Loading...
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/fileattr.h>
32#include <linux/mm.h>
33#include <linux/random.h>
34#include <linux/sched/signal.h>
35#include <linux/export.h>
36#include <linux/swap.h>
37#include <linux/uio.h>
38#include <linux/hugetlb.h>
39#include <linux/fs_parser.h>
40#include <linux/swapfile.h>
41#include <linux/iversion.h>
42#include "swap.h"
43
44static struct vfsmount *shm_mnt;
45
46#ifdef CONFIG_SHMEM
47/*
48 * This virtual memory filesystem is heavily based on the ramfs. It
49 * extends ramfs by the ability to use swap and honor resource limits
50 * which makes it a completely usable filesystem.
51 */
52
53#include <linux/xattr.h>
54#include <linux/exportfs.h>
55#include <linux/posix_acl.h>
56#include <linux/posix_acl_xattr.h>
57#include <linux/mman.h>
58#include <linux/string.h>
59#include <linux/slab.h>
60#include <linux/backing-dev.h>
61#include <linux/shmem_fs.h>
62#include <linux/writeback.h>
63#include <linux/pagevec.h>
64#include <linux/percpu_counter.h>
65#include <linux/falloc.h>
66#include <linux/splice.h>
67#include <linux/security.h>
68#include <linux/swapops.h>
69#include <linux/mempolicy.h>
70#include <linux/namei.h>
71#include <linux/ctype.h>
72#include <linux/migrate.h>
73#include <linux/highmem.h>
74#include <linux/seq_file.h>
75#include <linux/magic.h>
76#include <linux/syscalls.h>
77#include <linux/fcntl.h>
78#include <uapi/linux/memfd.h>
79#include <linux/userfaultfd_k.h>
80#include <linux/rmap.h>
81#include <linux/uuid.h>
82
83#include <linux/uaccess.h>
84
85#include "internal.h"
86
87#define BLOCKS_PER_PAGE (PAGE_SIZE/512)
88#define VM_ACCT(size) (PAGE_ALIGN(size) >> PAGE_SHIFT)
89
90/* Pretend that each entry is of this size in directory's i_size */
91#define BOGO_DIRENT_SIZE 20
92
93/* Symlink up to this size is kmalloc'ed instead of using a swappable page */
94#define SHORT_SYMLINK_LEN 128
95
96/*
97 * shmem_fallocate communicates with shmem_fault or shmem_writepage via
98 * inode->i_private (with i_rwsem making sure that it has only one user at
99 * a time): we would prefer not to enlarge the shmem inode just for that.
100 */
101struct shmem_falloc {
102 wait_queue_head_t *waitq; /* faults into hole wait for punch to end */
103 pgoff_t start; /* start of range currently being fallocated */
104 pgoff_t next; /* the next page offset to be fallocated */
105 pgoff_t nr_falloced; /* how many new pages have been fallocated */
106 pgoff_t nr_unswapped; /* how often writepage refused to swap out */
107};
108
109struct shmem_options {
110 unsigned long long blocks;
111 unsigned long long inodes;
112 struct mempolicy *mpol;
113 kuid_t uid;
114 kgid_t gid;
115 umode_t mode;
116 bool full_inums;
117 int huge;
118 int seen;
119#define SHMEM_SEEN_BLOCKS 1
120#define SHMEM_SEEN_INODES 2
121#define SHMEM_SEEN_HUGE 4
122#define SHMEM_SEEN_INUMS 8
123};
124
125#ifdef CONFIG_TMPFS
126static unsigned long shmem_default_max_blocks(void)
127{
128 return totalram_pages() / 2;
129}
130
131static unsigned long shmem_default_max_inodes(void)
132{
133 unsigned long nr_pages = totalram_pages();
134
135 return min(nr_pages - totalhigh_pages(), nr_pages / 2);
136}
137#endif
138
139static int shmem_swapin_folio(struct inode *inode, pgoff_t index,
140 struct folio **foliop, enum sgp_type sgp,
141 gfp_t gfp, struct vm_area_struct *vma,
142 vm_fault_t *fault_type);
143
144static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
145{
146 return sb->s_fs_info;
147}
148
149/*
150 * shmem_file_setup pre-accounts the whole fixed size of a VM object,
151 * for shared memory and for shared anonymous (/dev/zero) mappings
152 * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
153 * consistent with the pre-accounting of private mappings ...
154 */
155static inline int shmem_acct_size(unsigned long flags, loff_t size)
156{
157 return (flags & VM_NORESERVE) ?
158 0 : security_vm_enough_memory_mm(current->mm, VM_ACCT(size));
159}
160
161static inline void shmem_unacct_size(unsigned long flags, loff_t size)
162{
163 if (!(flags & VM_NORESERVE))
164 vm_unacct_memory(VM_ACCT(size));
165}
166
167static inline int shmem_reacct_size(unsigned long flags,
168 loff_t oldsize, loff_t newsize)
169{
170 if (!(flags & VM_NORESERVE)) {
171 if (VM_ACCT(newsize) > VM_ACCT(oldsize))
172 return security_vm_enough_memory_mm(current->mm,
173 VM_ACCT(newsize) - VM_ACCT(oldsize));
174 else if (VM_ACCT(newsize) < VM_ACCT(oldsize))
175 vm_unacct_memory(VM_ACCT(oldsize) - VM_ACCT(newsize));
176 }
177 return 0;
178}
179
180/*
181 * ... whereas tmpfs objects are accounted incrementally as
182 * pages are allocated, in order to allow large sparse files.
183 * shmem_get_folio reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
184 * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
185 */
186static inline int shmem_acct_block(unsigned long flags, long pages)
187{
188 if (!(flags & VM_NORESERVE))
189 return 0;
190
191 return security_vm_enough_memory_mm(current->mm,
192 pages * VM_ACCT(PAGE_SIZE));
193}
194
195static inline void shmem_unacct_blocks(unsigned long flags, long pages)
196{
197 if (flags & VM_NORESERVE)
198 vm_unacct_memory(pages * VM_ACCT(PAGE_SIZE));
199}
200
201static inline bool shmem_inode_acct_block(struct inode *inode, long pages)
202{
203 struct shmem_inode_info *info = SHMEM_I(inode);
204 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
205
206 if (shmem_acct_block(info->flags, pages))
207 return false;
208
209 if (sbinfo->max_blocks) {
210 if (percpu_counter_compare(&sbinfo->used_blocks,
211 sbinfo->max_blocks - pages) > 0)
212 goto unacct;
213 percpu_counter_add(&sbinfo->used_blocks, pages);
214 }
215
216 return true;
217
218unacct:
219 shmem_unacct_blocks(info->flags, pages);
220 return false;
221}
222
223static inline void shmem_inode_unacct_blocks(struct inode *inode, long pages)
224{
225 struct shmem_inode_info *info = SHMEM_I(inode);
226 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
227
228 if (sbinfo->max_blocks)
229 percpu_counter_sub(&sbinfo->used_blocks, pages);
230 shmem_unacct_blocks(info->flags, pages);
231}
232
233static const struct super_operations shmem_ops;
234const struct address_space_operations shmem_aops;
235static const struct file_operations shmem_file_operations;
236static const struct inode_operations shmem_inode_operations;
237static const struct inode_operations shmem_dir_inode_operations;
238static const struct inode_operations shmem_special_inode_operations;
239static const struct vm_operations_struct shmem_vm_ops;
240static const struct vm_operations_struct shmem_anon_vm_ops;
241static struct file_system_type shmem_fs_type;
242
243bool vma_is_anon_shmem(struct vm_area_struct *vma)
244{
245 return vma->vm_ops == &shmem_anon_vm_ops;
246}
247
248bool vma_is_shmem(struct vm_area_struct *vma)
249{
250 return vma_is_anon_shmem(vma) || vma->vm_ops == &shmem_vm_ops;
251}
252
253static LIST_HEAD(shmem_swaplist);
254static DEFINE_MUTEX(shmem_swaplist_mutex);
255
256/*
257 * shmem_reserve_inode() performs bookkeeping to reserve a shmem inode, and
258 * produces a novel ino for the newly allocated inode.
259 *
260 * It may also be called when making a hard link to permit the space needed by
261 * each dentry. However, in that case, no new inode number is needed since that
262 * internally draws from another pool of inode numbers (currently global
263 * get_next_ino()). This case is indicated by passing NULL as inop.
264 */
265#define SHMEM_INO_BATCH 1024
266static int shmem_reserve_inode(struct super_block *sb, ino_t *inop)
267{
268 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
269 ino_t ino;
270
271 if (!(sb->s_flags & SB_KERNMOUNT)) {
272 raw_spin_lock(&sbinfo->stat_lock);
273 if (sbinfo->max_inodes) {
274 if (!sbinfo->free_inodes) {
275 raw_spin_unlock(&sbinfo->stat_lock);
276 return -ENOSPC;
277 }
278 sbinfo->free_inodes--;
279 }
280 if (inop) {
281 ino = sbinfo->next_ino++;
282 if (unlikely(is_zero_ino(ino)))
283 ino = sbinfo->next_ino++;
284 if (unlikely(!sbinfo->full_inums &&
285 ino > UINT_MAX)) {
286 /*
287 * Emulate get_next_ino uint wraparound for
288 * compatibility
289 */
290 if (IS_ENABLED(CONFIG_64BIT))
291 pr_warn("%s: inode number overflow on device %d, consider using inode64 mount option\n",
292 __func__, MINOR(sb->s_dev));
293 sbinfo->next_ino = 1;
294 ino = sbinfo->next_ino++;
295 }
296 *inop = ino;
297 }
298 raw_spin_unlock(&sbinfo->stat_lock);
299 } else if (inop) {
300 /*
301 * __shmem_file_setup, one of our callers, is lock-free: it
302 * doesn't hold stat_lock in shmem_reserve_inode since
303 * max_inodes is always 0, and is called from potentially
304 * unknown contexts. As such, use a per-cpu batched allocator
305 * which doesn't require the per-sb stat_lock unless we are at
306 * the batch boundary.
307 *
308 * We don't need to worry about inode{32,64} since SB_KERNMOUNT
309 * shmem mounts are not exposed to userspace, so we don't need
310 * to worry about things like glibc compatibility.
311 */
312 ino_t *next_ino;
313
314 next_ino = per_cpu_ptr(sbinfo->ino_batch, get_cpu());
315 ino = *next_ino;
316 if (unlikely(ino % SHMEM_INO_BATCH == 0)) {
317 raw_spin_lock(&sbinfo->stat_lock);
318 ino = sbinfo->next_ino;
319 sbinfo->next_ino += SHMEM_INO_BATCH;
320 raw_spin_unlock(&sbinfo->stat_lock);
321 if (unlikely(is_zero_ino(ino)))
322 ino++;
323 }
324 *inop = ino;
325 *next_ino = ++ino;
326 put_cpu();
327 }
328
329 return 0;
330}
331
332static void shmem_free_inode(struct super_block *sb)
333{
334 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
335 if (sbinfo->max_inodes) {
336 raw_spin_lock(&sbinfo->stat_lock);
337 sbinfo->free_inodes++;
338 raw_spin_unlock(&sbinfo->stat_lock);
339 }
340}
341
342/**
343 * shmem_recalc_inode - recalculate the block usage of an inode
344 * @inode: inode to recalc
345 *
346 * We have to calculate the free blocks since the mm can drop
347 * undirtied hole pages behind our back.
348 *
349 * But normally info->alloced == inode->i_mapping->nrpages + info->swapped
350 * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
351 *
352 * It has to be called with the spinlock held.
353 */
354static void shmem_recalc_inode(struct inode *inode)
355{
356 struct shmem_inode_info *info = SHMEM_I(inode);
357 long freed;
358
359 freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
360 if (freed > 0) {
361 info->alloced -= freed;
362 inode->i_blocks -= freed * BLOCKS_PER_PAGE;
363 shmem_inode_unacct_blocks(inode, freed);
364 }
365}
366
367bool shmem_charge(struct inode *inode, long pages)
368{
369 struct shmem_inode_info *info = SHMEM_I(inode);
370 unsigned long flags;
371
372 if (!shmem_inode_acct_block(inode, pages))
373 return false;
374
375 /* nrpages adjustment first, then shmem_recalc_inode() when balanced */
376 inode->i_mapping->nrpages += pages;
377
378 spin_lock_irqsave(&info->lock, flags);
379 info->alloced += pages;
380 inode->i_blocks += pages * BLOCKS_PER_PAGE;
381 shmem_recalc_inode(inode);
382 spin_unlock_irqrestore(&info->lock, flags);
383
384 return true;
385}
386
387void shmem_uncharge(struct inode *inode, long pages)
388{
389 struct shmem_inode_info *info = SHMEM_I(inode);
390 unsigned long flags;
391
392 /* nrpages adjustment done by __filemap_remove_folio() or caller */
393
394 spin_lock_irqsave(&info->lock, flags);
395 info->alloced -= pages;
396 inode->i_blocks -= pages * BLOCKS_PER_PAGE;
397 shmem_recalc_inode(inode);
398 spin_unlock_irqrestore(&info->lock, flags);
399
400 shmem_inode_unacct_blocks(inode, pages);
401}
402
403/*
404 * Replace item expected in xarray by a new item, while holding xa_lock.
405 */
406static int shmem_replace_entry(struct address_space *mapping,
407 pgoff_t index, void *expected, void *replacement)
408{
409 XA_STATE(xas, &mapping->i_pages, index);
410 void *item;
411
412 VM_BUG_ON(!expected);
413 VM_BUG_ON(!replacement);
414 item = xas_load(&xas);
415 if (item != expected)
416 return -ENOENT;
417 xas_store(&xas, replacement);
418 return 0;
419}
420
421/*
422 * Sometimes, before we decide whether to proceed or to fail, we must check
423 * that an entry was not already brought back from swap by a racing thread.
424 *
425 * Checking page is not enough: by the time a SwapCache page is locked, it
426 * might be reused, and again be SwapCache, using the same swap as before.
427 */
428static bool shmem_confirm_swap(struct address_space *mapping,
429 pgoff_t index, swp_entry_t swap)
430{
431 return xa_load(&mapping->i_pages, index) == swp_to_radix_entry(swap);
432}
433
434/*
435 * Definitions for "huge tmpfs": tmpfs mounted with the huge= option
436 *
437 * SHMEM_HUGE_NEVER:
438 * disables huge pages for the mount;
439 * SHMEM_HUGE_ALWAYS:
440 * enables huge pages for the mount;
441 * SHMEM_HUGE_WITHIN_SIZE:
442 * only allocate huge pages if the page will be fully within i_size,
443 * also respect fadvise()/madvise() hints;
444 * SHMEM_HUGE_ADVISE:
445 * only allocate huge pages if requested with fadvise()/madvise();
446 */
447
448#define SHMEM_HUGE_NEVER 0
449#define SHMEM_HUGE_ALWAYS 1
450#define SHMEM_HUGE_WITHIN_SIZE 2
451#define SHMEM_HUGE_ADVISE 3
452
453/*
454 * Special values.
455 * Only can be set via /sys/kernel/mm/transparent_hugepage/shmem_enabled:
456 *
457 * SHMEM_HUGE_DENY:
458 * disables huge on shm_mnt and all mounts, for emergency use;
459 * SHMEM_HUGE_FORCE:
460 * enables huge on shm_mnt and all mounts, w/o needing option, for testing;
461 *
462 */
463#define SHMEM_HUGE_DENY (-1)
464#define SHMEM_HUGE_FORCE (-2)
465
466#ifdef CONFIG_TRANSPARENT_HUGEPAGE
467/* ifdef here to avoid bloating shmem.o when not necessary */
468
469static int shmem_huge __read_mostly = SHMEM_HUGE_NEVER;
470
471bool shmem_is_huge(struct vm_area_struct *vma, struct inode *inode,
472 pgoff_t index, bool shmem_huge_force)
473{
474 loff_t i_size;
475
476 if (!S_ISREG(inode->i_mode))
477 return false;
478 if (vma && ((vma->vm_flags & VM_NOHUGEPAGE) ||
479 test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags)))
480 return false;
481 if (shmem_huge == SHMEM_HUGE_DENY)
482 return false;
483 if (shmem_huge_force || shmem_huge == SHMEM_HUGE_FORCE)
484 return true;
485
486 switch (SHMEM_SB(inode->i_sb)->huge) {
487 case SHMEM_HUGE_ALWAYS:
488 return true;
489 case SHMEM_HUGE_WITHIN_SIZE:
490 index = round_up(index + 1, HPAGE_PMD_NR);
491 i_size = round_up(i_size_read(inode), PAGE_SIZE);
492 if (i_size >> PAGE_SHIFT >= index)
493 return true;
494 fallthrough;
495 case SHMEM_HUGE_ADVISE:
496 if (vma && (vma->vm_flags & VM_HUGEPAGE))
497 return true;
498 fallthrough;
499 default:
500 return false;
501 }
502}
503
504#if defined(CONFIG_SYSFS)
505static int shmem_parse_huge(const char *str)
506{
507 if (!strcmp(str, "never"))
508 return SHMEM_HUGE_NEVER;
509 if (!strcmp(str, "always"))
510 return SHMEM_HUGE_ALWAYS;
511 if (!strcmp(str, "within_size"))
512 return SHMEM_HUGE_WITHIN_SIZE;
513 if (!strcmp(str, "advise"))
514 return SHMEM_HUGE_ADVISE;
515 if (!strcmp(str, "deny"))
516 return SHMEM_HUGE_DENY;
517 if (!strcmp(str, "force"))
518 return SHMEM_HUGE_FORCE;
519 return -EINVAL;
520}
521#endif
522
523#if defined(CONFIG_SYSFS) || defined(CONFIG_TMPFS)
524static const char *shmem_format_huge(int huge)
525{
526 switch (huge) {
527 case SHMEM_HUGE_NEVER:
528 return "never";
529 case SHMEM_HUGE_ALWAYS:
530 return "always";
531 case SHMEM_HUGE_WITHIN_SIZE:
532 return "within_size";
533 case SHMEM_HUGE_ADVISE:
534 return "advise";
535 case SHMEM_HUGE_DENY:
536 return "deny";
537 case SHMEM_HUGE_FORCE:
538 return "force";
539 default:
540 VM_BUG_ON(1);
541 return "bad_val";
542 }
543}
544#endif
545
546static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
547 struct shrink_control *sc, unsigned long nr_to_split)
548{
549 LIST_HEAD(list), *pos, *next;
550 LIST_HEAD(to_remove);
551 struct inode *inode;
552 struct shmem_inode_info *info;
553 struct folio *folio;
554 unsigned long batch = sc ? sc->nr_to_scan : 128;
555 int split = 0;
556
557 if (list_empty(&sbinfo->shrinklist))
558 return SHRINK_STOP;
559
560 spin_lock(&sbinfo->shrinklist_lock);
561 list_for_each_safe(pos, next, &sbinfo->shrinklist) {
562 info = list_entry(pos, struct shmem_inode_info, shrinklist);
563
564 /* pin the inode */
565 inode = igrab(&info->vfs_inode);
566
567 /* inode is about to be evicted */
568 if (!inode) {
569 list_del_init(&info->shrinklist);
570 goto next;
571 }
572
573 /* Check if there's anything to gain */
574 if (round_up(inode->i_size, PAGE_SIZE) ==
575 round_up(inode->i_size, HPAGE_PMD_SIZE)) {
576 list_move(&info->shrinklist, &to_remove);
577 goto next;
578 }
579
580 list_move(&info->shrinklist, &list);
581next:
582 sbinfo->shrinklist_len--;
583 if (!--batch)
584 break;
585 }
586 spin_unlock(&sbinfo->shrinklist_lock);
587
588 list_for_each_safe(pos, next, &to_remove) {
589 info = list_entry(pos, struct shmem_inode_info, shrinklist);
590 inode = &info->vfs_inode;
591 list_del_init(&info->shrinklist);
592 iput(inode);
593 }
594
595 list_for_each_safe(pos, next, &list) {
596 int ret;
597 pgoff_t index;
598
599 info = list_entry(pos, struct shmem_inode_info, shrinklist);
600 inode = &info->vfs_inode;
601
602 if (nr_to_split && split >= nr_to_split)
603 goto move_back;
604
605 index = (inode->i_size & HPAGE_PMD_MASK) >> PAGE_SHIFT;
606 folio = filemap_get_folio(inode->i_mapping, index);
607 if (!folio)
608 goto drop;
609
610 /* No huge page at the end of the file: nothing to split */
611 if (!folio_test_large(folio)) {
612 folio_put(folio);
613 goto drop;
614 }
615
616 /*
617 * Move the inode on the list back to shrinklist if we failed
618 * to lock the page at this time.
619 *
620 * Waiting for the lock may lead to deadlock in the
621 * reclaim path.
622 */
623 if (!folio_trylock(folio)) {
624 folio_put(folio);
625 goto move_back;
626 }
627
628 ret = split_folio(folio);
629 folio_unlock(folio);
630 folio_put(folio);
631
632 /* If split failed move the inode on the list back to shrinklist */
633 if (ret)
634 goto move_back;
635
636 split++;
637drop:
638 list_del_init(&info->shrinklist);
639 goto put;
640move_back:
641 /*
642 * Make sure the inode is either on the global list or deleted
643 * from any local list before iput() since it could be deleted
644 * in another thread once we put the inode (then the local list
645 * is corrupted).
646 */
647 spin_lock(&sbinfo->shrinklist_lock);
648 list_move(&info->shrinklist, &sbinfo->shrinklist);
649 sbinfo->shrinklist_len++;
650 spin_unlock(&sbinfo->shrinklist_lock);
651put:
652 iput(inode);
653 }
654
655 return split;
656}
657
658static long shmem_unused_huge_scan(struct super_block *sb,
659 struct shrink_control *sc)
660{
661 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
662
663 if (!READ_ONCE(sbinfo->shrinklist_len))
664 return SHRINK_STOP;
665
666 return shmem_unused_huge_shrink(sbinfo, sc, 0);
667}
668
669static long shmem_unused_huge_count(struct super_block *sb,
670 struct shrink_control *sc)
671{
672 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
673 return READ_ONCE(sbinfo->shrinklist_len);
674}
675#else /* !CONFIG_TRANSPARENT_HUGEPAGE */
676
677#define shmem_huge SHMEM_HUGE_DENY
678
679bool shmem_is_huge(struct vm_area_struct *vma, struct inode *inode,
680 pgoff_t index, bool shmem_huge_force)
681{
682 return false;
683}
684
685static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
686 struct shrink_control *sc, unsigned long nr_to_split)
687{
688 return 0;
689}
690#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
691
692/*
693 * Like filemap_add_folio, but error if expected item has gone.
694 */
695static int shmem_add_to_page_cache(struct folio *folio,
696 struct address_space *mapping,
697 pgoff_t index, void *expected, gfp_t gfp,
698 struct mm_struct *charge_mm)
699{
700 XA_STATE_ORDER(xas, &mapping->i_pages, index, folio_order(folio));
701 long nr = folio_nr_pages(folio);
702 int error;
703
704 VM_BUG_ON_FOLIO(index != round_down(index, nr), folio);
705 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
706 VM_BUG_ON_FOLIO(!folio_test_swapbacked(folio), folio);
707 VM_BUG_ON(expected && folio_test_large(folio));
708
709 folio_ref_add(folio, nr);
710 folio->mapping = mapping;
711 folio->index = index;
712
713 if (!folio_test_swapcache(folio)) {
714 error = mem_cgroup_charge(folio, charge_mm, gfp);
715 if (error) {
716 if (folio_test_pmd_mappable(folio)) {
717 count_vm_event(THP_FILE_FALLBACK);
718 count_vm_event(THP_FILE_FALLBACK_CHARGE);
719 }
720 goto error;
721 }
722 }
723 folio_throttle_swaprate(folio, gfp);
724
725 do {
726 xas_lock_irq(&xas);
727 if (expected != xas_find_conflict(&xas)) {
728 xas_set_err(&xas, -EEXIST);
729 goto unlock;
730 }
731 if (expected && xas_find_conflict(&xas)) {
732 xas_set_err(&xas, -EEXIST);
733 goto unlock;
734 }
735 xas_store(&xas, folio);
736 if (xas_error(&xas))
737 goto unlock;
738 if (folio_test_pmd_mappable(folio)) {
739 count_vm_event(THP_FILE_ALLOC);
740 __lruvec_stat_mod_folio(folio, NR_SHMEM_THPS, nr);
741 }
742 mapping->nrpages += nr;
743 __lruvec_stat_mod_folio(folio, NR_FILE_PAGES, nr);
744 __lruvec_stat_mod_folio(folio, NR_SHMEM, nr);
745unlock:
746 xas_unlock_irq(&xas);
747 } while (xas_nomem(&xas, gfp));
748
749 if (xas_error(&xas)) {
750 error = xas_error(&xas);
751 goto error;
752 }
753
754 return 0;
755error:
756 folio->mapping = NULL;
757 folio_ref_sub(folio, nr);
758 return error;
759}
760
761/*
762 * Like delete_from_page_cache, but substitutes swap for @folio.
763 */
764static void shmem_delete_from_page_cache(struct folio *folio, void *radswap)
765{
766 struct address_space *mapping = folio->mapping;
767 long nr = folio_nr_pages(folio);
768 int error;
769
770 xa_lock_irq(&mapping->i_pages);
771 error = shmem_replace_entry(mapping, folio->index, folio, radswap);
772 folio->mapping = NULL;
773 mapping->nrpages -= nr;
774 __lruvec_stat_mod_folio(folio, NR_FILE_PAGES, -nr);
775 __lruvec_stat_mod_folio(folio, NR_SHMEM, -nr);
776 xa_unlock_irq(&mapping->i_pages);
777 folio_put(folio);
778 BUG_ON(error);
779}
780
781/*
782 * Remove swap entry from page cache, free the swap and its page cache.
783 */
784static int shmem_free_swap(struct address_space *mapping,
785 pgoff_t index, void *radswap)
786{
787 void *old;
788
789 old = xa_cmpxchg_irq(&mapping->i_pages, index, radswap, NULL, 0);
790 if (old != radswap)
791 return -ENOENT;
792 free_swap_and_cache(radix_to_swp_entry(radswap));
793 return 0;
794}
795
796/*
797 * Determine (in bytes) how many of the shmem object's pages mapped by the
798 * given offsets are swapped out.
799 *
800 * This is safe to call without i_rwsem or the i_pages lock thanks to RCU,
801 * as long as the inode doesn't go away and racy results are not a problem.
802 */
803unsigned long shmem_partial_swap_usage(struct address_space *mapping,
804 pgoff_t start, pgoff_t end)
805{
806 XA_STATE(xas, &mapping->i_pages, start);
807 struct page *page;
808 unsigned long swapped = 0;
809
810 rcu_read_lock();
811 xas_for_each(&xas, page, end - 1) {
812 if (xas_retry(&xas, page))
813 continue;
814 if (xa_is_value(page))
815 swapped++;
816
817 if (need_resched()) {
818 xas_pause(&xas);
819 cond_resched_rcu();
820 }
821 }
822
823 rcu_read_unlock();
824
825 return swapped << PAGE_SHIFT;
826}
827
828/*
829 * Determine (in bytes) how many of the shmem object's pages mapped by the
830 * given vma is swapped out.
831 *
832 * This is safe to call without i_rwsem or the i_pages lock thanks to RCU,
833 * as long as the inode doesn't go away and racy results are not a problem.
834 */
835unsigned long shmem_swap_usage(struct vm_area_struct *vma)
836{
837 struct inode *inode = file_inode(vma->vm_file);
838 struct shmem_inode_info *info = SHMEM_I(inode);
839 struct address_space *mapping = inode->i_mapping;
840 unsigned long swapped;
841
842 /* Be careful as we don't hold info->lock */
843 swapped = READ_ONCE(info->swapped);
844
845 /*
846 * The easier cases are when the shmem object has nothing in swap, or
847 * the vma maps it whole. Then we can simply use the stats that we
848 * already track.
849 */
850 if (!swapped)
851 return 0;
852
853 if (!vma->vm_pgoff && vma->vm_end - vma->vm_start >= inode->i_size)
854 return swapped << PAGE_SHIFT;
855
856 /* Here comes the more involved part */
857 return shmem_partial_swap_usage(mapping, vma->vm_pgoff,
858 vma->vm_pgoff + vma_pages(vma));
859}
860
861/*
862 * SysV IPC SHM_UNLOCK restore Unevictable pages to their evictable lists.
863 */
864void shmem_unlock_mapping(struct address_space *mapping)
865{
866 struct folio_batch fbatch;
867 pgoff_t index = 0;
868
869 folio_batch_init(&fbatch);
870 /*
871 * Minor point, but we might as well stop if someone else SHM_LOCKs it.
872 */
873 while (!mapping_unevictable(mapping) &&
874 filemap_get_folios(mapping, &index, ~0UL, &fbatch)) {
875 check_move_unevictable_folios(&fbatch);
876 folio_batch_release(&fbatch);
877 cond_resched();
878 }
879}
880
881static struct folio *shmem_get_partial_folio(struct inode *inode, pgoff_t index)
882{
883 struct folio *folio;
884
885 /*
886 * At first avoid shmem_get_folio(,,,SGP_READ): that fails
887 * beyond i_size, and reports fallocated pages as holes.
888 */
889 folio = __filemap_get_folio(inode->i_mapping, index,
890 FGP_ENTRY | FGP_LOCK, 0);
891 if (!xa_is_value(folio))
892 return folio;
893 /*
894 * But read a page back from swap if any of it is within i_size
895 * (although in some cases this is just a waste of time).
896 */
897 folio = NULL;
898 shmem_get_folio(inode, index, &folio, SGP_READ);
899 return folio;
900}
901
902/*
903 * Remove range of pages and swap entries from page cache, and free them.
904 * If !unfalloc, truncate or punch hole; if unfalloc, undo failed fallocate.
905 */
906static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
907 bool unfalloc)
908{
909 struct address_space *mapping = inode->i_mapping;
910 struct shmem_inode_info *info = SHMEM_I(inode);
911 pgoff_t start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
912 pgoff_t end = (lend + 1) >> PAGE_SHIFT;
913 struct folio_batch fbatch;
914 pgoff_t indices[PAGEVEC_SIZE];
915 struct folio *folio;
916 bool same_folio;
917 long nr_swaps_freed = 0;
918 pgoff_t index;
919 int i;
920
921 if (lend == -1)
922 end = -1; /* unsigned, so actually very big */
923
924 if (info->fallocend > start && info->fallocend <= end && !unfalloc)
925 info->fallocend = start;
926
927 folio_batch_init(&fbatch);
928 index = start;
929 while (index < end && find_lock_entries(mapping, &index, end - 1,
930 &fbatch, indices)) {
931 for (i = 0; i < folio_batch_count(&fbatch); i++) {
932 folio = fbatch.folios[i];
933
934 if (xa_is_value(folio)) {
935 if (unfalloc)
936 continue;
937 nr_swaps_freed += !shmem_free_swap(mapping,
938 indices[i], folio);
939 continue;
940 }
941
942 if (!unfalloc || !folio_test_uptodate(folio))
943 truncate_inode_folio(mapping, folio);
944 folio_unlock(folio);
945 }
946 folio_batch_remove_exceptionals(&fbatch);
947 folio_batch_release(&fbatch);
948 cond_resched();
949 }
950
951 /*
952 * When undoing a failed fallocate, we want none of the partial folio
953 * zeroing and splitting below, but shall want to truncate the whole
954 * folio when !uptodate indicates that it was added by this fallocate,
955 * even when [lstart, lend] covers only a part of the folio.
