         if (IS_ERR_VALUE(inflated_addr))
2106                 return addr;
2107         if (inflated_addr & ~PAGE_MASK)
2108                 return addr;
2109
2110         inflated_offset = inflated_addr & (HPAGE_PMD_SIZE-1);
2111         inflated_addr += offset - inflated_offset;
2112         if (inflated_offset > offset)
2113                 inflated_addr += HPAGE_PMD_SIZE;
2114
2115         if (inflated_addr > TASK_SIZE - len)
2116                 return addr;
2117         return inflated_addr;
2118 }
2119
2120 #ifdef CONFIG_NUMA
2121 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
2122 {
2123         struct inode *inode = file_inode(vma->vm_file);
2124         return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
2125 }
2126
2127 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
2128                                           unsigned long addr)
2129 {
2130         struct inode *inode = file_inode(vma->vm_file);
2131         pgoff_t index;
2132
2133         index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2134         return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
2135 }
2136 #endif
2137
2138 int shmem_lock(struct file *file, int lock, struct user_struct *user)
2139 {
2140         struct inode *inode = file_inode(file);
2141         struct shmem_inode_info *info = SHMEM_I(inode);
2142         int retval = -ENOMEM;
2143
2144         spin_lock_irq(&info->lock);
2145         if (lock && !(info->flags & VM_LOCKED)) {
2146                 if (!user_shm_lock(inode->i_size, user))
2147                         goto out_nomem;
2148                 info->flags |= VM_LOCKED;
2149                 mapping_set_unevictable(file->f_mapping);
2150         }
2151         if (!lock && (info->flags & VM_LOCKED) && user) {
2152                 user_shm_unlock(inode->i_size, user);
2153                 info->flags &= ~VM_LOCKED;
2154                 mapping_clear_unevictable(file->f_mapping);
2155         }
2156         retval = 0;
2157
2158 out_nomem:
2159         spin_unlock_irq(&info->lock);
2160         return retval;
2161 }
2162
2163 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
2164 {
2165         file_accessed(file);
2166         vma->vm_ops = &shmem_vm_ops;
2167         if (IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE) &&
2168                         ((vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK) <
2169                         (vma->vm_end & HPAGE_PMD_MASK)) {
2170                 khugepaged_enter(vma, vma->vm_flags);
2171         }
2172         return 0;
2173 }
2174
2175 static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
2176                                      umode_t mode, dev_t dev, unsigned long flags)
2177 {
2178         struct inode *inode;
2179         struct shmem_inode_info *info;
2180         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2181
2182         if (shmem_reserve_inode(sb))
2183                 return NULL;
2184
2185         inode = new_inode(sb);
2186         if (inode) {
2187                 inode->i_ino = get_next_ino();
2188                 inode_init_owner(inode, dir, mode);
2189                 inode->i_blocks = 0;
2190                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
2191                 inode->i_generation = prandom_u32();
2192                 info = SHMEM_I(inode);
2193                 memset(info, 0, (char *)inode - (char *)info);
2194                 spin_lock_init(&info->lock);
2195                 info->seals = F_SEAL_SEAL;
2196                 info->flags = flags & VM_NORESERVE;
2197                 INIT_LIST_HEAD(&info->shrinklist);
2198                 INIT_LIST_HEAD(&info->swaplist);
2199                 simple_xattrs_init(&info->xattrs);
2200                 cache_no_acl(inode);
2201
2202                 switch (mode & S_IFMT) {
2203                 default:
2204                         inode->i_op = &shmem_special_inode_operations;
2205                         init_special_inode(inode, mode, dev);
2206                         break;
2207                 case S_IFREG:
2208                         inode->i_mapping->a_ops = &shmem_aops;
2209                         inode->i_op = &shmem_inode_operations;
2210                         inode->i_fop = &shmem_file_operations;
2211                         mpol_shared_policy_init(&info->policy,
2212                                                  shmem_get_sbmpol(sbinfo));
2213                         break;
2214                 case S_IFDIR:
2215                         inc_nlink(inode);
2216                         /* Some things misbehave if size == 0 on a directory */
2217                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
2218                         inode->i_op = &shmem_dir_inode_operations;
2219                         inode->i_fop = &simple_dir_operations;
2220                         break;
2221                 case S_IFLNK:
2222                         /*
2223                          * Must not load anything in the rbtree,
2224                          * mpol_free_shared_policy will not be called.
2225                          */
2226                         mpol_shared_policy_init(&info->policy, NULL);
2227                         break;
2228                 }
2229         } else
2230                 shmem_free_inode(sb);
2231         return inode;
2232 }
2233
2234 bool shmem_mapping(struct address_space *mapping)
2235 {
2236         return mapping->a_ops == &shmem_aops;
2237 }
2238
2239 static int shmem_mfill_atomic_pte(struct mm_struct *dst_mm,
2240                                   pmd_t *dst_pmd,
2241                                   struct vm_area_struct *dst_vma,
2242                                   unsigned long dst_addr,
2243                                   unsigned long src_addr,
2244                                   bool zeropage,
2245                                   struct page **pagep)
2246 {
2247         struct inode *inode = file_inode(dst_vma->vm_file);
2248         struct shmem_inode_info *info = SHMEM_I(inode);
2249         struct address_space *mapping = inode->i_mapping;
2250         gfp_t gfp = mapping_gfp_mask(mapping);
2251         pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
2252         struct mem_cgroup *memcg;
2253         spinlock_t *ptl;
2254         void *page_kaddr;
2255         struct page *page;
2256         pte_t _dst_pte, *dst_pte;
2257         int ret;
2258
2259         ret = -ENOMEM;
2260         if (!shmem_inode_acct_block(inode, 1))
2261                 goto out;
2262
2263         if (!*pagep) {
2264                 page = shmem_alloc_page(gfp, info, pgoff);
2265                 if (!page)
2266                         goto out_unacct_blocks;
2267
2268                 if (!zeropage) {        /* mcopy_atomic */
2269                         page_kaddr = kmap_atomic(page);
2270                         ret = copy_from_user(page_kaddr,
2271                                              (const void __user *)src_addr,
2272                                              PAGE_SIZE);
2273                         kunmap_atomic(page_kaddr);
2274
2275                         /* fallback to copy_from_user outside mmap_sem */
2276                         if (unlikely(ret)) {
2277                                 *pagep = page;
2278                                 shmem_inode_unacct_blocks(inode, 1);
2279                                 /* don't free the page */
2280                                 return -EFAULT;
2281                         }
2282                 } else {                /* mfill_zeropage_atomic */
2283                         clear_highpage(page);
2284                 }
2285         } else {
2286                 page = *pagep;
2287                 *pagep = NULL;
2288         }
2289
2290         VM_BUG_ON(PageLocked(page) || PageSwapBacked(page));
2291         __SetPageLocked(page);
2292         __SetPageSwapBacked(page);
2293         __SetPageUptodate(page);
2294
2295         ret = mem_cgroup_try_charge_delay(page, dst_mm, gfp, &memcg, false);
2296         if (ret)
2297                 goto out_release;
2298
2299         ret = radix_tree_maybe_preload(gfp & GFP_RECLAIM_MASK);
2300         if (!ret) {
2301                 ret = shmem_add_to_page_cache(page, mapping, pgoff, NULL);
2302                 radix_tree_preload_end();
2303         }
2304         if (ret)
2305                 goto out_release_uncharge;
2306
2307         mem_cgroup_commit_charge(page, memcg, false, false);
2308
2309         _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
2310         if (dst_vma->vm_flags & VM_WRITE)
2311                 _dst_pte = pte_mkwrite(pte_mkdirty(_dst_pte));
2312
2313         ret = -EEXIST;
2314         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
2315         if (!pte_none(*dst_pte))
2316                 goto out_release_uncharge_unlock;
2317
2318         lru_cache_add_anon(page);
2319
2320         spin_lock(&info->lock);
2321         info->alloced++;
