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