956 */
957 if (unfalloc)
958 goto whole_folios;
959
960 same_folio = (lstart >> PAGE_SHIFT) == (lend >> PAGE_SHIFT);
961 folio = shmem_get_partial_folio(inode, lstart >> PAGE_SHIFT);
962 if (folio) {
963 same_folio = lend < folio_pos(folio) + folio_size(folio);
964 folio_mark_dirty(folio);
965 if (!truncate_inode_partial_folio(folio, lstart, lend)) {
966 start = folio->index + folio_nr_pages(folio);
967 if (same_folio)
968 end = folio->index;
969 }
970 folio_unlock(folio);
971 folio_put(folio);
972 folio = NULL;
973 }
974
975 if (!same_folio)
976 folio = shmem_get_partial_folio(inode, lend >> PAGE_SHIFT);
977 if (folio) {
978 folio_mark_dirty(folio);
979 if (!truncate_inode_partial_folio(folio, lstart, lend))
980 end = folio->index;
981 folio_unlock(folio);
982 folio_put(folio);
983 }
984
985whole_folios:
986
987 index = start;
988 while (index < end) {
989 cond_resched();
990
991 if (!find_get_entries(mapping, &index, end - 1, &fbatch,
992 indices)) {
993 /* If all gone or hole-punch or unfalloc, we're done */
994 if (index == start || end != -1)
995 break;
996 /* But if truncating, restart to make sure all gone */
997 index = start;
998 continue;
999 }
1000 for (i = 0; i < folio_batch_count(&fbatch); i++) {
1001 folio = fbatch.folios[i];
1002
1003 if (xa_is_value(folio)) {
1004 if (unfalloc)
1005 continue;
1006 if (shmem_free_swap(mapping, indices[i], folio)) {
1007 /* Swap was replaced by page: retry */
1008 index = indices[i];
1009 break;
1010 }
1011 nr_swaps_freed++;
1012 continue;
1013 }
1014
1015 folio_lock(folio);
1016
1017 if (!unfalloc || !folio_test_uptodate(folio)) {
1018 if (folio_mapping(folio) != mapping) {
1019 /* Page was replaced by swap: retry */
1020 folio_unlock(folio);
1021 index = indices[i];
1022 break;
1023 }
1024 VM_BUG_ON_FOLIO(folio_test_writeback(folio),
1025 folio);
1026 truncate_inode_folio(mapping, folio);
1027 }
1028 folio_unlock(folio);
1029 }
1030 folio_batch_remove_exceptionals(&fbatch);
1031 folio_batch_release(&fbatch);
1032 }
1033
1034 spin_lock_irq(&info->lock);
1035 info->swapped -= nr_swaps_freed;
1036 shmem_recalc_inode(inode);
1037 spin_unlock_irq(&info->lock);
1038}
1039
1040void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
1041{
1042 shmem_undo_range(inode, lstart, lend, false);
1043 inode->i_ctime = inode->i_mtime = current_time(inode);
1044 inode_inc_iversion(inode);
1045}
1046EXPORT_SYMBOL_GPL(shmem_truncate_range);
1047
1048static int shmem_getattr(struct user_namespace *mnt_userns,
1049 const struct path *path, struct kstat *stat,
1050 u32 request_mask, unsigned int query_flags)
1051{
1052 struct inode *inode = path->dentry->d_inode;
1053 struct shmem_inode_info *info = SHMEM_I(inode);
1054
1055 if (info->alloced - info->swapped != inode->i_mapping->nrpages) {
1056 spin_lock_irq(&info->lock);
1057 shmem_recalc_inode(inode);
1058 spin_unlock_irq(&info->lock);
1059 }
1060 if (info->fsflags & FS_APPEND_FL)
1061 stat->attributes |= STATX_ATTR_APPEND;
1062 if (info->fsflags & FS_IMMUTABLE_FL)
1063 stat->attributes |= STATX_ATTR_IMMUTABLE;
1064 if (info->fsflags & FS_NODUMP_FL)
1065 stat->attributes |= STATX_ATTR_NODUMP;
1066 stat->attributes_mask |= (STATX_ATTR_APPEND |
1067 STATX_ATTR_IMMUTABLE |
1068 STATX_ATTR_NODUMP);
1069 generic_fillattr(&init_user_ns, inode, stat);
1070
1071 if (shmem_is_huge(NULL, inode, 0, false))
1072 stat->blksize = HPAGE_PMD_SIZE;
1073
1074 if (request_mask & STATX_BTIME) {
1075 stat->result_mask |= STATX_BTIME;
1076 stat->btime.tv_sec = info->i_crtime.tv_sec;
1077 stat->btime.tv_nsec = info->i_crtime.tv_nsec;
1078 }
1079
1080 return 0;
1081}
1082
1083static int shmem_setattr(struct user_namespace *mnt_userns,
1084 struct dentry *dentry, struct iattr *attr)
1085{
1086 struct inode *inode = d_inode(dentry);
1087 struct shmem_inode_info *info = SHMEM_I(inode);
1088 int error;
1089 bool update_mtime = false;
1090 bool update_ctime = true;
1091
1092 error = setattr_prepare(&init_user_ns, dentry, attr);
1093 if (error)
1094 return error;
1095
1096 if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
1097 loff_t oldsize = inode->i_size;
1098 loff_t newsize = attr->ia_size;
1099
1100 /* protected by i_rwsem */
1101 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
1102 (newsize > oldsize && (info->seals & F_SEAL_GROW)))
1103 return -EPERM;
1104
1105 if (newsize != oldsize) {
1106 error = shmem_reacct_size(SHMEM_I(inode)->flags,
1107 oldsize, newsize);
1108 if (error)
1109 return error;
1110 i_size_write(inode, newsize);
1111 update_mtime = true;
1112 } else {
1113 update_ctime = false;
1114 }
1115 if (newsize <= oldsize) {
1116 loff_t holebegin = round_up(newsize, PAGE_SIZE);
1117 if (oldsize > holebegin)
1118 unmap_mapping_range(inode->i_mapping,
1119 holebegin, 0, 1);
1120 if (info->alloced)
1121 shmem_truncate_range(inode,
1122 newsize, (loff_t)-1);
1123 /* unmap again to remove racily COWed private pages */
1124 if (oldsize > holebegin)
1125 unmap_mapping_range(inode->i_mapping,
1126 holebegin, 0, 1);
1127 }
1128 }
1129
1130 setattr_copy(&init_user_ns, inode, attr);
1131 if (attr->ia_valid & ATTR_MODE)
1132 error = posix_acl_chmod(&init_user_ns, dentry, inode->i_mode);
1133 if (!error && update_ctime) {
1134 inode->i_ctime = current_time(inode);
1135 if (update_mtime)
1136 inode->i_mtime = inode->i_ctime;
1137 inode_inc_iversion(inode);
1138 }
1139 return error;
1140}
1141
1142static void shmem_evict_inode(struct inode *inode)
1143{
1144 struct shmem_inode_info *info = SHMEM_I(inode);
1145 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1146
1147 if (shmem_mapping(inode->i_mapping)) {
1148 shmem_unacct_size(info->flags, inode->i_size);
1149 inode->i_size = 0;
1150 mapping_set_exiting(inode->i_mapping);
1151 shmem_truncate_range(inode, 0, (loff_t)-1);
1152 if (!list_empty(&info->shrinklist)) {
1153 spin_lock(&sbinfo->shrinklist_lock);
1154 if (!list_empty(&info->shrinklist)) {
1155 list_del_init(&info->shrinklist);
1156 sbinfo->shrinklist_len--;
1157 }
1158 spin_unlock(&sbinfo->shrinklist_lock);
1159 }
1160 while (!list_empty(&info->swaplist)) {
1161 /* Wait while shmem_unuse() is scanning this inode... */
1162 wait_var_event(&info->stop_eviction,
1163 !atomic_read(&info->stop_eviction));
1164 mutex_lock(&shmem_swaplist_mutex);
1165 /* ...but beware of the race if we peeked too early */
1166 if (!atomic_read(&info->stop_eviction))
1167 list_del_init(&info->swaplist);
1168 mutex_unlock(&shmem_swaplist_mutex);
1169 }
1170 }
1171
1172 simple_xattrs_free(&info->xattrs);
1173 WARN_ON(inode->i_blocks);
1174 shmem_free_inode(inode->i_sb);
1175 clear_inode(inode);
1176}
1177
1178static int shmem_find_swap_entries(struct address_space *mapping,
1179 pgoff_t start, struct folio_batch *fbatch,
1180 pgoff_t *indices, unsigned int type)
1181{
1182 XA_STATE(xas, &mapping->i_pages, start);
1183 struct folio *folio;
1184 swp_entry_t entry;
1185
1186 rcu_read_lock();
1187 xas_for_each(&xas, folio, ULONG_MAX) {
1188 if (xas_retry(&xas, folio))
1189 continue;
1190
1191 if (!xa_is_value(folio))
1192 continue;
1193
1194 entry = radix_to_swp_entry(folio);
1195 /*
1196 * swapin error entries can be found in the mapping. But they're
1197 * deliberately ignored here as we've done everything we can do.
1198 */
1199 if (swp_type(entry) != type)
1200 continue;
1201
1202 indices[folio_batch_count(fbatch)] = xas.xa_index;
1203 if (!folio_batch_add(fbatch, folio))
1204 break;
1205
1206 if (need_resched()) {
1207 xas_pause(&xas);
1208 cond_resched_rcu();
1209 }
1210 }
1211 rcu_read_unlock();
1212
1213 return xas.xa_index;
1214}
1215
1216/*
1217 * Move the swapped pages for an inode to page cache. Returns the count
1218 * of pages swapped in, or the error in case of failure.
1219 */
1220static int shmem_unuse_swap_entries(struct inode *inode,
1221 struct folio_batch *fbatch, pgoff_t *indices)
1222{
1223 int i = 0;
1224 int ret = 0;
1225 int error = 0;
1226 struct address_space *mapping = inode->i_mapping;
1227
1228 for (i = 0; i < folio_batch_count(fbatch); i++) {
1229 struct folio *folio = fbatch->folios[i];
1230
1231 if (!xa_is_value(folio))
1232 continue;
1233 error = shmem_swapin_folio(inode, indices[i],
1234 &folio, SGP_CACHE,
1235 mapping_gfp_mask(mapping),
1236 NULL, NULL);
1237 if (error == 0) {
1238 folio_unlock(folio);
1239 folio_put(folio);
1240 ret++;
1241 }
1242 if (error == -ENOMEM)
1243 break;
1244 error = 0;
1245 }
1246 return error ? error : ret;
1247}
1248
1249/*
1250 * If swap found in inode, free it and move page from swapcache to filecache.
1251 */
1252static int shmem_unuse_inode(struct inode *inode, unsigned int type)
1253{
1254 struct address_space *mapping = inode->i_mapping;
1255 pgoff_t start = 0;
1256 struct folio_batch fbatch;
1257 pgoff_t indices[PAGEVEC_SIZE];
1258 int ret = 0;
1259
1260 do {
1261 folio_batch_init(&fbatch);
1262 shmem_find_swap_entries(mapping, start, &fbatch, indices, type);
1263 if (folio_batch_count(&fbatch) == 0) {
1264 ret = 0;
1265 break;
1266 }
1267
1268 ret = shmem_unuse_swap_entries(inode, &fbatch, indices);
1269 if (ret < 0)
1270 break;
1271
1272 start = indices[folio_batch_count(&fbatch) - 1];
1273 } while (true);
1274
1275 return ret;
1276}
1277
1278/*
1279 * Read all the shared memory data that resides in the swap
1280 * device 'type' back into memory, so the swap device can be
1281 * unused.
1282 */
1283int shmem_unuse(unsigned int type)
1284{
1285 struct shmem_inode_info *info, *next;
1286 int error = 0;
1287
1288 if (list_empty(&shmem_swaplist))
1289 return 0;
1290
1291 mutex_lock(&shmem_swaplist_mutex);
1292 list_for_each_entry_safe(info, next, &shmem_swaplist, swaplist) {
1293 if (!info->swapped) {
1294 list_del_init(&info->swaplist);
1295 continue;
1296 }
1297 /*
1298 * Drop the swaplist mutex while searching the inode for swap;
1299 * but before doing so, make sure shmem_evict_inode() will not
1300 * remove placeholder inode from swaplist, nor let it be freed
1301 * (igrab() would protect from unlink, but not from unmount).
1302 */
1303 atomic_inc(&info->stop_eviction);
1304 mutex_unlock(&shmem_swaplist_mutex);
1305
1306 error = shmem_unuse_inode(&info->vfs_inode, type);
1307 cond_resched();
1308
1309 mutex_lock(&shmem_swaplist_mutex);
1310 next = list_next_entry(info, swaplist);
1311 if (!info->swapped)
1312 list_del_init(&info->swaplist);
1313 if (atomic_dec_and_test(&info->stop_eviction))
1314 wake_up_var(&info->stop_eviction);
1315 if (error)
1316 break;
1317 }
1318 mutex_unlock(&shmem_swaplist_mutex);
1319
1320 return error;
1321}
1322
1323/*
1324 * Move the page from the page cache to the swap cache.
1325 */
1326static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1327{
1328 struct folio *folio = page_folio(page);
1329 struct shmem_inode_info *info;
1330 struct address_space *mapping;
1331 struct inode *inode;
1332 swp_entry_t swap;
1333 pgoff_t index;
1334
1335 /*
1336 * If /sys/kernel/mm/transparent_hugepage/shmem_enabled is "always" or
1337 * "force", drivers/gpu/drm/i915/gem/i915_gem_shmem.c gets huge pages,
1338 * and its shmem_writeback() needs them to be split when swapping.
1339 */
1340 if (folio_test_large(folio)) {
1341 /* Ensure the subpages are still dirty */
1342 folio_test_set_dirty(folio);
1343 if (split_huge_page(page) < 0)
1344 goto redirty;
1345 folio = page_folio(page);
1346 folio_clear_dirty(folio);
1347 }
1348
1349 BUG_ON(!folio_test_locked(folio));
1350 mapping = folio->mapping;
1351 index = folio->index;
1352 inode = mapping->host;
1353 info = SHMEM_I(inode);
1354 if (info->flags & VM_LOCKED)
1355 goto redirty;
1356 if (!total_swap_pages)
1357 goto redirty;
1358
1359 /*
1360 * Our capabilities prevent regular writeback or sync from ever calling
1361 * shmem_writepage; but a stacking filesystem might use ->writepage of
1362 * its underlying filesystem, in which case tmpfs should write out to
1363 * swap only in response to memory pressure, and not for the writeback
1364 * threads or sync.
1365 */
1366 if (!wbc->for_reclaim) {
1367 WARN_ON_ONCE(1); /* Still happens? Tell us about it! */
1368 goto redirty;
1369 }
1370
1371 /*
1372 * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
1373 * value into swapfile.c, the only way we can correctly account for a
1374 * fallocated folio arriving here is now to initialize it and write it.
1375 *
1376 * That's okay for a folio already fallocated earlier, but if we have
1377 * not yet completed the fallocation, then (a) we want to keep track
1378 * of this folio in case we have to undo it, and (b) it may not be a
1379 * good idea to continue anyway, once we're pushing into swap. So
1380 * reactivate the folio, and let shmem_fallocate() quit when too many.
1381 */
1382 if (!folio_test_uptodate(folio)) {
1383 if (inode->i_private) {
1384 struct shmem_falloc *shmem_falloc;
1385 spin_lock(&inode->i_lock);
1386 shmem_falloc = inode->i_private;
1387 if (shmem_falloc &&
1388 !shmem_falloc->waitq &&
1389 index >= shmem_falloc->start &&
1390 index < shmem_falloc->next)
1391 shmem_falloc->nr_unswapped++;
1392 else
1393 shmem_falloc = NULL;
1394 spin_unlock(&inode->i_lock);
1395 if (shmem_falloc)
1396 goto redirty;
1397 }
1398 folio_zero_range(folio, 0, folio_size(folio));
1399 flush_dcache_folio(folio);
1400 folio_mark_uptodate(folio);
1401 }
1402
1403 swap = folio_alloc_swap(folio);
1404 if (!swap.val)
1405 goto redirty;
1406
1407 /*
1408 * Add inode to shmem_unuse()'s list of swapped-out inodes,
1409 * if it's not already there. Do it now before the folio is
1410 * moved to swap cache, when its pagelock no longer protects
1411 * the inode from eviction. But don't unlock the mutex until
1412 * we've incremented swapped, because shmem_unuse_inode() will
1413 * prune a !swapped inode from the swaplist under this mutex.
1414 */
1415 mutex_lock(&shmem_swaplist_mutex);
1416 if (list_empty(&info->swaplist))
1417 list_add(&info->swaplist, &shmem_swaplist);
1418
1419 if (add_to_swap_cache(folio, swap,
1420 __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN,
1421 NULL) == 0) {
1422 spin_lock_irq(&info->lock);
1423 shmem_recalc_inode(inode);
1424 info->swapped++;
1425 spin_unlock_irq(&info->lock);
1426
1427 swap_shmem_alloc(swap);
1428 shmem_delete_from_page_cache(folio, swp_to_radix_entry(swap));
1429
1430 mutex_unlock(&shmem_swaplist_mutex);
1431 BUG_ON(folio_mapped(folio));
1432 swap_writepage(&folio->page, wbc);
1433 return 0;
1434 }
1435
1436 mutex_unlock(&shmem_swaplist_mutex);
1437 put_swap_folio(folio, swap);
1438redirty:
1439 folio_mark_dirty(folio);
1440 if (wbc->for_reclaim)
1441 return AOP_WRITEPAGE_ACTIVATE; /* Return with folio locked */
1442 folio_unlock(folio);
1443 return 0;
1444}
1445
1446#if defined(CONFIG_NUMA) && defined(CONFIG_TMPFS)
1447static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1448{
1449 char buffer[64];
1450
1451 if (!mpol || mpol->mode == MPOL_DEFAULT)
1452 return; /* show nothing */
1453
1454 mpol_to_str(buffer, sizeof(buffer), mpol);
1455
1456 seq_printf(seq, ",mpol=%s", buffer);
1457}
1458
1459static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1460{
1461 struct mempolicy *mpol = NULL;
1462 if (sbinfo->mpol) {
1463 raw_spin_lock(&sbinfo->stat_lock); /* prevent replace/use races */
1464 mpol = sbinfo->mpol;
1465 mpol_get(mpol);
1466 raw_spin_unlock(&sbinfo->stat_lock);
1467 }
1468 return mpol;
1469}
1470#else /* !CONFIG_NUMA || !CONFIG_TMPFS */
1471static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1472{
1473}
1474static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1475{
1476 return NULL;
1477}
1478#endif /* CONFIG_NUMA && CONFIG_TMPFS */
1479#ifndef CONFIG_NUMA
1480#define vm_policy vm_private_data
1481#endif
1482
1483static void shmem_pseudo_vma_init(struct vm_area_struct *vma,
1484 struct shmem_inode_info *info, pgoff_t index)
1485{
1486 /* Create a pseudo vma that just contains the policy */
1487 vma_init(vma, NULL);
1488 /* Bias interleave by inode number to distribute better across nodes */
1489 vma->vm_pgoff = index + info->vfs_inode.i_ino;
1490 vma->vm_policy = mpol_shared_policy_lookup(&info->policy, index);
1491}
1492
1493static void shmem_pseudo_vma_destroy(struct vm_area_struct *vma)
1494{
1495 /* Drop reference taken by mpol_shared_policy_lookup() */
1496 mpol_cond_put(vma->vm_policy);
1497}
1498
1499static struct folio *shmem_swapin(swp_entry_t swap, gfp_t gfp,
1500 struct shmem_inode_info *info, pgoff_t index)
1501{
1502 struct vm_area_struct pvma;
1503 struct page *page;
1504 struct vm_fault vmf = {
1505 .vma = &pvma,
1506 };
1507
1508 shmem_pseudo_vma_init(&pvma, info, index);
1509 page = swap_cluster_readahead(swap, gfp, &vmf);
1510 shmem_pseudo_vma_destroy(&pvma);
1511
1512 if (!page)
1513 return NULL;
1514 return page_folio(page);
1515}
1516
1517/*
1518 * Make sure huge_gfp is always more limited than limit_gfp.
1519 * Some of the flags set permissions, while others set limitations.
1520 */
1521static gfp_t limit_gfp_mask(gfp_t huge_gfp, gfp_t limit_gfp)
1522{
1523 gfp_t allowflags = __GFP_IO | __GFP_FS | __GFP_RECLAIM;
1524 gfp_t denyflags = __GFP_NOWARN | __GFP_NORETRY;
1525 gfp_t zoneflags = limit_gfp & GFP_ZONEMASK;
1526 gfp_t result = huge_gfp & ~(allowflags | GFP_ZONEMASK);
1527
1528 /* Allow allocations only from the originally specified zones. */
1529 result |= zoneflags;
1530
1531 /*
1532 * Minimize the result gfp by taking the union with the deny flags,
1533 * and the intersection of the allow flags.
1534 */
1535 result |= (limit_gfp & denyflags);
1536 result |= (huge_gfp & limit_gfp) & allowflags;
1537
1538 return result;
1539}
1540
1541static struct folio *shmem_alloc_hugefolio(gfp_t gfp,
1542 struct shmem_inode_info *info, pgoff_t index)
1543{
1544 struct vm_area_struct pvma;
1545 struct address_space *mapping = info->vfs_inode.i_mapping;
1546 pgoff_t hindex;
1547 struct folio *folio;
1548
1549 hindex = round_down(index, HPAGE_PMD_NR);
1550 if (xa_find(&mapping->i_pages, &hindex, hindex + HPAGE_PMD_NR - 1,
1551 XA_PRESENT))
1552 return NULL;
1553
1554 shmem_pseudo_vma_init(&pvma, info, hindex);
1555 folio = vma_alloc_folio(gfp, HPAGE_PMD_ORDER, &pvma, 0, true);
1556 shmem_pseudo_vma_destroy(&pvma);
1557 if (!folio)
1558 count_vm_event(THP_FILE_FALLBACK);
1559 return folio;
1560}
1561
1562static struct folio *shmem_alloc_folio(gfp_t gfp,
1563 struct shmem_inode_info *info, pgoff_t index)
1564{
1565 struct vm_area_struct pvma;
1566 struct folio *folio;
1567
1568 shmem_pseudo_vma_init(&pvma, info, index);
1569 folio = vma_alloc_folio(gfp, 0, &pvma, 0, false);
1570 shmem_pseudo_vma_destroy(&pvma);
1571
1572 return folio;
1573}
1574
1575static struct folio *shmem_alloc_and_acct_folio(gfp_t gfp, struct inode *inode,
1576 pgoff_t index, bool huge)
1577{
1578 struct shmem_inode_info *info = SHMEM_I(inode);
1579 struct folio *folio;
1580 int nr;
1581 int err = -ENOSPC;
1582
1583 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1584 huge = false;
1585 nr = huge ? HPAGE_PMD_NR : 1;
1586
1587 if (!shmem_inode_acct_block(inode, nr))
1588 goto failed;
1589
1590 if (huge)
1591 folio = shmem_alloc_hugefolio(gfp, info, index);
1592 else
1593 folio = shmem_alloc_folio(gfp, info, index);
1594 if (folio) {
1595 __folio_set_locked(folio);
1596 __folio_set_swapbacked(folio);
1597 return folio;
1598 }
1599
1600 err = -ENOMEM;
1601 shmem_inode_unacct_blocks(inode, nr);
1602failed:
1603 return ERR_PTR(err);
1604}
1605
1606/*
1607 * When a page is moved from swapcache to shmem filecache (either by the
1608 * usual swapin of shmem_get_folio_gfp(), or by the less common swapoff of
1609 * shmem_unuse_inode()), it may have been read in earlier from swap, in
1610 * ignorance of the mapping it belongs to. If that mapping has special
1611 * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
1612 * we may need to copy to a suitable page before moving to filecache.
1613 *
1614 * In a future release, this may well be extended to respect cpuset and
1615 * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
1616 * but for now it is a simple matter of zone.
1617 */
1618static bool shmem_should_replace_folio(struct folio *folio, gfp_t gfp)
1619{
1620 return folio_zonenum(folio) > gfp_zone(gfp);
1621}
1622
1623static int shmem_replace_folio(struct folio **foliop, gfp_t gfp,
1624 struct shmem_inode_info *info, pgoff_t index)
1625{
1626 struct folio *old, *new;
1627 struct address_space *swap_mapping;
1628 swp_entry_t entry;
1629 pgoff_t swap_index;
1630 int error;
1631
1632 old = *foliop;
1633 entry = folio_swap_entry(old);
1634 swap_index = swp_offset(entry);
1635 swap_mapping = swap_address_space(entry);
1636
1637 /*
1638 * We have arrived here because our zones are constrained, so don't
1639 * limit chance of success by further cpuset and node constraints.
1640 */
1641 gfp &= ~GFP_CONSTRAINT_MASK;
1642 VM_BUG_ON_FOLIO(folio_test_large(old), old);
1643 new = shmem_alloc_folio(gfp, info, index);
1644 if (!new)
1645 return -ENOMEM;
1646
1647 folio_get(new);
1648 folio_copy(new, old);
1649 flush_dcache_folio(new);
1650
1651 __folio_set_locked(new);
1652 __folio_set_swapbacked(new);
1653 folio_mark_uptodate(new);
1654 folio_set_swap_entry(new, entry);
1655 folio_set_swapcache(new);
1656
1657 /*
1658 * Our caller will very soon move newpage out of swapcache, but it's
1659 * a nice clean interface for us to replace oldpage by newpage there.
1660 */
1661 xa_lock_irq(&swap_mapping->i_pages);
1662 error = shmem_replace_entry(swap_mapping, swap_index, old, new);
1663 if (!error) {
1664 mem_cgroup_migrate(old, new);
1665 __lruvec_stat_mod_folio(new, NR_FILE_PAGES, 1);
1666 __lruvec_stat_mod_folio(new, NR_SHMEM, 1);
1667 __lruvec_stat_mod_folio(old, NR_FILE_PAGES, -1);
1668 __lruvec_stat_mod_folio(old, NR_SHMEM, -1);
1669 }
1670 xa_unlock_irq(&swap_mapping->i_pages);
1671
1672 if (unlikely(error)) {
1673 /*
1674 * Is this possible? I think not, now that our callers check
1675 * both PageSwapCache and page_private after getting page lock;
1676 * but be defensive. Reverse old to newpage for clear and free.
1677 */
1678 old = new;
1679 } else {
1680 folio_add_lru(new);
1681 *foliop = new;
1682 }
1683
1684 folio_clear_swapcache(old);
1685 old->private = NULL;
1686
1687 folio_unlock(old);
1688 folio_put_refs(old, 2);
1689 return error;
1690}
1691
1692static void shmem_set_folio_swapin_error(struct inode *inode, pgoff_t index,
1693 struct folio *folio, swp_entry_t swap)
1694{
1695 struct address_space *mapping = inode->i_mapping;
1696 struct shmem_inode_info *info = SHMEM_I(inode);
1697 swp_entry_t swapin_error;
1698 void *old;
1699
1700 swapin_error = make_swapin_error_entry();
1701 old = xa_cmpxchg_irq(&mapping->i_pages, index,
1702 swp_to_radix_entry(swap),
1703 swp_to_radix_entry(swapin_error), 0);
1704 if (old != swp_to_radix_entry(swap))
1705 return;
1706
1707 folio_wait_writeback(folio);
1708 delete_from_swap_cache(folio);
1709 spin_lock_irq(&info->lock);
1710 /*
1711 * Don't treat swapin error folio as alloced. Otherwise inode->i_blocks won't
1712 * be 0 when inode is released and thus trigger WARN_ON(inode->i_blocks) in
1713 * shmem_evict_inode.
1714 */
1715 info->alloced--;
1716 info->swapped--;
1717 shmem_recalc_inode(inode);
1718 spin_unlock_irq(&info->lock);
1719 swap_free(swap);
1720}
1721
1722/*
1723 * Swap in the folio pointed to by *foliop.
1724 * Caller has to make sure that *foliop contains a valid swapped folio.
1725 * Returns 0 and the folio in foliop if success. On failure, returns the
1726 * error code and NULL in *foliop.
1727 */
1728static int shmem_swapin_folio(struct inode *inode, pgoff_t index,
1729 struct folio **foliop, enum sgp_type sgp,
1730 gfp_t gfp, struct vm_area_struct *vma,
1731 vm_fault_t *fault_type)
1732{
1733 struct address_space *mapping = inode->i_mapping;
1734 struct shmem_inode_info *info = SHMEM_I(inode);
1735 struct mm_struct *charge_mm = vma ? vma->vm_mm : NULL;
1736 struct folio *folio = NULL;
1737 swp_entry_t swap;
1738 int error;
1739
1740 VM_BUG_ON(!*foliop || !xa_is_value(*foliop));
1741 swap = radix_to_swp_entry(*foliop);
1742 *foliop = NULL;
1743
1744 if (is_swapin_error_entry(swap))
1745 return -EIO;
1746
1747 /* Look it up and read it in.. */
1748 folio = swap_cache_get_folio(swap, NULL, 0);
1749 if (!folio) {
1750 /* Or update major stats only when swapin succeeds?? */
1751 if (fault_type) {
1752 *fault_type |= VM_FAULT_MAJOR;
1753 count_vm_event(PGMAJFAULT);
1754 count_memcg_event_mm(charge_mm, PGMAJFAULT);
1755 }
1756 /* Here we actually start the io */
1757 folio = shmem_swapin(swap, gfp, info, index);
1758 if (!folio) {
1759 error = -ENOMEM;
1760 goto failed;
1761 }
1762 }
1763
1764 /* We have to do this with folio locked to prevent races */
1765 folio_lock(folio);
1766 if (!folio_test_swapcache(folio) ||
1767 folio_swap_entry(folio).val != swap.val ||
1768 !shmem_confirm_swap(mapping, index, swap)) {
1769 error = -EEXIST;
1770 goto unlock;
1771 }
1772 if (!folio_test_uptodate(folio)) {
1773 error = -EIO;
1774 goto failed;
1775 }
1776 folio_wait_writeback(folio);
1777
1778 /*
1779 * Some architectures may have to restore extra metadata to the
1780 * folio after reading from swap.
1781 */
1782 arch_swap_restore(swap, folio);
1783
1784 if (shmem_should_replace_folio(folio, gfp)) {
1785 error = shmem_replace_folio(&folio, gfp, info, index);
1786 if (error)
1787 goto failed;
1788 }
1789
1790 error = shmem_add_to_page_cache(folio, mapping, index,
1791 swp_to_radix_entry(swap), gfp,
1792 charge_mm);
1793 if (error)
1794 goto failed;
1795
1796 spin_lock_irq(&info->lock);
1797 info->swapped--;
1798 shmem_recalc_inode(inode);
1799 spin_unlock_irq(&info->lock);
1800
1801 if (sgp == SGP_WRITE)
1802 folio_mark_accessed(folio);
1803
1804 delete_from_swap_cache(folio);
1805 folio_mark_dirty(folio);
1806 swap_free(swap);
1807
1808 *foliop = folio;
1809 return 0;
1810failed:
1811 if (!shmem_confirm_swap(mapping, index, swap))
1812 error = -EEXIST;
1813 if (error == -EIO)
1814 shmem_set_folio_swapin_error(inode, index, folio, swap);
1815unlock:
1816 if (folio) {
1817 folio_unlock(folio);
1818 folio_put(folio);
1819 }
1820
1821 return error;
1822}
1823
1824/*
1825 * shmem_get_folio_gfp - find page in cache, or get from swap, or allocate
1826 *
1827 * If we allocate a new one we do not mark it dirty. That's up to the
1828 * vm. If we swap it in we mark it dirty since we also free the swap
1829 * entry since a page cannot live in both the swap and page cache.
1830 *
1831 * vma, vmf, and fault_type are only supplied by shmem_fault:
1832 * otherwise they are NULL.
1833 */
1834static int shmem_get_folio_gfp(struct inode *inode, pgoff_t index,
1835 struct folio **foliop, enum sgp_type sgp, gfp_t gfp,
1836 struct vm_area_struct *vma, struct vm_fault *vmf,
1837 vm_fault_t *fault_type)
1838{
1839 struct address_space *mapping = inode->i_mapping;
1840 struct shmem_inode_info *info = SHMEM_I(inode);
1841 struct shmem_sb_info *sbinfo;
1842 struct mm_struct *charge_mm;
1843 struct folio *folio;
1844 pgoff_t hindex;
1845 gfp_t huge_gfp;
1846 int error;
1847 int once = 0;
1848 int alloced = 0;
1849
1850 if (index > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
1851 return -EFBIG;
1852repeat:
1853 if (sgp <= SGP_CACHE &&
1854 ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1855 return -EINVAL;
1856 }
1857
1858 sbinfo = SHMEM_SB(inode->i_sb);
1859 charge_mm = vma ? vma->vm_mm : NULL;
1860
1861 folio = __filemap_get_folio(mapping, index, FGP_ENTRY | FGP_LOCK, 0);
1862 if (folio && vma && userfaultfd_minor(vma)) {
1863 if (!xa_is_value(folio)) {
1864 folio_unlock(folio);
1865 folio_put(folio);
1866 }
1867 *fault_type = handle_userfault(vmf, VM_UFFD_MINOR);
1868 return 0;
1869 }
1870
1871 if (xa_is_value(folio)) {
1872 error = shmem_swapin_folio(inode, index, &folio,
1873 sgp, gfp, vma, fault_type);
1874 if (error == -EEXIST)
1875 goto repeat;
1876
1877 *foliop = folio;
1878 return error;
1879 }
1880
1881 if (folio) {
1882 if (sgp == SGP_WRITE)
1883 folio_mark_accessed(folio);
1884 if (folio_test_uptodate(folio))
1885 goto out;
1886 /* fallocated folio */
1887 if (sgp != SGP_READ)
1888 goto clear;
1889 folio_unlock(folio);
1890 folio_put(folio);
1891 }
1892
1893 /*
1894 * SGP_READ: succeed on hole, with NULL folio, letting caller zero.
1895 * SGP_NOALLOC: fail on hole, with NULL folio, letting caller fail.
1896 */
1897 *foliop = NULL;
1898 if (sgp == SGP_READ)
1899 return 0;
1900 if (sgp == SGP_NOALLOC)
1901 return -ENOENT;
1902
1903 /*
1904 * Fast cache lookup and swap lookup did not find it: allocate.