2322         inode->i_blocks += BLOCKS_PER_PAGE;
2323         shmem_recalc_inode(inode);
2324         spin_unlock(&info->lock);
2325
2326         inc_mm_counter(dst_mm, mm_counter_file(page));
2327         page_add_file_rmap(page, false);
2328         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
2329
2330         /* No need to invalidate - it was non-present before */
2331         update_mmu_cache(dst_vma, dst_addr, dst_pte);
2332         unlock_page(page);
2333         pte_unmap_unlock(dst_pte, ptl);
2334         ret = 0;
2335 out:
2336         return ret;
2337 out_release_uncharge_unlock:
2338         pte_unmap_unlock(dst_pte, ptl);
2339 out_release_uncharge:
2340         mem_cgroup_cancel_charge(page, memcg, false);
2341 out_release:
2342         unlock_page(page);
2343         put_page(page);
2344 out_unacct_blocks:
2345         shmem_inode_unacct_blocks(inode, 1);
2346         goto out;
2347 }
2348
2349 int shmem_mcopy_atomic_pte(struct mm_struct *dst_mm,
2350                            pmd_t *dst_pmd,
2351                            struct vm_area_struct *dst_vma,
2352                            unsigned long dst_addr,
2353                            unsigned long src_addr,
2354                            struct page **pagep)
2355 {
2356         return shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
2357                                       dst_addr, src_addr, false, pagep);
2358 }
2359
2360 int shmem_mfill_zeropage_pte(struct mm_struct *dst_mm,
2361                              pmd_t *dst_pmd,
2362                              struct vm_area_struct *dst_vma,
2363                              unsigned long dst_addr)
2364 {
2365         struct page *page = NULL;
2366
2367         return shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
2368                                       dst_addr, 0, true, &page);
2369 }
2370
2371 #ifdef CONFIG_TMPFS
2372 static const struct inode_operations shmem_symlink_inode_operations;
2373 static const struct inode_operations shmem_short_symlink_operations;
2374
2375 #ifdef CONFIG_TMPFS_XATTR
2376 static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
2377 #else
2378 #define shmem_initxattrs NULL
2379 #endif
2380
2381 static int
2382 shmem_write_begin(struct file *file, struct address_space *mapping,
2383                         loff_t pos, unsigned len, unsigned flags,
2384                         struct page **pagep, void **fsdata)
2385 {
2386         struct inode *inode = mapping->host;
2387         struct shmem_inode_info *info = SHMEM_I(inode);
2388         pgoff_t index = pos >> PAGE_SHIFT;
2389
2390         /* i_mutex is held by caller */
2391         if (unlikely(info->seals & (F_SEAL_WRITE | F_SEAL_GROW))) {
2392                 if (info->seals & F_SEAL_WRITE)
2393                         return -EPERM;
2394                 if ((info->seals & F_SEAL_GROW) && pos + len > inode->i_size)
2395                         return -EPERM;
2396         }
2397
2398         return shmem_getpage(inode, index, pagep, SGP_WRITE);
2399 }
2400
2401 static int
2402 shmem_write_end(struct file *file, struct address_space *mapping,
2403                         loff_t pos, unsigned len, unsigned copied,
2404                         struct page *page, void *fsdata)
2405 {
2406         struct inode *inode = mapping->host;
2407
2408         if (pos + copied > inode->i_size)
2409                 i_size_write(inode, pos + copied);
2410
2411         if (!PageUptodate(page)) {
2412                 struct page *head = compound_head(page);
2413                 if (PageTransCompound(page)) {
2414                         int i;
2415
2416                         for (i = 0; i < HPAGE_PMD_NR; i++) {
2417                                 if (head + i == page)
2418                                         continue;
2419                                 clear_highpage(head + i);
2420                                 flush_dcache_page(head + i);
2421                         }
2422                 }
2423                 if (copied < PAGE_SIZE) {
2424                         unsigned from = pos & (PAGE_SIZE - 1);
2425                         zero_user_segments(page, 0, from,
2426                                         from + copied, PAGE_SIZE);
2427                 }
2428                 SetPageUptodate(head);
2429         }
2430         set_page_dirty(page);
2431         unlock_page(page);
2432         put_page(page);
2433
2434         return copied;
2435 }
2436
2437 static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
2438 {
2439         struct file *file = iocb->ki_filp;
2440         struct inode *inode = file_inode(file);
2441         struct address_space *mapping = inode->i_mapping;
2442         pgoff_t index;
2443         unsigned long offset;
2444         enum sgp_type sgp = SGP_READ;
2445         int error = 0;
2446         ssize_t retval = 0;
2447         loff_t *ppos = &iocb->ki_pos;
2448
2449         /*
2450          * Might this read be for a stacking filesystem?  Then when reading
2451          * holes of a sparse file, we actually need to allocate those pages,
2452          * and even mark them dirty, so it cannot exceed the max_blocks limit.
2453          */
2454         if (!iter_is_iovec(to))
2455                 sgp = SGP_CACHE;
2456
2457         index = *ppos >> PAGE_SHIFT;
2458         offset = *ppos & ~PAGE_MASK;
2459
2460         for (;;) {
2461                 struct page *page = NULL;
2462                 pgoff_t end_index;
2463                 unsigned long nr, ret;
2464                 loff_t i_size = i_size_read(inode);
2465
2466                 end_index = i_size >> PAGE_SHIFT;
2467                 if (index > end_index)
2468                         break;
2469                 if (index == end_index) {
2470                         nr = i_size & ~PAGE_MASK;
2471                         if (nr <= offset)
2472                                 break;
2473                 }
2474
2475                 error = shmem_getpage(inode, index, &page, sgp);
2476                 if (error) {
2477                         if (error == -EINVAL)
2478                                 error = 0;
2479                         break;
2480                 }
2481                 if (page) {
2482                         if (sgp == SGP_CACHE)
2483                                 set_page_dirty(page);
2484                         unlock_page(page);
2485                 }
2486
2487                 /*
2488                  * We must evaluate after, since reads (unlike writes)
2489                  * are called without i_mutex protection against truncate
2490                  */
2491                 nr = PAGE_SIZE;
2492                 i_size = i_size_read(inode);
2493                 end_index = i_size >> PAGE_SHIFT;
2494                 if (index == end_index) {
2495                         nr = i_size & ~PAGE_MASK;
2496                         if (nr <= offset) {
2497                                 if (page)
2498                                         put_page(page);
2499                                 break;
2500                         }
2501                 }
2502                 nr -= offset;
2503
2504                 if (page) {
2505                         /*
2506                          * If users can be writing to this page using arbitrary
2507                          * virtual addresses, take care about potential aliasing
2508                          * before reading the page on the kernel side.
2509                          */
2510                         if (mapping_writably_mapped(mapping))
2511                                 flush_dcache_page(page);
2512                         /*
2513                          * Mark the page accessed if we read the beginning.
2514                          */
2515                         if (!offset)
2516                                 mark_page_accessed(page);
2517                 } else {
2518                         page = ZERO_PAGE(0);
2519                         get_page(page);
2520                 }
2521
2522                 /*
2523                  * Ok, we have the page, and it's up-to-date, so
2524                  * now we can copy it to user space...
2525                  */
2526                 ret = copy_page_to_iter(page, offset, nr, to);
2527                 retval += ret;
2528                 offset += ret;
2529                 index += offset >> PAGE_SHIFT;
2530                 offset &= ~PAGE_MASK;
2531
2532                 put_page(page);
2533                 if (!iov_iter_count(to))
2534                         break;
2535                 if (ret < nr) {
2536                         error = -EFAULT;
2537                         break;
2538                 }
2539                 cond_resched();
2540         }
2541
2542         *ppos = ((loff_t) index << PAGE_SHIFT) + offset;
2543         file_accessed(file);
2544         return retval ? retval : error;
2545 }
2546
2547 /*
2548  * llseek SEEK_DATA or SEEK_HOLE through the radix_tree.
