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