1905 */
1906
1907 if (vma && userfaultfd_missing(vma)) {
1908 *fault_type = handle_userfault(vmf, VM_UFFD_MISSING);
1909 return 0;
1910 }
1911
1912 if (!shmem_is_huge(vma, inode, index, false))
1913 goto alloc_nohuge;
1914
1915 huge_gfp = vma_thp_gfp_mask(vma);
1916 huge_gfp = limit_gfp_mask(huge_gfp, gfp);
1917 folio = shmem_alloc_and_acct_folio(huge_gfp, inode, index, true);
1918 if (IS_ERR(folio)) {
1919alloc_nohuge:
1920 folio = shmem_alloc_and_acct_folio(gfp, inode, index, false);
1921 }
1922 if (IS_ERR(folio)) {
1923 int retry = 5;
1924
1925 error = PTR_ERR(folio);
1926 folio = NULL;
1927 if (error != -ENOSPC)
1928 goto unlock;
1929 /*
1930 * Try to reclaim some space by splitting a large folio
1931 * beyond i_size on the filesystem.
1932 */
1933 while (retry--) {
1934 int ret;
1935
1936 ret = shmem_unused_huge_shrink(sbinfo, NULL, 1);
1937 if (ret == SHRINK_STOP)
1938 break;
1939 if (ret)
1940 goto alloc_nohuge;
1941 }
1942 goto unlock;
1943 }
1944
1945 hindex = round_down(index, folio_nr_pages(folio));
1946
1947 if (sgp == SGP_WRITE)
1948 __folio_set_referenced(folio);
1949
1950 error = shmem_add_to_page_cache(folio, mapping, hindex,
1951 NULL, gfp & GFP_RECLAIM_MASK,
1952 charge_mm);
1953 if (error)
1954 goto unacct;
1955 folio_add_lru(folio);
1956
1957 spin_lock_irq(&info->lock);
1958 info->alloced += folio_nr_pages(folio);
1959 inode->i_blocks += (blkcnt_t)BLOCKS_PER_PAGE << folio_order(folio);
1960 shmem_recalc_inode(inode);
1961 spin_unlock_irq(&info->lock);
1962 alloced = true;
1963
1964 if (folio_test_pmd_mappable(folio) &&
1965 DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE) <
1966 folio_next_index(folio) - 1) {
1967 /*
1968 * Part of the large folio is beyond i_size: subject
1969 * to shrink under memory pressure.
1970 */
1971 spin_lock(&sbinfo->shrinklist_lock);
1972 /*
1973 * _careful to defend against unlocked access to
1974 * ->shrink_list in shmem_unused_huge_shrink()
1975 */
1976 if (list_empty_careful(&info->shrinklist)) {
1977 list_add_tail(&info->shrinklist,
1978 &sbinfo->shrinklist);
1979 sbinfo->shrinklist_len++;
1980 }
1981 spin_unlock(&sbinfo->shrinklist_lock);
1982 }
1983
1984 /*
1985 * Let SGP_FALLOC use the SGP_WRITE optimization on a new folio.
1986 */
1987 if (sgp == SGP_FALLOC)
1988 sgp = SGP_WRITE;
1989clear:
1990 /*
1991 * Let SGP_WRITE caller clear ends if write does not fill folio;
1992 * but SGP_FALLOC on a folio fallocated earlier must initialize
1993 * it now, lest undo on failure cancel our earlier guarantee.
1994 */
1995 if (sgp != SGP_WRITE && !folio_test_uptodate(folio)) {
1996 long i, n = folio_nr_pages(folio);
1997
1998 for (i = 0; i < n; i++)
1999 clear_highpage(folio_page(folio, i));
2000 flush_dcache_folio(folio);
2001 folio_mark_uptodate(folio);
2002 }
2003
2004 /* Perhaps the file has been truncated since we checked */
2005 if (sgp <= SGP_CACHE &&
2006 ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
2007 if (alloced) {
2008 folio_clear_dirty(folio);
2009 filemap_remove_folio(folio);
2010 spin_lock_irq(&info->lock);
2011 shmem_recalc_inode(inode);
2012 spin_unlock_irq(&info->lock);
2013 }
2014 error = -EINVAL;
2015 goto unlock;
2016 }
2017out:
2018 *foliop = folio;
2019 return 0;
2020
2021 /*
2022 * Error recovery.
2023 */
2024unacct:
2025 shmem_inode_unacct_blocks(inode, folio_nr_pages(folio));
2026
2027 if (folio_test_large(folio)) {
2028 folio_unlock(folio);
2029 folio_put(folio);
2030 goto alloc_nohuge;
2031 }
2032unlock:
2033 if (folio) {
2034 folio_unlock(folio);
2035 folio_put(folio);
2036 }
2037 if (error == -ENOSPC && !once++) {
2038 spin_lock_irq(&info->lock);
2039 shmem_recalc_inode(inode);
2040 spin_unlock_irq(&info->lock);
2041 goto repeat;
2042 }
2043 if (error == -EEXIST)
2044 goto repeat;
2045 return error;
2046}
2047
2048int shmem_get_folio(struct inode *inode, pgoff_t index, struct folio **foliop,
2049 enum sgp_type sgp)
2050{
2051 return shmem_get_folio_gfp(inode, index, foliop, sgp,
2052 mapping_gfp_mask(inode->i_mapping), NULL, NULL, NULL);
2053}
2054
2055/*
2056 * This is like autoremove_wake_function, but it removes the wait queue
2057 * entry unconditionally - even if something else had already woken the
2058 * target.
2059 */
2060static int synchronous_wake_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
2061{
2062 int ret = default_wake_function(wait, mode, sync, key);
2063 list_del_init(&wait->entry);
2064 return ret;
2065}
2066
2067static vm_fault_t shmem_fault(struct vm_fault *vmf)
2068{
2069 struct vm_area_struct *vma = vmf->vma;
2070 struct inode *inode = file_inode(vma->vm_file);
2071 gfp_t gfp = mapping_gfp_mask(inode->i_mapping);
2072 struct folio *folio = NULL;
2073 int err;
2074 vm_fault_t ret = VM_FAULT_LOCKED;
2075
2076 /*
2077 * Trinity finds that probing a hole which tmpfs is punching can
2078 * prevent the hole-punch from ever completing: which in turn
2079 * locks writers out with its hold on i_rwsem. So refrain from
2080 * faulting pages into the hole while it's being punched. Although
2081 * shmem_undo_range() does remove the additions, it may be unable to
2082 * keep up, as each new page needs its own unmap_mapping_range() call,
2083 * and the i_mmap tree grows ever slower to scan if new vmas are added.
2084 *
2085 * It does not matter if we sometimes reach this check just before the
2086 * hole-punch begins, so that one fault then races with the punch:
2087 * we just need to make racing faults a rare case.
2088 *
2089 * The implementation below would be much simpler if we just used a
2090 * standard mutex or completion: but we cannot take i_rwsem in fault,
2091 * and bloating every shmem inode for this unlikely case would be sad.
2092 */
2093 if (unlikely(inode->i_private)) {
2094 struct shmem_falloc *shmem_falloc;
2095
2096 spin_lock(&inode->i_lock);
2097 shmem_falloc = inode->i_private;
2098 if (shmem_falloc &&
2099 shmem_falloc->waitq &&
2100 vmf->pgoff >= shmem_falloc->start &&
2101 vmf->pgoff < shmem_falloc->next) {
2102 struct file *fpin;
2103 wait_queue_head_t *shmem_falloc_waitq;
2104 DEFINE_WAIT_FUNC(shmem_fault_wait, synchronous_wake_function);
2105
2106 ret = VM_FAULT_NOPAGE;
2107 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2108 if (fpin)
2109 ret = VM_FAULT_RETRY;
2110
2111 shmem_falloc_waitq = shmem_falloc->waitq;
2112 prepare_to_wait(shmem_falloc_waitq, &shmem_fault_wait,
2113 TASK_UNINTERRUPTIBLE);
2114 spin_unlock(&inode->i_lock);
2115 schedule();
2116
2117 /*
2118 * shmem_falloc_waitq points into the shmem_fallocate()
2119 * stack of the hole-punching task: shmem_falloc_waitq
2120 * is usually invalid by the time we reach here, but
2121 * finish_wait() does not dereference it in that case;
2122 * though i_lock needed lest racing with wake_up_all().
2123 */
2124 spin_lock(&inode->i_lock);
2125 finish_wait(shmem_falloc_waitq, &shmem_fault_wait);
2126 spin_unlock(&inode->i_lock);
2127
2128 if (fpin)
2129 fput(fpin);
2130 return ret;
2131 }
2132 spin_unlock(&inode->i_lock);
2133 }
2134
2135 err = shmem_get_folio_gfp(inode, vmf->pgoff, &folio, SGP_CACHE,
2136 gfp, vma, vmf, &ret);
2137 if (err)
2138 return vmf_error(err);
2139 if (folio)
2140 vmf->page = folio_file_page(folio, vmf->pgoff);
2141 return ret;
2142}
2143
2144unsigned long shmem_get_unmapped_area(struct file *file,
2145 unsigned long uaddr, unsigned long len,
2146 unsigned long pgoff, unsigned long flags)
2147{
2148 unsigned long (*get_area)(struct file *,
2149 unsigned long, unsigned long, unsigned long, unsigned long);
2150 unsigned long addr;
2151 unsigned long offset;
2152 unsigned long inflated_len;
2153 unsigned long inflated_addr;
2154 unsigned long inflated_offset;
2155
2156 if (len > TASK_SIZE)
2157 return -ENOMEM;
2158
2159 get_area = current->mm->get_unmapped_area;
2160 addr = get_area(file, uaddr, len, pgoff, flags);
2161
2162 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
2163 return addr;
2164 if (IS_ERR_VALUE(addr))
2165 return addr;
2166 if (addr & ~PAGE_MASK)
2167 return addr;
2168 if (addr > TASK_SIZE - len)
2169 return addr;
2170
2171 if (shmem_huge == SHMEM_HUGE_DENY)
2172 return addr;
2173 if (len < HPAGE_PMD_SIZE)
2174 return addr;
2175 if (flags & MAP_FIXED)
2176 return addr;
2177 /*
2178 * Our priority is to support MAP_SHARED mapped hugely;
2179 * and support MAP_PRIVATE mapped hugely too, until it is COWed.
2180 * But if caller specified an address hint and we allocated area there
2181 * successfully, respect that as before.
2182 */
2183 if (uaddr == addr)
2184 return addr;
2185
2186 if (shmem_huge != SHMEM_HUGE_FORCE) {
2187 struct super_block *sb;
2188
2189 if (file) {
2190 VM_BUG_ON(file->f_op != &shmem_file_operations);
2191 sb = file_inode(file)->i_sb;
2192 } else {
2193 /*
2194 * Called directly from mm/mmap.c, or drivers/char/mem.c
2195 * for "/dev/zero", to create a shared anonymous object.
2196 */
2197 if (IS_ERR(shm_mnt))
2198 return addr;
2199 sb = shm_mnt->mnt_sb;
2200 }
2201 if (SHMEM_SB(sb)->huge == SHMEM_HUGE_NEVER)
2202 return addr;
2203 }
2204
2205 offset = (pgoff << PAGE_SHIFT) & (HPAGE_PMD_SIZE-1);
2206 if (offset && offset + len < 2 * HPAGE_PMD_SIZE)
2207 return addr;
2208 if ((addr & (HPAGE_PMD_SIZE-1)) == offset)
2209 return addr;
2210
2211 inflated_len = len + HPAGE_PMD_SIZE - PAGE_SIZE;
2212 if (inflated_len > TASK_SIZE)
2213 return addr;
2214 if (inflated_len < len)
2215 return addr;
2216
2217 inflated_addr = get_area(NULL, uaddr, inflated_len, 0, flags);
2218 if (IS_ERR_VALUE(inflated_addr))
2219 return addr;
2220 if (inflated_addr & ~PAGE_MASK)
2221 return addr;
2222
2223 inflated_offset = inflated_addr & (HPAGE_PMD_SIZE-1);
2224 inflated_addr += offset - inflated_offset;
2225 if (inflated_offset > offset)
2226 inflated_addr += HPAGE_PMD_SIZE;
2227
2228 if (inflated_addr > TASK_SIZE - len)
2229 return addr;
2230 return inflated_addr;
2231}
2232
2233#ifdef CONFIG_NUMA
2234static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
2235{
2236 struct inode *inode = file_inode(vma->vm_file);
2237 return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
2238}
2239
2240static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
2241 unsigned long addr)
2242{
2243 struct inode *inode = file_inode(vma->vm_file);
2244 pgoff_t index;
2245
2246 index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2247 return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
2248}
2249#endif
2250
2251int shmem_lock(struct file *file, int lock, struct ucounts *ucounts)
2252{
2253 struct inode *inode = file_inode(file);
2254 struct shmem_inode_info *info = SHMEM_I(inode);
2255 int retval = -ENOMEM;
2256
2257 /*
2258 * What serializes the accesses to info->flags?
2259 * ipc_lock_object() when called from shmctl_do_lock(),
2260 * no serialization needed when called from shm_destroy().
2261 */
2262 if (lock && !(info->flags & VM_LOCKED)) {
2263 if (!user_shm_lock(inode->i_size, ucounts))
2264 goto out_nomem;
2265 info->flags |= VM_LOCKED;
2266 mapping_set_unevictable(file->f_mapping);
2267 }
2268 if (!lock && (info->flags & VM_LOCKED) && ucounts) {
2269 user_shm_unlock(inode->i_size, ucounts);
2270 info->flags &= ~VM_LOCKED;
2271 mapping_clear_unevictable(file->f_mapping);
2272 }
2273 retval = 0;
2274
2275out_nomem:
2276 return retval;
2277}
2278
2279static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
2280{
2281 struct inode *inode = file_inode(file);
2282 struct shmem_inode_info *info = SHMEM_I(inode);
2283 int ret;
2284
2285 ret = seal_check_future_write(info->seals, vma);
2286 if (ret)
2287 return ret;
2288
2289 /* arm64 - allow memory tagging on RAM-based files */
2290 vma->vm_flags |= VM_MTE_ALLOWED;
2291
2292 file_accessed(file);
2293 /* This is anonymous shared memory if it is unlinked at the time of mmap */
2294 if (inode->i_nlink)
2295 vma->vm_ops = &shmem_vm_ops;
2296 else
2297 vma->vm_ops = &shmem_anon_vm_ops;
2298 return 0;
2299}
2300
2301#ifdef CONFIG_TMPFS_XATTR
2302static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
2303
2304/*
2305 * chattr's fsflags are unrelated to extended attributes,
2306 * but tmpfs has chosen to enable them under the same config option.
2307 */
2308static void shmem_set_inode_flags(struct inode *inode, unsigned int fsflags)
2309{
2310 unsigned int i_flags = 0;
2311
2312 if (fsflags & FS_NOATIME_FL)
2313 i_flags |= S_NOATIME;
2314 if (fsflags & FS_APPEND_FL)
2315 i_flags |= S_APPEND;
2316 if (fsflags & FS_IMMUTABLE_FL)
2317 i_flags |= S_IMMUTABLE;
2318 /*
2319 * But FS_NODUMP_FL does not require any action in i_flags.
2320 */
2321 inode_set_flags(inode, i_flags, S_NOATIME | S_APPEND | S_IMMUTABLE);
2322}
2323#else
2324static void shmem_set_inode_flags(struct inode *inode, unsigned int fsflags)
2325{
2326}
2327#define shmem_initxattrs NULL
2328#endif
2329
2330static struct inode *shmem_get_inode(struct super_block *sb, struct inode *dir,
2331 umode_t mode, dev_t dev, unsigned long flags)
2332{
2333 struct inode *inode;
2334 struct shmem_inode_info *info;
2335 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2336 ino_t ino;
2337
2338 if (shmem_reserve_inode(sb, &ino))
2339 return NULL;
2340
2341 inode = new_inode(sb);
2342 if (inode) {
2343 inode->i_ino = ino;
2344 inode_init_owner(&init_user_ns, inode, dir, mode);
2345 inode->i_blocks = 0;
2346 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
2347 inode->i_generation = get_random_u32();
2348 info = SHMEM_I(inode);
2349 memset(info, 0, (char *)inode - (char *)info);
2350 spin_lock_init(&info->lock);
2351 atomic_set(&info->stop_eviction, 0);
2352 info->seals = F_SEAL_SEAL;
2353 info->flags = flags & VM_NORESERVE;
2354 info->i_crtime = inode->i_mtime;
2355 info->fsflags = (dir == NULL) ? 0 :
2356 SHMEM_I(dir)->fsflags & SHMEM_FL_INHERITED;
2357 if (info->fsflags)
2358 shmem_set_inode_flags(inode, info->fsflags);
2359 INIT_LIST_HEAD(&info->shrinklist);
2360 INIT_LIST_HEAD(&info->swaplist);
2361 simple_xattrs_init(&info->xattrs);
2362 cache_no_acl(inode);
2363 mapping_set_large_folios(inode->i_mapping);
2364
2365 switch (mode & S_IFMT) {
2366 default:
2367 inode->i_op = &shmem_special_inode_operations;
2368 init_special_inode(inode, mode, dev);
2369 break;
2370 case S_IFREG:
2371 inode->i_mapping->a_ops = &shmem_aops;
2372 inode->i_op = &shmem_inode_operations;
2373 inode->i_fop = &shmem_file_operations;
2374 mpol_shared_policy_init(&info->policy,
2375 shmem_get_sbmpol(sbinfo));
2376 break;
2377 case S_IFDIR:
2378 inc_nlink(inode);
2379 /* Some things misbehave if size == 0 on a directory */
2380 inode->i_size = 2 * BOGO_DIRENT_SIZE;
2381 inode->i_op = &shmem_dir_inode_operations;
2382 inode->i_fop = &simple_dir_operations;
2383 break;
2384 case S_IFLNK:
2385 /*
2386 * Must not load anything in the rbtree,
2387 * mpol_free_shared_policy will not be called.
2388 */
2389 mpol_shared_policy_init(&info->policy, NULL);
2390 break;
2391 }
2392
2393 lockdep_annotate_inode_mutex_key(inode);
2394 } else
2395 shmem_free_inode(sb);
2396 return inode;
2397}
2398
2399#ifdef CONFIG_USERFAULTFD
2400int shmem_mfill_atomic_pte(struct mm_struct *dst_mm,
2401 pmd_t *dst_pmd,
2402 struct vm_area_struct *dst_vma,
2403 unsigned long dst_addr,
2404 unsigned long src_addr,
2405 bool zeropage, bool wp_copy,
2406 struct page **pagep)
2407{
2408 struct inode *inode = file_inode(dst_vma->vm_file);
2409 struct shmem_inode_info *info = SHMEM_I(inode);
2410 struct address_space *mapping = inode->i_mapping;
2411 gfp_t gfp = mapping_gfp_mask(mapping);
2412 pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
2413 void *page_kaddr;
2414 struct folio *folio;
2415 int ret;
2416 pgoff_t max_off;
2417
2418 if (!shmem_inode_acct_block(inode, 1)) {
2419 /*
2420 * We may have got a page, returned -ENOENT triggering a retry,
2421 * and now we find ourselves with -ENOMEM. Release the page, to
2422 * avoid a BUG_ON in our caller.
2423 */
2424 if (unlikely(*pagep)) {
2425 put_page(*pagep);
2426 *pagep = NULL;
2427 }
2428 return -ENOMEM;
2429 }
2430
2431 if (!*pagep) {
2432 ret = -ENOMEM;
2433 folio = shmem_alloc_folio(gfp, info, pgoff);
2434 if (!folio)
2435 goto out_unacct_blocks;
2436
2437 if (!zeropage) { /* COPY */
2438 page_kaddr = kmap_local_folio(folio, 0);
2439 /*
2440 * The read mmap_lock is held here. Despite the
2441 * mmap_lock being read recursive a deadlock is still
2442 * possible if a writer has taken a lock. For example:
2443 *
2444 * process A thread 1 takes read lock on own mmap_lock
2445 * process A thread 2 calls mmap, blocks taking write lock
2446 * process B thread 1 takes page fault, read lock on own mmap lock
2447 * process B thread 2 calls mmap, blocks taking write lock
2448 * process A thread 1 blocks taking read lock on process B
2449 * process B thread 1 blocks taking read lock on process A
2450 *
2451 * Disable page faults to prevent potential deadlock
2452 * and retry the copy outside the mmap_lock.
2453 */
2454 pagefault_disable();
2455 ret = copy_from_user(page_kaddr,
2456 (const void __user *)src_addr,
2457 PAGE_SIZE);
2458 pagefault_enable();
2459 kunmap_local(page_kaddr);
2460
2461 /* fallback to copy_from_user outside mmap_lock */
2462 if (unlikely(ret)) {
2463 *pagep = &folio->page;
2464 ret = -ENOENT;
2465 /* don't free the page */
2466 goto out_unacct_blocks;
2467 }
2468
2469 flush_dcache_folio(folio);
2470 } else { /* ZEROPAGE */
2471 clear_user_highpage(&folio->page, dst_addr);
2472 }
2473 } else {
2474 folio = page_folio(*pagep);
2475 VM_BUG_ON_FOLIO(folio_test_large(folio), folio);
2476 *pagep = NULL;
2477 }
2478
2479 VM_BUG_ON(folio_test_locked(folio));
2480 VM_BUG_ON(folio_test_swapbacked(folio));
2481 __folio_set_locked(folio);
2482 __folio_set_swapbacked(folio);
2483 __folio_mark_uptodate(folio);
2484
2485 ret = -EFAULT;
2486 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2487 if (unlikely(pgoff >= max_off))
2488 goto out_release;
2489
2490 ret = shmem_add_to_page_cache(folio, mapping, pgoff, NULL,
2491 gfp & GFP_RECLAIM_MASK, dst_mm);
2492 if (ret)
2493 goto out_release;
2494
2495 ret = mfill_atomic_install_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
2496 &folio->page, true, wp_copy);
2497 if (ret)
2498 goto out_delete_from_cache;
2499
2500 spin_lock_irq(&info->lock);
2501 info->alloced++;
2502 inode->i_blocks += BLOCKS_PER_PAGE;
2503 shmem_recalc_inode(inode);
2504 spin_unlock_irq(&info->lock);
2505
2506 folio_unlock(folio);
2507 return 0;
2508out_delete_from_cache:
2509 filemap_remove_folio(folio);
2510out_release:
2511 folio_unlock(folio);
2512 folio_put(folio);
2513out_unacct_blocks:
2514 shmem_inode_unacct_blocks(inode, 1);
2515 return ret;
2516}
2517#endif /* CONFIG_USERFAULTFD */
2518
2519#ifdef CONFIG_TMPFS
2520static const struct inode_operations shmem_symlink_inode_operations;
2521static const struct inode_operations shmem_short_symlink_operations;
2522
2523static int
2524shmem_write_begin(struct file *file, struct address_space *mapping,
2525 loff_t pos, unsigned len,
2526 struct page **pagep, void **fsdata)
2527{
2528 struct inode *inode = mapping->host;
2529 struct shmem_inode_info *info = SHMEM_I(inode);
2530 pgoff_t index = pos >> PAGE_SHIFT;
2531 struct folio *folio;
2532 int ret = 0;
2533
2534 /* i_rwsem is held by caller */
2535 if (unlikely(info->seals & (F_SEAL_GROW |
2536 F_SEAL_WRITE | F_SEAL_FUTURE_WRITE))) {
2537 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE))
2538 return -EPERM;
2539 if ((info->seals & F_SEAL_GROW) && pos + len > inode->i_size)
2540 return -EPERM;
2541 }
2542
2543 ret = shmem_get_folio(inode, index, &folio, SGP_WRITE);
2544
2545 if (ret)
2546 return ret;
2547
2548 *pagep = folio_file_page(folio, index);
2549 if (PageHWPoison(*pagep)) {
2550 folio_unlock(folio);
2551 folio_put(folio);
2552 *pagep = NULL;
2553 return -EIO;
2554 }
2555
2556 return 0;
2557}
2558
2559static int
2560shmem_write_end(struct file *file, struct address_space *mapping,
2561 loff_t pos, unsigned len, unsigned copied,
2562 struct page *page, void *fsdata)
2563{
2564 struct inode *inode = mapping->host;
2565
2566 if (pos + copied > inode->i_size)
2567 i_size_write(inode, pos + copied);
2568
2569 if (!PageUptodate(page)) {
2570 struct page *head = compound_head(page);
2571 if (PageTransCompound(page)) {
2572 int i;
2573
2574 for (i = 0; i < HPAGE_PMD_NR; i++) {
2575 if (head + i == page)
2576 continue;
2577 clear_highpage(head + i);
2578 flush_dcache_page(head + i);
2579 }
2580 }
2581 if (copied < PAGE_SIZE) {
2582 unsigned from = pos & (PAGE_SIZE - 1);
2583 zero_user_segments(page, 0, from,
2584 from + copied, PAGE_SIZE);
2585 }
2586 SetPageUptodate(head);
2587 }
2588 set_page_dirty(page);
2589 unlock_page(page);
2590 put_page(page);
2591
2592 return copied;
2593}
2594
2595static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
2596{
2597 struct file *file = iocb->ki_filp;
2598 struct inode *inode = file_inode(file);
2599 struct address_space *mapping = inode->i_mapping;
2600 pgoff_t index;
2601 unsigned long offset;
2602 int error = 0;
2603 ssize_t retval = 0;
2604 loff_t *ppos = &iocb->ki_pos;
2605
2606 index = *ppos >> PAGE_SHIFT;
2607 offset = *ppos & ~PAGE_MASK;
2608
2609 for (;;) {
2610 struct folio *folio = NULL;
2611 struct page *page = NULL;
2612 pgoff_t end_index;
2613 unsigned long nr, ret;
2614 loff_t i_size = i_size_read(inode);
2615
2616 end_index = i_size >> PAGE_SHIFT;
2617 if (index > end_index)
2618 break;
2619 if (index == end_index) {
2620 nr = i_size & ~PAGE_MASK;
2621 if (nr <= offset)
2622 break;
2623 }
2624
2625 error = shmem_get_folio(inode, index, &folio, SGP_READ);
2626 if (error) {
2627 if (error == -EINVAL)
2628 error = 0;
2629 break;
2630 }
2631 if (folio) {
2632 folio_unlock(folio);
2633
2634 page = folio_file_page(folio, index);
2635 if (PageHWPoison(page)) {
2636 folio_put(folio);
2637 error = -EIO;
2638 break;
2639 }
2640 }
2641
2642 /*
2643 * We must evaluate after, since reads (unlike writes)
2644 * are called without i_rwsem protection against truncate
2645 */
2646 nr = PAGE_SIZE;
2647 i_size = i_size_read(inode);
2648 end_index = i_size >> PAGE_SHIFT;
2649 if (index == end_index) {
2650 nr = i_size & ~PAGE_MASK;
2651 if (nr <= offset) {
2652 if (folio)
2653 folio_put(folio);
2654 break;
2655 }
2656 }
2657 nr -= offset;
2658
2659 if (folio) {
2660 /*
2661 * If users can be writing to this page using arbitrary
2662 * virtual addresses, take care about potential aliasing
2663 * before reading the page on the kernel side.
2664 */
2665 if (mapping_writably_mapped(mapping))
2666 flush_dcache_page(page);
2667 /*
2668 * Mark the page accessed if we read the beginning.
2669 */
2670 if (!offset)
2671 folio_mark_accessed(folio);
2672 /*
2673 * Ok, we have the page, and it's up-to-date, so
2674 * now we can copy it to user space...
2675 */
2676 ret = copy_page_to_iter(page, offset, nr, to);
2677 folio_put(folio);
2678
2679 } else if (user_backed_iter(to)) {
2680 /*
2681 * Copy to user tends to be so well optimized, but
2682 * clear_user() not so much, that it is noticeably
2683 * faster to copy the zero page instead of clearing.
2684 */
2685 ret = copy_page_to_iter(ZERO_PAGE(0), offset, nr, to);
2686 } else {
2687 /*
2688 * But submitting the same page twice in a row to
2689 * splice() - or others? - can result in confusion:
2690 * so don't attempt that optimization on pipes etc.
2691 */
2692 ret = iov_iter_zero(nr, to);
2693 }
2694
2695 retval += ret;
2696 offset += ret;
2697 index += offset >> PAGE_SHIFT;
2698 offset &= ~PAGE_MASK;
2699
2700 if (!iov_iter_count(to))
2701 break;
2702 if (ret < nr) {
2703 error = -EFAULT;
2704 break;
2705 }
2706 cond_resched();
2707 }
2708
2709 *ppos = ((loff_t) index << PAGE_SHIFT) + offset;
2710 file_accessed(file);
2711 return retval ? retval : error;
2712}
2713
2714static loff_t shmem_file_llseek(struct file *file, loff_t offset, int whence)
2715{
2716 struct address_space *mapping = file->f_mapping;
2717 struct inode *inode = mapping->host;
2718
2719 if (whence != SEEK_DATA && whence != SEEK_HOLE)
2720 return generic_file_llseek_size(file, offset, whence,
2721 MAX_LFS_FILESIZE, i_size_read(inode));
2722 if (offset < 0)
2723 return -ENXIO;
2724
2725 inode_lock(inode);
2726 /* We're holding i_rwsem so we can access i_size directly */
2727 offset = mapping_seek_hole_data(mapping, offset, inode->i_size, whence);
2728 if (offset >= 0)
2729 offset = vfs_setpos(file, offset, MAX_LFS_FILESIZE);
2730 inode_unlock(inode);
2731 return offset;
2732}
2733
2734static long shmem_fallocate(struct file *file, int mode, loff_t offset,
2735 loff_t len)
2736{
2737 struct inode *inode = file_inode(file);
2738 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
2739 struct shmem_inode_info *info = SHMEM_I(inode);
2740 struct shmem_falloc shmem_falloc;
2741 pgoff_t start, index, end, undo_fallocend;
2742 int error;
2743
2744 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2745 return -EOPNOTSUPP;
2746
2747 inode_lock(inode);
2748
2749 if (mode & FALLOC_FL_PUNCH_HOLE) {
2750 struct address_space *mapping = file->f_mapping;
2751 loff_t unmap_start = round_up(offset, PAGE_SIZE);
2752 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
2753 DECLARE_WAIT_QUEUE_HEAD_ONSTACK(shmem_falloc_waitq);
2754
2755 /* protected by i_rwsem */
2756 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
2757 error = -EPERM;
2758 goto out;
2759 }
2760
2761 shmem_falloc.waitq = &shmem_falloc_waitq;
2762 shmem_falloc.start = (u64)unmap_start >> PAGE_SHIFT;
2763 shmem_falloc.next = (unmap_end + 1) >> PAGE_SHIFT;
2764 spin_lock(&inode->i_lock);
2765 inode->i_private = &shmem_falloc;
2766 spin_unlock(&inode->i_lock);
2767
2768 if ((u64)unmap_end > (u64)unmap_start)
2769 unmap_mapping_range(mapping, unmap_start,
2770 1 + unmap_end - unmap_start, 0);
2771 shmem_truncate_range(inode, offset, offset + len - 1);
2772 /* No need to unmap again: hole-punching leaves COWed pages */
2773
2774 spin_lock(&inode->i_lock);
2775 inode->i_private = NULL;
2776 wake_up_all(&shmem_falloc_waitq);
2777 WARN_ON_ONCE(!list_empty(&shmem_falloc_waitq.head));
2778 spin_unlock(&inode->i_lock);
2779 error = 0;
2780 goto out;
2781 }
2782
2783 /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
2784 error = inode_newsize_ok(inode, offset + len);
2785 if (error)
2786 goto out;
2787
2788 if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
2789 error = -EPERM;
2790 goto out;
2791 }
2792
2793 start = offset >> PAGE_SHIFT;
2794 end = (offset + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
2795 /* Try to avoid a swapstorm if len is impossible to satisfy */
2796 if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
2797 error = -ENOSPC;
2798 goto out;
2799 }
2800
2801 shmem_falloc.waitq = NULL;
2802 shmem_falloc.start = start;
2803 shmem_falloc.next = start;
2804 shmem_falloc.nr_falloced = 0;
2805 shmem_falloc.nr_unswapped = 0;
2806 spin_lock(&inode->i_lock);
2807 inode->i_private = &shmem_falloc;
2808 spin_unlock(&inode->i_lock);
2809
2810 /*
2811 * info->fallocend is only relevant when huge pages might be
2812 * involved: to prevent split_huge_page() freeing fallocated
2813 * pages when FALLOC_FL_KEEP_SIZE committed beyond i_size.
2814 */
2815 undo_fallocend = info->fallocend;
2816 if (info->fallocend < end)
2817 info->fallocend = end;
2818
2819 for (index = start; index < end; ) {
2820 struct folio *folio;
2821
2822 /*
2823 * Good, the fallocate(2) manpage permits EINTR: we may have
2824 * been interrupted because we are using up too much memory.
2825 */
2826 if (signal_pending(current))
2827 error = -EINTR;
2828 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
2829 error = -ENOMEM;
2830 else
2831 error = shmem_get_folio(inode, index, &folio,
2832 SGP_FALLOC);
2833 if (error) {
2834 info->fallocend = undo_fallocend;
2835 /* Remove the !uptodate folios we added */
2836 if (index > start) {
2837 shmem_undo_range(inode,
2838 (loff_t)start << PAGE_SHIFT,
2839 ((loff_t)index << PAGE_SHIFT) - 1, true);
2840 }
2841 goto undone;
2842 }
2843
2844 /*
2845 * Here is a more important optimization than it appears:
2846 * a second SGP_FALLOC on the same large folio will clear it,
2847 * making it uptodate and un-undoable if we fail later.