2549  */
2550 static pgoff_t shmem_seek_hole_data(struct address_space *mapping,
2551                                     pgoff_t index, pgoff_t end, int whence)
2552 {
2553         struct page *page;
2554         struct pagevec pvec;
2555         pgoff_t indices[PAGEVEC_SIZE];
2556         bool done = false;
2557         int i;
2558
2559         pagevec_init(&pvec);
2560         pvec.nr = 1;            /* start small: we may be there already */
2561         while (!done) {
2562                 pvec.nr = find_get_entries(mapping, index,
2563                                         pvec.nr, pvec.pages, indices);
2564                 if (!pvec.nr) {
2565                         if (whence == SEEK_DATA)
2566                                 index = end;
2567                         break;
2568                 }
2569                 for (i = 0; i < pvec.nr; i++, index++) {
2570                         if (index < indices[i]) {
2571                                 if (whence == SEEK_HOLE) {
2572                                         done = true;
2573                                         break;
2574                                 }
2575                                 index = indices[i];
2576                         }
2577                         page = pvec.pages[i];
2578                         if (page && !radix_tree_exceptional_entry(page)) {
2579                                 if (!PageUptodate(page))
2580                                         page = NULL;
2581                         }
2582                         if (index >= end ||
2583                             (page && whence == SEEK_DATA) ||
2584                             (!page && whence == SEEK_HOLE)) {
2585                                 done = true;
2586                                 break;
2587                         }
2588                 }
2589                 pagevec_remove_exceptionals(&pvec);
2590                 pagevec_release(&pvec);
2591                 pvec.nr = PAGEVEC_SIZE;
2592                 cond_resched();
2593         }
2594         return index;
2595 }
2596
2597 static loff_t shmem_file_llseek(struct file *file, loff_t offset, int whence)
2598 {
2599         struct address_space *mapping = file->f_mapping;
2600         struct inode *inode = mapping->host;
2601         pgoff_t start, end;
2602         loff_t new_offset;
2603
2604         if (whence != SEEK_DATA && whence != SEEK_HOLE)
2605                 return generic_file_llseek_size(file, offset, whence,
2606                                         MAX_LFS_FILESIZE, i_size_read(inode));
2607         inode_lock(inode);
2608         /* We're holding i_mutex so we can access i_size directly */
2609
2610         if (offset < 0)
2611                 offset = -EINVAL;
2612         else if (offset >= inode->i_size)
2613                 offset = -ENXIO;
2614         else {
2615                 start = offset >> PAGE_SHIFT;
2616                 end = (inode->i_size + PAGE_SIZE - 1) >> PAGE_SHIFT;
2617                 new_offset = shmem_seek_hole_data(mapping, start, end, whence);
2618                 new_offset <<= PAGE_SHIFT;
2619                 if (new_offset > offset) {
2620                         if (new_offset < inode->i_size)
2621                                 offset = new_offset;
2622                         else if (whence == SEEK_DATA)
2623                                 offset = -ENXIO;
2624                         else
2625                                 offset = inode->i_size;
2626                 }
2627         }
2628
2629         if (offset >= 0)
2630                 offset = vfs_setpos(file, offset, MAX_LFS_FILESIZE);
2631         inode_unlock(inode);
2632         return offset;
2633 }
2634
2635 static long shmem_fallocate(struct file *file, int mode, loff_t offset,
2636                                                          loff_t len)
2637 {
2638         struct inode *inode = file_inode(file);
2639         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
2640         struct shmem_inode_info *info = SHMEM_I(inode);
2641         struct shmem_falloc shmem_falloc;
2642         pgoff_t start, index, end;
2643         int error;
2644
2645         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2646                 return -EOPNOTSUPP;
2647
2648         inode_lock(inode);
2649
2650         if (mode & FALLOC_FL_PUNCH_HOLE) {
2651                 struct address_space *mapping = file->f_mapping;
2652                 loff_t unmap_start = round_up(offset, PAGE_SIZE);
2653                 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
2654                 DECLARE_WAIT_QUEUE_HEAD_ONSTACK(shmem_falloc_waitq);
2655
2656                 /* protected by i_mutex */
2657                 if (info->seals & F_SEAL_WRITE) {
2658                         error = -EPERM;
2659                         goto out;
2660                 }
2661
2662                 shmem_falloc.waitq = &shmem_falloc_waitq;
2663                 shmem_falloc.start = unmap_start >> PAGE_SHIFT;
2664                 shmem_falloc.next = (unmap_end + 1) >> PAGE_SHIFT;
2665                 spin_lock(&inode->i_lock);
2666                 inode->i_private = &shmem_falloc;
2667                 spin_unlock(&inode->i_lock);
2668
2669                 if ((u64)unmap_end > (u64)unmap_start)
2670                         unmap_mapping_range(mapping, unmap_start,
2671                                             1 + unmap_end - unmap_start, 0);
2672                 shmem_truncate_range(inode, offset, offset + len - 1);
2673                 /* No need to unmap again: hole-punching leaves COWed pages */
2674
2675                 spin_lock(&inode->i_lock);
2676                 inode->i_private = NULL;
2677                 wake_up_all(&shmem_falloc_waitq);
2678                 WARN_ON_ONCE(!list_empty(&shmem_falloc_waitq.head));
2679                 spin_unlock(&inode->i_lock);
2680                 error = 0;
2681                 goto out;
2682         }
2683
2684         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
2685         error = inode_newsize_ok(inode, offset + len);
2686         if (error)
2687                 goto out;
2688
2689         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
2690                 error = -EPERM;
2691                 goto out;
2692         }
2693
2694         start = offset >> PAGE_SHIFT;
2695         end = (offset + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
2696         /* Try to avoid a swapstorm if len is impossible to satisfy */
2697         if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
2698                 error = -ENOSPC;
2699                 goto out;
2700         }
2701
2702         shmem_falloc.waitq = NULL;
2703         shmem_falloc.start = start;
2704         shmem_falloc.next  = start;
2705         shmem_falloc.nr_falloced = 0;
2706         shmem_falloc.nr_unswapped = 0;
2707         spin_lock(&inode->i_lock);
2708         inode->i_private = &shmem_falloc;
2709         spin_unlock(&inode->i_lock);
2710
2711         for (index = start; index < end; index++) {
2712                 struct page *page;
2713
2714                 /*
2715                  * Good, the fallocate(2) manpage permits EINTR: we may have
2716                  * been interrupted because we are using up too much memory.
2717                  */
2718                 if (signal_pending(current))
2719                         error = -EINTR;
2720                 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
2721                         error = -ENOMEM;
2722                 else
2723                         error = shmem_getpage(inode, index, &page, SGP_FALLOC);
2724                 if (error) {
2725                         /* Remove the !PageUptodate pages we added */
2726                         if (index > start) {
2727                                 shmem_undo_range(inode,
2728                                     (loff_t)start << PAGE_SHIFT,
2729                                     ((loff_t)index << PAGE_SHIFT) - 1, true);
2730                         }
2731                         goto undone;
2732                 }
2733
2734                 /*
2735                  * Inform shmem_writepage() how far we have reached.
2736                  * No need for lock or barrier: we have the page lock.
2737                  */
2738                 shmem_falloc.next++;
2739                 if (!PageUptodate(page))
2740                         shmem_falloc.nr_falloced++;
2741
2742                 /*
2743                  * If !PageUptodate, leave it that way so that freeable pages
2744                  * can be recognized if we need to rollback on error later.
2745                  * But set_page_dirty so that memory pressure will swap rather
2746                  * than free the pages we are allocating (and SGP_CACHE pages
2747                  * might still be clean: we now need to mark those dirty too).
2748                  */
2749                 set_page_dirty(page);
2750                 unlock_page(page);
2751                 put_page(page);
2752                 cond_resched();
2753         }
2754
2755         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
2756                 i_size_write(inode, offset + len);
2757         inode->i_ctime = current_time(inode);
2758 undone:
2759         spin_lock(&inode->i_lock);
2760         inode->i_private = NULL;
2761         spin_unlock(&inode->i_lock);
2762 out:
2763         inode_unlock(inode);
2764         return error;
2765 }
2766
2767 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
2768 {
2769         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
2770
2771         buf->f_type = TMPFS_MAGIC;
2772         buf->f_bsize = PAGE_SIZE;
2773         buf->f_namelen = NAME_MAX;
2774         if (sbinfo->max_blocks) {
2775                 buf->f_blocks = sbinfo->max_blocks;
2776                 buf->f_bavail =
2777                 buf->f_bfree  = sbinfo->max_blocks -
2778                                 percpu_counter_sum(&sbinfo->used_blocks);
2779         }
2780         if (sbinfo->max_inodes) {
2781                 buf->f_files = sbinfo->max_inodes;
2782                 buf->f_ffree = sbinfo->free_inodes;
2783         }
2784         /* else leave those fields 0 like simple_statfs */
2785         return 0;
2786 }
2787
2788 /*
2789  * File creation. Allocate an inode, and we're done..
2790  */
2791 static int
2792 shmem_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
2793 {
2794         struct inode *inode;
2795         int error = -ENOSPC;
2796
2797         inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
2798         if (inode) {
2799                 error = simple_acl_create(dir, inode);
2800                 if (error)
2801                         goto out_iput;
2802                 error = security_inode_init_security(inode, dir,
2803                                                      &dentry->d_name,
2804                                                      shmem_initxattrs, NULL);
2805                 if (error && error != -EOPNOTSUPP)
2806                         goto out_iput;
2807
2808                 error = 0;
2809                 dir->i_size += BOGO_DIRENT_SIZE;
2810                 dir->i_ctime = dir->i_mtime = current_time(dir);
2811                 d_instantiate(dentry, inode);
2812                 dget(dentry); /* Extra count - pin the dentry in core */
2813         }
2814         return error;
2815 out_iput:
2816         iput(inode);
2817         return error;
2818 }
2819
2820 static int
2821 shmem_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
2822 {
2823         struct inode *inode;
2824         int error = -ENOSPC;
2825
2826         inode = shmem_get_inode(dir->i_sb, dir, mode, 0, VM_NORESERVE);
2827         if (inode) {
2828                 error = security_inode_init_security(inode, dir,
2829                                                      NULL,
2830                                                      shmem_initxattrs, NULL);
2831                 if (error && error != -EOPNOTSUPP)
2832                         goto out_iput;
2833                 error = simple_acl_create(dir, inode);
2834                 if (error)
2835                         goto out_iput;
2836                 d_tmpfile(dentry, inode);
2837         }
2838         return error;
2839 out_iput:
2840         iput(inode);
2841         return error;
2842 }
2843
2844 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
2845 {
2846         int error;
2847
2848         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
2849                 return error;
2850         inc_nlink(dir);
2851         return 0;
2852 }
2853
2854 static int shmem_create(struct inode *dir, struct dentry *dentry, umode_t mode,
2855                 bool excl)
2856 {
2857         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
2858 }
2859
2860 /*
2861  * Link a file..