2848 */
2849 index = folio_next_index(folio);
2850 /* Beware 32-bit wraparound */
2851 if (!index)
2852 index--;
2853
2854 /*
2855 * Inform shmem_writepage() how far we have reached.
2856 * No need for lock or barrier: we have the page lock.
2857 */
2858 if (!folio_test_uptodate(folio))
2859 shmem_falloc.nr_falloced += index - shmem_falloc.next;
2860 shmem_falloc.next = index;
2861
2862 /*
2863 * If !uptodate, leave it that way so that freeable folios
2864 * can be recognized if we need to rollback on error later.
2865 * But mark it dirty so that memory pressure will swap rather
2866 * than free the folios we are allocating (and SGP_CACHE folios
2867 * might still be clean: we now need to mark those dirty too).
2868 */
2869 folio_mark_dirty(folio);
2870 folio_unlock(folio);
2871 folio_put(folio);
2872 cond_resched();
2873 }
2874
2875 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
2876 i_size_write(inode, offset + len);
2877undone:
2878 spin_lock(&inode->i_lock);
2879 inode->i_private = NULL;
2880 spin_unlock(&inode->i_lock);
2881out:
2882 if (!error)
2883 file_modified(file);
2884 inode_unlock(inode);
2885 return error;
2886}
2887
2888static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
2889{
2890 struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
2891
2892 buf->f_type = TMPFS_MAGIC;
2893 buf->f_bsize = PAGE_SIZE;
2894 buf->f_namelen = NAME_MAX;
2895 if (sbinfo->max_blocks) {
2896 buf->f_blocks = sbinfo->max_blocks;
2897 buf->f_bavail =
2898 buf->f_bfree = sbinfo->max_blocks -
2899 percpu_counter_sum(&sbinfo->used_blocks);
2900 }
2901 if (sbinfo->max_inodes) {
2902 buf->f_files = sbinfo->max_inodes;
2903 buf->f_ffree = sbinfo->free_inodes;
2904 }
2905 /* else leave those fields 0 like simple_statfs */
2906
2907 buf->f_fsid = uuid_to_fsid(dentry->d_sb->s_uuid.b);
2908
2909 return 0;
2910}
2911
2912/*
2913 * File creation. Allocate an inode, and we're done..
2914 */
2915static int
2916shmem_mknod(struct user_namespace *mnt_userns, struct inode *dir,
2917 struct dentry *dentry, umode_t mode, dev_t dev)
2918{
2919 struct inode *inode;
2920 int error = -ENOSPC;
2921
2922 inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
2923 if (inode) {
2924 error = simple_acl_create(dir, inode);
2925 if (error)
2926 goto out_iput;
2927 error = security_inode_init_security(inode, dir,
2928 &dentry->d_name,
2929 shmem_initxattrs, NULL);
2930 if (error && error != -EOPNOTSUPP)
2931 goto out_iput;
2932
2933 error = 0;
2934 dir->i_size += BOGO_DIRENT_SIZE;
2935 dir->i_ctime = dir->i_mtime = current_time(dir);
2936 inode_inc_iversion(dir);
2937 d_instantiate(dentry, inode);
2938 dget(dentry); /* Extra count - pin the dentry in core */
2939 }
2940 return error;
2941out_iput:
2942 iput(inode);
2943 return error;
2944}
2945
2946static int
2947shmem_tmpfile(struct user_namespace *mnt_userns, struct inode *dir,
2948 struct file *file, umode_t mode)
2949{
2950 struct inode *inode;
2951 int error = -ENOSPC;
2952
2953 inode = shmem_get_inode(dir->i_sb, dir, mode, 0, VM_NORESERVE);
2954 if (inode) {
2955 error = security_inode_init_security(inode, dir,
2956 NULL,
2957 shmem_initxattrs, NULL);
2958 if (error && error != -EOPNOTSUPP)
2959 goto out_iput;
2960 error = simple_acl_create(dir, inode);
2961 if (error)
2962 goto out_iput;
2963 d_tmpfile(file, inode);
2964 }
2965 return finish_open_simple(file, error);
2966out_iput:
2967 iput(inode);
2968 return error;
2969}
2970
2971static int shmem_mkdir(struct user_namespace *mnt_userns, struct inode *dir,
2972 struct dentry *dentry, umode_t mode)
2973{
2974 int error;
2975
2976 if ((error = shmem_mknod(&init_user_ns, dir, dentry,
2977 mode | S_IFDIR, 0)))
2978 return error;
2979 inc_nlink(dir);
2980 return 0;
2981}
2982
2983static int shmem_create(struct user_namespace *mnt_userns, struct inode *dir,
2984 struct dentry *dentry, umode_t mode, bool excl)
2985{
2986 return shmem_mknod(&init_user_ns, dir, dentry, mode | S_IFREG, 0);
2987}
2988
2989/*
2990 * Link a file..
2991 */
2992static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
2993{
2994 struct inode *inode = d_inode(old_dentry);
2995 int ret = 0;
2996
2997 /*
2998 * No ordinary (disk based) filesystem counts links as inodes;
2999 * but each new link needs a new dentry, pinning lowmem, and
3000 * tmpfs dentries cannot be pruned until they are unlinked.
3001 * But if an O_TMPFILE file is linked into the tmpfs, the
3002 * first link must skip that, to get the accounting right.
3003 */
3004 if (inode->i_nlink) {
3005 ret = shmem_reserve_inode(inode->i_sb, NULL);
3006 if (ret)
3007 goto out;
3008 }
3009
3010 dir->i_size += BOGO_DIRENT_SIZE;
3011 inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
3012 inode_inc_iversion(dir);
3013 inc_nlink(inode);
3014 ihold(inode); /* New dentry reference */
3015 dget(dentry); /* Extra pinning count for the created dentry */
3016 d_instantiate(dentry, inode);
3017out:
3018 return ret;
3019}
3020
3021static int shmem_unlink(struct inode *dir, struct dentry *dentry)
3022{
3023 struct inode *inode = d_inode(dentry);
3024
3025 if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
3026 shmem_free_inode(inode->i_sb);
3027
3028 dir->i_size -= BOGO_DIRENT_SIZE;
3029 inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
3030 inode_inc_iversion(dir);
3031 drop_nlink(inode);
3032 dput(dentry); /* Undo the count from "create" - this does all the work */
3033 return 0;
3034}
3035
3036static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
3037{
3038 if (!simple_empty(dentry))
3039 return -ENOTEMPTY;
3040
3041 drop_nlink(d_inode(dentry));
3042 drop_nlink(dir);
3043 return shmem_unlink(dir, dentry);
3044}
3045
3046static int shmem_whiteout(struct user_namespace *mnt_userns,
3047 struct inode *old_dir, struct dentry *old_dentry)
3048{
3049 struct dentry *whiteout;
3050 int error;
3051
3052 whiteout = d_alloc(old_dentry->d_parent, &old_dentry->d_name);
3053 if (!whiteout)
3054 return -ENOMEM;
3055
3056 error = shmem_mknod(&init_user_ns, old_dir, whiteout,
3057 S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
3058 dput(whiteout);
3059 if (error)
3060 return error;
3061
3062 /*
3063 * Cheat and hash the whiteout while the old dentry is still in
3064 * place, instead of playing games with FS_RENAME_DOES_D_MOVE.
3065 *
3066 * d_lookup() will consistently find one of them at this point,
3067 * not sure which one, but that isn't even important.
3068 */
3069 d_rehash(whiteout);
3070 return 0;
3071}
3072
3073/*
3074 * The VFS layer already does all the dentry stuff for rename,
3075 * we just have to decrement the usage count for the target if
3076 * it exists so that the VFS layer correctly free's it when it
3077 * gets overwritten.
3078 */
3079static int shmem_rename2(struct user_namespace *mnt_userns,
3080 struct inode *old_dir, struct dentry *old_dentry,
3081 struct inode *new_dir, struct dentry *new_dentry,
3082 unsigned int flags)
3083{
3084 struct inode *inode = d_inode(old_dentry);
3085 int they_are_dirs = S_ISDIR(inode->i_mode);
3086
3087 if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
3088 return -EINVAL;
3089
3090 if (flags & RENAME_EXCHANGE)
3091 return simple_rename_exchange(old_dir, old_dentry, new_dir, new_dentry);
3092
3093 if (!simple_empty(new_dentry))
3094 return -ENOTEMPTY;
3095
3096 if (flags & RENAME_WHITEOUT) {
3097 int error;
3098
3099 error = shmem_whiteout(&init_user_ns, old_dir, old_dentry);
3100 if (error)
3101 return error;
3102 }
3103
3104 if (d_really_is_positive(new_dentry)) {
3105 (void) shmem_unlink(new_dir, new_dentry);
3106 if (they_are_dirs) {
3107 drop_nlink(d_inode(new_dentry));
3108 drop_nlink(old_dir);
3109 }
3110 } else if (they_are_dirs) {
3111 drop_nlink(old_dir);
3112 inc_nlink(new_dir);
3113 }
3114
3115 old_dir->i_size -= BOGO_DIRENT_SIZE;
3116 new_dir->i_size += BOGO_DIRENT_SIZE;
3117 old_dir->i_ctime = old_dir->i_mtime =
3118 new_dir->i_ctime = new_dir->i_mtime =
3119 inode->i_ctime = current_time(old_dir);
3120 inode_inc_iversion(old_dir);
3121 inode_inc_iversion(new_dir);
3122 return 0;
3123}
3124
3125static int shmem_symlink(struct user_namespace *mnt_userns, struct inode *dir,
3126 struct dentry *dentry, const char *symname)
3127{
3128 int error;
3129 int len;
3130 struct inode *inode;
3131 struct folio *folio;
3132
3133 len = strlen(symname) + 1;
3134 if (len > PAGE_SIZE)
3135 return -ENAMETOOLONG;
3136
3137 inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK | 0777, 0,
3138 VM_NORESERVE);
3139 if (!inode)
3140 return -ENOSPC;
3141
3142 error = security_inode_init_security(inode, dir, &dentry->d_name,
3143 shmem_initxattrs, NULL);
3144 if (error && error != -EOPNOTSUPP) {
3145 iput(inode);
3146 return error;
3147 }
3148
3149 inode->i_size = len-1;
3150 if (len <= SHORT_SYMLINK_LEN) {
3151 inode->i_link = kmemdup(symname, len, GFP_KERNEL);
3152 if (!inode->i_link) {
3153 iput(inode);
3154 return -ENOMEM;
3155 }
3156 inode->i_op = &shmem_short_symlink_operations;
3157 } else {
3158 inode_nohighmem(inode);
3159 error = shmem_get_folio(inode, 0, &folio, SGP_WRITE);
3160 if (error) {
3161 iput(inode);
3162 return error;
3163 }
3164 inode->i_mapping->a_ops = &shmem_aops;
3165 inode->i_op = &shmem_symlink_inode_operations;
3166 memcpy(folio_address(folio), symname, len);
3167 folio_mark_uptodate(folio);
3168 folio_mark_dirty(folio);
3169 folio_unlock(folio);
3170 folio_put(folio);
3171 }
3172 dir->i_size += BOGO_DIRENT_SIZE;
3173 dir->i_ctime = dir->i_mtime = current_time(dir);
3174 inode_inc_iversion(dir);
3175 d_instantiate(dentry, inode);
3176 dget(dentry);
3177 return 0;
3178}
3179
3180static void shmem_put_link(void *arg)
3181{
3182 folio_mark_accessed(arg);
3183 folio_put(arg);
3184}
3185
3186static const char *shmem_get_link(struct dentry *dentry,
3187 struct inode *inode,
3188 struct delayed_call *done)
3189{
3190 struct folio *folio = NULL;
3191 int error;
3192
3193 if (!dentry) {
3194 folio = filemap_get_folio(inode->i_mapping, 0);
3195 if (!folio)
3196 return ERR_PTR(-ECHILD);
3197 if (PageHWPoison(folio_page(folio, 0)) ||
3198 !folio_test_uptodate(folio)) {
3199 folio_put(folio);
3200 return ERR_PTR(-ECHILD);
3201 }
3202 } else {
3203 error = shmem_get_folio(inode, 0, &folio, SGP_READ);
3204 if (error)
3205 return ERR_PTR(error);
3206 if (!folio)
3207 return ERR_PTR(-ECHILD);
3208 if (PageHWPoison(folio_page(folio, 0))) {
3209 folio_unlock(folio);
3210 folio_put(folio);
3211 return ERR_PTR(-ECHILD);
3212 }
3213 folio_unlock(folio);
3214 }
3215 set_delayed_call(done, shmem_put_link, folio);
3216 return folio_address(folio);
3217}
3218
3219#ifdef CONFIG_TMPFS_XATTR
3220
3221static int shmem_fileattr_get(struct dentry *dentry, struct fileattr *fa)
3222{
3223 struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
3224
3225 fileattr_fill_flags(fa, info->fsflags & SHMEM_FL_USER_VISIBLE);
3226
3227 return 0;
3228}
3229
3230static int shmem_fileattr_set(struct user_namespace *mnt_userns,
3231 struct dentry *dentry, struct fileattr *fa)
3232{
3233 struct inode *inode = d_inode(dentry);
3234 struct shmem_inode_info *info = SHMEM_I(inode);
3235
3236 if (fileattr_has_fsx(fa))
3237 return -EOPNOTSUPP;
3238 if (fa->flags & ~SHMEM_FL_USER_MODIFIABLE)
3239 return -EOPNOTSUPP;
3240
3241 info->fsflags = (info->fsflags & ~SHMEM_FL_USER_MODIFIABLE) |
3242 (fa->flags & SHMEM_FL_USER_MODIFIABLE);
3243
3244 shmem_set_inode_flags(inode, info->fsflags);
3245 inode->i_ctime = current_time(inode);
3246 inode_inc_iversion(inode);
3247 return 0;
3248}
3249
3250/*
3251 * Superblocks without xattr inode operations may get some security.* xattr
3252 * support from the LSM "for free". As soon as we have any other xattrs
3253 * like ACLs, we also need to implement the security.* handlers at
3254 * filesystem level, though.
3255 */
3256
3257/*
3258 * Callback for security_inode_init_security() for acquiring xattrs.
3259 */
3260static int shmem_initxattrs(struct inode *inode,
3261 const struct xattr *xattr_array,
3262 void *fs_info)
3263{
3264 struct shmem_inode_info *info = SHMEM_I(inode);
3265 const struct xattr *xattr;
3266 struct simple_xattr *new_xattr;
3267 size_t len;
3268
3269 for (xattr = xattr_array; xattr->name != NULL; xattr++) {
3270 new_xattr = simple_xattr_alloc(xattr->value, xattr->value_len);
3271 if (!new_xattr)
3272 return -ENOMEM;
3273
3274 len = strlen(xattr->name) + 1;
3275 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
3276 GFP_KERNEL);
3277 if (!new_xattr->name) {
3278 kvfree(new_xattr);
3279 return -ENOMEM;
3280 }
3281
3282 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
3283 XATTR_SECURITY_PREFIX_LEN);
3284 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
3285 xattr->name, len);
3286
3287 simple_xattr_add(&info->xattrs, new_xattr);
3288 }
3289
3290 return 0;
3291}
3292
3293static int shmem_xattr_handler_get(const struct xattr_handler *handler,
3294 struct dentry *unused, struct inode *inode,
3295 const char *name, void *buffer, size_t size)
3296{
3297 struct shmem_inode_info *info = SHMEM_I(inode);
3298
3299 name = xattr_full_name(handler, name);
3300 return simple_xattr_get(&info->xattrs, name, buffer, size);
3301}
3302
3303static int shmem_xattr_handler_set(const struct xattr_handler *handler,
3304 struct user_namespace *mnt_userns,
3305 struct dentry *unused, struct inode *inode,
3306 const char *name, const void *value,
3307 size_t size, int flags)
3308{
3309 struct shmem_inode_info *info = SHMEM_I(inode);
3310 int err;
3311
3312 name = xattr_full_name(handler, name);
3313 err = simple_xattr_set(&info->xattrs, name, value, size, flags, NULL);
3314 if (!err) {
3315 inode->i_ctime = current_time(inode);
3316 inode_inc_iversion(inode);
3317 }
3318 return err;
3319}
3320
3321static const struct xattr_handler shmem_security_xattr_handler = {
3322 .prefix = XATTR_SECURITY_PREFIX,
3323 .get = shmem_xattr_handler_get,
3324 .set = shmem_xattr_handler_set,
3325};
3326
3327static const struct xattr_handler shmem_trusted_xattr_handler = {
3328 .prefix = XATTR_TRUSTED_PREFIX,
3329 .get = shmem_xattr_handler_get,
3330 .set = shmem_xattr_handler_set,
3331};
3332
3333static const struct xattr_handler *shmem_xattr_handlers[] = {
3334#ifdef CONFIG_TMPFS_POSIX_ACL
3335 &posix_acl_access_xattr_handler,
3336 &posix_acl_default_xattr_handler,
3337#endif
3338 &shmem_security_xattr_handler,
3339 &shmem_trusted_xattr_handler,
3340 NULL
3341};
3342
3343static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
3344{
3345 struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
3346 return simple_xattr_list(d_inode(dentry), &info->xattrs, buffer, size);
3347}
3348#endif /* CONFIG_TMPFS_XATTR */
3349
3350static const struct inode_operations shmem_short_symlink_operations = {
3351 .getattr = shmem_getattr,
3352 .get_link = simple_get_link,
3353#ifdef CONFIG_TMPFS_XATTR
3354 .listxattr = shmem_listxattr,
3355#endif
3356};
3357
3358static const struct inode_operations shmem_symlink_inode_operations = {
3359 .getattr = shmem_getattr,
3360 .get_link = shmem_get_link,
3361#ifdef CONFIG_TMPFS_XATTR
3362 .listxattr = shmem_listxattr,
3363#endif
3364};
3365
3366static struct dentry *shmem_get_parent(struct dentry *child)
3367{
3368 return ERR_PTR(-ESTALE);
3369}
3370
3371static int shmem_match(struct inode *ino, void *vfh)
3372{
3373 __u32 *fh = vfh;
3374 __u64 inum = fh[2];
3375 inum = (inum << 32) | fh[1];
3376 return ino->i_ino == inum && fh[0] == ino->i_generation;
3377}
3378
3379/* Find any alias of inode, but prefer a hashed alias */
3380static struct dentry *shmem_find_alias(struct inode *inode)
3381{
3382 struct dentry *alias = d_find_alias(inode);
3383
3384 return alias ?: d_find_any_alias(inode);
3385}
3386
3387
3388static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
3389 struct fid *fid, int fh_len, int fh_type)
3390{
3391 struct inode *inode;
3392 struct dentry *dentry = NULL;
3393 u64 inum;
3394
3395 if (fh_len < 3)
3396 return NULL;
3397
3398 inum = fid->raw[2];
3399 inum = (inum << 32) | fid->raw[1];
3400
3401 inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
3402 shmem_match, fid->raw);
3403 if (inode) {
3404 dentry = shmem_find_alias(inode);
3405 iput(inode);
3406 }
3407
3408 return dentry;
3409}
3410
3411static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
3412 struct inode *parent)
3413{
3414 if (*len < 3) {
3415 *len = 3;
3416 return FILEID_INVALID;
3417 }
3418
3419 if (inode_unhashed(inode)) {
3420 /* Unfortunately insert_inode_hash is not idempotent,
3421 * so as we hash inodes here rather than at creation
3422 * time, we need a lock to ensure we only try
3423 * to do it once
3424 */
3425 static DEFINE_SPINLOCK(lock);
3426 spin_lock(&lock);
3427 if (inode_unhashed(inode))
3428 __insert_inode_hash(inode,
3429 inode->i_ino + inode->i_generation);
3430 spin_unlock(&lock);
3431 }
3432
3433 fh[0] = inode->i_generation;
3434 fh[1] = inode->i_ino;
3435 fh[2] = ((__u64)inode->i_ino) >> 32;
3436
3437 *len = 3;
3438 return 1;
3439}
3440
3441static const struct export_operations shmem_export_ops = {
3442 .get_parent = shmem_get_parent,
3443 .encode_fh = shmem_encode_fh,
3444 .fh_to_dentry = shmem_fh_to_dentry,
3445};
3446
3447enum shmem_param {
3448 Opt_gid,
3449 Opt_huge,
3450 Opt_mode,
3451 Opt_mpol,
3452 Opt_nr_blocks,
3453 Opt_nr_inodes,
3454 Opt_size,
3455 Opt_uid,
3456 Opt_inode32,
3457 Opt_inode64,
3458};
3459
3460static const struct constant_table shmem_param_enums_huge[] = {
3461 {"never", SHMEM_HUGE_NEVER },
3462 {"always", SHMEM_HUGE_ALWAYS },
3463 {"within_size", SHMEM_HUGE_WITHIN_SIZE },
3464 {"advise", SHMEM_HUGE_ADVISE },
3465 {}
3466};
3467
3468const struct fs_parameter_spec shmem_fs_parameters[] = {
3469 fsparam_u32 ("gid", Opt_gid),
3470 fsparam_enum ("huge", Opt_huge, shmem_param_enums_huge),
3471 fsparam_u32oct("mode", Opt_mode),
3472 fsparam_string("mpol", Opt_mpol),
3473 fsparam_string("nr_blocks", Opt_nr_blocks),
3474 fsparam_string("nr_inodes", Opt_nr_inodes),
3475 fsparam_string("size", Opt_size),
3476 fsparam_u32 ("uid", Opt_uid),
3477 fsparam_flag ("inode32", Opt_inode32),
3478 fsparam_flag ("inode64", Opt_inode64),
3479 {}
3480};
3481
3482static int shmem_parse_one(struct fs_context *fc, struct fs_parameter *param)
3483{
3484 struct shmem_options *ctx = fc->fs_private;
3485 struct fs_parse_result result;
3486 unsigned long long size;
3487 char *rest;
3488 int opt;
3489
3490 opt = fs_parse(fc, shmem_fs_parameters, param, &result);
3491 if (opt < 0)
3492 return opt;
3493
3494 switch (opt) {
3495 case Opt_size:
3496 size = memparse(param->string, &rest);
3497 if (*rest == '%') {
3498 size <<= PAGE_SHIFT;
3499 size *= totalram_pages();
3500 do_div(size, 100);
3501 rest++;
3502 }
3503 if (*rest)
3504 goto bad_value;
3505 ctx->blocks = DIV_ROUND_UP(size, PAGE_SIZE);
3506 ctx->seen |= SHMEM_SEEN_BLOCKS;
3507 break;
3508 case Opt_nr_blocks:
3509 ctx->blocks = memparse(param->string, &rest);
3510 if (*rest || ctx->blocks > S64_MAX)
3511 goto bad_value;
3512 ctx->seen |= SHMEM_SEEN_BLOCKS;
3513 break;
3514 case Opt_nr_inodes:
3515 ctx->inodes = memparse(param->string, &rest);
3516 if (*rest)
3517 goto bad_value;
3518 ctx->seen |= SHMEM_SEEN_INODES;
3519 break;
3520 case Opt_mode:
3521 ctx->mode = result.uint_32 & 07777;
3522 break;
3523 case Opt_uid:
3524 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
3525 if (!uid_valid(ctx->uid))
3526 goto bad_value;
3527 break;
3528 case Opt_gid:
3529 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
3530 if (!gid_valid(ctx->gid))
3531 goto bad_value;
3532 break;
3533 case Opt_huge:
3534 ctx->huge = result.uint_32;
3535 if (ctx->huge != SHMEM_HUGE_NEVER &&
3536 !(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
3537 has_transparent_hugepage()))
3538 goto unsupported_parameter;
3539 ctx->seen |= SHMEM_SEEN_HUGE;
3540 break;
3541 case Opt_mpol:
3542 if (IS_ENABLED(CONFIG_NUMA)) {
3543 mpol_put(ctx->mpol);
3544 ctx->mpol = NULL;
3545 if (mpol_parse_str(param->string, &ctx->mpol))
3546 goto bad_value;
3547 break;
3548 }
3549 goto unsupported_parameter;
3550 case Opt_inode32:
3551 ctx->full_inums = false;
3552 ctx->seen |= SHMEM_SEEN_INUMS;
3553 break;
3554 case Opt_inode64:
3555 if (sizeof(ino_t) < 8) {
3556 return invalfc(fc,
3557 "Cannot use inode64 with <64bit inums in kernel\n");
3558 }
3559 ctx->full_inums = true;
3560 ctx->seen |= SHMEM_SEEN_INUMS;
3561 break;
3562 }
3563 return 0;
3564
3565unsupported_parameter:
3566 return invalfc(fc, "Unsupported parameter '%s'", param->key);
3567bad_value:
3568 return invalfc(fc, "Bad value for '%s'", param->key);
3569}
3570
3571static int shmem_parse_options(struct fs_context *fc, void *data)
3572{
3573 char *options = data;
3574
3575 if (options) {
3576 int err = security_sb_eat_lsm_opts(options, &fc->security);
3577 if (err)
3578 return err;
3579 }
3580
3581 while (options != NULL) {
3582 char *this_char = options;
3583 for (;;) {
3584 /*
3585 * NUL-terminate this option: unfortunately,
3586 * mount options form a comma-separated list,
3587 * but mpol's nodelist may also contain commas.
3588 */
3589 options = strchr(options, ',');
3590 if (options == NULL)
3591 break;
3592 options++;
3593 if (!isdigit(*options)) {
3594 options[-1] = '\0';
3595 break;
3596 }
3597 }
3598 if (*this_char) {
3599 char *value = strchr(this_char, '=');
3600 size_t len = 0;
3601 int err;
3602
3603 if (value) {
3604 *value++ = '\0';
3605 len = strlen(value);
3606 }
3607 err = vfs_parse_fs_string(fc, this_char, value, len);
3608 if (err < 0)
3609 return err;
3610 }
3611 }
3612 return 0;
3613}
3614
3615/*
3616 * Reconfigure a shmem filesystem.
3617 *
3618 * Note that we disallow change from limited->unlimited blocks/inodes while any
3619 * are in use; but we must separately disallow unlimited->limited, because in
3620 * that case we have no record of how much is already in use.
3621 */
3622static int shmem_reconfigure(struct fs_context *fc)
3623{
3624 struct shmem_options *ctx = fc->fs_private;
3625 struct shmem_sb_info *sbinfo = SHMEM_SB(fc->root->d_sb);
3626 unsigned long inodes;
3627 struct mempolicy *mpol = NULL;
3628 const char *err;
3629
3630 raw_spin_lock(&sbinfo->stat_lock);
3631 inodes = sbinfo->max_inodes - sbinfo->free_inodes;
3632
3633 if ((ctx->seen & SHMEM_SEEN_BLOCKS) && ctx->blocks) {
3634 if (!sbinfo->max_blocks) {
3635 err = "Cannot retroactively limit size";
3636 goto out;
3637 }
3638 if (percpu_counter_compare(&sbinfo->used_blocks,
3639 ctx->blocks) > 0) {
3640 err = "Too small a size for current use";
3641 goto out;
3642 }
3643 }
3644 if ((ctx->seen & SHMEM_SEEN_INODES) && ctx->inodes) {
3645 if (!sbinfo->max_inodes) {
3646 err = "Cannot retroactively limit inodes";
3647 goto out;
3648 }
3649 if (ctx->inodes < inodes) {
3650 err = "Too few inodes for current use";
3651 goto out;
3652 }
3653 }
3654
3655 if ((ctx->seen & SHMEM_SEEN_INUMS) && !ctx->full_inums &&
3656 sbinfo->next_ino > UINT_MAX) {
3657 err = "Current inum too high to switch to 32-bit inums";
3658 goto out;
3659 }
3660
3661 if (ctx->seen & SHMEM_SEEN_HUGE)
3662 sbinfo->huge = ctx->huge;
3663 if (ctx->seen & SHMEM_SEEN_INUMS)
3664 sbinfo->full_inums = ctx->full_inums;
3665 if (ctx->seen & SHMEM_SEEN_BLOCKS)
3666 sbinfo->max_blocks = ctx->blocks;
3667 if (ctx->seen & SHMEM_SEEN_INODES) {
3668 sbinfo->max_inodes = ctx->inodes;
3669 sbinfo->free_inodes = ctx->inodes - inodes;
3670 }
3671
3672 /*
3673 * Preserve previous mempolicy unless mpol remount option was specified.
3674 */
3675 if (ctx->mpol) {
3676 mpol = sbinfo->mpol;
3677 sbinfo->mpol = ctx->mpol; /* transfers initial ref */
3678 ctx->mpol = NULL;
3679 }
3680 raw_spin_unlock(&sbinfo->stat_lock);
3681 mpol_put(mpol);
3682 return 0;
3683out:
3684 raw_spin_unlock(&sbinfo->stat_lock);
3685 return invalfc(fc, "%s", err);
3686}
3687
3688static int shmem_show_options(struct seq_file *seq, struct dentry *root)
3689{
3690 struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
3691
3692 if (sbinfo->max_blocks != shmem_default_max_blocks())
3693 seq_printf(seq, ",size=%luk",
3694 sbinfo->max_blocks << (PAGE_SHIFT - 10));
3695 if (sbinfo->max_inodes != shmem_default_max_inodes())
3696 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
3697 if (sbinfo->mode != (0777 | S_ISVTX))
3698 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
3699 if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
3700 seq_printf(seq, ",uid=%u",
3701 from_kuid_munged(&init_user_ns, sbinfo->uid));
3702 if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
3703 seq_printf(seq, ",gid=%u",
3704 from_kgid_munged(&init_user_ns, sbinfo->gid));
3705
3706 /*
3707 * Showing inode{64,32} might be useful even if it's the system default,
3708 * since then people don't have to resort to checking both here and
3709 * /proc/config.gz to confirm 64-bit inums were successfully applied
3710 * (which may not even exist if IKCONFIG_PROC isn't enabled).
3711 *
3712 * We hide it when inode64 isn't the default and we are using 32-bit
3713 * inodes, since that probably just means the feature isn't even under
3714 * consideration.
3715 *
3716 * As such:
3717 *
3718 * +-----------------+-----------------+
3719 * | TMPFS_INODE64=y | TMPFS_INODE64=n |
3720 * +------------------+-----------------+-----------------+
3721 * | full_inums=true | show | show |
3722 * | full_inums=false | show | hide |
3723 * +------------------+-----------------+-----------------+
3724 *
3725 */
3726 if (IS_ENABLED(CONFIG_TMPFS_INODE64) || sbinfo->full_inums)
3727 seq_printf(seq, ",inode%d", (sbinfo->full_inums ? 64 : 32));
3728#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3729 /* Rightly or wrongly, show huge mount option unmasked by shmem_huge */
3730 if (sbinfo->huge)
3731 seq_printf(seq, ",huge=%s", shmem_format_huge(sbinfo->huge));
3732#endif
3733 shmem_show_mpol(seq, sbinfo->mpol);
3734 return 0;
3735}
3736
3737#endif /* CONFIG_TMPFS */
3738
3739static void shmem_put_super(struct super_block *sb)
3740{
3741 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
3742
3743 free_percpu(sbinfo->ino_batch);
3744 percpu_counter_destroy(&sbinfo->used_blocks);
3745 mpol_put(sbinfo->mpol);
3746 kfree(sbinfo);
3747 sb->s_fs_info = NULL;
3748}
3749
3750static int shmem_fill_super(struct super_block *sb, struct fs_context *fc)
3751{
3752 struct shmem_options *ctx = fc->fs_private;
3753 struct inode *inode;
3754 struct shmem_sb_info *sbinfo;
3755
3756 /* Round up to L1_CACHE_BYTES to resist false sharing */
3757 sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
3758 L1_CACHE_BYTES), GFP_KERNEL);
3759 if (!sbinfo)
3760 return -ENOMEM;
3761
3762 sb->s_fs_info = sbinfo;
3763
3764#ifdef CONFIG_TMPFS
3765 /*
3766 * Per default we only allow half of the physical ram per
3767 * tmpfs instance, limiting inodes to one per page of lowmem;
3768 * but the internal instance is left unlimited.
3769 */
3770 if (!(sb->s_flags & SB_KERNMOUNT)) {
3771 if (!(ctx->seen & SHMEM_SEEN_BLOCKS))
3772 ctx->blocks = shmem_default_max_blocks();
3773 if (!(ctx->seen & SHMEM_SEEN_INODES))
3774 ctx->inodes = shmem_default_max_inodes();
3775 if (!(ctx->seen & SHMEM_SEEN_INUMS))
3776 ctx->full_inums = IS_ENABLED(CONFIG_TMPFS_INODE64);
3777 } else {
3778 sb->s_flags |= SB_NOUSER;
3779 }
3780 sb->s_export_op = &shmem_export_ops;
3781 sb->s_flags |= SB_NOSEC | SB_I_VERSION;
3782#else
3783 sb->s_flags |= SB_NOUSER;
3784#endif
3785 sbinfo->max_blocks = ctx->blocks;
3786 sbinfo->free_inodes = sbinfo->max_inodes = ctx->inodes;
3787 if (sb->s_flags & SB_KERNMOUNT) {
3788 sbinfo->ino_batch = alloc_percpu(ino_t);
3789 if (!sbinfo->ino_batch)
3790 goto failed;
3791 }
3792 sbinfo->uid = ctx->uid;
3793 sbinfo->gid = ctx->gid;
3794 sbinfo->full_inums = ctx->full_inums;
3795 sbinfo->mode = ctx->mode;
3796 sbinfo->huge = ctx->huge;
3797 sbinfo->mpol = ctx->mpol;
3798 ctx->mpol = NULL;
3799
3800 raw_spin_lock_init(&sbinfo->stat_lock);
3801 if (percpu_counter_init(&sbinfo->used_blocks, 0, GFP_KERNEL))
3802 goto failed;
3803 spin_lock_init(&sbinfo->shrinklist_lock);
3804 INIT_LIST_HEAD(&sbinfo->shrinklist);
3805
3806 sb->s_maxbytes = MAX_LFS_FILESIZE;
3807 sb->s_blocksize = PAGE_SIZE;
3808 sb->s_blocksize_bits = PAGE_SHIFT;
3809 sb->s_magic = TMPFS_MAGIC;
3810 sb->s_op = &shmem_ops;
3811 sb->s_time_gran = 1;
3812#ifdef CONFIG_TMPFS_XATTR
3813 sb->s_xattr = shmem_xattr_handlers;
3814#endif
3815#ifdef CONFIG_TMPFS_POSIX_ACL
3816 sb->s_flags |= SB_POSIXACL;
3817#endif
3818 uuid_gen(&sb->s_uuid);
3819
3820 inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
3821 if (!inode)
3822 goto failed;
3823 inode->i_uid = sbinfo->uid;
3824 inode->i_gid = sbinfo->gid;
3825 sb->s_root = d_make_root(inode);
3826 if (!sb->s_root)
3827 goto failed;
3828 return 0;
3829
3830failed:
3831 shmem_put_super(sb);
3832 return -ENOMEM;
3833}
3834
3835static int shmem_get_tree(struct fs_context *fc)
3836{
3837 return get_tree_nodev(fc, shmem_fill_super);
3838}
3839
3840static void shmem_free_fc(struct fs_context *fc)
3841{
3842 struct shmem_options *ctx = fc->fs_private;
3843
3844 if (ctx) {
3845 mpol_put(ctx->mpol);
3846 kfree(ctx);
3847 }
3848}
3849
3850static const struct fs_context_operations shmem_fs_context_ops = {
3851 .free = shmem_free_fc,
3852 .get_tree = shmem_get_tree,
3853#ifdef CONFIG_TMPFS
3854 .parse_monolithic = shmem_parse_options,
3855 .parse_param = shmem_parse_one,
3856 .reconfigure = shmem_reconfigure,
3857#endif
3858};
3859
3860static struct kmem_cache *shmem_inode_cachep;
3861
3862static struct inode *shmem_alloc_inode(struct super_block *sb)
3863{
3864 struct shmem_inode_info *info;
3865 info = alloc_inode_sb(sb, shmem_inode_cachep, GFP_KERNEL);
3866 if (!info)
3867 return NULL;
3868 return &info->vfs_inode;
3869}
3870
3871static void shmem_free_in_core_inode(struct inode *inode)
3872{
3873 if (S_ISLNK(inode->i_mode))
3874 kfree(inode->i_link);
3875 kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
3876}
3877
3878static void shmem_destroy_inode(struct inode *inode)
3879{
3880 if (S_ISREG(inode->i_mode))
3881 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
3882}
3883
3884static void shmem_init_inode(void *foo)
3885{
3886 struct shmem_inode_info *info = foo;
3887 inode_init_once(&info->vfs_inode);
3888}
3889
3890static void shmem_init_inodecache(void)
3891{
3892 shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
3893 sizeof(struct shmem_inode_info),
3894 0, SLAB_PANIC|SLAB_ACCOUNT, shmem_init_inode);
3895}
3896
3897static void shmem_destroy_inodecache(void)
3898{
3899 kmem_cache_destroy(shmem_inode_cachep);
3900}
3901
3902/* Keep the page in page cache instead of truncating it */
3903static int shmem_error_remove_page(struct address_space *mapping,
3904 struct page *page)
3905{
3906 return 0;
3907}
3908
3909const struct address_space_operations shmem_aops = {
3910 .writepage = shmem_writepage,
3911 .dirty_folio = noop_dirty_folio,
3912#ifdef CONFIG_TMPFS
3913 .write_begin = shmem_write_begin,
3914 .write_end = shmem_write_end,
3915#endif
3916#ifdef CONFIG_MIGRATION
3917 .migrate_folio = migrate_folio,
3918#endif
3919 .error_remove_page = shmem_error_remove_page,
3920};
3921EXPORT_SYMBOL(shmem_aops);
3922
3923static const struct file_operations shmem_file_operations = {
3924 .mmap = shmem_mmap,
3925 .open = generic_file_open,
3926 .get_unmapped_area = shmem_get_unmapped_area,
3927#ifdef CONFIG_TMPFS
3928 .llseek = shmem_file_llseek,
3929 .read_iter = shmem_file_read_iter,
3930 .write_iter = generic_file_write_iter,
3931 .fsync = noop_fsync,
3932 .splice_read = generic_file_splice_read,
3933 .splice_write = iter_file_splice_write,
3934 .fallocate = shmem_fallocate,
3935#endif
3936};
3937
3938static const struct inode_operations shmem_inode_operations = {
3939 .getattr = shmem_getattr,
3940 .setattr = shmem_setattr,
3941#ifdef CONFIG_TMPFS_XATTR
3942 .listxattr = shmem_listxattr,
3943 .set_acl = simple_set_acl,
3944 .fileattr_get = shmem_fileattr_get,
3945 .fileattr_set = shmem_fileattr_set,
3946#endif
3947};
3948
3949static const struct inode_operations shmem_dir_inode_operations = {
3950#ifdef CONFIG_TMPFS
3951 .getattr = shmem_getattr,
3952 .create = shmem_create,
3953 .lookup = simple_lookup,
3954 .link = shmem_link,
3955 .unlink = shmem_unlink,
3956 .symlink = shmem_symlink,
3957 .mkdir = shmem_mkdir,
3958 .rmdir = shmem_rmdir,
3959 .mknod = shmem_mknod,
3960 .rename = shmem_rename2,
3961 .tmpfile = shmem_tmpfile,
3962#endif
3963#ifdef CONFIG_TMPFS_XATTR
3964 .listxattr = shmem_listxattr,
3965 .fileattr_get = shmem_fileattr_get,
3966 .fileattr_set = shmem_fileattr_set,
3967#endif
3968#ifdef CONFIG_TMPFS_POSIX_ACL
3969 .setattr = shmem_setattr,
3970 .set_acl = simple_set_acl,
3971#endif
3972};
3973
3974static const struct inode_operations shmem_special_inode_operations = {
3975 .getattr = shmem_getattr,
3976#ifdef CONFIG_TMPFS_XATTR
3977 .listxattr = shmem_listxattr,
3978#endif
3979#ifdef CONFIG_TMPFS_POSIX_ACL
3980 .setattr = shmem_setattr,
3981 .set_acl = simple_set_acl,
3982#endif
3983};
3984
3985static const struct super_operations shmem_ops = {
3986 .alloc_inode = shmem_alloc_inode,
3987 .free_inode = shmem_free_in_core_inode,
3988 .destroy_inode = shmem_destroy_inode,
3989#ifdef CONFIG_TMPFS
3990 .statfs = shmem_statfs,
3991 .show_options = shmem_show_options,
3992#endif
3993 .evict_inode = shmem_evict_inode,
3994 .drop_inode = generic_delete_inode,
3995 .put_super = shmem_put_super,
3996#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3997 .nr_cached_objects = shmem_unused_huge_count,
3998 .free_cached_objects = shmem_unused_huge_scan,
3999#endif
4000};
4001
4002static const struct vm_operations_struct shmem_vm_ops = {
4003 .fault = shmem_fault,
4004 .map_pages = filemap_map_pages,
4005#ifdef CONFIG_NUMA
4006 .set_policy = shmem_set_policy,
4007 .get_policy = shmem_get_policy,
4008#endif
4009};
4010
4011static const struct vm_operations_struct shmem_anon_vm_ops = {
4012 .fault = shmem_fault,
4013 .map_pages = filemap_map_pages,
4014#ifdef CONFIG_NUMA
4015 .set_policy = shmem_set_policy,
4016 .get_policy = shmem_get_policy,
4017#endif
4018};
4019
4020int shmem_init_fs_context(struct fs_context *fc)
4021{
4022 struct shmem_options *ctx;
4023
4024 ctx = kzalloc(sizeof(struct shmem_options), GFP_KERNEL);
4025 if (!ctx)
4026 return -ENOMEM;
4027
4028 ctx->mode = 0777 | S_ISVTX;
4029 ctx->uid = current_fsuid();
4030 ctx->gid = current_fsgid();
4031
4032 fc->fs_private = ctx;
4033 fc->ops = &shmem_fs_context_ops;
4034 return 0;
4035}
4036
4037static struct file_system_type shmem_fs_type = {
4038 .owner = THIS_MODULE,
4039 .name = "tmpfs",
4040 .init_fs_context = shmem_init_fs_context,
4041#ifdef CONFIG_TMPFS
4042 .parameters = shmem_fs_parameters,
4043#endif
4044 .kill_sb = kill_litter_super,
4045 .fs_flags = FS_USERNS_MOUNT,
4046};
4047
4048void __init shmem_init(void)
4049{
4050 int error;
4051
4052 shmem_init_inodecache();
4053
4054 error = register_filesystem(&shmem_fs_type);
4055 if (error) {
4056 pr_err("Could not register tmpfs\n");
4057 goto out2;
4058 }
4059
4060 shm_mnt = kern_mount(&shmem_fs_type);
4061 if (IS_ERR(shm_mnt)) {
4062 error = PTR_ERR(shm_mnt);
4063 pr_err("Could not kern_mount tmpfs\n");
4064 goto out1;
4065 }
4066
4067#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4068 if (has_transparent_hugepage() && shmem_huge > SHMEM_HUGE_DENY)
4069 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
4070 else
4071 shmem_huge = SHMEM_HUGE_NEVER; /* just in case it was patched */
4072#endif
4073 return;
4074
4075out1:
4076 unregister_filesystem(&shmem_fs_type);
4077out2:
4078 shmem_destroy_inodecache();
4079 shm_mnt = ERR_PTR(error);
4080}
4081
4082#if defined(CONFIG_TRANSPARENT_HUGEPAGE) && defined(CONFIG_SYSFS)
4083static ssize_t shmem_enabled_show(struct kobject *kobj,
4084 struct kobj_attribute *attr, char *buf)
4085{
4086 static const int values[] = {
4087 SHMEM_HUGE_ALWAYS,
4088 SHMEM_HUGE_WITHIN_SIZE,
4089 SHMEM_HUGE_ADVISE,
4090 SHMEM_HUGE_NEVER,
4091 SHMEM_HUGE_DENY,
4092 SHMEM_HUGE_FORCE,
4093 };
4094 int len = 0;
4095 int i;
4096
4097 for (i = 0; i < ARRAY_SIZE(values); i++) {
4098 len += sysfs_emit_at(buf, len,
4099 shmem_huge == values[i] ? "%s[%s]" : "%s%s",
4100 i ? " " : "",
4101 shmem_format_huge(values[i]));
4102 }
4103
4104 len += sysfs_emit_at(buf, len, "\n");
4105
4106 return len;
4107}
4108
4109static ssize_t shmem_enabled_store(struct kobject *kobj,
4110 struct kobj_attribute *attr, const char *buf, size_t count)
4111{
4112 char tmp[16];
4113 int huge;
4114
4115 if (count + 1 > sizeof(tmp))
4116 return -EINVAL;
4117 memcpy(tmp, buf, count);
4118 tmp[count] = '\0';
4119 if (count && tmp[count - 1] == '\n')
4120 tmp[count - 1] = '\0';
4121
4122 huge = shmem_parse_huge(tmp);
4123 if (huge == -EINVAL)
4124 return -EINVAL;
4125 if (!has_transparent_hugepage() &&
4126 huge != SHMEM_HUGE_NEVER && huge != SHMEM_HUGE_DENY)
4127 return -EINVAL;
4128
4129 shmem_huge = huge;
4130 if (shmem_huge > SHMEM_HUGE_DENY)
4131 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
4132 return count;
4133}
4134
4135struct kobj_attribute shmem_enabled_attr = __ATTR_RW(shmem_enabled);
4136#endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_SYSFS */
4137
4138#else /* !CONFIG_SHMEM */
4139
4140/*
4141 * tiny-shmem: simple shmemfs and tmpfs using ramfs code
4142 *
4143 * This is intended for small system where the benefits of the full
4144 * shmem code (swap-backed and resource-limited) are outweighed by
4145 * their complexity. On systems without swap this code should be
4146 * effectively equivalent, but much lighter weight.
4147 */
4148
4149static struct file_system_type shmem_fs_type = {
4150 .name = "tmpfs",
4151 .init_fs_context = ramfs_init_fs_context,
4152 .parameters = ramfs_fs_parameters,
4153 .kill_sb = kill_litter_super,
4154 .fs_flags = FS_USERNS_MOUNT,
4155};
4156
4157void __init shmem_init(void)
4158{
4159 BUG_ON(register_filesystem(&shmem_fs_type) != 0);
4160
4161 shm_mnt = kern_mount(&shmem_fs_type);
4162 BUG_ON(IS_ERR(shm_mnt));
4163}
4164
4165int shmem_unuse(unsigned int type)
4166{
4167 return 0;
4168}
4169
4170int shmem_lock(struct file *file, int lock, struct ucounts *ucounts)
4171{
4172 return 0;
4173}
4174
4175void shmem_unlock_mapping(struct address_space *mapping)
4176{
4177}
4178
4179#ifdef CONFIG_MMU
4180unsigned long shmem_get_unmapped_area(struct file *file,
4181 unsigned long addr, unsigned long len,
4182 unsigned long pgoff, unsigned long flags)
4183{
4184 return current->mm->get_unmapped_area(file, addr, len, pgoff, flags);
4185}
4186#endif
4187
4188void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
4189{
4190 truncate_inode_pages_range(inode->i_mapping, lstart, lend);
4191}
4192EXPORT_SYMBOL_GPL(shmem_truncate_range);
4193
4194#define shmem_vm_ops generic_file_vm_ops
4195#define shmem_anon_vm_ops generic_file_vm_ops
4196#define shmem_file_operations ramfs_file_operations
4197#define shmem_get_inode(sb, dir, mode, dev, flags) ramfs_get_inode(sb, dir, mode, dev)
4198#define shmem_acct_size(flags, size) 0
4199#define shmem_unacct_size(flags, size) do {} while (0)
4200
4201#endif /* CONFIG_SHMEM */
4202
4203/* common code */
4204
4205static struct file *__shmem_file_setup(struct vfsmount *mnt, const char *name, loff_t size,
4206 unsigned long flags, unsigned int i_flags)
4207{
4208 struct inode *inode;
4209 struct file *res;
4210
4211 if (IS_ERR(mnt))
4212 return ERR_CAST(mnt);
4213
4214 if (size < 0 || size > MAX_LFS_FILESIZE)
4215 return ERR_PTR(-EINVAL);
4216
4217 if (shmem_acct_size(flags, size))
4218 return ERR_PTR(-ENOMEM);
4219
4220 inode = shmem_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0,
4221 flags);
4222 if (unlikely(!inode)) {
4223 shmem_unacct_size(flags, size);
4224 return ERR_PTR(-ENOSPC);
4225 }
4226 inode->i_flags |= i_flags;
4227 inode->i_size = size;
4228 clear_nlink(inode); /* It is unlinked */
4229 res = ERR_PTR(ramfs_nommu_expand_for_mapping(inode, size));
4230 if (!IS_ERR(res))
4231 res = alloc_file_pseudo(inode, mnt, name, O_RDWR,
4232 &shmem_file_operations);
4233 if (IS_ERR(res))
4234 iput(inode);
4235 return res;
4236}
4237
4238/**
4239 * shmem_kernel_file_setup - get an unlinked file living in tmpfs which must be
4240 * kernel internal. There will be NO LSM permission checks against the
4241 * underlying inode. So users of this interface must do LSM checks at a
4242 * higher layer. The users are the big_key and shm implementations. LSM
4243 * checks are provided at the key or shm level rather than the inode.
4244 * @name: name for dentry (to be seen in /proc/<pid>/maps
4245 * @size: size to be set for the file
4246 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4247 */
4248struct file *shmem_kernel_file_setup(const char *name, loff_t size, unsigned long flags)
4249{
4250 return __shmem_file_setup(shm_mnt, name, size, flags, S_PRIVATE);
4251}
4252
4253/**
4254 * shmem_file_setup - get an unlinked file living in tmpfs
4255 * @name: name for dentry (to be seen in /proc/<pid>/maps
4256 * @size: size to be set for the file
4257 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4258 */
4259struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
4260{
4261 return __shmem_file_setup(shm_mnt, name, size, flags, 0);
4262}
4263EXPORT_SYMBOL_GPL(shmem_file_setup);
4264
4265/**
4266 * shmem_file_setup_with_mnt - get an unlinked file living in tmpfs
4267 * @mnt: the tmpfs mount where the file will be created
4268 * @name: name for dentry (to be seen in /proc/<pid>/maps
4269 * @size: size to be set for the file
4270 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4271 */
4272struct file *shmem_file_setup_with_mnt(struct vfsmount *mnt, const char *name,
4273 loff_t size, unsigned long flags)
4274{
4275 return __shmem_file_setup(mnt, name, size, flags, 0);
4276}
4277EXPORT_SYMBOL_GPL(shmem_file_setup_with_mnt);
4278
4279/**
4280 * shmem_zero_setup - setup a shared anonymous mapping
4281 * @vma: the vma to be mmapped is prepared by do_mmap
4282 */
4283int shmem_zero_setup(struct vm_area_struct *vma)
4284{
4285 struct file *file;
4286 loff_t size = vma->vm_end - vma->vm_start;
4287
4288 /*
4289 * Cloning a new file under mmap_lock leads to a lock ordering conflict
4290 * between XFS directory reading and selinux: since this file is only
4291 * accessible to the user through its mapping, use S_PRIVATE flag to
4292 * bypass file security, in the same way as shmem_kernel_file_setup().
4293 */
4294 file = shmem_kernel_file_setup("dev/zero", size, vma->vm_flags);
4295 if (IS_ERR(file))
4296 return PTR_ERR(file);
4297
4298 if (vma->vm_file)
4299 fput(vma->vm_file);
4300 vma->vm_file = file;
4301 vma->vm_ops = &shmem_anon_vm_ops;
4302
4303 return 0;
4304}
4305
4306/**
4307 * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
4308 * @mapping: the page's address_space
4309 * @index: the page index
4310 * @gfp: the page allocator flags to use if allocating
4311 *
4312 * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
4313 * with any new page allocations done using the specified allocation flags.
4314 * But read_cache_page_gfp() uses the ->read_folio() method: which does not
4315 * suit tmpfs, since it may have pages in swapcache, and needs to find those
4316 * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
4317 *
4318 * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
4319 * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
4320 */
4321struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
4322 pgoff_t index, gfp_t gfp)
4323{
4324#ifdef CONFIG_SHMEM
4325 struct inode *inode = mapping->host;
4326 struct folio *folio;
4327 struct page *page;
4328 int error;
4329
4330 BUG_ON(!shmem_mapping(mapping));
4331 error = shmem_get_folio_gfp(inode, index, &folio, SGP_CACHE,
4332 gfp, NULL, NULL, NULL);
4333 if (error)
4334 return ERR_PTR(error);
4335
4336 folio_unlock(folio);
4337 page = folio_file_page(folio, index);
4338 if (PageHWPoison(page)) {
4339 folio_put(folio);
4340 return ERR_PTR(-EIO);
4341 }
4342
4343 return page;
4344#else
4345 /*
4346 * The tiny !SHMEM case uses ramfs without swap
4347 */
4348 return read_cache_page_gfp(mapping, index, gfp);
4349#endif
4350}
4351EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);
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/export.h>
33#include <linux/swap.h>
34#include <linux/uio.h>
35
36static struct vfsmount *shm_mnt;
37
38#ifdef CONFIG_SHMEM
39/*
40 * This virtual memory filesystem is heavily based on the ramfs. It
41 * extends ramfs by the ability to use swap and honor resource limits
42 * which makes it a completely usable filesystem.
43 */
44
45#include <linux/xattr.h>
46#include <linux/exportfs.h>
47#include <linux/posix_acl.h>
48#include <linux/posix_acl_xattr.h>
49#include <linux/mman.h>
50#include <linux/string.h>
51#include <linux/slab.h>
52#include <linux/backing-dev.h>
53#include <linux/shmem_fs.h>
54#include <linux/writeback.h>
55#include <linux/blkdev.h>
56#include <linux/pagevec.h>
57#include <linux/percpu_counter.h>
58#include <linux/falloc.h>
59#include <linux/splice.h>
60#include <linux/security.h>
61#include <linux/swapops.h>
62#include <linux/mempolicy.h>
63#include <linux/namei.h>
64#include <linux/ctype.h>
65#include <linux/migrate.h>
66#include <linux/highmem.h>
67#include <linux/seq_file.h>
68#include <linux/magic.h>
69#include <linux/syscalls.h>
70#include <linux/fcntl.h>
71#include <uapi/linux/memfd.h>
72
73#include <asm/uaccess.h>
74#include <asm/pgtable.h>
75
76#include "internal.h"
77
78#define BLOCKS_PER_PAGE (PAGE_SIZE/512)
79#define VM_ACCT(size) (PAGE_ALIGN(size) >> PAGE_SHIFT)
80
81/* Pretend that each entry is of this size in directory's i_size */
82#define BOGO_DIRENT_SIZE 20
83
84/* Symlink up to this size is kmalloc'ed instead of using a swappable page */
85#define SHORT_SYMLINK_LEN 128
86
87/*
88 * shmem_fallocate communicates with shmem_fault or shmem_writepage via
89 * inode->i_private (with i_mutex making sure that it has only one user at
90 * a time): we would prefer not to enlarge the shmem inode just for that.
91 */
92struct shmem_falloc {
93 wait_queue_head_t *waitq; /* faults into hole wait for punch to end */
94 pgoff_t start; /* start of range currently being fallocated */
95 pgoff_t next; /* the next page offset to be fallocated */
96 pgoff_t nr_falloced; /* how many new pages have been fallocated */
97 pgoff_t nr_unswapped; /* how often writepage refused to swap out */
98};
99
100/* Flag allocation requirements to shmem_getpage */
101enum sgp_type {
102 SGP_READ, /* don't exceed i_size, don't allocate page */
103 SGP_CACHE, /* don't exceed i_size, may allocate page */
104 SGP_DIRTY, /* like SGP_CACHE, but set new page dirty */
105 SGP_WRITE, /* may exceed i_size, may allocate !Uptodate page */
106 SGP_FALLOC, /* like SGP_WRITE, but make existing page Uptodate */
107};
108
109#ifdef CONFIG_TMPFS
110static unsigned long shmem_default_max_blocks(void)
111{
112 return totalram_pages / 2;
113}
114
115static unsigned long shmem_default_max_inodes(void)
116{
117 return min(totalram_pages - totalhigh_pages, totalram_pages / 2);
118}
119#endif
120
121static bool shmem_should_replace_page(struct page *page, gfp_t gfp);
122static int shmem_replace_page(struct page **pagep, gfp_t gfp,
123 struct shmem_inode_info *info, pgoff_t index);
124static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
125 struct page **pagep, enum sgp_type sgp, gfp_t gfp, int *fault_type);
126
127static inline int shmem_getpage(struct inode *inode, pgoff_t index,
128 struct page **pagep, enum sgp_type sgp, int *fault_type)
129{
130 return shmem_getpage_gfp(inode, index, pagep, sgp,
131 mapping_gfp_mask(inode->i_mapping), fault_type);
132}
133
134static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
135{
136 return sb->s_fs_info;
137}
138
139/*
140 * shmem_file_setup pre-accounts the whole fixed size of a VM object,
141 * for shared memory and for shared anonymous (/dev/zero) mappings
142 * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
143 * consistent with the pre-accounting of private mappings ...
144 */
145static inline int shmem_acct_size(unsigned long flags, loff_t size)
146{
147 return (flags & VM_NORESERVE) ?
148 0 : security_vm_enough_memory_mm(current->mm, VM_ACCT(size));
149}
150
151static inline void shmem_unacct_size(unsigned long flags, loff_t size)
152{
153 if (!(flags & VM_NORESERVE))
154 vm_unacct_memory(VM_ACCT(size));
155}
156
157static inline int shmem_reacct_size(unsigned long flags,
158 loff_t oldsize, loff_t newsize)
159{
160 if (!(flags & VM_NORESERVE)) {
161 if (VM_ACCT(newsize) > VM_ACCT(oldsize))
162 return security_vm_enough_memory_mm(current->mm,
163 VM_ACCT(newsize) - VM_ACCT(oldsize));
164 else if (VM_ACCT(newsize) < VM_ACCT(oldsize))
165 vm_unacct_memory(VM_ACCT(oldsize) - VM_ACCT(newsize));
166 }
167 return 0;
168}
169
170/*
171 * ... whereas tmpfs objects are accounted incrementally as
172 * pages are allocated, in order to allow huge sparse files.
173 * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
174 * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
175 */
176static inline int shmem_acct_block(unsigned long flags)
177{
178 return (flags & VM_NORESERVE) ?
179 security_vm_enough_memory_mm(current->mm, VM_ACCT(PAGE_SIZE)) : 0;
180}
181
182static inline void shmem_unacct_blocks(unsigned long flags, long pages)
183{
184 if (flags & VM_NORESERVE)
185 vm_unacct_memory(pages * VM_ACCT(PAGE_SIZE));
186}
187
188static const struct super_operations shmem_ops;
189static const struct address_space_operations shmem_aops;
190static const struct file_operations shmem_file_operations;
191static const struct inode_operations shmem_inode_operations;
192static const struct inode_operations shmem_dir_inode_operations;
193static const struct inode_operations shmem_special_inode_operations;
194static const struct vm_operations_struct shmem_vm_ops;
195
196static LIST_HEAD(shmem_swaplist);
197static DEFINE_MUTEX(shmem_swaplist_mutex);
198
199static int shmem_reserve_inode(struct super_block *sb)
200{
201 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
202 if (sbinfo->max_inodes) {
203 spin_lock(&sbinfo->stat_lock);
204 if (!sbinfo->free_inodes) {
205 spin_unlock(&sbinfo->stat_lock);
206 return -ENOSPC;
207 }
208 sbinfo->free_inodes--;
209 spin_unlock(&sbinfo->stat_lock);
210 }
211 return 0;
212}
213
214static void shmem_free_inode(struct super_block *sb)
215{
216 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
217 if (sbinfo->max_inodes) {
218 spin_lock(&sbinfo->stat_lock);
219 sbinfo->free_inodes++;
220 spin_unlock(&sbinfo->stat_lock);
221 }
222}
223
224/**
225 * shmem_recalc_inode - recalculate the block usage of an inode
226 * @inode: inode to recalc
227 *
228 * We have to calculate the free blocks since the mm can drop
229 * undirtied hole pages behind our back.
230 *
231 * But normally info->alloced == inode->i_mapping->nrpages + info->swapped
232 * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
233 *
234 * It has to be called with the spinlock held.
235 */
236static void shmem_recalc_inode(struct inode *inode)
237{
238 struct shmem_inode_info *info = SHMEM_I(inode);
239 long freed;
240
241 freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
242 if (freed > 0) {
243 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
244 if (sbinfo->max_blocks)
245 percpu_counter_add(&sbinfo->used_blocks, -freed);
246 info->alloced -= freed;
247 inode->i_blocks -= freed * BLOCKS_PER_PAGE;
248 shmem_unacct_blocks(info->flags, freed);
249 }
250}
251
252/*
253 * Replace item expected in radix tree by a new item, while holding tree lock.
254 */
255static int shmem_radix_tree_replace(struct address_space *mapping,
256 pgoff_t index, void *expected, void *replacement)
257{
258 void **pslot;
259 void *item;
260
261 VM_BUG_ON(!expected);
262 VM_BUG_ON(!replacement);
263 pslot = radix_tree_lookup_slot(&mapping->page_tree, index);
264 if (!pslot)
265 return -ENOENT;
266 item = radix_tree_deref_slot_protected(pslot, &mapping->tree_lock);
267 if (item != expected)
268 return -ENOENT;
269 radix_tree_replace_slot(pslot, replacement);
270 return 0;
271}
272
273/*
274 * Sometimes, before we decide whether to proceed or to fail, we must check
275 * that an entry was not already brought back from swap by a racing thread.
276 *
277 * Checking page is not enough: by the time a SwapCache page is locked, it
278 * might be reused, and again be SwapCache, using the same swap as before.
279 */
280static bool shmem_confirm_swap(struct address_space *mapping,
281 pgoff_t index, swp_entry_t swap)
282{
283 void *item;
284
285 rcu_read_lock();
286 item = radix_tree_lookup(&mapping->page_tree, index);
287 rcu_read_unlock();
288 return item == swp_to_radix_entry(swap);
289}
290
291/*
292 * Like add_to_page_cache_locked, but error if expected item has gone.
293 */
294static int shmem_add_to_page_cache(struct page *page,
295 struct address_space *mapping,
296 pgoff_t index, void *expected)
297{
298 int error;
299
300 VM_BUG_ON_PAGE(!PageLocked(page), page);
301 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
302
303 get_page(page);
304 page->mapping = mapping;
305 page->index = index;
306
307 spin_lock_irq(&mapping->tree_lock);
308 if (!expected)
309 error = radix_tree_insert(&mapping->page_tree, index, page);
310 else
311 error = shmem_radix_tree_replace(mapping, index, expected,
312 page);
313 if (!error) {
314 mapping->nrpages++;
315 __inc_zone_page_state(page, NR_FILE_PAGES);
316 __inc_zone_page_state(page, NR_SHMEM);
317 spin_unlock_irq(&mapping->tree_lock);
318 } else {
319 page->mapping = NULL;
320 spin_unlock_irq(&mapping->tree_lock);
321 put_page(page);
322 }
323 return error;
324}
325
326/*
327 * Like delete_from_page_cache, but substitutes swap for page.
328 */
329static void shmem_delete_from_page_cache(struct page *page, void *radswap)
330{
331 struct address_space *mapping = page->mapping;
332 int error;
333
334 spin_lock_irq(&mapping->tree_lock);
335 error = shmem_radix_tree_replace(mapping, page->index, page, radswap);
336 page->mapping = NULL;
337 mapping->nrpages--;
338 __dec_zone_page_state(page, NR_FILE_PAGES);
339 __dec_zone_page_state(page, NR_SHMEM);
340 spin_unlock_irq(&mapping->tree_lock);
341 put_page(page);
342 BUG_ON(error);
343}
344
345/*
346 * Remove swap entry from radix tree, free the swap and its page cache.
347 */
348static int shmem_free_swap(struct address_space *mapping,
349 pgoff_t index, void *radswap)
350{
351 void *old;
352
353 spin_lock_irq(&mapping->tree_lock);
354 old = radix_tree_delete_item(&mapping->page_tree, index, radswap);
355 spin_unlock_irq(&mapping->tree_lock);
356 if (old != radswap)
357 return -ENOENT;
358 free_swap_and_cache(radix_to_swp_entry(radswap));
359 return 0;
360}
361
362/*
363 * Determine (in bytes) how many of the shmem object's pages mapped by the
364 * given offsets are swapped out.
365 *
366 * This is safe to call without i_mutex or mapping->tree_lock thanks to RCU,
367 * as long as the inode doesn't go away and racy results are not a problem.
368 */
369unsigned long shmem_partial_swap_usage(struct address_space *mapping,
370 pgoff_t start, pgoff_t end)
371{
372 struct radix_tree_iter iter;
373 void **slot;
374 struct page *page;
375 unsigned long swapped = 0;
376
377 rcu_read_lock();
378
379 radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
380 if (iter.index >= end)
381 break;
382
383 page = radix_tree_deref_slot(slot);
384
385 if (radix_tree_deref_retry(page)) {
386 slot = radix_tree_iter_retry(&iter);
387 continue;
388 }
389
390 if (radix_tree_exceptional_entry(page))
391 swapped++;
392
393 if (need_resched()) {
394 cond_resched_rcu();
395 slot = radix_tree_iter_next(&iter);
396 }
397 }
398
399 rcu_read_unlock();
400
401 return swapped << PAGE_SHIFT;
402}
403
404/*
405 * Determine (in bytes) how many of the shmem object's pages mapped by the
406 * given vma is swapped out.
407 *
408 * This is safe to call without i_mutex or mapping->tree_lock thanks to RCU,
409 * as long as the inode doesn't go away and racy results are not a problem.
410 */
411unsigned long shmem_swap_usage(struct vm_area_struct *vma)
412{
413 struct inode *inode = file_inode(vma->vm_file);
414 struct shmem_inode_info *info = SHMEM_I(inode);
415 struct address_space *mapping = inode->i_mapping;
416 unsigned long swapped;
417
418 /* Be careful as we don't hold info->lock */
419 swapped = READ_ONCE(info->swapped);
420
421 /*
422 * The easier cases are when the shmem object has nothing in swap, or
423 * the vma maps it whole. Then we can simply use the stats that we
424 * already track.