2862  */
2863 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
2864 {
2865         struct inode *inode = d_inode(old_dentry);
2866         int ret;
2867
2868         /*
2869          * No ordinary (disk based) filesystem counts links as inodes;
2870          * but each new link needs a new dentry, pinning lowmem, and
2871          * tmpfs dentries cannot be pruned until they are unlinked.
2872          */
2873         ret = shmem_reserve_inode(inode->i_sb);
2874         if (ret)
2875                 goto out;
2876
2877         dir->i_size += BOGO_DIRENT_SIZE;
2878         inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
2879         inc_nlink(inode);
2880         ihold(inode);   /* New dentry reference */
2881         dget(dentry);           /* Extra pinning count for the created dentry */
2882         d_instantiate(dentry, inode);
2883 out:
2884         return ret;
2885 }
2886
2887 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
2888 {
2889         struct inode *inode = d_inode(dentry);
2890
2891         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
2892                 shmem_free_inode(inode->i_sb);
2893
2894         dir->i_size -= BOGO_DIRENT_SIZE;
2895         inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
2896         drop_nlink(inode);
2897         dput(dentry);   /* Undo the count from "create" - this does all the work */
2898         return 0;
2899 }
2900
2901 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
2902 {
2903         if (!simple_empty(dentry))
2904                 return -ENOTEMPTY;
2905
2906         drop_nlink(d_inode(dentry));
2907         drop_nlink(dir);
2908         return shmem_unlink(dir, dentry);
2909 }
2910
2911 static int shmem_exchange(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
2912 {
2913         bool old_is_dir = d_is_dir(old_dentry);
2914         bool new_is_dir = d_is_dir(new_dentry);
2915
2916         if (old_dir != new_dir && old_is_dir != new_is_dir) {
2917                 if (old_is_dir) {
2918                         drop_nlink(old_dir);
2919                         inc_nlink(new_dir);
2920                 } else {
2921                         drop_nlink(new_dir);
2922                         inc_nlink(old_dir);
2923                 }
2924         }
2925         old_dir->i_ctime = old_dir->i_mtime =
2926         new_dir->i_ctime = new_dir->i_mtime =
2927         d_inode(old_dentry)->i_ctime =
2928         d_inode(new_dentry)->i_ctime = current_time(old_dir);
2929
2930         return 0;
2931 }
2932
2933 static int shmem_whiteout(struct inode *old_dir, struct dentry *old_dentry)
2934 {
2935         struct dentry *whiteout;
2936         int error;
2937
2938         whiteout = d_alloc(old_dentry->d_parent, &old_dentry->d_name);
2939         if (!whiteout)
2940                 return -ENOMEM;
2941
2942         error = shmem_mknod(old_dir, whiteout,
2943                             S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
2944         dput(whiteout);
2945         if (error)
2946                 return error;
2947
2948         /*
2949          * Cheat and hash the whiteout while the old dentry is still in
2950          * place, instead of playing games with FS_RENAME_DOES_D_MOVE.
2951          *
2952          * d_lookup() will consistently find one of them at this point,
2953          * not sure which one, but that isn't even important.
2954          */
2955         d_rehash(whiteout);
2956         return 0;
2957 }
2958
2959 /*
2960  * The VFS layer already does all the dentry stuff for rename,
2961  * we just have to decrement the usage count for the target if
2962  * it exists so that the VFS layer correctly free's it when it
2963  * gets overwritten.
2964  */
2965 static int shmem_rename2(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry, unsigned int flags)
2966 {
2967         struct inode *inode = d_inode(old_dentry);
2968         int they_are_dirs = S_ISDIR(inode->i_mode);
2969
2970         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
2971                 return -EINVAL;
2972
2973         if (flags & RENAME_EXCHANGE)
2974                 return shmem_exchange(old_dir, old_dentry, new_dir, new_dentry);
2975
2976         if (!simple_empty(new_dentry))
2977                 return -ENOTEMPTY;
2978
2979         if (flags & RENAME_WHITEOUT) {
2980                 int error;
2981
2982                 error = shmem_whiteout(old_dir, old_dentry);
2983                 if (error)
2984                         return error;
2985         }
2986
2987         if (d_really_is_positive(new_dentry)) {
2988                 (void) shmem_unlink(new_dir, new_dentry);
2989                 if (they_are_dirs) {
2990                         drop_nlink(d_inode(new_dentry));
2991                         drop_nlink(old_dir);
2992                 }
2993         } else if (they_are_dirs) {
2994                 drop_nlink(old_dir);
2995                 inc_nlink(new_dir);
2996         }
2997
2998         old_dir->i_size -= BOGO_DIRENT_SIZE;
2999         new_dir->i_size += BOGO_DIRENT_SIZE;
3000         old_dir->i_ctime = old_dir->i_mtime =
3001         new_dir->i_ctime = new_dir->i_mtime =
3002         inode->i_ctime = current_time(old_dir);
3003         return 0;
3004 }
3005
3006 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
3007 {
3008         int error;
3009         int len;
3010         struct inode *inode;
3011         struct page *page;
3012
3013         len = strlen(symname) + 1;
3014         if (len > PAGE_SIZE)
3015                 return -ENAMETOOLONG;
3016
3017         inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK | 0777, 0,
3018                                 VM_NORESERVE);
3019         if (!inode)
3020                 return -ENOSPC;
3021
3022         error = security_inode_init_security(inode, dir, &dentry->d_name,
3023                                              shmem_initxattrs, NULL);
3024         if (error) {
3025                 if (error != -EOPNOTSUPP) {
3026                         iput(inode);
3027                         return error;
3028                 }
3029                 error = 0;
3030         }
3031
3032         inode->i_size = len-1;
3033         if (len <= SHORT_SYMLINK_LEN) {
3034                 inode->i_link = kmemdup(symname, len, GFP_KERNEL);
3035                 if (!inode->i_link) {
3036                         iput(inode);
3037                         return -ENOMEM;
3038                 }
3039                 inode->i_op = &shmem_short_symlink_operations;
3040         } else {
3041                 inode_nohighmem(inode);
3042                 error = shmem_getpage(inode, 0, &page, SGP_WRITE);
3043                 if (error) {
3044                         iput(inode);
3045                         return error;
3046                 }
3047                 inode->i_mapping->a_ops = &shmem_aops;
3048                 inode->i_op = &shmem_symlink_inode_operations;
3049                 memcpy(page_address(page), symname, len);
3050                 SetPageUptodate(page);
3051                 set_page_dirty(page);
3052                 unlock_page(page);
3053                 put_page(page);
3054         }
3055         dir->i_size += BOGO_DIRENT_SIZE;
3056         dir->i_ctime = dir->i_mtime = current_time(dir);
3057         d_instantiate(dentry, inode);
3058         dget(dentry);
3059         return 0;
3060 }
3061
3062 static void shmem_put_link(void *arg)
3063 {
3064         mark_page_accessed(arg);
3065         put_page(arg);
3066 }
3067
3068 static const char *shmem_get_link(struct dentry *dentry,
3069                                   struct inode *inode,
3070                                   struct delayed_call *done)
3071 {
3072         struct page *page = NULL;
3073         int error;
3074         if (!dentry) {
3075                 page = find_get_page(inode->i_mapping, 0);
3076                 if (!page)
3077                         return ERR_PTR(-ECHILD);
3078                 if (!PageUptodate(page)) {
3079                         put_page(page);
3080                         return ERR_PTR(-ECHILD);
3081                 }
3082         } else {
3083                 error = shmem_getpage(inode, 0, &page, SGP_READ);
3084                 if (error)
3085                         return ERR_PTR(error);
3086                 unlock_page(page);
3087         }
3088         set_delayed_call(done, shmem_put_link, page);
3089         return page_address(page);
3090 }
3091
3092 #ifdef CONFIG_TMPFS_XATTR
3093 /*
3094  * Superblocks without xattr inode operations may get some security.* xattr
3095  * support from the LSM "for free". As soon as we have any other xattrs
3096  * like ACLs, we also need to implement the security.* handlers at
3097  * filesystem level, though.