425 */
426 if (!swapped)
427 return 0;
428
429 if (!vma->vm_pgoff && vma->vm_end - vma->vm_start >= inode->i_size)
430 return swapped << PAGE_SHIFT;
431
432 /* Here comes the more involved part */
433 return shmem_partial_swap_usage(mapping,
434 linear_page_index(vma, vma->vm_start),
435 linear_page_index(vma, vma->vm_end));
436}
437
438/*
439 * SysV IPC SHM_UNLOCK restore Unevictable pages to their evictable lists.
440 */
441void shmem_unlock_mapping(struct address_space *mapping)
442{
443 struct pagevec pvec;
444 pgoff_t indices[PAGEVEC_SIZE];
445 pgoff_t index = 0;
446
447 pagevec_init(&pvec, 0);
448 /*
449 * Minor point, but we might as well stop if someone else SHM_LOCKs it.
450 */
451 while (!mapping_unevictable(mapping)) {
452 /*
453 * Avoid pagevec_lookup(): find_get_pages() returns 0 as if it
454 * has finished, if it hits a row of PAGEVEC_SIZE swap entries.
455 */
456 pvec.nr = find_get_entries(mapping, index,
457 PAGEVEC_SIZE, pvec.pages, indices);
458 if (!pvec.nr)
459 break;
460 index = indices[pvec.nr - 1] + 1;
461 pagevec_remove_exceptionals(&pvec);
462 check_move_unevictable_pages(pvec.pages, pvec.nr);
463 pagevec_release(&pvec);
464 cond_resched();
465 }
466}
467
468/*
469 * Remove range of pages and swap entries from radix tree, and free them.
470 * If !unfalloc, truncate or punch hole; if unfalloc, undo failed fallocate.
471 */
472static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
473 bool unfalloc)
474{
475 struct address_space *mapping = inode->i_mapping;
476 struct shmem_inode_info *info = SHMEM_I(inode);
477 pgoff_t start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
478 pgoff_t end = (lend + 1) >> PAGE_SHIFT;
479 unsigned int partial_start = lstart & (PAGE_SIZE - 1);
480 unsigned int partial_end = (lend + 1) & (PAGE_SIZE - 1);
481 struct pagevec pvec;
482 pgoff_t indices[PAGEVEC_SIZE];
483 long nr_swaps_freed = 0;
484 pgoff_t index;
485 int i;
486
487 if (lend == -1)
488 end = -1; /* unsigned, so actually very big */
489
490 pagevec_init(&pvec, 0);
491 index = start;
492 while (index < end) {
493 pvec.nr = find_get_entries(mapping, index,
494 min(end - index, (pgoff_t)PAGEVEC_SIZE),
495 pvec.pages, indices);
496 if (!pvec.nr)
497 break;
498 for (i = 0; i < pagevec_count(&pvec); i++) {
499 struct page *page = pvec.pages[i];
500
501 index = indices[i];
502 if (index >= end)
503 break;
504
505 if (radix_tree_exceptional_entry(page)) {
506 if (unfalloc)
507 continue;
508 nr_swaps_freed += !shmem_free_swap(mapping,
509 index, page);
510 continue;
511 }
512
513 if (!trylock_page(page))
514 continue;
515 if (!unfalloc || !PageUptodate(page)) {
516 if (page->mapping == mapping) {
517 VM_BUG_ON_PAGE(PageWriteback(page), page);
518 truncate_inode_page(mapping, page);
519 }
520 }
521 unlock_page(page);
522 }
523 pagevec_remove_exceptionals(&pvec);
524 pagevec_release(&pvec);
525 cond_resched();
526 index++;
527 }
528
529 if (partial_start) {
530 struct page *page = NULL;
531 shmem_getpage(inode, start - 1, &page, SGP_READ, NULL);
532 if (page) {
533 unsigned int top = PAGE_SIZE;
534 if (start > end) {
535 top = partial_end;
536 partial_end = 0;
537 }
538 zero_user_segment(page, partial_start, top);
539 set_page_dirty(page);
540 unlock_page(page);
541 put_page(page);
542 }
543 }
544 if (partial_end) {
545 struct page *page = NULL;
546 shmem_getpage(inode, end, &page, SGP_READ, NULL);
547 if (page) {
548 zero_user_segment(page, 0, partial_end);
549 set_page_dirty(page);
550 unlock_page(page);
551 put_page(page);
552 }
553 }
554 if (start >= end)
555 return;
556
557 index = start;
558 while (index < end) {
559 cond_resched();
560
561 pvec.nr = find_get_entries(mapping, index,
562 min(end - index, (pgoff_t)PAGEVEC_SIZE),
563 pvec.pages, indices);
564 if (!pvec.nr) {
565 /* If all gone or hole-punch or unfalloc, we're done */
566 if (index == start || end != -1)
567 break;
568 /* But if truncating, restart to make sure all gone */
569 index = start;
570 continue;
571 }
572 for (i = 0; i < pagevec_count(&pvec); i++) {
573 struct page *page = pvec.pages[i];
574
575 index = indices[i];
576 if (index >= end)
577 break;
578
579 if (radix_tree_exceptional_entry(page)) {
580 if (unfalloc)
581 continue;
582 if (shmem_free_swap(mapping, index, page)) {
583 /* Swap was replaced by page: retry */
584 index--;
585 break;
586 }
587 nr_swaps_freed++;
588 continue;
589 }
590
591 lock_page(page);
592 if (!unfalloc || !PageUptodate(page)) {
593 if (page->mapping == mapping) {
594 VM_BUG_ON_PAGE(PageWriteback(page), page);
595 truncate_inode_page(mapping, page);
596 } else {
597 /* Page was replaced by swap: retry */
598 unlock_page(page);
599 index--;
600 break;
601 }
602 }
603 unlock_page(page);
604 }
605 pagevec_remove_exceptionals(&pvec);
606 pagevec_release(&pvec);
607 index++;
608 }
609
610 spin_lock(&info->lock);
611 info->swapped -= nr_swaps_freed;
612 shmem_recalc_inode(inode);
613 spin_unlock(&info->lock);
614}
615
616void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
617{
618 shmem_undo_range(inode, lstart, lend, false);
619 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
620}
621EXPORT_SYMBOL_GPL(shmem_truncate_range);
622
623static int shmem_getattr(struct vfsmount *mnt, struct dentry *dentry,
624 struct kstat *stat)
625{
626 struct inode *inode = dentry->d_inode;
627 struct shmem_inode_info *info = SHMEM_I(inode);
628
629 if (info->alloced - info->swapped != inode->i_mapping->nrpages) {
630 spin_lock(&info->lock);
631 shmem_recalc_inode(inode);
632 spin_unlock(&info->lock);
633 }
634 generic_fillattr(inode, stat);
635 return 0;
636}
637
638static int shmem_setattr(struct dentry *dentry, struct iattr *attr)
639{
640 struct inode *inode = d_inode(dentry);
641 struct shmem_inode_info *info = SHMEM_I(inode);
642 int error;
643
644 error = inode_change_ok(inode, attr);
645 if (error)
646 return error;
647
648 if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
649 loff_t oldsize = inode->i_size;
650 loff_t newsize = attr->ia_size;
651
652 /* protected by i_mutex */
653 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
654 (newsize > oldsize && (info->seals & F_SEAL_GROW)))
655 return -EPERM;
656
657 if (newsize != oldsize) {
658 error = shmem_reacct_size(SHMEM_I(inode)->flags,
659 oldsize, newsize);
660 if (error)
661 return error;
662 i_size_write(inode, newsize);
663 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
664 }
665 if (newsize <= oldsize) {
666 loff_t holebegin = round_up(newsize, PAGE_SIZE);
667 if (oldsize > holebegin)
668 unmap_mapping_range(inode->i_mapping,
669 holebegin, 0, 1);
670 if (info->alloced)
671 shmem_truncate_range(inode,
672 newsize, (loff_t)-1);
673 /* unmap again to remove racily COWed private pages */
674 if (oldsize > holebegin)
675 unmap_mapping_range(inode->i_mapping,
676 holebegin, 0, 1);
677 }
678 }
679
680 setattr_copy(inode, attr);
681 if (attr->ia_valid & ATTR_MODE)
682 error = posix_acl_chmod(inode, inode->i_mode);
683 return error;
684}
685
686static void shmem_evict_inode(struct inode *inode)
687{
688 struct shmem_inode_info *info = SHMEM_I(inode);
689
690 if (inode->i_mapping->a_ops == &shmem_aops) {
691 shmem_unacct_size(info->flags, inode->i_size);
692 inode->i_size = 0;
693 shmem_truncate_range(inode, 0, (loff_t)-1);
694 if (!list_empty(&info->swaplist)) {
695 mutex_lock(&shmem_swaplist_mutex);
696 list_del_init(&info->swaplist);
697 mutex_unlock(&shmem_swaplist_mutex);
698 }
699 }
700
701 simple_xattrs_free(&info->xattrs);
702 WARN_ON(inode->i_blocks);
703 shmem_free_inode(inode->i_sb);
704 clear_inode(inode);
705}
706
707/*
708 * If swap found in inode, free it and move page from swapcache to filecache.
709 */
710static int shmem_unuse_inode(struct shmem_inode_info *info,
711 swp_entry_t swap, struct page **pagep)
712{
713 struct address_space *mapping = info->vfs_inode.i_mapping;
714 void *radswap;
715 pgoff_t index;
716 gfp_t gfp;
717 int error = 0;
718
719 radswap = swp_to_radix_entry(swap);
720 index = radix_tree_locate_item(&mapping->page_tree, radswap);
721 if (index == -1)
722 return -EAGAIN; /* tell shmem_unuse we found nothing */
723
724 /*
725 * Move _head_ to start search for next from here.
726 * But be careful: shmem_evict_inode checks list_empty without taking
727 * mutex, and there's an instant in list_move_tail when info->swaplist
728 * would appear empty, if it were the only one on shmem_swaplist.
729 */
730 if (shmem_swaplist.next != &info->swaplist)
731 list_move_tail(&shmem_swaplist, &info->swaplist);
732
733 gfp = mapping_gfp_mask(mapping);
734 if (shmem_should_replace_page(*pagep, gfp)) {
735 mutex_unlock(&shmem_swaplist_mutex);
736 error = shmem_replace_page(pagep, gfp, info, index);
737 mutex_lock(&shmem_swaplist_mutex);
738 /*
739 * We needed to drop mutex to make that restrictive page
740 * allocation, but the inode might have been freed while we
741 * dropped it: although a racing shmem_evict_inode() cannot
742 * complete without emptying the radix_tree, our page lock
743 * on this swapcache page is not enough to prevent that -
744 * free_swap_and_cache() of our swap entry will only
745 * trylock_page(), removing swap from radix_tree whatever.
746 *
747 * We must not proceed to shmem_add_to_page_cache() if the
748 * inode has been freed, but of course we cannot rely on
749 * inode or mapping or info to check that. However, we can
750 * safely check if our swap entry is still in use (and here
751 * it can't have got reused for another page): if it's still
752 * in use, then the inode cannot have been freed yet, and we
753 * can safely proceed (if it's no longer in use, that tells
754 * nothing about the inode, but we don't need to unuse swap).
755 */
756 if (!page_swapcount(*pagep))
757 error = -ENOENT;
758 }
759
760 /*
761 * We rely on shmem_swaplist_mutex, not only to protect the swaplist,
762 * but also to hold up shmem_evict_inode(): so inode cannot be freed
763 * beneath us (pagelock doesn't help until the page is in pagecache).
764 */
765 if (!error)
766 error = shmem_add_to_page_cache(*pagep, mapping, index,
767 radswap);
768 if (error != -ENOMEM) {
769 /*
770 * Truncation and eviction use free_swap_and_cache(), which
771 * only does trylock page: if we raced, best clean up here.
772 */
773 delete_from_swap_cache(*pagep);
774 set_page_dirty(*pagep);
775 if (!error) {
776 spin_lock(&info->lock);
777 info->swapped--;
778 spin_unlock(&info->lock);
779 swap_free(swap);
780 }
781 }
782 return error;
783}
784
785/*
786 * Search through swapped inodes to find and replace swap by page.
787 */
788int shmem_unuse(swp_entry_t swap, struct page *page)
789{
790 struct list_head *this, *next;
791 struct shmem_inode_info *info;
792 struct mem_cgroup *memcg;
793 int error = 0;
794
795 /*
796 * There's a faint possibility that swap page was replaced before
797 * caller locked it: caller will come back later with the right page.
798 */
799 if (unlikely(!PageSwapCache(page) || page_private(page) != swap.val))
800 goto out;
801
802 /*
803 * Charge page using GFP_KERNEL while we can wait, before taking
804 * the shmem_swaplist_mutex which might hold up shmem_writepage().
805 * Charged back to the user (not to caller) when swap account is used.
806 */
807 error = mem_cgroup_try_charge(page, current->mm, GFP_KERNEL, &memcg,
808 false);
809 if (error)
810 goto out;
811 /* No radix_tree_preload: swap entry keeps a place for page in tree */
812 error = -EAGAIN;
813
814 mutex_lock(&shmem_swaplist_mutex);
815 list_for_each_safe(this, next, &shmem_swaplist) {
816 info = list_entry(this, struct shmem_inode_info, swaplist);
817 if (info->swapped)
818 error = shmem_unuse_inode(info, swap, &page);
819 else
820 list_del_init(&info->swaplist);
821 cond_resched();
822 if (error != -EAGAIN)
823 break;
824 /* found nothing in this: move on to search the next */
825 }
826 mutex_unlock(&shmem_swaplist_mutex);
827
828 if (error) {
829 if (error != -ENOMEM)
830 error = 0;
831 mem_cgroup_cancel_charge(page, memcg, false);
832 } else
833 mem_cgroup_commit_charge(page, memcg, true, false);
834out:
835 unlock_page(page);
836 put_page(page);
837 return error;
838}
839
840/*
841 * Move the page from the page cache to the swap cache.
842 */
843static int shmem_writepage(struct page *page, struct writeback_control *wbc)
844{
845 struct shmem_inode_info *info;
846 struct address_space *mapping;
847 struct inode *inode;
848 swp_entry_t swap;
849 pgoff_t index;
850
851 BUG_ON(!PageLocked(page));
852 mapping = page->mapping;
853 index = page->index;
854 inode = mapping->host;
855 info = SHMEM_I(inode);
856 if (info->flags & VM_LOCKED)
857 goto redirty;
858 if (!total_swap_pages)
859 goto redirty;
860
861 /*
862 * Our capabilities prevent regular writeback or sync from ever calling
863 * shmem_writepage; but a stacking filesystem might use ->writepage of
864 * its underlying filesystem, in which case tmpfs should write out to
865 * swap only in response to memory pressure, and not for the writeback
866 * threads or sync.
867 */
868 if (!wbc->for_reclaim) {
869 WARN_ON_ONCE(1); /* Still happens? Tell us about it! */
870 goto redirty;
871 }
872
873 /*
874 * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
875 * value into swapfile.c, the only way we can correctly account for a
876 * fallocated page arriving here is now to initialize it and write it.
877 *
878 * That's okay for a page already fallocated earlier, but if we have
879 * not yet completed the fallocation, then (a) we want to keep track
880 * of this page in case we have to undo it, and (b) it may not be a
881 * good idea to continue anyway, once we're pushing into swap. So
882 * reactivate the page, and let shmem_fallocate() quit when too many.
883 */
884 if (!PageUptodate(page)) {
885 if (inode->i_private) {
886 struct shmem_falloc *shmem_falloc;
887 spin_lock(&inode->i_lock);
888 shmem_falloc = inode->i_private;
889 if (shmem_falloc &&
890 !shmem_falloc->waitq &&
891 index >= shmem_falloc->start &&
892 index < shmem_falloc->next)
893 shmem_falloc->nr_unswapped++;
894 else
895 shmem_falloc = NULL;
896 spin_unlock(&inode->i_lock);
897 if (shmem_falloc)
898 goto redirty;
899 }
900 clear_highpage(page);
901 flush_dcache_page(page);
902 SetPageUptodate(page);
903 }
904
905 swap = get_swap_page();
906 if (!swap.val)
907 goto redirty;
908
909 if (mem_cgroup_try_charge_swap(page, swap))
910 goto free_swap;
911
912 /*
913 * Add inode to shmem_unuse()'s list of swapped-out inodes,
914 * if it's not already there. Do it now before the page is
915 * moved to swap cache, when its pagelock no longer protects
916 * the inode from eviction. But don't unlock the mutex until
917 * we've incremented swapped, because shmem_unuse_inode() will
918 * prune a !swapped inode from the swaplist under this mutex.
919 */
920 mutex_lock(&shmem_swaplist_mutex);
921 if (list_empty(&info->swaplist))
922 list_add_tail(&info->swaplist, &shmem_swaplist);
923
924 if (add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
925 spin_lock(&info->lock);
926 shmem_recalc_inode(inode);
927 info->swapped++;
928 spin_unlock(&info->lock);
929
930 swap_shmem_alloc(swap);
931 shmem_delete_from_page_cache(page, swp_to_radix_entry(swap));
932
933 mutex_unlock(&shmem_swaplist_mutex);
934 BUG_ON(page_mapped(page));
935 swap_writepage(page, wbc);
936 return 0;
937 }
938
939 mutex_unlock(&shmem_swaplist_mutex);
940free_swap:
941 swapcache_free(swap);
942redirty:
943 set_page_dirty(page);
944 if (wbc->for_reclaim)
945 return AOP_WRITEPAGE_ACTIVATE; /* Return with page locked */
946 unlock_page(page);
947 return 0;
948}
949
950#ifdef CONFIG_NUMA
951#ifdef CONFIG_TMPFS
952static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
953{
954 char buffer[64];
955
956 if (!mpol || mpol->mode == MPOL_DEFAULT)
957 return; /* show nothing */
958
959 mpol_to_str(buffer, sizeof(buffer), mpol);
960
961 seq_printf(seq, ",mpol=%s", buffer);
962}
963
964static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
965{
966 struct mempolicy *mpol = NULL;
967 if (sbinfo->mpol) {
968 spin_lock(&sbinfo->stat_lock); /* prevent replace/use races */
969 mpol = sbinfo->mpol;
970 mpol_get(mpol);
971 spin_unlock(&sbinfo->stat_lock);
972 }
973 return mpol;
974}
975#endif /* CONFIG_TMPFS */
976
977static struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
978 struct shmem_inode_info *info, pgoff_t index)
979{
980 struct vm_area_struct pvma;
981 struct page *page;
982
983 /* Create a pseudo vma that just contains the policy */
984 pvma.vm_start = 0;
985 /* Bias interleave by inode number to distribute better across nodes */
986 pvma.vm_pgoff = index + info->vfs_inode.i_ino;
987 pvma.vm_ops = NULL;
988 pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, index);
989
990 page = swapin_readahead(swap, gfp, &pvma, 0);
991
992 /* Drop reference taken by mpol_shared_policy_lookup() */
993 mpol_cond_put(pvma.vm_policy);
994
995 return page;
996}
997
998static struct page *shmem_alloc_page(gfp_t gfp,
999 struct shmem_inode_info *info, pgoff_t index)
1000{
1001 struct vm_area_struct pvma;
1002 struct page *page;
1003
1004 /* Create a pseudo vma that just contains the policy */
1005 pvma.vm_start = 0;
1006 /* Bias interleave by inode number to distribute better across nodes */
1007 pvma.vm_pgoff = index + info->vfs_inode.i_ino;
1008 pvma.vm_ops = NULL;
1009 pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, index);
1010
1011 page = alloc_page_vma(gfp, &pvma, 0);
1012
1013 /* Drop reference taken by mpol_shared_policy_lookup() */
1014 mpol_cond_put(pvma.vm_policy);
1015
1016 return page;
1017}
1018#else /* !CONFIG_NUMA */
1019#ifdef CONFIG_TMPFS
1020static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1021{
1022}
1023#endif /* CONFIG_TMPFS */
1024
1025static inline struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
1026 struct shmem_inode_info *info, pgoff_t index)
1027{
1028 return swapin_readahead(swap, gfp, NULL, 0);
1029}
1030
1031static inline struct page *shmem_alloc_page(gfp_t gfp,
1032 struct shmem_inode_info *info, pgoff_t index)
1033{
1034 return alloc_page(gfp);
1035}
1036#endif /* CONFIG_NUMA */
1037
1038#if !defined(CONFIG_NUMA) || !defined(CONFIG_TMPFS)
1039static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1040{
1041 return NULL;
1042}
1043#endif
1044
1045/*
1046 * When a page is moved from swapcache to shmem filecache (either by the
1047 * usual swapin of shmem_getpage_gfp(), or by the less common swapoff of
1048 * shmem_unuse_inode()), it may have been read in earlier from swap, in
1049 * ignorance of the mapping it belongs to. If that mapping has special
1050 * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
1051 * we may need to copy to a suitable page before moving to filecache.
1052 *
1053 * In a future release, this may well be extended to respect cpuset and
1054 * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
1055 * but for now it is a simple matter of zone.
1056 */
1057static bool shmem_should_replace_page(struct page *page, gfp_t gfp)
1058{
1059 return page_zonenum(page) > gfp_zone(gfp);
1060}
1061
1062static int shmem_replace_page(struct page **pagep, gfp_t gfp,
1063 struct shmem_inode_info *info, pgoff_t index)
1064{
1065 struct page *oldpage, *newpage;
1066 struct address_space *swap_mapping;
1067 pgoff_t swap_index;
1068 int error;
1069
1070 oldpage = *pagep;
1071 swap_index = page_private(oldpage);
1072 swap_mapping = page_mapping(oldpage);
1073
1074 /*
1075 * We have arrived here because our zones are constrained, so don't
1076 * limit chance of success by further cpuset and node constraints.
1077 */
1078 gfp &= ~GFP_CONSTRAINT_MASK;
1079 newpage = shmem_alloc_page(gfp, info, index);
1080 if (!newpage)
1081 return -ENOMEM;
1082
1083 get_page(newpage);
1084 copy_highpage(newpage, oldpage);
1085 flush_dcache_page(newpage);
1086
1087 __SetPageLocked(newpage);
1088 SetPageUptodate(newpage);
1089 SetPageSwapBacked(newpage);
1090 set_page_private(newpage, swap_index);
1091 SetPageSwapCache(newpage);
1092
1093 /*
1094 * Our caller will very soon move newpage out of swapcache, but it's
1095 * a nice clean interface for us to replace oldpage by newpage there.
1096 */
1097 spin_lock_irq(&swap_mapping->tree_lock);
1098 error = shmem_radix_tree_replace(swap_mapping, swap_index, oldpage,
1099 newpage);
1100 if (!error) {
1101 __inc_zone_page_state(newpage, NR_FILE_PAGES);
1102 __dec_zone_page_state(oldpage, NR_FILE_PAGES);
1103 }
1104 spin_unlock_irq(&swap_mapping->tree_lock);
1105
1106 if (unlikely(error)) {
1107 /*
1108 * Is this possible? I think not, now that our callers check
1109 * both PageSwapCache and page_private after getting page lock;
1110 * but be defensive. Reverse old to newpage for clear and free.
1111 */
1112 oldpage = newpage;
1113 } else {
1114 mem_cgroup_migrate(oldpage, newpage);
1115 lru_cache_add_anon(newpage);
1116 *pagep = newpage;
1117 }
1118
1119 ClearPageSwapCache(oldpage);
1120 set_page_private(oldpage, 0);
1121
1122 unlock_page(oldpage);
1123 put_page(oldpage);
1124 put_page(oldpage);
1125 return error;
1126}
1127
1128/*
1129 * shmem_getpage_gfp - find page in cache, or get from swap, or allocate
1130 *
1131 * If we allocate a new one we do not mark it dirty. That's up to the
1132 * vm. If we swap it in we mark it dirty since we also free the swap
1133 * entry since a page cannot live in both the swap and page cache
1134 */
1135static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
1136 struct page **pagep, enum sgp_type sgp, gfp_t gfp, int *fault_type)
1137{
1138 struct address_space *mapping = inode->i_mapping;
1139 struct shmem_inode_info *info;
1140 struct shmem_sb_info *sbinfo;
1141 struct mem_cgroup *memcg;
1142 struct page *page;
1143 swp_entry_t swap;
1144 int error;
1145 int once = 0;
1146 int alloced = 0;
1147
1148 if (index > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
1149 return -EFBIG;
1150repeat:
1151 swap.val = 0;
1152 page = find_lock_entry(mapping, index);
1153 if (radix_tree_exceptional_entry(page)) {
1154 swap = radix_to_swp_entry(page);
1155 page = NULL;
1156 }
1157
1158 if (sgp != SGP_WRITE && sgp != SGP_FALLOC &&
1159 ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1160 error = -EINVAL;
1161 goto unlock;
1162 }
1163
1164 if (page && sgp == SGP_WRITE)
1165 mark_page_accessed(page);
1166
1167 /* fallocated page? */
1168 if (page && !PageUptodate(page)) {
1169 if (sgp != SGP_READ)
1170 goto clear;
1171 unlock_page(page);
1172 put_page(page);
1173 page = NULL;
1174 }
1175 if (page || (sgp == SGP_READ && !swap.val)) {
1176 *pagep = page;
1177 return 0;
1178 }
1179
1180 /*
1181 * Fast cache lookup did not find it:
1182 * bring it back from swap or allocate.
1183 */
1184 info = SHMEM_I(inode);
1185 sbinfo = SHMEM_SB(inode->i_sb);
1186
1187 if (swap.val) {
1188 /* Look it up and read it in.. */
1189 page = lookup_swap_cache(swap);
1190 if (!page) {
1191 /* here we actually do the io */
1192 if (fault_type)
1193 *fault_type |= VM_FAULT_MAJOR;
1194 page = shmem_swapin(swap, gfp, info, index);
1195 if (!page) {
1196 error = -ENOMEM;
1197 goto failed;
1198 }
1199 }
1200
1201 /* We have to do this with page locked to prevent races */
1202 lock_page(page);
1203 if (!PageSwapCache(page) || page_private(page) != swap.val ||
1204 !shmem_confirm_swap(mapping, index, swap)) {
1205 error = -EEXIST; /* try again */
1206 goto unlock;
1207 }
1208 if (!PageUptodate(page)) {
1209 error = -EIO;
1210 goto failed;
1211 }
1212 wait_on_page_writeback(page);
1213
1214 if (shmem_should_replace_page(page, gfp)) {
1215 error = shmem_replace_page(&page, gfp, info, index);
1216 if (error)
1217 goto failed;
1218 }
1219
1220 error = mem_cgroup_try_charge(page, current->mm, gfp, &memcg,
1221 false);
1222 if (!error) {
1223 error = shmem_add_to_page_cache(page, mapping, index,
1224 swp_to_radix_entry(swap));
1225 /*
1226 * We already confirmed swap under page lock, and make
1227 * no memory allocation here, so usually no possibility
1228 * of error; but free_swap_and_cache() only trylocks a
1229 * page, so it is just possible that the entry has been
1230 * truncated or holepunched since swap was confirmed.
1231 * shmem_undo_range() will have done some of the
1232 * unaccounting, now delete_from_swap_cache() will do
1233 * the rest.
1234 * Reset swap.val? No, leave it so "failed" goes back to
1235 * "repeat": reading a hole and writing should succeed.
1236 */
1237 if (error) {
1238 mem_cgroup_cancel_charge(page, memcg, false);
1239 delete_from_swap_cache(page);
1240 }
1241 }
1242 if (error)
1243 goto failed;
1244
1245 mem_cgroup_commit_charge(page, memcg, true, false);
1246
1247 spin_lock(&info->lock);
1248 info->swapped--;
1249 shmem_recalc_inode(inode);
1250 spin_unlock(&info->lock);
1251
1252 if (sgp == SGP_WRITE)
1253 mark_page_accessed(page);
1254
1255 delete_from_swap_cache(page);
1256 set_page_dirty(page);
1257 swap_free(swap);
1258
1259 } else {
1260 if (shmem_acct_block(info->flags)) {
1261 error = -ENOSPC;
1262 goto failed;
1263 }
1264 if (sbinfo->max_blocks) {
1265 if (percpu_counter_compare(&sbinfo->used_blocks,
1266 sbinfo->max_blocks) >= 0) {
1267 error = -ENOSPC;
1268 goto unacct;
1269 }
1270 percpu_counter_inc(&sbinfo->used_blocks);
1271 }
1272
1273 page = shmem_alloc_page(gfp, info, index);
1274 if (!page) {
1275 error = -ENOMEM;
1276 goto decused;
1277 }
1278
1279 __SetPageSwapBacked(page);
1280 __SetPageLocked(page);
1281 if (sgp == SGP_WRITE)
1282 __SetPageReferenced(page);
1283
1284 error = mem_cgroup_try_charge(page, current->mm, gfp, &memcg,
1285 false);
1286 if (error)
1287 goto decused;
1288 error = radix_tree_maybe_preload(gfp & GFP_RECLAIM_MASK);
1289 if (!error) {
1290 error = shmem_add_to_page_cache(page, mapping, index,
1291 NULL);
1292 radix_tree_preload_end();
1293 }
1294 if (error) {
1295 mem_cgroup_cancel_charge(page, memcg, false);
1296 goto decused;
1297 }
1298 mem_cgroup_commit_charge(page, memcg, false, false);
1299 lru_cache_add_anon(page);
1300
1301 spin_lock(&info->lock);
1302 info->alloced++;
1303 inode->i_blocks += BLOCKS_PER_PAGE;
1304 shmem_recalc_inode(inode);
1305 spin_unlock(&info->lock);
1306 alloced = true;
1307
1308 /*
1309 * Let SGP_FALLOC use the SGP_WRITE optimization on a new page.
1310 */
1311 if (sgp == SGP_FALLOC)
1312 sgp = SGP_WRITE;
1313clear:
1314 /*
1315 * Let SGP_WRITE caller clear ends if write does not fill page;
1316 * but SGP_FALLOC on a page fallocated earlier must initialize
1317 * it now, lest undo on failure cancel our earlier guarantee.
1318 */
1319 if (sgp != SGP_WRITE) {
1320 clear_highpage(page);
1321 flush_dcache_page(page);
1322 SetPageUptodate(page);
1323 }
1324 if (sgp == SGP_DIRTY)
1325 set_page_dirty(page);
1326 }
1327
1328 /* Perhaps the file has been truncated since we checked */
1329 if (sgp != SGP_WRITE && sgp != SGP_FALLOC &&
1330 ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1331 if (alloced) {
1332 ClearPageDirty(page);
1333 delete_from_page_cache(page);
1334 spin_lock(&info->lock);
1335 shmem_recalc_inode(inode);
1336 spin_unlock(&info->lock);
1337 }
1338 error = -EINVAL;
1339 goto unlock;
1340 }
1341 *pagep = page;
1342 return 0;
1343
1344 /*
1345 * Error recovery.
1346 */
1347decused:
1348 if (sbinfo->max_blocks)
1349 percpu_counter_add(&sbinfo->used_blocks, -1);
1350unacct:
1351 shmem_unacct_blocks(info->flags, 1);
1352failed:
1353 if (swap.val && !shmem_confirm_swap(mapping, index, swap))
1354 error = -EEXIST;
1355unlock:
1356 if (page) {
1357 unlock_page(page);
1358 put_page(page);
1359 }
1360 if (error == -ENOSPC && !once++) {
1361 info = SHMEM_I(inode);
1362 spin_lock(&info->lock);
1363 shmem_recalc_inode(inode);
1364 spin_unlock(&info->lock);
1365 goto repeat;
1366 }
1367 if (error == -EEXIST) /* from above or from radix_tree_insert */
1368 goto repeat;
1369 return error;
1370}
1371
1372static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1373{
1374 struct inode *inode = file_inode(vma->vm_file);
1375 int error;
1376 int ret = VM_FAULT_LOCKED;
1377
1378 /*
1379 * Trinity finds that probing a hole which tmpfs is punching can
1380 * prevent the hole-punch from ever completing: which in turn
1381 * locks writers out with its hold on i_mutex. So refrain from
1382 * faulting pages into the hole while it's being punched. Although
1383 * shmem_undo_range() does remove the additions, it may be unable to
1384 * keep up, as each new page needs its own unmap_mapping_range() call,
1385 * and the i_mmap tree grows ever slower to scan if new vmas are added.
1386 *
1387 * It does not matter if we sometimes reach this check just before the
1388 * hole-punch begins, so that one fault then races with the punch:
1389 * we just need to make racing faults a rare case.
1390 *
1391 * The implementation below would be much simpler if we just used a
1392 * standard mutex or completion: but we cannot take i_mutex in fault,
1393 * and bloating every shmem inode for this unlikely case would be sad.
1394 */
1395 if (unlikely(inode->i_private)) {
1396 struct shmem_falloc *shmem_falloc;
1397
1398 spin_lock(&inode->i_lock);
1399 shmem_falloc = inode->i_private;
1400 if (shmem_falloc &&
1401 shmem_falloc->waitq &&
1402 vmf->pgoff >= shmem_falloc->start &&
1403 vmf->pgoff < shmem_falloc->next) {
1404 wait_queue_head_t *shmem_falloc_waitq;
1405 DEFINE_WAIT(shmem_fault_wait);
1406
1407 ret = VM_FAULT_NOPAGE;
1408 if ((vmf->flags & FAULT_FLAG_ALLOW_RETRY) &&
1409 !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
1410 /* It's polite to up mmap_sem if we can */
1411 up_read(&vma->vm_mm->mmap_sem);
1412 ret = VM_FAULT_RETRY;
1413 }
1414
1415 shmem_falloc_waitq = shmem_falloc->waitq;
1416 prepare_to_wait(shmem_falloc_waitq, &shmem_fault_wait,
1417 TASK_UNINTERRUPTIBLE);
1418 spin_unlock(&inode->i_lock);
1419 schedule();
1420
1421 /*
1422 * shmem_falloc_waitq points into the shmem_fallocate()
1423 * stack of the hole-punching task: shmem_falloc_waitq
1424 * is usually invalid by the time we reach here, but
1425 * finish_wait() does not dereference it in that case;
1426 * though i_lock needed lest racing with wake_up_all().
1427 */
1428 spin_lock(&inode->i_lock);
1429 finish_wait(shmem_falloc_waitq, &shmem_fault_wait);
1430 spin_unlock(&inode->i_lock);
1431 return ret;
1432 }
1433 spin_unlock(&inode->i_lock);
1434 }
1435
1436 error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_CACHE, &ret);
1437 if (error)
1438 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1439
1440 if (ret & VM_FAULT_MAJOR) {
1441 count_vm_event(PGMAJFAULT);
1442 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1443 }
1444 return ret;
1445}
1446
1447#ifdef CONFIG_NUMA
1448static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
1449{
1450 struct inode *inode = file_inode(vma->vm_file);
1451 return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
1452}
1453
1454static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
1455 unsigned long addr)
1456{
1457 struct inode *inode = file_inode(vma->vm_file);
1458 pgoff_t index;
1459
1460 index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1461 return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
1462}
1463#endif
1464
1465int shmem_lock(struct file *file, int lock, struct user_struct *user)
1466{
1467 struct inode *inode = file_inode(file);
1468 struct shmem_inode_info *info = SHMEM_I(inode);
1469 int retval = -ENOMEM;
1470
1471 spin_lock(&info->lock);
1472 if (lock && !(info->flags & VM_LOCKED)) {
1473 if (!user_shm_lock(inode->i_size, user))
1474 goto out_nomem;
1475 info->flags |= VM_LOCKED;
1476 mapping_set_unevictable(file->f_mapping);
1477 }
1478 if (!lock && (info->flags & VM_LOCKED) && user) {
1479 user_shm_unlock(inode->i_size, user);
1480 info->flags &= ~VM_LOCKED;
1481 mapping_clear_unevictable(file->f_mapping);
1482 }
1483 retval = 0;
1484
1485out_nomem:
1486 spin_unlock(&info->lock);
1487 return retval;
1488}
1489
1490static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1491{
1492 file_accessed(file);
1493 vma->vm_ops = &shmem_vm_ops;
1494 return 0;
1495}
1496
1497static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
1498 umode_t mode, dev_t dev, unsigned long flags)
1499{
1500 struct inode *inode;
1501 struct shmem_inode_info *info;
1502 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1503
1504 if (shmem_reserve_inode(sb))
1505 return NULL;
1506
1507 inode = new_inode(sb);
1508 if (inode) {
1509 inode->i_ino = get_next_ino();
1510 inode_init_owner(inode, dir, mode);
1511 inode->i_blocks = 0;
1512 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1513 inode->i_generation = get_seconds();
1514 info = SHMEM_I(inode);
1515 memset(info, 0, (char *)inode - (char *)info);
1516 spin_lock_init(&info->lock);
1517 info->seals = F_SEAL_SEAL;
1518 info->flags = flags & VM_NORESERVE;
1519 INIT_LIST_HEAD(&info->swaplist);
1520 simple_xattrs_init(&info->xattrs);
1521 cache_no_acl(inode);
1522
1523 switch (mode & S_IFMT) {
1524 default:
1525 inode->i_op = &shmem_special_inode_operations;
1526 init_special_inode(inode, mode, dev);
1527 break;
1528 case S_IFREG:
1529 inode->i_mapping->a_ops = &shmem_aops;
1530 inode->i_op = &shmem_inode_operations;
1531 inode->i_fop = &shmem_file_operations;
1532 mpol_shared_policy_init(&info->policy,
1533 shmem_get_sbmpol(sbinfo));
1534 break;
1535 case S_IFDIR:
1536 inc_nlink(inode);
1537 /* Some things misbehave if size == 0 on a directory */
1538 inode->i_size = 2 * BOGO_DIRENT_SIZE;
1539 inode->i_op = &shmem_dir_inode_operations;
1540 inode->i_fop = &simple_dir_operations;
1541 break;
1542 case S_IFLNK:
1543 /*
1544 * Must not load anything in the rbtree,
1545 * mpol_free_shared_policy will not be called.
1546 */
1547 mpol_shared_policy_init(&info->policy, NULL);
1548 break;
1549 }
1550 } else
1551 shmem_free_inode(sb);
1552 return inode;
1553}
1554
1555bool shmem_mapping(struct address_space *mapping)
1556{
1557 if (!mapping->host)
1558 return false;
1559
1560 return mapping->host->i_sb->s_op == &shmem_ops;
1561}
1562
1563#ifdef CONFIG_TMPFS
1564static const struct inode_operations shmem_symlink_inode_operations;
1565static const struct inode_operations shmem_short_symlink_operations;
1566
1567#ifdef CONFIG_TMPFS_XATTR
1568static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
1569#else
1570#define shmem_initxattrs NULL
1571#endif
1572
1573static int
1574shmem_write_begin(struct file *file, struct address_space *mapping,
1575 loff_t pos, unsigned len, unsigned flags,
1576 struct page **pagep, void **fsdata)
1577{
1578 struct inode *inode = mapping->host;
1579 struct shmem_inode_info *info = SHMEM_I(inode);
1580 pgoff_t index = pos >> PAGE_SHIFT;
1581
1582 /* i_mutex is held by caller */
1583 if (unlikely(info->seals)) {
1584 if (info->seals & F_SEAL_WRITE)
1585 return -EPERM;
1586 if ((info->seals & F_SEAL_GROW) && pos + len > inode->i_size)
1587 return -EPERM;
1588 }
1589
1590 return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1591}
1592
1593static int
1594shmem_write_end(struct file *file, struct address_space *mapping,
1595 loff_t pos, unsigned len, unsigned copied,
1596 struct page *page, void *fsdata)
1597{
1598 struct inode *inode = mapping->host;
1599
1600 if (pos + copied > inode->i_size)
1601 i_size_write(inode, pos + copied);
1602
1603 if (!PageUptodate(page)) {
1604 if (copied < PAGE_SIZE) {
1605 unsigned from = pos & (PAGE_SIZE - 1);
1606 zero_user_segments(page, 0, from,
1607 from + copied, PAGE_SIZE);
1608 }
1609 SetPageUptodate(page);
1610 }
1611 set_page_dirty(page);
1612 unlock_page(page);
1613 put_page(page);
1614
1615 return copied;
1616}
1617
1618static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
1619{
1620 struct file *file = iocb->ki_filp;
1621 struct inode *inode = file_inode(file);
1622 struct address_space *mapping = inode->i_mapping;
1623 pgoff_t index;
1624 unsigned long offset;
1625 enum sgp_type sgp = SGP_READ;
1626 int error = 0;
1627 ssize_t retval = 0;
1628 loff_t *ppos = &iocb->ki_pos;
1629
1630 /*
1631 * Might this read be for a stacking filesystem? Then when reading
1632 * holes of a sparse file, we actually need to allocate those pages,
1633 * and even mark them dirty, so it cannot exceed the max_blocks limit.
1634 */
1635 if (!iter_is_iovec(to))
1636 sgp = SGP_DIRTY;
1637
1638 index = *ppos >> PAGE_SHIFT;
1639 offset = *ppos & ~PAGE_MASK;
1640
1641 for (;;) {
1642 struct page *page = NULL;
1643 pgoff_t end_index;
1644 unsigned long nr, ret;
1645 loff_t i_size = i_size_read(inode);
1646
1647 end_index = i_size >> PAGE_SHIFT;
1648 if (index > end_index)
1649 break;
1650 if (index == end_index) {
1651 nr = i_size & ~PAGE_MASK;
1652 if (nr <= offset)
1653 break;
1654 }
1655
1656 error = shmem_getpage(inode, index, &page, sgp, NULL);
1657 if (error) {
1658 if (error == -EINVAL)
1659 error = 0;
1660 break;
1661 }
1662 if (page)
1663 unlock_page(page);
1664
1665 /*
1666 * We must evaluate after, since reads (unlike writes)
1667 * are called without i_mutex protection against truncate
1668 */
1669 nr = PAGE_SIZE;
1670 i_size = i_size_read(inode);
1671 end_index = i_size >> PAGE_SHIFT;
1672 if (index == end_index) {
1673 nr = i_size & ~PAGE_MASK;
1674 if (nr <= offset) {
1675 if (page)
1676 put_page(page);
1677 break;
1678 }
1679 }
1680 nr -= offset;
1681
1682 if (page) {
1683 /*
1684 * If users can be writing to this page using arbitrary
1685 * virtual addresses, take care about potential aliasing
1686 * before reading the page on the kernel side.
1687 */
1688 if (mapping_writably_mapped(mapping))
1689 flush_dcache_page(page);
1690 /*
1691 * Mark the page accessed if we read the beginning.
1692 */
1693 if (!offset)
1694 mark_page_accessed(page);
1695 } else {
1696 page = ZERO_PAGE(0);
1697 get_page(page);
1698 }
1699
1700 /*
1701 * Ok, we have the page, and it's up-to-date, so
1702 * now we can copy it to user space...
1703 */
1704 ret = copy_page_to_iter(page, offset, nr, to);
1705 retval += ret;
1706 offset += ret;
1707 index += offset >> PAGE_SHIFT;
1708 offset &= ~PAGE_MASK;
1709
1710 put_page(page);
1711 if (!iov_iter_count(to))
1712 break;
1713 if (ret < nr) {
1714 error = -EFAULT;
1715 break;
1716 }
1717 cond_resched();
1718 }
1719
1720 *ppos = ((loff_t) index << PAGE_SHIFT) + offset;
1721 file_accessed(file);
1722 return retval ? retval : error;
1723}
1724
1725static ssize_t shmem_file_splice_read(struct file *in, loff_t *ppos,
1726 struct pipe_inode_info *pipe, size_t len,
1727 unsigned int flags)
1728{
1729 struct address_space *mapping = in->f_mapping;
1730 struct inode *inode = mapping->host;
1731 unsigned int loff, nr_pages, req_pages;
1732 struct page *pages[PIPE_DEF_BUFFERS];
1733 struct partial_page partial[PIPE_DEF_BUFFERS];
1734 struct page *page;
1735 pgoff_t index, end_index;
1736 loff_t isize, left;
1737 int error, page_nr;
1738 struct splice_pipe_desc spd = {
1739 .pages = pages,
1740 .partial = partial,
1741 .nr_pages_max = PIPE_DEF_BUFFERS,
1742 .flags = flags,
1743 .ops = &page_cache_pipe_buf_ops,
1744 .spd_release = spd_release_page,
1745 };
1746
1747 isize = i_size_read(inode);
1748 if (unlikely(*ppos >= isize))
1749 return 0;
1750
1751 left = isize - *ppos;
1752 if (unlikely(left < len))
1753 len = left;
1754
1755 if (splice_grow_spd(pipe, &spd))
1756 return -ENOMEM;
1757
1758 index = *ppos >> PAGE_SHIFT;
1759 loff = *ppos & ~PAGE_MASK;
1760 req_pages = (len + loff + PAGE_SIZE - 1) >> PAGE_SHIFT;
1761 nr_pages = min(req_pages, spd.nr_pages_max);
1762
1763 spd.nr_pages = find_get_pages_contig(mapping, index,
1764 nr_pages, spd.pages);
1765 index += spd.nr_pages;
1766 error = 0;
1767
1768 while (spd.nr_pages < nr_pages) {
1769 error = shmem_getpage(inode, index, &page, SGP_CACHE, NULL);
1770 if (error)
1771 break;
1772 unlock_page(page);
1773 spd.pages[spd.nr_pages++] = page;
1774 index++;
1775 }
1776
1777 index = *ppos >> PAGE_SHIFT;
1778 nr_pages = spd.nr_pages;
1779 spd.nr_pages = 0;
1780
1781 for (page_nr = 0; page_nr < nr_pages; page_nr++) {
1782 unsigned int this_len;
1783
1784 if (!len)
1785 break;
1786
1787 this_len = min_t(unsigned long, len, PAGE_SIZE - loff);
1788 page = spd.pages[page_nr];
1789
1790 if (!PageUptodate(page) || page->mapping != mapping) {
1791 error = shmem_getpage(inode, index, &page,
1792 SGP_CACHE, NULL);
1793 if (error)
1794 break;
1795 unlock_page(page);
1796 put_page(spd.pages[page_nr]);
1797 spd.pages[page_nr] = page;
1798 }
1799
1800 isize = i_size_read(inode);
1801 end_index = (isize - 1) >> PAGE_SHIFT;
1802 if (unlikely(!isize || index > end_index))
1803 break;
1804
1805 if (end_index == index) {
1806 unsigned int plen;
1807
1808 plen = ((isize - 1) & ~PAGE_MASK) + 1;
1809 if (plen <= loff)
1810 break;
1811
1812 this_len = min(this_len, plen - loff);
1813 len = this_len;
1814 }
1815
1816 spd.partial[page_nr].offset = loff;
1817 spd.partial[page_nr].len = this_len;
1818 len -= this_len;
1819 loff = 0;
1820 spd.nr_pages++;
1821 index++;
1822 }
1823
1824 while (page_nr < nr_pages)
1825 put_page(spd.pages[page_nr++]);
1826
1827 if (spd.nr_pages)
1828 error = splice_to_pipe(pipe, &spd);
1829
1830 splice_shrink_spd(&spd);
1831
1832 if (error > 0) {
1833 *ppos += error;
1834 file_accessed(in);
1835 }
1836 return error;
1837}
1838
1839/*
1840 * llseek SEEK_DATA or SEEK_HOLE through the radix_tree.
1841 */
1842static pgoff_t shmem_seek_hole_data(struct address_space *mapping,
1843 pgoff_t index, pgoff_t end, int whence)
1844{
1845 struct page *page;
1846 struct pagevec pvec;
1847 pgoff_t indices[PAGEVEC_SIZE];
1848 bool done = false;
1849 int i;
1850
1851 pagevec_init(&pvec, 0);
1852 pvec.nr = 1; /* start small: we may be there already */
1853 while (!done) {
1854 pvec.nr = find_get_entries(mapping, index,
1855 pvec.nr, pvec.pages, indices);
1856 if (!pvec.nr) {
1857 if (whence == SEEK_DATA)
1858 index = end;
1859 break;
1860 }
1861 for (i = 0; i < pvec.nr; i++, index++) {
1862 if (index < indices[i]) {
1863 if (whence == SEEK_HOLE) {
1864 done = true;
1865 break;
1866 }
1867 index = indices[i];
1868 }
1869 page = pvec.pages[i];
1870 if (page && !radix_tree_exceptional_entry(page)) {
1871 if (!PageUptodate(page))
1872 page = NULL;
1873 }
1874 if (index >= end ||
1875 (page && whence == SEEK_DATA) ||
1876 (!page && whence == SEEK_HOLE)) {
1877 done = true;
1878 break;
1879 }
1880 }
1881 pagevec_remove_exceptionals(&pvec);
1882 pagevec_release(&pvec);
1883 pvec.nr = PAGEVEC_SIZE;
1884 cond_resched();
1885 }
1886 return index;
1887}
1888
1889static loff_t shmem_file_llseek(struct file *file, loff_t offset, int whence)
1890{
1891 struct address_space *mapping = file->f_mapping;
1892 struct inode *inode = mapping->host;
1893 pgoff_t start, end;
1894 loff_t new_offset;
1895
1896 if (whence != SEEK_DATA && whence != SEEK_HOLE)
1897 return generic_file_llseek_size(file, offset, whence,
1898 MAX_LFS_FILESIZE, i_size_read(inode));
1899 inode_lock(inode);
1900 /* We're holding i_mutex so we can access i_size directly */
1901
1902 if (offset < 0)
1903 offset = -EINVAL;
1904 else if (offset >= inode->i_size)
1905 offset = -ENXIO;
1906 else {
1907 start = offset >> PAGE_SHIFT;
1908 end = (inode->i_size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1909 new_offset = shmem_seek_hole_data(mapping, start, end, whence);
1910 new_offset <<= PAGE_SHIFT;
1911 if (new_offset > offset) {
1912 if (new_offset < inode->i_size)
1913 offset = new_offset;
1914 else if (whence == SEEK_DATA)
1915 offset = -ENXIO;
1916 else
1917 offset = inode->i_size;
1918 }
1919 }
1920
1921 if (offset >= 0)
1922 offset = vfs_setpos(file, offset, MAX_LFS_FILESIZE);
1923 inode_unlock(inode);
1924 return offset;
1925}
1926
1927/*
1928 * We need a tag: a new tag would expand every radix_tree_node by 8 bytes,
1929 * so reuse a tag which we firmly believe is never set or cleared on shmem.
1930 */
1931#define SHMEM_TAG_PINNED PAGECACHE_TAG_TOWRITE
1932#define LAST_SCAN 4 /* about 150ms max */
1933
1934static void shmem_tag_pins(struct address_space *mapping)
1935{
1936 struct radix_tree_iter iter;
1937 void **slot;
1938 pgoff_t start;
1939 struct page *page;
1940
1941 lru_add_drain();
1942 start = 0;
1943 rcu_read_lock();
1944
1945 radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1946 page = radix_tree_deref_slot(slot);
1947 if (!page || radix_tree_exception(page)) {
1948 if (radix_tree_deref_retry(page)) {
1949 slot = radix_tree_iter_retry(&iter);
1950 continue;
1951 }
1952 } else if (page_count(page) - page_mapcount(page) > 1) {
1953 spin_lock_irq(&mapping->tree_lock);
1954 radix_tree_tag_set(&mapping->page_tree, iter.index,
1955 SHMEM_TAG_PINNED);
1956 spin_unlock_irq(&mapping->tree_lock);
1957 }
1958
1959 if (need_resched()) {
1960 cond_resched_rcu();
1961 slot = radix_tree_iter_next(&iter);
1962 }
1963 }
1964 rcu_read_unlock();
1965}
1966
1967/*
1968 * Setting SEAL_WRITE requires us to verify there's no pending writer. However,
1969 * via get_user_pages(), drivers might have some pending I/O without any active
1970 * user-space mappings (eg., direct-IO, AIO). Therefore, we look at all pages
1971 * and see whether it has an elevated ref-count. If so, we tag them and wait for
1972 * them to be dropped.
1973 * The caller must guarantee that no new user will acquire writable references
1974 * to those pages to avoid races.
1975 */
1976static int shmem_wait_for_pins(struct address_space *mapping)
1977{
1978 struct radix_tree_iter iter;
1979 void **slot;
1980 pgoff_t start;
1981 struct page *page;
1982 int error, scan;
1983
1984 shmem_tag_pins(mapping);
1985
1986 error = 0;
1987 for (scan = 0; scan <= LAST_SCAN; scan++) {
1988 if (!radix_tree_tagged(&mapping->page_tree, SHMEM_TAG_PINNED))
1989 break;
1990
1991 if (!scan)
1992 lru_add_drain_all();
1993 else if (schedule_timeout_killable((HZ << scan) / 200))
1994 scan = LAST_SCAN;
1995
1996 start = 0;
1997 rcu_read_lock();
1998 radix_tree_for_each_tagged(slot, &mapping->page_tree, &iter,
1999 start, SHMEM_TAG_PINNED) {
2000
2001 page = radix_tree_deref_slot(slot);
2002 if (radix_tree_exception(page)) {
2003 if (radix_tree_deref_retry(page)) {
2004 slot = radix_tree_iter_retry(&iter);
2005 continue;
2006 }
2007
2008 page = NULL;
2009 }
2010
2011 if (page &&
2012 page_count(page) - page_mapcount(page) != 1) {
2013 if (scan < LAST_SCAN)
2014 goto continue_resched;
2015
2016 /*
2017 * On the last scan, we clean up all those tags
2018 * we inserted; but make a note that we still
2019 * found pages pinned.
2020 */
2021 error = -EBUSY;
2022 }
2023
2024 spin_lock_irq(&mapping->tree_lock);
2025 radix_tree_tag_clear(&mapping->page_tree,
2026 iter.index, SHMEM_TAG_PINNED);
2027 spin_unlock_irq(&mapping->tree_lock);
2028continue_resched:
2029 if (need_resched()) {
2030 cond_resched_rcu();
2031 slot = radix_tree_iter_next(&iter);
2032 }
2033 }
2034 rcu_read_unlock();
2035 }
2036
2037 return error;
2038}
2039
2040#define F_ALL_SEALS (F_SEAL_SEAL | \
2041 F_SEAL_SHRINK | \
2042 F_SEAL_GROW | \
2043 F_SEAL_WRITE)
2044
2045int shmem_add_seals(struct file *file, unsigned int seals)
2046{
2047 struct inode *inode = file_inode(file);
2048 struct shmem_inode_info *info = SHMEM_I(inode);
2049 int error;
2050
2051 /*
2052 * SEALING
2053 * Sealing allows multiple parties to share a shmem-file but restrict
2054 * access to a specific subset of file operations. Seals can only be
2055 * added, but never removed. This way, mutually untrusted parties can
2056 * share common memory regions with a well-defined policy. A malicious
2057 * peer can thus never perform unwanted operations on a shared object.
2058 *
2059 * Seals are only supported on special shmem-files and always affect
2060 * the whole underlying inode. Once a seal is set, it may prevent some
2061 * kinds of access to the file. Currently, the following seals are
2062 * defined:
2063 * SEAL_SEAL: Prevent further seals from being set on this file
2064 * SEAL_SHRINK: Prevent the file from shrinking
2065 * SEAL_GROW: Prevent the file from growing
2066 * SEAL_WRITE: Prevent write access to the file
2067 *
2068 * As we don't require any trust relationship between two parties, we
2069 * must prevent seals from being removed. Therefore, sealing a file
2070 * only adds a given set of seals to the file, it never touches
2071 * existing seals. Furthermore, the "setting seals"-operation can be
2072 * sealed itself, which basically prevents any further seal from being
2073 * added.
2074 *
2075 * Semantics of sealing are only defined on volatile files. Only
2076 * anonymous shmem files support sealing. More importantly, seals are
2077 * never written to disk. Therefore, there's no plan to support it on
2078 * other file types.
2079 */
2080
2081 if (file->f_op != &shmem_file_operations)
2082 return -EINVAL;
2083 if (!(file->f_mode & FMODE_WRITE))
2084 return -EPERM;
2085 if (seals & ~(unsigned int)F_ALL_SEALS)
2086 return -EINVAL;
2087
2088 inode_lock(inode);
2089
2090 if (info->seals & F_SEAL_SEAL) {
2091 error = -EPERM;
2092 goto unlock;
2093 }
2094
2095 if ((seals & F_SEAL_WRITE) && !(info->seals & F_SEAL_WRITE)) {
2096 error = mapping_deny_writable(file->f_mapping);
2097 if (error)
2098 goto unlock;
2099
2100 error = shmem_wait_for_pins(file->f_mapping);
2101 if (error) {
2102 mapping_allow_writable(file->f_mapping);
2103 goto unlock;
2104 }
2105 }
2106
2107 info->seals |= seals;
2108 error = 0;
2109
2110unlock:
2111 inode_unlock(inode);
2112 return error;
2113}
2114EXPORT_SYMBOL_GPL(shmem_add_seals);
2115
2116int shmem_get_seals(struct file *file)
2117{
2118 if (file->f_op != &shmem_file_operations)
2119 return -EINVAL;
2120
2121 return SHMEM_I(file_inode(file))->seals;
2122}
2123EXPORT_SYMBOL_GPL(shmem_get_seals);
2124
2125long shmem_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
2126{
2127 long error;
2128
2129 switch (cmd) {
2130 case F_ADD_SEALS:
2131 /* disallow upper 32bit */
2132 if (arg > UINT_MAX)
2133 return -EINVAL;
2134
2135 error = shmem_add_seals(file, arg);
2136 break;
2137 case F_GET_SEALS:
2138 error = shmem_get_seals(file);
2139 break;
2140 default:
2141 error = -EINVAL;
2142 break;
2143 }
2144
2145 return error;
2146}
2147
2148static long shmem_fallocate(struct file *file, int mode, loff_t offset,
2149 loff_t len)
2150{
2151 struct inode *inode = file_inode(file);
2152 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
2153 struct shmem_inode_info *info = SHMEM_I(inode);
2154 struct shmem_falloc shmem_falloc;
2155 pgoff_t start, index, end;
2156 int error;
2157
2158 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2159 return -EOPNOTSUPP;
2160
2161 inode_lock(inode);
2162
2163 if (mode & FALLOC_FL_PUNCH_HOLE) {
2164 struct address_space *mapping = file->f_mapping;
2165 loff_t unmap_start = round_up(offset, PAGE_SIZE);
2166 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
2167 DECLARE_WAIT_QUEUE_HEAD_ONSTACK(shmem_falloc_waitq);
2168
2169 /* protected by i_mutex */
2170 if (info->seals & F_SEAL_WRITE) {
2171 error = -EPERM;
2172 goto out;
2173 }
2174
2175 shmem_falloc.waitq = &shmem_falloc_waitq;
2176 shmem_falloc.start = unmap_start >> PAGE_SHIFT;
2177 shmem_falloc.next = (unmap_end + 1) >> PAGE_SHIFT;
2178 spin_lock(&inode->i_lock);
2179 inode->i_private = &shmem_falloc;
2180 spin_unlock(&inode->i_lock);
2181
2182 if ((u64)unmap_end > (u64)unmap_start)
2183 unmap_mapping_range(mapping, unmap_start,
2184 1 + unmap_end - unmap_start, 0);
2185 shmem_truncate_range(inode, offset, offset + len - 1);
2186 /* No need to unmap again: hole-punching leaves COWed pages */
2187
2188 spin_lock(&inode->i_lock);
2189 inode->i_private = NULL;
2190 wake_up_all(&shmem_falloc_waitq);
2191 spin_unlock(&inode->i_lock);
2192 error = 0;
2193 goto out;
2194 }
2195
2196 /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
2197 error = inode_newsize_ok(inode, offset + len);
2198 if (error)
2199 goto out;
2200
2201 if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
2202 error = -EPERM;
2203 goto out;
2204 }
2205
2206 start = offset >> PAGE_SHIFT;
2207 end = (offset + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
2208 /* Try to avoid a swapstorm if len is impossible to satisfy */
2209 if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
2210 error = -ENOSPC;
2211 goto out;
2212 }
2213
2214 shmem_falloc.waitq = NULL;
2215 shmem_falloc.start = start;
2216 shmem_falloc.next = start;
2217 shmem_falloc.nr_falloced = 0;
2218 shmem_falloc.nr_unswapped = 0;
2219 spin_lock(&inode->i_lock);
2220 inode->i_private = &shmem_falloc;
2221 spin_unlock(&inode->i_lock);
2222
2223 for (index = start; index < end; index++) {
2224 struct page *page;
2225
2226 /*
2227 * Good, the fallocate(2) manpage permits EINTR: we may have
2228 * been interrupted because we are using up too much memory.
2229 */
2230 if (signal_pending(current))
2231 error = -EINTR;
2232 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
2233 error = -ENOMEM;
2234 else
2235 error = shmem_getpage(inode, index, &page, SGP_FALLOC,
2236 NULL);
2237 if (error) {
2238 /* Remove the !PageUptodate pages we added */
2239 shmem_undo_range(inode,
2240 (loff_t)start << PAGE_SHIFT,
2241 (loff_t)index << PAGE_SHIFT, true);
2242 goto undone;
2243 }
2244
2245 /*
2246 * Inform shmem_writepage() how far we have reached.
2247 * No need for lock or barrier: we have the page lock.
2248 */
2249 shmem_falloc.next++;
2250 if (!PageUptodate(page))
2251 shmem_falloc.nr_falloced++;
2252
2253 /*
2254 * If !PageUptodate, leave it that way so that freeable pages
2255 * can be recognized if we need to rollback on error later.
2256 * But set_page_dirty so that memory pressure will swap rather
2257 * than free the pages we are allocating (and SGP_CACHE pages
2258 * might still be clean: we now need to mark those dirty too).
2259 */
2260 set_page_dirty(page);
2261 unlock_page(page);
2262 put_page(page);
2263 cond_resched();
2264 }
2265
2266 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
2267 i_size_write(inode, offset + len);
2268 inode->i_ctime = CURRENT_TIME;
2269undone:
2270 spin_lock(&inode->i_lock);
2271 inode->i_private = NULL;
2272 spin_unlock(&inode->i_lock);
2273out:
2274 inode_unlock(inode);
2275 return error;
2276}
2277
2278static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
2279{
2280 struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
2281
2282 buf->f_type = TMPFS_MAGIC;
2283 buf->f_bsize = PAGE_SIZE;
2284 buf->f_namelen = NAME_MAX;
2285 if (sbinfo->max_blocks) {
2286 buf->f_blocks = sbinfo->max_blocks;
2287 buf->f_bavail =
2288 buf->f_bfree = sbinfo->max_blocks -
2289 percpu_counter_sum(&sbinfo->used_blocks);
2290 }
2291 if (sbinfo->max_inodes) {
2292 buf->f_files = sbinfo->max_inodes;
2293 buf->f_ffree = sbinfo->free_inodes;
2294 }
2295 /* else leave those fields 0 like simple_statfs */
2296 return 0;
2297}
2298
2299/*
2300 * File creation. Allocate an inode, and we're done..
2301 */
2302static int
2303shmem_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
2304{
2305 struct inode *inode;
2306 int error = -ENOSPC;
2307
2308 inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
2309 if (inode) {
2310 error = simple_acl_create(dir, inode);
2311 if (error)
2312 goto out_iput;
2313 error = security_inode_init_security(inode, dir,
2314 &dentry->d_name,
2315 shmem_initxattrs, NULL);
2316 if (error && error != -EOPNOTSUPP)
2317 goto out_iput;
2318
2319 error = 0;
2320 dir->i_size += BOGO_DIRENT_SIZE;
2321 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2322 d_instantiate(dentry, inode);
2323 dget(dentry); /* Extra count - pin the dentry in core */
2324 }
2325 return error;
2326out_iput:
2327 iput(inode);
2328 return error;
2329}
2330
2331static int
2332shmem_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
2333{
2334 struct inode *inode;
2335 int error = -ENOSPC;
2336
2337 inode = shmem_get_inode(dir->i_sb, dir, mode, 0, VM_NORESERVE);
2338 if (inode) {
2339 error = security_inode_init_security(inode, dir,
2340 NULL,
2341 shmem_initxattrs, NULL);
2342 if (error && error != -EOPNOTSUPP)
2343 goto out_iput;
2344 error = simple_acl_create(dir, inode);
2345 if (error)
2346 goto out_iput;
2347 d_tmpfile(dentry, inode);
2348 }
2349 return error;
2350out_iput:
2351 iput(inode);
2352 return error;
2353}
2354
2355static int shmem_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
2356{
2357 int error;
2358
2359 if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
2360 return error;
2361 inc_nlink(dir);
2362 return 0;
2363}
2364
2365static int shmem_create(struct inode *dir, struct dentry *dentry, umode_t mode,
2366 bool excl)
2367{
2368 return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
2369}
2370
2371/*
2372 * Link a file..
2373 */
2374static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
2375{
2376 struct inode *inode = d_inode(old_dentry);
2377 int ret;
2378
2379 /*
2380 * No ordinary (disk based) filesystem counts links as inodes;
2381 * but each new link needs a new dentry, pinning lowmem, and
2382 * tmpfs dentries cannot be pruned until they are unlinked.