3098  */
3099
3100 /*
3101  * Callback for security_inode_init_security() for acquiring xattrs.
3102  */
3103 static int shmem_initxattrs(struct inode *inode,
3104                             const struct xattr *xattr_array,
3105                             void *fs_info)
3106 {
3107         struct shmem_inode_info *info = SHMEM_I(inode);
3108         const struct xattr *xattr;
3109         struct simple_xattr *new_xattr;
3110         size_t len;
3111
3112         for (xattr = xattr_array; xattr->name != NULL; xattr++) {
3113                 new_xattr = simple_xattr_alloc(xattr->value, xattr->value_len);
3114                 if (!new_xattr)
3115                         return -ENOMEM;
3116
3117                 len = strlen(xattr->name) + 1;
3118                 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
3119                                           GFP_KERNEL);
3120                 if (!new_xattr->name) {
3121                         kfree(new_xattr);
3122                         return -ENOMEM;
3123                 }
3124
3125                 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
3126                        XATTR_SECURITY_PREFIX_LEN);
3127                 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
3128                        xattr->name, len);
3129
3130                 simple_xattr_list_add(&info->xattrs, new_xattr);
3131         }
3132
3133         return 0;
3134 }
3135
3136 static int shmem_xattr_handler_get(const struct xattr_handler *handler,
3137                                    struct dentry *unused, struct inode *inode,
3138                                    const char *name, void *buffer, size_t size)
3139 {
3140         struct shmem_inode_info *info = SHMEM_I(inode);
3141
3142         name = xattr_full_name(handler, name);
3143         return simple_xattr_get(&info->xattrs, name, buffer, size);
3144 }
3145
3146 static int shmem_xattr_handler_set(const struct xattr_handler *handler,
3147                                    struct dentry *unused, struct inode *inode,
3148                                    const char *name, const void *value,
3149                                    size_t size, int flags)
3150 {
3151         struct shmem_inode_info *info = SHMEM_I(inode);
3152
3153         name = xattr_full_name(handler, name);
3154         return simple_xattr_set(&info->xattrs, name, value, size, flags);
3155 }
3156
3157 static const struct xattr_handler shmem_security_xattr_handler = {
3158         .prefix = XATTR_SECURITY_PREFIX,
3159         .get = shmem_xattr_handler_get,
3160         .set = shmem_xattr_handler_set,
3161 };
3162
3163 static const struct xattr_handler shmem_trusted_xattr_handler = {
3164         .prefix = XATTR_TRUSTED_PREFIX,
3165         .get = shmem_xattr_handler_get,
3166         .set = shmem_xattr_handler_set,
3167 };
3168
3169 static const struct xattr_handler *shmem_xattr_handlers[] = {
3170 #ifdef CONFIG_TMPFS_POSIX_ACL
3171         &posix_acl_access_xattr_handler,
3172         &posix_acl_default_xattr_handler,
3173 #endif
3174         &shmem_security_xattr_handler,
3175         &shmem_trusted_xattr_handler,
3176         NULL
3177 };
3178
3179 static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
3180 {
3181         struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
3182         return simple_xattr_list(d_inode(dentry), &info->xattrs, buffer, size);
3183 }
3184 #endif /* CONFIG_TMPFS_XATTR */
3185
3186 static const struct inode_operations shmem_short_symlink_operations = {
3187         .get_link       = simple_get_link,
3188 #ifdef CONFIG_TMPFS_XATTR
3189         .listxattr      = shmem_listxattr,
3190 #endif
3191 };
3192
3193 static const struct inode_operations shmem_symlink_inode_operations = {
3194         .get_link       = shmem_get_link,
3195 #ifdef CONFIG_TMPFS_XATTR
3196         .listxattr      = shmem_listxattr,
3197 #endif
3198 };
3199
3200 static struct dentry *shmem_get_parent(struct dentry *child)
3201 {
3202         return ERR_PTR(-ESTALE);
3203 }
3204
3205 static int shmem_match(struct inode *ino, void *vfh)
3206 {
3207         __u32 *fh = vfh;
3208         __u64 inum = fh[2];
3209         inum = (inum << 32) | fh[1];
3210         return ino->i_ino == inum && fh[0] == ino->i_generation;
3211 }
3212
3213 /* Find any alias of inode, but prefer a hashed alias */
3214 static struct dentry *shmem_find_alias(struct inode *inode)
3215 {
3216         struct dentry *alias = d_find_alias(inode);
3217
3218         return alias ?: d_find_any_alias(inode);
3219 }
3220
3221
3222 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
3223                 struct fid *fid, int fh_len, int fh_type)
3224 {
3225         struct inode *inode;
3226         struct dentry *dentry = NULL;
3227         u64 inum;
3228
3229         if (fh_len < 3)
3230                 return NULL;
3231
3232         inum = fid->raw[2];
3233         inum = (inum << 32) | fid->raw[1];
3234
3235         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
3236                         shmem_match, fid->raw);
3237         if (inode) {
3238                 dentry = shmem_find_alias(inode);
3239                 iput(inode);
3240         }
3241
3242         return dentry;
3243 }
3244
3245 static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
3246                                 struct inode *parent)
3247 {
3248         if (*len < 3) {
3249                 *len = 3;
3250                 return FILEID_INVALID;
3251         }
3252
3253         if (inode_unhashed(inode)) {
3254                 /* Unfortunately insert_inode_hash is not idempotent,
3255                  * so as we hash inodes here rather than at creation
3256                  * time, we need a lock to ensure we only try
3257                  * to do it once
3258                  */
3259                 static DEFINE_SPINLOCK(lock);
3260                 spin_lock(&lock);
3261                 if (inode_unhashed(inode))
3262                         __insert_inode_hash(inode,
3263                                             inode->i_ino + inode->i_generation);
3264                 spin_unlock(&lock);
3265         }
3266
3267         fh[0] = inode->i_generation;
3268         fh[1] = inode->i_ino;
3269         fh[2] = ((__u64)inode->i_ino) >> 32;
3270
3271         *len = 3;
3272         return 1;
3273 }
3274
3275 static const struct export_operations shmem_export_ops = {
3276         .get_parent     = shmem_get_parent,
3277         .encode_fh      = shmem_encode_fh,
3278         .fh_to_dentry   = shmem_fh_to_dentry,
3279 };
3280
3281 static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
3282                                bool remount)
3283 {
3284         char *this_char, *value, *rest;
3285         struct mempolicy *mpol = NULL;
3286         uid_t uid;
3287         gid_t gid;
3288
3289         while (options != NULL) {
3290                 this_char = options;
3291                 for (;;) {
3292                         /*
3293                          * NUL-terminate this option: unfortunately,
3294                          * mount options form a comma-separated list,
3295                          * but mpol's nodelist may also contain commas.
3296                          */
3297                         options = strchr(options, ',');
3298                         if (options == NULL)
3299                                 break;
3300                         options++;
3301                         if (!isdigit(*options)) {
3302                                 options[-1] = '\0';
3303                                 break;
3304                         }
3305                 }
3306                 if (!*this_char)
3307                         continue;
3308                 if ((value = strchr(this_char,'=')) != NULL) {
3309                         *value++ = 0;
3310                 } else {
3311                         pr_err("tmpfs: No value for mount option '%s'\n",
3312                                this_char);
3313                         goto error;
3314                 }
3315
3316                 if (!strcmp(this_char,"size")) {
3317                         unsigned long long size;
3318                         size = memparse(value,&rest);
3319                         if (*rest == '%') {
3320                                 size <<= PAGE_SHIFT;
3321                                 size *= totalram_pages;
3322                                 do_div(size, 100);
3323                                 rest++;
3324                         }
3325                         if (*rest)
3326                                 goto bad_val;
3327                         sbinfo->max_blocks =
3328                                 DIV_ROUND_UP(size, PAGE_SIZE);
3329                 } else if (!strcmp(this_char,"nr_blocks")) {
3330                         sbinfo->max_blocks = memparse(value, &rest);
3331                         if (*rest)
3332                                 goto bad_val;
3333                 } else if (!strcmp(this_char,"nr_inodes")) {
3334                         sbinfo->max_inodes = memparse(value, &rest);
3335                         if (*rest)
3336                                 goto bad_val;
3337                 } else if (!strcmp(this_char,"mode")) {
3338                         if (remount)
3339                                 continue;
3340                         sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
3341                         if (*rest)
3342                                 goto bad_val;
3343                 } else if (!strcmp(this_char,"uid")) {
3344                         if (remount)
3345                                 continue;
3346                         uid = simple_strtoul(value, &rest, 0);
3347                         if (*rest)
3348                                 goto bad_val;
3349                         sbinfo->uid = make_kuid(current_user_ns(), uid);
3350                         if (!uid_valid(sbinfo->uid))
3351                                 goto bad_val;
3352                 } else if (!strcmp(this_char,"gid")) {
3353                         if (remount)
3354                                 continue;
3355                         gid = simple_strtoul(value, &rest, 0);
3356                         if (*rest)
3357                                 goto bad_val;
3358                         sbinfo->gid = make_kgid(current_user_ns(), gid);
3359                         if (!gid_valid(sbinfo->gid))
3360                                 goto bad_val;
3361 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3362                 } else if (!strcmp(this_char, "huge")) {
3363                         int huge;
3364                         huge = shmem_parse_huge(value);
3365                         if (huge < 0)
3366                                 goto bad_val;
3367                         if (!has_transparent_hugepage() &&
3368                                         huge != SHMEM_HUGE_NEVER)
3369                                 goto bad_val;
3370                         sbinfo->huge = huge;
3371 #endif
3372 #ifdef CONFIG_NUMA
3373                 } else if (!strcmp(this_char,"mpol")) {
3374                         mpol_put(mpol);
3375                         mpol = NULL;
3376                         if (mpol_parse_str(value, &mpol))
3377                                 goto bad_val;
3378 #endif
3379                 } else {
3380                         pr_err("tmpfs: Bad mount option %s\n", this_char);
3381                         goto error;
3382                 }
3383         }
3384         sbinfo->mpol = mpol;
3385         return 0;
3386
3387 bad_val:
3388         pr_err("tmpfs: Bad value '%s' for mount option '%s'\n",
3389                value, this_char);
3390 error:
3391         mpol_put(mpol);
3392         return 1;
3393
3394 }
3395
3396 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
3397 {
3398         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
3399         struct shmem_sb_info config = *sbinfo;
3400         unsigned long inodes;
3401         int error = -EINVAL;
3402
3403         config.mpol = NULL;
3404         if (shmem_parse_options(data, &config, true))
3405                 return error;
3406
3407         spin_lock(&sbinfo->stat_lock);
3408         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
3409         if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
3410                 goto out;
3411         if (config.max_inodes < inodes)
3412                 goto out;
3413         /*
3414          * Those tests disallow limited->unlimited while any are in use;
3415          * but we must separately disallow unlimited->limited, because
3416          * in that case we have no record of how much is already in use.
3417          */
3418         if (config.max_blocks && !sbinfo->max_blocks)
3419                 goto out;
3420         if (config.max_inodes && !sbinfo->max_inodes)
3421                 goto out;
3422
3423         error = 0;
3424         sbinfo->huge = config.huge;
3425         sbinfo->max_blocks  = config.max_blocks;
3426         sbinfo->max_inodes  = config.max_inodes;
3427         sbinfo->free_inodes = config.max_inodes - inodes;
3428
3429         /*
3430          * Preserve previous mempolicy unless mpol remount option was specified.
3431          */
3432         if (config.mpol) {
3433                 mpol_put(sbinfo->mpol);
3434                 sbinfo->mpol = config.mpol;     /* transfers initial ref */
3435         }
3436 out:
3437         spin_unlock(&sbinfo->stat_lock);
3438         return error;
3439 }
3440
3441 static int shmem_show_options(struct seq_file *seq, struct dentry *root)
3442 {
3443         struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
3444
3445         if (sbinfo->max_blocks != shmem_default_max_blocks())
3446                 seq_printf(seq, ",size=%luk",
3447                         sbinfo->max_blocks << (PAGE_SHIFT - 10));
3448         if (sbinfo->max_inodes != shmem_default_max_inodes())
3449                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
3450         if (sbinfo->mode != (0777 | S_ISVTX))
3451                 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
3452         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
3453                 seq_printf(seq, ",uid=%u",
3454                                 from_kuid_munged(&init_user_ns, sbinfo->uid));
3455         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
3456                 seq_printf(seq, ",gid=%u",
3457                                 from_kgid_munged(&init_user_ns, sbinfo->gid));
3458 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3459         /* Rightly or wrongly, show huge mount option unmasked by shmem_huge */
3460         if (sbinfo->huge)
3461                 seq_printf(seq, ",huge=%s", shmem_format_huge(sbinfo->huge));
3462 #endif
3463         shmem_show_mpol(seq, sbinfo->mpol);
3464         return 0;
3465 }
3466
3467 #endif /* CONFIG_TMPFS */
3468
3469 static void shmem_put_super(struct super_block *sb)
3470 {
3471         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
3472
3473         percpu_counter_destroy(&sbinfo->used_blocks);
3474         mpol_put(sbinfo->mpol);
3475         kfree(sbinfo);
3476         sb->s_fs_info = NULL;
3477 }
3478
3479 int shmem_fill_super(struct super_block *sb, void *data, int silent)
3480 {
3481         struct inode *inode;
3482         struct shmem_sb_info *sbinfo;
3483         int err = -ENOMEM;
3484
3485         /* Round up to L1_CACHE_BYTES to resist false sharing */
3486         sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
3487                                 L1_CACHE_BYTES), GFP_KERNEL);
3488         if (!sbinfo)
3489                 return -ENOMEM;
3490
3491         sbinfo->mode = 0777 | S_ISVTX;
3492         sbinfo->uid = current_fsuid();
3493         sbinfo->gid = current_fsgid();
3494         sb->s_fs_info = sbinfo;
3495
3496 #ifdef CONFIG_TMPFS
3497         /*
3498          * Per default we only allow half of the physical ram per
3499          * tmpfs instance, limiting inodes to one per page of lowmem;
3500          * but the internal instance is left unlimited.
3501          */
3502         if (!(sb->s_flags & SB_KERNMOUNT)) {
3503                 sbinfo->max_blocks = shmem_default_max_blocks();
3504                 sbinfo->max_inodes = shmem_default_max_inodes();
3505                 if (shmem_parse_options(data, sbinfo, false)) {
3506                         err = -EINVAL;
3507                         goto failed;
3508                 }
3509         } else {
3510                 sb->s_flags |= SB_NOUSER;
3511         }
3512         sb->s_export_op = &shmem_export_ops;
3513         sb->s_flags |= SB_NOSEC;
3514 #else
3515         sb->s_flags |= SB_NOUSER;
3516 #endif
3517
3518         spin_lock_init(&sbinfo->stat_lock);
3519         if (percpu_counter_init(&sbinfo->used_blocks, 0, GFP_KERNEL))
3520                 goto failed;
3521         sbinfo->free_inodes = sbinfo->max_inodes;
3522         spin_lock_init(&sbinfo->shrinklist_lock);
3523         INIT_LIST_HEAD(&sbinfo->shrinklist);
3524
3525         sb->s_maxbytes = MAX_LFS_FILESIZE;
3526         sb->s_blocksize = PAGE_SIZE;
3527         sb->s_blocksize_bits = PAGE_SHIFT;
3528         sb->s_magic = TMPFS_MAGIC;
3529         sb->s_op = &shmem_ops;
3530         sb->s_time_gran = 1;
3531 #ifdef CONFIG_TMPFS_XATTR
3532         sb->s_xattr = shmem_xattr_handlers;
3533 #endif
3534 #ifdef CONFIG_TMPFS_POSIX_ACL
3535         sb->s_flags |= SB_POSIXACL;
3536 #endif
3537         uuid_gen(&sb->s_uuid);
3538
3539         inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
3540         if (!inode)
3541                 goto failed;
3542         inode->i_uid = sbinfo->uid;
3543         inode->i_gid = sbinfo->gid;
3544         sb->s_root = d_make_root(inode);
3545         if (!sb->s_root)
3546                 goto failed;
3547         return 0;
3548
3549 failed:
3550         shmem_put_super(sb);
3551         return err;
3552 }
3553
3554 static struct kmem_cache *shmem_inode_cachep;
3555
3556 static struct inode *shmem_alloc_inode(struct super_block *sb)
3557 {
3558         struct shmem_inode_info *info;
3559         info = kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
3560         if (!info)
3561                 return NULL;
3562         return &info->vfs_inode;
3563 }
3564
3565 static void shmem_destroy_callback(struct rcu_head *head)
3566 {
3567         struct inode *inode = container_of(head, struct inode, i_rcu);
3568         if (S_ISLNK(inode->i_mode))
3569                 kfree(inode->i_link);
3570         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
3571 }
3572
3573 static void shmem_destroy_inode(struct inode *inode)
3574 {
3575         if (S_ISREG(inode->i_mode))
3576                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
3577         call_rcu(&inode->i_rcu, shmem_destroy_callback);
3578 }
3579
3580 static void shmem_init_inode(void *foo)
3581 {
3582         struct shmem_inode_info *info = foo;
3583         inode_init_once(&info->vfs_inode);
3584 }
3585
3586 static void shmem_init_inodecache(void)
3587 {
3588         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
3589                                 sizeof(struct shmem_inode_info),
3590                                 0, SLAB_PANIC|SLAB_ACCOUNT, shmem_init_inode);
3591 }
3592
3593 static void shmem_destroy_inodecache(void)
3594 {
3595         kmem_cache_destroy(shmem_inode_cachep);
3596 }
3597
3598 static const struct address_space_operations shmem_aops = {
3599         .writepage      = shmem_writepage,
3600         .set_page_dirty = __set_page_dirty_no_writeback,
3601 #ifdef CONFIG_TMPFS
3602         .write_begin    = shmem_write_begin,
3603         .write_end      = shmem_write_end,
3604 #endif
3605 #ifdef CONFIG_MIGRATION
3606         .migratepage    = migrate_page,
3607 #endif
3608         .error_remove_page = generic_error_remove_page,
3609 };
3610
3611 static const struct file_operations shmem_file_operations = {
3612         .mmap           = shmem_mmap,
3613         .get_unmapped_area = shmem_get_unmapped_area,
3614 #ifdef CONFIG_TMPFS
3615         .llseek         = shmem_file_llseek,
3616         .read_iter      = shmem_file_read_iter,
3617         .write_iter     = generic_file_write_iter,
3618         .fsync          = noop_fsync,
3619         .splice_read    = generic_file_splice_read,
3620         .splice_write   = iter_file_splice_write,
3621         .fallocate      = shmem_fallocate,
3622 #endif
3623 };
3624
3625 static const struct inode_operations shmem_inode_operations = {
3626         .getattr        = shmem_getattr,
3627         .setattr        = shmem_setattr,
3628 #ifdef CONFIG_TMPFS_XATTR
3629         .listxattr      = shmem_listxattr,
3630         .set_acl        = simple_set_acl,
3631 #endif
3632 };
3633
3634 static const struct inode_operations shmem_dir_inode_operations = {
3635 #ifdef CONFIG_TMPFS
3636         .create         = shmem_create,
3637         .lookup         = simple_lookup,
3638         .link           = shmem_link,
3639         .unlink         = shmem_unlink,
3640         .symlink        = shmem_symlink,
3641         .mkdir          = shmem_mkdir,
3642         .rmdir          = shmem_rmdir,
3643         .mknod          = shmem_mknod,
3644         .rename         = shmem_rename2,
3645         .tmpfile        = shmem_tmpfile,
3646 #endif
3647 #ifdef CONFIG_TMPFS_XATTR
3648         .listxattr      = shmem_listxattr,
3649 #endif
3650 #ifdef CONFIG_TMPFS_POSIX_ACL
3651         .setattr        = shmem_setattr,
3652         .set_acl        = simple_set_acl,
3653 #endif
3654 };
3655
3656 static const struct inode_operations shmem_special_inode_operations = {
3657 #ifdef CONFIG_TMPFS_XATTR
3658         .listxattr      = shmem_listxattr,
3659 #endif
3660 #ifdef CONFIG_TMPFS_POSIX_ACL
3661         .setattr        = shmem_setattr,
3662         .set_acl        = simple_set_acl,
3663 #endif
3664 };
3665
3666 static const struct super_operations shmem_ops = {
3667         .alloc_inode    = shmem_alloc_inode,
3668         .destroy_inode  = shmem_destroy_inode,
3669 #ifdef CONFIG_TMPFS
3670         .statfs         = shmem_statfs,
3671         .remount_fs     = shmem_remount_fs,
3672         .show_options   = shmem_show_options,
3673 #endif
3674         .evict_inode    = shmem_evict_inode,
3675         .drop_inode     = generic_delete_inode,
3676         .put_super      = shmem_put_super,
3677 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3678         .nr_cached_objects      = shmem_unused_huge_count,
3679         .free_cached_objects    = shmem_unused_huge_scan,
3680 #endif
3681 };
3682
3683 static const struct vm_operations_struct shmem_vm_ops = {
3684         .fault          = shmem_fault,
3685         .map_pages      = filemap_map_pages,
3686 #ifdef CONFIG_NUMA
3687         .set_policy     = shmem_set_policy,
3688         .get_policy     = shmem_get_policy,
3689 #endif
3690 };
3691
3692 static struct dentry *shmem_mount(struct file_system_type *fs_type,
3693         int flags, const char *dev_name, void *data)
3694 {
3695         return mount_nodev(fs_type, flags, data, shmem_fill_super);
3696 }
3697
3698 static struct file_system_type shmem_fs_type = {
3699         .owner          = THIS_MODULE,
3700         .name           = "tmpfs",
3701         .mount          = shmem_mount,
3702         .kill_sb        = kill_litter_super,
3703         .fs_flags       = FS_USERNS_MOUNT,
3704 };
3705
3706 int __init shmem_init(void)
3707 {
3708         int error;
3709
3710         /* If rootfs called this, don't re-init */
3711         if (shmem_inode_cachep)
3712                 return 0;
3713
3714         shmem_init_inodecache();
3715
3716         error = register_filesystem(&shmem_fs_type);
3717         if (error) {
3718                 pr_err("Could not register tmpfs\n");
3719                 goto out2;
3720         }
3721
3722         shm_mnt = kern_mount(&shmem_fs_type);
3723         if (IS_ERR(shm_mnt)) {
3724                 error = PTR_ERR(shm_mnt);
3725                 pr_err("Could not kern_mount tmpfs\n");
3726                 goto out1;
3727         }
3728
3729 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3730         if (has_transparent_hugepage() && shmem_huge > SHMEM_HUGE_DENY)
3731                 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
3732         else
3733                 shmem_huge = 0; /* just in case it was patched */
3734 #endif
3735         return 0;
3736
3737 out1:
3738         unregister_filesystem(&shmem_fs_type);
3739 out2:
3740         shmem_destroy_inodecache();
3741         shm_mnt = ERR_PTR(error);
3742         return error;
3743 }
3744
3745 #if defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE) && defined(CONFIG_SYSFS)
3746 static ssize_t shmem_enabled_show(struct kobject *kobj,
3747                 struct kobj_attribute *attr, char *buf)
3748 {
3749         int values[] = {
3750                 SHMEM_HUGE_ALWAYS,
3751                 SHMEM_HUGE_WITHIN_SIZE,
3752                 SHMEM_HUGE_ADVISE,
3753                 SHMEM_HUGE_NEVER,
3754                 SHMEM_HUGE_DENY,
3755                 SHMEM_HUGE_FORCE,
3756         };
3757         int i, count;
3758
3759         for (i = 0, count = 0; i < ARRAY_SIZE(values); i++) {
3760                 const char *fmt = shmem_huge == values[i] ? "[%s] " : "%s ";
3761
3762                 count += sprintf(buf + count, fmt,
3763                                 shmem_format_huge(values[i]));
3764         }
3765         buf[count - 1] = '\n';
3766         return count;
3767 }
3768
3769 static ssize_t shmem_enabled_store(struct kobject *kobj,
3770                 struct kobj_attribute *attr, const char *buf, size_t count)
3771 {
3772         char tmp[16];
3773         int huge;
3774
3775         if (count + 1 > sizeof(tmp))
3776                 return -EINVAL;
3777         memcpy(tmp, buf, count);
3778         tmp[count] = '\0';
3779         if (count && tmp[count - 1] == '\n')
3780                 tmp[count - 1] = '\0';
3781
3782         huge = shmem_parse_huge(tmp);
3783         if (huge == -EINVAL)
3784                 return -EINVAL;
3785         if (!has_transparent_hugepage() &&
3786                         huge != SHMEM_HUGE_NEVER && huge != SHMEM_HUGE_DENY)
3787                 return -EINVAL;
3788
3789         shmem_huge = huge;
3790         if (shmem_huge > SHMEM_HUGE_DENY)
3791                 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
3792         return count;
3793 }
3794
3795 struct kobj_attribute shmem_enabled_attr =
3796         __ATTR(shmem_enabled, 0644, shmem_enabled_show, shmem_enabled_store);
3797 #endif /* CONFIG_TRANSPARENT_HUGE_PAGECACHE && CONFIG_SYSFS */
3798
3799 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3800 bool shmem_huge_enabled(struct vm_area_struct *vma)
3801 {
3802         struct inode *inode = file_inode(vma->vm_file);
3803         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
3804         loff_t i_size;
3805         pgoff_t off;
3806
3807         if (shmem_huge == SHMEM_HUGE_FORCE)
3808                 return true;
3809         if (shmem_huge == SHMEM_HUGE_DENY)
3810                 return false;
3811         switch (sbinfo->huge) {
3812                 case SHMEM_HUGE_NEVER:
3813                         return false;
3814                 case SHMEM_HUGE_ALWAYS:
3815                         return true;
3816                 case SHMEM_HUGE_WITHIN_SIZE:
3817                         off = round_up(vma->vm_pgoff, HPAGE_PMD_NR);
3818                         i_size = round_up(i_size_read(inode), PAGE_SIZE);
3819                         if (i_size >= HPAGE_PMD_SIZE &&
3820                                         i_size >> PAGE_SHIFT >= off)
3821                                 return true;
3822                         /* fall through */
3823                 case SHMEM_HUGE_ADVISE:
3824                         /* TODO: implement fadvise() hints */
3825                         return (vma->vm_flags & VM_HUGEPAGE);
3826                 default:
3827                         VM_BUG_ON(1);
3828                         return false;
3829         }
3830 }
3831 #endif /* CONFIG_TRANSPARENT_HUGE_PAGECACHE */
3832
3833 #else /* !CONFIG_SHMEM */
3834
3835 /*
3836  * tiny-shmem: simple shmemfs and tmpfs using ramfs code
3837  *
3838  * This is intended for small system where the benefits of the full
3839  * shmem code (swap-backed and resource-limited) are outweighed by
3840  * their complexity. On systems without swap this code should be
3841  * effectively equivalent, but much lighter weight.
3842  */
3843
3844 static struct file_system_type shmem_fs_type = {
3845         .name           = "tmpfs",
3846         .mount          = ramfs_mount,
3847         .kill_sb        = kill_litter_super,
3848         .fs_flags       = FS_USERNS_MOUNT,
3849 };
3850
3851 int __init shmem_init(void)
3852 {
3853         BUG_ON(register_filesystem(&shmem_fs_type) != 0);
3854
3855         shm_mnt = kern_mount(&shmem_fs_type);
3856         BUG_ON(IS_ERR(shm_mnt));
3857
3858         return 0;
3859 }
3860
3861 int shmem_unuse(swp_entry_t swap, struct page *page)
3862 {
3863         return 0;
3864 }
3865
3866 int shmem_lock(struct file *file, int lock, struct user_struct *user)
3867 {
3868         return 0;
3869 }
3870
3871 void shmem_unlock_mapping(struct address_space *mapping)
3872 {
3873 }
3874
3875 #ifdef CONFIG_MMU
3876 unsigned long shmem_get_unmapped_area(struct file *file,
3877                                       unsigned long addr, unsigned long len,
3878                                       unsigned long pgoff, unsigned long flags)
3879 {
3880         return current->mm->get_unmapped_area(file, addr, len, pgoff, flags);
3881 }
3882 #endif
3883
3884 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
3885 {
3886         truncate_inode_pages_range(inode->i_mapping, lstart, lend);
3887 }
3888 EXPORT_SYMBOL_GPL(shmem_truncate_range);
3889
3890 #define shmem_vm_ops                            generic_file_vm_ops
3891 #define shmem_file_operations                   ramfs_file_operations
3892 #define shmem_get_inode(sb, dir, mode, dev, flags)      ramfs_get_inode(sb, dir, mode, dev)
3893 #define shmem_acct_size(flags, size)            0
3894 #define shmem_unacct_size(flags, size)          do {} while (0)
3895
3896 #endif /* CONFIG_SHMEM */
3897
3898 /* common code */
3899
3900 static struct file *__shmem_file_setup(struct vfsmount *mnt, const char *name, loff_t size,
3901                                        unsigned long flags, unsigned int i_flags)
3902 {
3903         struct inode *inode;
3904         struct file *res;
3905
3906         if (IS_ERR(mnt))
3907                 return ERR_CAST(mnt);
3908
3909         if (size < 0 || size > MAX_LFS_FILESIZE)
3910                 return ERR_PTR(-EINVAL);
3911
3912         if (shmem_acct_size(flags, size))
3913                 return ERR_PTR(-ENOMEM);
3914
3915         inode = shmem_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0,
3916                                 flags);
3917         if (unlikely(!inode)) {
3918                 shmem_unacct_size(flags, size);
3919                 return ERR_PTR(-ENOSPC);
3920         }
3921         inode->i_flags |= i_flags;
3922         inode->i_size = size;
3923         clear_nlink(inode);     /* It is unlinked */
3924         res = ERR_PTR(ramfs_nommu_expand_for_mapping(inode, size));
3925         if (!IS_ERR(res))
3926                 res = alloc_file_pseudo(inode, mnt, name, O_RDWR,
3927                                 &shmem_file_operations);
3928         if (IS_ERR(res))
3929                 iput(inode);
3930         return res;
3931 }
3932
3933 /**
3934  * shmem_kernel_file_setup - get an unlinked file living in tmpfs which must be
3935  *      kernel internal.  There will be NO LSM permission checks against the
3936  *      underlying inode.  So users of this interface must do LSM checks at a
3937  *      higher layer.  The users are the big_key and shm implementations.  LSM
3938  *      checks are provided at the key or shm level rather than the inode.
3939  * @name: name for dentry (to be seen in /proc/<pid>/maps
3940  * @size: size to be set for the file
3941  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
3942  */
3943 struct file *shmem_kernel_file_setup(const char *name, loff_t size, unsigned long flags)
3944 {
3945         return __shmem_file_setup(shm_mnt, name, size, flags, S_PRIVATE);
3946 }
3947
3948 /**
3949  * shmem_file_setup - get an unlinked file living in tmpfs
3950  * @name: name for dentry (to be seen in /proc/<pid>/maps
3951  * @size: size to be set for the file
3952  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
3953  */
3954 struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
3955 {
3956         return __shmem_file_setup(shm_mnt, name, size, flags, 0);
3957 }
3958 EXPORT_SYMBOL_GPL(shmem_file_setup);
3959
3960 /**
3961  * shmem_file_setup_with_mnt - get an unlinked file living in tmpfs
3962  * @mnt: the tmpfs mount where the file will be created
3963  * @name: name for dentry (to be seen in /proc/<pid>/maps
3964  * @size: size to be set for the file
3965  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
3966  */
3967 struct file *shmem_file_setup_with_mnt(struct vfsmount *mnt, const char *name,
3968                                        loff_t size, unsigned long flags)
3969 {
3970         return __shmem_file_setup(mnt, name, size, flags, 0);
3971 }
3972 EXPORT_SYMBOL_GPL(shmem_file_setup_with_mnt);
3973
3974 /**
3975  * shmem_zero_setup - setup a shared anonymous mapping
3976  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
3977  */
3978 int shmem_zero_setup(struct vm_area_struct *vma)
3979 {
3980         struct file *file;
3981         loff_t size = vma->vm_end - vma->vm_start;
3982
3983         /*
3984          * Cloning a new file under mmap_sem leads to a lock ordering conflict
3985          * between XFS directory reading and selinux: since this file is only
3986          * accessible to the user through its mapping, use S_PRIVATE flag to
3987          * bypass file security, in the same way as shmem_kernel_file_setup().
3988          */
3989         file = shmem_kernel_file_setup("dev/zero", size, vma->vm_flags);
3990         if (IS_ERR(file))
3991                 return PTR_ERR(file);
3992
3993         if (vma->vm_file)
3994                 fput(vma->vm_file);
3995         vma->vm_file = file;
3996         vma->vm_ops = &shmem_vm_ops;
3997
3998         if (IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE) &&
3999                         ((vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK) <
4000                         (vma->vm_end & HPAGE_PMD_MASK)) {
4001                 khugepaged_enter(vma, vma->vm_flags);
4002         }
4003
4004         return 0;
4005 }
4006
4007 /**
4008  * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
4009  * @mapping:    the page's address_space
4010  * @index:      the page index
4011  * @gfp:        the page allocator flags to use if allocating
4012  *
4013  * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
4014  * with any new page allocations done using the specified allocation flags.
4015  * But read_cache_page_gfp() uses the ->readpage() method: which does not
4016  * suit tmpfs, since it may have pages in swapcache, and needs to find those
4017  * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
4018  *
4019  * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
4020  * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
4021  */
4022 struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
4023                                          pgoff_t index, gfp_t gfp)
4024 {
4025 #ifdef CONFIG_SHMEM
4026         struct inode *inode = mapping->host;
4027         struct page *page;
4028         int error;
4029
4030         BUG_ON(mapping->a_ops != &shmem_aops);
4031         error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE,
4032                                   gfp, NULL, NULL, NULL);
4033         if (error)
4034                 page = ERR_PTR(error);
4035         else
4036                 unlock_page(page);
4037         return page;
4038 #else
4039         /*
4040          * The tiny !SHMEM case uses ramfs without swap
4041          */
4042         return read_cache_page_gfp(mapping, index, gfp);
4043 #endif
4044 }
4045 EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);