2383 */
2384 ret = shmem_reserve_inode(inode->i_sb);
2385 if (ret)
2386 goto out;
2387
2388 dir->i_size += BOGO_DIRENT_SIZE;
2389 inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2390 inc_nlink(inode);
2391 ihold(inode); /* New dentry reference */
2392 dget(dentry); /* Extra pinning count for the created dentry */
2393 d_instantiate(dentry, inode);
2394out:
2395 return ret;
2396}
2397
2398static int shmem_unlink(struct inode *dir, struct dentry *dentry)
2399{
2400 struct inode *inode = d_inode(dentry);
2401
2402 if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
2403 shmem_free_inode(inode->i_sb);
2404
2405 dir->i_size -= BOGO_DIRENT_SIZE;
2406 inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2407 drop_nlink(inode);
2408 dput(dentry); /* Undo the count from "create" - this does all the work */
2409 return 0;
2410}
2411
2412static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
2413{
2414 if (!simple_empty(dentry))
2415 return -ENOTEMPTY;
2416
2417 drop_nlink(d_inode(dentry));
2418 drop_nlink(dir);
2419 return shmem_unlink(dir, dentry);
2420}
2421
2422static int shmem_exchange(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
2423{
2424 bool old_is_dir = d_is_dir(old_dentry);
2425 bool new_is_dir = d_is_dir(new_dentry);
2426
2427 if (old_dir != new_dir && old_is_dir != new_is_dir) {
2428 if (old_is_dir) {
2429 drop_nlink(old_dir);
2430 inc_nlink(new_dir);
2431 } else {
2432 drop_nlink(new_dir);
2433 inc_nlink(old_dir);
2434 }
2435 }
2436 old_dir->i_ctime = old_dir->i_mtime =
2437 new_dir->i_ctime = new_dir->i_mtime =
2438 d_inode(old_dentry)->i_ctime =
2439 d_inode(new_dentry)->i_ctime = CURRENT_TIME;
2440
2441 return 0;
2442}
2443
2444static int shmem_whiteout(struct inode *old_dir, struct dentry *old_dentry)
2445{
2446 struct dentry *whiteout;
2447 int error;
2448
2449 whiteout = d_alloc(old_dentry->d_parent, &old_dentry->d_name);
2450 if (!whiteout)
2451 return -ENOMEM;
2452
2453 error = shmem_mknod(old_dir, whiteout,
2454 S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
2455 dput(whiteout);
2456 if (error)
2457 return error;
2458
2459 /*
2460 * Cheat and hash the whiteout while the old dentry is still in
2461 * place, instead of playing games with FS_RENAME_DOES_D_MOVE.
2462 *
2463 * d_lookup() will consistently find one of them at this point,
2464 * not sure which one, but that isn't even important.
2465 */
2466 d_rehash(whiteout);
2467 return 0;
2468}
2469
2470/*
2471 * The VFS layer already does all the dentry stuff for rename,
2472 * we just have to decrement the usage count for the target if
2473 * it exists so that the VFS layer correctly free's it when it
2474 * gets overwritten.
2475 */
2476static int shmem_rename2(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry, unsigned int flags)
2477{
2478 struct inode *inode = d_inode(old_dentry);
2479 int they_are_dirs = S_ISDIR(inode->i_mode);
2480
2481 if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
2482 return -EINVAL;
2483
2484 if (flags & RENAME_EXCHANGE)
2485 return shmem_exchange(old_dir, old_dentry, new_dir, new_dentry);
2486
2487 if (!simple_empty(new_dentry))
2488 return -ENOTEMPTY;
2489
2490 if (flags & RENAME_WHITEOUT) {
2491 int error;
2492
2493 error = shmem_whiteout(old_dir, old_dentry);
2494 if (error)
2495 return error;
2496 }
2497
2498 if (d_really_is_positive(new_dentry)) {
2499 (void) shmem_unlink(new_dir, new_dentry);
2500 if (they_are_dirs) {
2501 drop_nlink(d_inode(new_dentry));
2502 drop_nlink(old_dir);
2503 }
2504 } else if (they_are_dirs) {
2505 drop_nlink(old_dir);
2506 inc_nlink(new_dir);
2507 }
2508
2509 old_dir->i_size -= BOGO_DIRENT_SIZE;
2510 new_dir->i_size += BOGO_DIRENT_SIZE;
2511 old_dir->i_ctime = old_dir->i_mtime =
2512 new_dir->i_ctime = new_dir->i_mtime =
2513 inode->i_ctime = CURRENT_TIME;
2514 return 0;
2515}
2516
2517static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
2518{
2519 int error;
2520 int len;
2521 struct inode *inode;
2522 struct page *page;
2523 struct shmem_inode_info *info;
2524
2525 len = strlen(symname) + 1;
2526 if (len > PAGE_SIZE)
2527 return -ENAMETOOLONG;
2528
2529 inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0, VM_NORESERVE);
2530 if (!inode)
2531 return -ENOSPC;
2532
2533 error = security_inode_init_security(inode, dir, &dentry->d_name,
2534 shmem_initxattrs, NULL);
2535 if (error) {
2536 if (error != -EOPNOTSUPP) {
2537 iput(inode);
2538 return error;
2539 }
2540 error = 0;
2541 }
2542
2543 info = SHMEM_I(inode);
2544 inode->i_size = len-1;
2545 if (len <= SHORT_SYMLINK_LEN) {
2546 inode->i_link = kmemdup(symname, len, GFP_KERNEL);
2547 if (!inode->i_link) {
2548 iput(inode);
2549 return -ENOMEM;
2550 }
2551 inode->i_op = &shmem_short_symlink_operations;
2552 } else {
2553 inode_nohighmem(inode);
2554 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
2555 if (error) {
2556 iput(inode);
2557 return error;
2558 }
2559 inode->i_mapping->a_ops = &shmem_aops;
2560 inode->i_op = &shmem_symlink_inode_operations;
2561 memcpy(page_address(page), symname, len);
2562 SetPageUptodate(page);
2563 set_page_dirty(page);
2564 unlock_page(page);
2565 put_page(page);
2566 }
2567 dir->i_size += BOGO_DIRENT_SIZE;
2568 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2569 d_instantiate(dentry, inode);
2570 dget(dentry);
2571 return 0;
2572}
2573
2574static void shmem_put_link(void *arg)
2575{
2576 mark_page_accessed(arg);
2577 put_page(arg);
2578}
2579
2580static const char *shmem_get_link(struct dentry *dentry,
2581 struct inode *inode,
2582 struct delayed_call *done)
2583{
2584 struct page *page = NULL;
2585 int error;
2586 if (!dentry) {
2587 page = find_get_page(inode->i_mapping, 0);
2588 if (!page)
2589 return ERR_PTR(-ECHILD);
2590 if (!PageUptodate(page)) {
2591 put_page(page);
2592 return ERR_PTR(-ECHILD);
2593 }
2594 } else {
2595 error = shmem_getpage(inode, 0, &page, SGP_READ, NULL);
2596 if (error)
2597 return ERR_PTR(error);
2598 unlock_page(page);
2599 }
2600 set_delayed_call(done, shmem_put_link, page);
2601 return page_address(page);
2602}
2603
2604#ifdef CONFIG_TMPFS_XATTR
2605/*
2606 * Superblocks without xattr inode operations may get some security.* xattr
2607 * support from the LSM "for free". As soon as we have any other xattrs
2608 * like ACLs, we also need to implement the security.* handlers at
2609 * filesystem level, though.
2610 */
2611
2612/*
2613 * Callback for security_inode_init_security() for acquiring xattrs.
2614 */
2615static int shmem_initxattrs(struct inode *inode,
2616 const struct xattr *xattr_array,
2617 void *fs_info)
2618{
2619 struct shmem_inode_info *info = SHMEM_I(inode);
2620 const struct xattr *xattr;
2621 struct simple_xattr *new_xattr;
2622 size_t len;
2623
2624 for (xattr = xattr_array; xattr->name != NULL; xattr++) {
2625 new_xattr = simple_xattr_alloc(xattr->value, xattr->value_len);
2626 if (!new_xattr)
2627 return -ENOMEM;
2628
2629 len = strlen(xattr->name) + 1;
2630 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
2631 GFP_KERNEL);
2632 if (!new_xattr->name) {
2633 kfree(new_xattr);
2634 return -ENOMEM;
2635 }
2636
2637 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
2638 XATTR_SECURITY_PREFIX_LEN);
2639 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
2640 xattr->name, len);
2641
2642 simple_xattr_list_add(&info->xattrs, new_xattr);
2643 }
2644
2645 return 0;
2646}
2647
2648static int shmem_xattr_handler_get(const struct xattr_handler *handler,
2649 struct dentry *dentry, const char *name,
2650 void *buffer, size_t size)
2651{
2652 struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
2653
2654 name = xattr_full_name(handler, name);
2655 return simple_xattr_get(&info->xattrs, name, buffer, size);
2656}
2657
2658static int shmem_xattr_handler_set(const struct xattr_handler *handler,
2659 struct dentry *dentry, const char *name,
2660 const void *value, size_t size, int flags)
2661{
2662 struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
2663
2664 name = xattr_full_name(handler, name);
2665 return simple_xattr_set(&info->xattrs, name, value, size, flags);
2666}
2667
2668static const struct xattr_handler shmem_security_xattr_handler = {
2669 .prefix = XATTR_SECURITY_PREFIX,
2670 .get = shmem_xattr_handler_get,
2671 .set = shmem_xattr_handler_set,
2672};
2673
2674static const struct xattr_handler shmem_trusted_xattr_handler = {
2675 .prefix = XATTR_TRUSTED_PREFIX,
2676 .get = shmem_xattr_handler_get,
2677 .set = shmem_xattr_handler_set,
2678};
2679
2680static const struct xattr_handler *shmem_xattr_handlers[] = {
2681#ifdef CONFIG_TMPFS_POSIX_ACL
2682 &posix_acl_access_xattr_handler,
2683 &posix_acl_default_xattr_handler,
2684#endif
2685 &shmem_security_xattr_handler,
2686 &shmem_trusted_xattr_handler,
2687 NULL
2688};
2689
2690static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
2691{
2692 struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
2693 return simple_xattr_list(d_inode(dentry), &info->xattrs, buffer, size);
2694}
2695#endif /* CONFIG_TMPFS_XATTR */
2696
2697static const struct inode_operations shmem_short_symlink_operations = {
2698 .readlink = generic_readlink,
2699 .get_link = simple_get_link,
2700#ifdef CONFIG_TMPFS_XATTR
2701 .setxattr = generic_setxattr,
2702 .getxattr = generic_getxattr,
2703 .listxattr = shmem_listxattr,
2704 .removexattr = generic_removexattr,
2705#endif
2706};
2707
2708static const struct inode_operations shmem_symlink_inode_operations = {
2709 .readlink = generic_readlink,
2710 .get_link = shmem_get_link,
2711#ifdef CONFIG_TMPFS_XATTR
2712 .setxattr = generic_setxattr,
2713 .getxattr = generic_getxattr,
2714 .listxattr = shmem_listxattr,
2715 .removexattr = generic_removexattr,
2716#endif
2717};
2718
2719static struct dentry *shmem_get_parent(struct dentry *child)
2720{
2721 return ERR_PTR(-ESTALE);
2722}
2723
2724static int shmem_match(struct inode *ino, void *vfh)
2725{
2726 __u32 *fh = vfh;
2727 __u64 inum = fh[2];
2728 inum = (inum << 32) | fh[1];
2729 return ino->i_ino == inum && fh[0] == ino->i_generation;
2730}
2731
2732static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
2733 struct fid *fid, int fh_len, int fh_type)
2734{
2735 struct inode *inode;
2736 struct dentry *dentry = NULL;
2737 u64 inum;
2738
2739 if (fh_len < 3)
2740 return NULL;
2741
2742 inum = fid->raw[2];
2743 inum = (inum << 32) | fid->raw[1];
2744
2745 inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
2746 shmem_match, fid->raw);
2747 if (inode) {
2748 dentry = d_find_alias(inode);
2749 iput(inode);
2750 }
2751
2752 return dentry;
2753}
2754
2755static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
2756 struct inode *parent)
2757{
2758 if (*len < 3) {
2759 *len = 3;
2760 return FILEID_INVALID;
2761 }
2762
2763 if (inode_unhashed(inode)) {
2764 /* Unfortunately insert_inode_hash is not idempotent,
2765 * so as we hash inodes here rather than at creation
2766 * time, we need a lock to ensure we only try
2767 * to do it once
2768 */
2769 static DEFINE_SPINLOCK(lock);
2770 spin_lock(&lock);
2771 if (inode_unhashed(inode))
2772 __insert_inode_hash(inode,
2773 inode->i_ino + inode->i_generation);
2774 spin_unlock(&lock);
2775 }
2776
2777 fh[0] = inode->i_generation;
2778 fh[1] = inode->i_ino;
2779 fh[2] = ((__u64)inode->i_ino) >> 32;
2780
2781 *len = 3;
2782 return 1;
2783}
2784
2785static const struct export_operations shmem_export_ops = {
2786 .get_parent = shmem_get_parent,
2787 .encode_fh = shmem_encode_fh,
2788 .fh_to_dentry = shmem_fh_to_dentry,
2789};
2790
2791static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
2792 bool remount)
2793{
2794 char *this_char, *value, *rest;
2795 struct mempolicy *mpol = NULL;
2796 uid_t uid;
2797 gid_t gid;
2798
2799 while (options != NULL) {
2800 this_char = options;
2801 for (;;) {
2802 /*
2803 * NUL-terminate this option: unfortunately,
2804 * mount options form a comma-separated list,
2805 * but mpol's nodelist may also contain commas.
2806 */
2807 options = strchr(options, ',');
2808 if (options == NULL)
2809 break;
2810 options++;
2811 if (!isdigit(*options)) {
2812 options[-1] = '\0';
2813 break;
2814 }
2815 }
2816 if (!*this_char)
2817 continue;
2818 if ((value = strchr(this_char,'=')) != NULL) {
2819 *value++ = 0;
2820 } else {
2821 pr_err("tmpfs: No value for mount option '%s'\n",
2822 this_char);
2823 goto error;
2824 }
2825
2826 if (!strcmp(this_char,"size")) {
2827 unsigned long long size;
2828 size = memparse(value,&rest);
2829 if (*rest == '%') {
2830 size <<= PAGE_SHIFT;
2831 size *= totalram_pages;
2832 do_div(size, 100);
2833 rest++;
2834 }
2835 if (*rest)
2836 goto bad_val;
2837 sbinfo->max_blocks =
2838 DIV_ROUND_UP(size, PAGE_SIZE);
2839 } else if (!strcmp(this_char,"nr_blocks")) {
2840 sbinfo->max_blocks = memparse(value, &rest);
2841 if (*rest)
2842 goto bad_val;
2843 } else if (!strcmp(this_char,"nr_inodes")) {
2844 sbinfo->max_inodes = memparse(value, &rest);
2845 if (*rest)
2846 goto bad_val;
2847 } else if (!strcmp(this_char,"mode")) {
2848 if (remount)
2849 continue;
2850 sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
2851 if (*rest)
2852 goto bad_val;
2853 } else if (!strcmp(this_char,"uid")) {
2854 if (remount)
2855 continue;
2856 uid = simple_strtoul(value, &rest, 0);
2857 if (*rest)
2858 goto bad_val;
2859 sbinfo->uid = make_kuid(current_user_ns(), uid);
2860 if (!uid_valid(sbinfo->uid))
2861 goto bad_val;
2862 } else if (!strcmp(this_char,"gid")) {
2863 if (remount)
2864 continue;
2865 gid = simple_strtoul(value, &rest, 0);
2866 if (*rest)
2867 goto bad_val;
2868 sbinfo->gid = make_kgid(current_user_ns(), gid);
2869 if (!gid_valid(sbinfo->gid))
2870 goto bad_val;
2871 } else if (!strcmp(this_char,"mpol")) {
2872 mpol_put(mpol);
2873 mpol = NULL;
2874 if (mpol_parse_str(value, &mpol))
2875 goto bad_val;
2876 } else {
2877 pr_err("tmpfs: Bad mount option %s\n", this_char);
2878 goto error;
2879 }
2880 }
2881 sbinfo->mpol = mpol;
2882 return 0;
2883
2884bad_val:
2885 pr_err("tmpfs: Bad value '%s' for mount option '%s'\n",
2886 value, this_char);
2887error:
2888 mpol_put(mpol);
2889 return 1;
2890
2891}
2892
2893static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2894{
2895 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2896 struct shmem_sb_info config = *sbinfo;
2897 unsigned long inodes;
2898 int error = -EINVAL;
2899
2900 config.mpol = NULL;
2901 if (shmem_parse_options(data, &config, true))
2902 return error;
2903
2904 spin_lock(&sbinfo->stat_lock);
2905 inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2906 if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
2907 goto out;
2908 if (config.max_inodes < inodes)
2909 goto out;
2910 /*
2911 * Those tests disallow limited->unlimited while any are in use;
2912 * but we must separately disallow unlimited->limited, because
2913 * in that case we have no record of how much is already in use.
2914 */
2915 if (config.max_blocks && !sbinfo->max_blocks)
2916 goto out;
2917 if (config.max_inodes && !sbinfo->max_inodes)
2918 goto out;
2919
2920 error = 0;
2921 sbinfo->max_blocks = config.max_blocks;
2922 sbinfo->max_inodes = config.max_inodes;
2923 sbinfo->free_inodes = config.max_inodes - inodes;
2924
2925 /*
2926 * Preserve previous mempolicy unless mpol remount option was specified.
2927 */
2928 if (config.mpol) {
2929 mpol_put(sbinfo->mpol);
2930 sbinfo->mpol = config.mpol; /* transfers initial ref */
2931 }
2932out:
2933 spin_unlock(&sbinfo->stat_lock);
2934 return error;
2935}
2936
2937static int shmem_show_options(struct seq_file *seq, struct dentry *root)
2938{
2939 struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
2940
2941 if (sbinfo->max_blocks != shmem_default_max_blocks())
2942 seq_printf(seq, ",size=%luk",
2943 sbinfo->max_blocks << (PAGE_SHIFT - 10));
2944 if (sbinfo->max_inodes != shmem_default_max_inodes())
2945 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
2946 if (sbinfo->mode != (S_IRWXUGO | S_ISVTX))
2947 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
2948 if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
2949 seq_printf(seq, ",uid=%u",
2950 from_kuid_munged(&init_user_ns, sbinfo->uid));
2951 if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
2952 seq_printf(seq, ",gid=%u",
2953 from_kgid_munged(&init_user_ns, sbinfo->gid));
2954 shmem_show_mpol(seq, sbinfo->mpol);
2955 return 0;
2956}
2957
2958#define MFD_NAME_PREFIX "memfd:"
2959#define MFD_NAME_PREFIX_LEN (sizeof(MFD_NAME_PREFIX) - 1)
2960#define MFD_NAME_MAX_LEN (NAME_MAX - MFD_NAME_PREFIX_LEN)
2961
2962#define MFD_ALL_FLAGS (MFD_CLOEXEC | MFD_ALLOW_SEALING)
2963
2964SYSCALL_DEFINE2(memfd_create,
2965 const char __user *, uname,
2966 unsigned int, flags)
2967{
2968 struct shmem_inode_info *info;
2969 struct file *file;
2970 int fd, error;
2971 char *name;
2972 long len;
2973
2974 if (flags & ~(unsigned int)MFD_ALL_FLAGS)
2975 return -EINVAL;
2976
2977 /* length includes terminating zero */
2978 len = strnlen_user(uname, MFD_NAME_MAX_LEN + 1);
2979 if (len <= 0)
2980 return -EFAULT;
2981 if (len > MFD_NAME_MAX_LEN + 1)
2982 return -EINVAL;
2983
2984 name = kmalloc(len + MFD_NAME_PREFIX_LEN, GFP_TEMPORARY);
2985 if (!name)
2986 return -ENOMEM;
2987
2988 strcpy(name, MFD_NAME_PREFIX);
2989 if (copy_from_user(&name[MFD_NAME_PREFIX_LEN], uname, len)) {
2990 error = -EFAULT;
2991 goto err_name;
2992 }
2993
2994 /* terminating-zero may have changed after strnlen_user() returned */
2995 if (name[len + MFD_NAME_PREFIX_LEN - 1]) {
2996 error = -EFAULT;
2997 goto err_name;
2998 }
2999
3000 fd = get_unused_fd_flags((flags & MFD_CLOEXEC) ? O_CLOEXEC : 0);
3001 if (fd < 0) {
3002 error = fd;
3003 goto err_name;
3004 }
3005
3006 file = shmem_file_setup(name, 0, VM_NORESERVE);
3007 if (IS_ERR(file)) {
3008 error = PTR_ERR(file);
3009 goto err_fd;
3010 }
3011 info = SHMEM_I(file_inode(file));
3012 file->f_mode |= FMODE_LSEEK | FMODE_PREAD | FMODE_PWRITE;
3013 file->f_flags |= O_RDWR | O_LARGEFILE;
3014 if (flags & MFD_ALLOW_SEALING)
3015 info->seals &= ~F_SEAL_SEAL;
3016
3017 fd_install(fd, file);
3018 kfree(name);
3019 return fd;
3020
3021err_fd:
3022 put_unused_fd(fd);
3023err_name:
3024 kfree(name);
3025 return error;
3026}
3027
3028#endif /* CONFIG_TMPFS */
3029
3030static void shmem_put_super(struct super_block *sb)
3031{
3032 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
3033
3034 percpu_counter_destroy(&sbinfo->used_blocks);
3035 mpol_put(sbinfo->mpol);
3036 kfree(sbinfo);
3037 sb->s_fs_info = NULL;
3038}
3039
3040int shmem_fill_super(struct super_block *sb, void *data, int silent)
3041{
3042 struct inode *inode;
3043 struct shmem_sb_info *sbinfo;
3044 int err = -ENOMEM;
3045
3046 /* Round up to L1_CACHE_BYTES to resist false sharing */
3047 sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
3048 L1_CACHE_BYTES), GFP_KERNEL);
3049 if (!sbinfo)
3050 return -ENOMEM;
3051
3052 sbinfo->mode = S_IRWXUGO | S_ISVTX;
3053 sbinfo->uid = current_fsuid();
3054 sbinfo->gid = current_fsgid();
3055 sb->s_fs_info = sbinfo;
3056
3057#ifdef CONFIG_TMPFS
3058 /*
3059 * Per default we only allow half of the physical ram per
3060 * tmpfs instance, limiting inodes to one per page of lowmem;
3061 * but the internal instance is left unlimited.
3062 */
3063 if (!(sb->s_flags & MS_KERNMOUNT)) {
3064 sbinfo->max_blocks = shmem_default_max_blocks();
3065 sbinfo->max_inodes = shmem_default_max_inodes();
3066 if (shmem_parse_options(data, sbinfo, false)) {
3067 err = -EINVAL;
3068 goto failed;
3069 }
3070 } else {
3071 sb->s_flags |= MS_NOUSER;
3072 }
3073 sb->s_export_op = &shmem_export_ops;
3074 sb->s_flags |= MS_NOSEC;
3075#else
3076 sb->s_flags |= MS_NOUSER;
3077#endif
3078
3079 spin_lock_init(&sbinfo->stat_lock);
3080 if (percpu_counter_init(&sbinfo->used_blocks, 0, GFP_KERNEL))
3081 goto failed;
3082 sbinfo->free_inodes = sbinfo->max_inodes;
3083
3084 sb->s_maxbytes = MAX_LFS_FILESIZE;
3085 sb->s_blocksize = PAGE_SIZE;
3086 sb->s_blocksize_bits = PAGE_SHIFT;
3087 sb->s_magic = TMPFS_MAGIC;
3088 sb->s_op = &shmem_ops;
3089 sb->s_time_gran = 1;
3090#ifdef CONFIG_TMPFS_XATTR
3091 sb->s_xattr = shmem_xattr_handlers;
3092#endif
3093#ifdef CONFIG_TMPFS_POSIX_ACL
3094 sb->s_flags |= MS_POSIXACL;
3095#endif
3096
3097 inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
3098 if (!inode)
3099 goto failed;
3100 inode->i_uid = sbinfo->uid;
3101 inode->i_gid = sbinfo->gid;
3102 sb->s_root = d_make_root(inode);
3103 if (!sb->s_root)
3104 goto failed;
3105 return 0;
3106
3107failed:
3108 shmem_put_super(sb);
3109 return err;
3110}
3111
3112static struct kmem_cache *shmem_inode_cachep;
3113
3114static struct inode *shmem_alloc_inode(struct super_block *sb)
3115{
3116 struct shmem_inode_info *info;
3117 info = kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
3118 if (!info)
3119 return NULL;
3120 return &info->vfs_inode;
3121}
3122
3123static void shmem_destroy_callback(struct rcu_head *head)
3124{
3125 struct inode *inode = container_of(head, struct inode, i_rcu);
3126 kfree(inode->i_link);
3127 kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
3128}
3129
3130static void shmem_destroy_inode(struct inode *inode)
3131{
3132 if (S_ISREG(inode->i_mode))
3133 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
3134 call_rcu(&inode->i_rcu, shmem_destroy_callback);
3135}
3136
3137static void shmem_init_inode(void *foo)
3138{
3139 struct shmem_inode_info *info = foo;
3140 inode_init_once(&info->vfs_inode);
3141}
3142
3143static int shmem_init_inodecache(void)
3144{
3145 shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
3146 sizeof(struct shmem_inode_info),
3147 0, SLAB_PANIC|SLAB_ACCOUNT, shmem_init_inode);
3148 return 0;
3149}
3150
3151static void shmem_destroy_inodecache(void)
3152{
3153 kmem_cache_destroy(shmem_inode_cachep);
3154}
3155
3156static const struct address_space_operations shmem_aops = {
3157 .writepage = shmem_writepage,
3158 .set_page_dirty = __set_page_dirty_no_writeback,
3159#ifdef CONFIG_TMPFS
3160 .write_begin = shmem_write_begin,
3161 .write_end = shmem_write_end,
3162#endif
3163#ifdef CONFIG_MIGRATION
3164 .migratepage = migrate_page,
3165#endif
3166 .error_remove_page = generic_error_remove_page,
3167};
3168
3169static const struct file_operations shmem_file_operations = {
3170 .mmap = shmem_mmap,
3171#ifdef CONFIG_TMPFS
3172 .llseek = shmem_file_llseek,
3173 .read_iter = shmem_file_read_iter,
3174 .write_iter = generic_file_write_iter,
3175 .fsync = noop_fsync,
3176 .splice_read = shmem_file_splice_read,
3177 .splice_write = iter_file_splice_write,
3178 .fallocate = shmem_fallocate,
3179#endif
3180};
3181
3182static const struct inode_operations shmem_inode_operations = {
3183 .getattr = shmem_getattr,
3184 .setattr = shmem_setattr,
3185#ifdef CONFIG_TMPFS_XATTR
3186 .setxattr = generic_setxattr,
3187 .getxattr = generic_getxattr,
3188 .listxattr = shmem_listxattr,
3189 .removexattr = generic_removexattr,
3190 .set_acl = simple_set_acl,
3191#endif
3192};
3193
3194static const struct inode_operations shmem_dir_inode_operations = {
3195#ifdef CONFIG_TMPFS
3196 .create = shmem_create,
3197 .lookup = simple_lookup,
3198 .link = shmem_link,
3199 .unlink = shmem_unlink,
3200 .symlink = shmem_symlink,
3201 .mkdir = shmem_mkdir,
3202 .rmdir = shmem_rmdir,
3203 .mknod = shmem_mknod,
3204 .rename2 = shmem_rename2,
3205 .tmpfile = shmem_tmpfile,
3206#endif
3207#ifdef CONFIG_TMPFS_XATTR
3208 .setxattr = generic_setxattr,
3209 .getxattr = generic_getxattr,
3210 .listxattr = shmem_listxattr,
3211 .removexattr = generic_removexattr,
3212#endif
3213#ifdef CONFIG_TMPFS_POSIX_ACL
3214 .setattr = shmem_setattr,
3215 .set_acl = simple_set_acl,
3216#endif
3217};
3218
3219static const struct inode_operations shmem_special_inode_operations = {
3220#ifdef CONFIG_TMPFS_XATTR
3221 .setxattr = generic_setxattr,
3222 .getxattr = generic_getxattr,
3223 .listxattr = shmem_listxattr,
3224 .removexattr = generic_removexattr,
3225#endif
3226#ifdef CONFIG_TMPFS_POSIX_ACL
3227 .setattr = shmem_setattr,
3228 .set_acl = simple_set_acl,
3229#endif
3230};
3231
3232static const struct super_operations shmem_ops = {
3233 .alloc_inode = shmem_alloc_inode,
3234 .destroy_inode = shmem_destroy_inode,
3235#ifdef CONFIG_TMPFS
3236 .statfs = shmem_statfs,
3237 .remount_fs = shmem_remount_fs,
3238 .show_options = shmem_show_options,
3239#endif
3240 .evict_inode = shmem_evict_inode,
3241 .drop_inode = generic_delete_inode,
3242 .put_super = shmem_put_super,
3243};
3244
3245static const struct vm_operations_struct shmem_vm_ops = {
3246 .fault = shmem_fault,
3247 .map_pages = filemap_map_pages,
3248#ifdef CONFIG_NUMA
3249 .set_policy = shmem_set_policy,
3250 .get_policy = shmem_get_policy,
3251#endif
3252};
3253
3254static struct dentry *shmem_mount(struct file_system_type *fs_type,
3255 int flags, const char *dev_name, void *data)
3256{
3257 return mount_nodev(fs_type, flags, data, shmem_fill_super);
3258}
3259
3260static struct file_system_type shmem_fs_type = {
3261 .owner = THIS_MODULE,
3262 .name = "tmpfs",
3263 .mount = shmem_mount,
3264 .kill_sb = kill_litter_super,
3265 .fs_flags = FS_USERNS_MOUNT,
3266};
3267
3268int __init shmem_init(void)
3269{
3270 int error;
3271
3272 /* If rootfs called this, don't re-init */
3273 if (shmem_inode_cachep)
3274 return 0;
3275
3276 error = shmem_init_inodecache();
3277 if (error)
3278 goto out3;
3279
3280 error = register_filesystem(&shmem_fs_type);
3281 if (error) {
3282 pr_err("Could not register tmpfs\n");
3283 goto out2;
3284 }
3285
3286 shm_mnt = kern_mount(&shmem_fs_type);
3287 if (IS_ERR(shm_mnt)) {
3288 error = PTR_ERR(shm_mnt);
3289 pr_err("Could not kern_mount tmpfs\n");
3290 goto out1;
3291 }
3292 return 0;
3293
3294out1:
3295 unregister_filesystem(&shmem_fs_type);
3296out2:
3297 shmem_destroy_inodecache();
3298out3:
3299 shm_mnt = ERR_PTR(error);
3300 return error;
3301}
3302
3303#else /* !CONFIG_SHMEM */
3304
3305/*
3306 * tiny-shmem: simple shmemfs and tmpfs using ramfs code
3307 *
3308 * This is intended for small system where the benefits of the full
3309 * shmem code (swap-backed and resource-limited) are outweighed by
3310 * their complexity. On systems without swap this code should be
3311 * effectively equivalent, but much lighter weight.
3312 */
3313
3314static struct file_system_type shmem_fs_type = {
3315 .name = "tmpfs",
3316 .mount = ramfs_mount,
3317 .kill_sb = kill_litter_super,
3318 .fs_flags = FS_USERNS_MOUNT,
3319};
3320
3321int __init shmem_init(void)
3322{
3323 BUG_ON(register_filesystem(&shmem_fs_type) != 0);
3324
3325 shm_mnt = kern_mount(&shmem_fs_type);
3326 BUG_ON(IS_ERR(shm_mnt));
3327
3328 return 0;
3329}
3330
3331int shmem_unuse(swp_entry_t swap, struct page *page)
3332{
3333 return 0;
3334}
3335
3336int shmem_lock(struct file *file, int lock, struct user_struct *user)
3337{
3338 return 0;
3339}
3340
3341void shmem_unlock_mapping(struct address_space *mapping)
3342{
3343}
3344
3345void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
3346{
3347 truncate_inode_pages_range(inode->i_mapping, lstart, lend);
3348}
3349EXPORT_SYMBOL_GPL(shmem_truncate_range);
3350
3351#define shmem_vm_ops generic_file_vm_ops
3352#define shmem_file_operations ramfs_file_operations
3353#define shmem_get_inode(sb, dir, mode, dev, flags) ramfs_get_inode(sb, dir, mode, dev)
3354#define shmem_acct_size(flags, size) 0
3355#define shmem_unacct_size(flags, size) do {} while (0)
3356
3357#endif /* CONFIG_SHMEM */
3358
3359/* common code */
3360
3361static struct dentry_operations anon_ops = {
3362 .d_dname = simple_dname
3363};
3364
3365static struct file *__shmem_file_setup(const char *name, loff_t size,
3366 unsigned long flags, unsigned int i_flags)
3367{
3368 struct file *res;
3369 struct inode *inode;
3370 struct path path;
3371 struct super_block *sb;
3372 struct qstr this;
3373
3374 if (IS_ERR(shm_mnt))
3375 return ERR_CAST(shm_mnt);
3376
3377 if (size < 0 || size > MAX_LFS_FILESIZE)
3378 return ERR_PTR(-EINVAL);
3379
3380 if (shmem_acct_size(flags, size))
3381 return ERR_PTR(-ENOMEM);
3382
3383 res = ERR_PTR(-ENOMEM);
3384 this.name = name;
3385 this.len = strlen(name);
3386 this.hash = 0; /* will go */
3387 sb = shm_mnt->mnt_sb;
3388 path.mnt = mntget(shm_mnt);
3389 path.dentry = d_alloc_pseudo(sb, &this);
3390 if (!path.dentry)
3391 goto put_memory;
3392 d_set_d_op(path.dentry, &anon_ops);
3393
3394 res = ERR_PTR(-ENOSPC);
3395 inode = shmem_get_inode(sb, NULL, S_IFREG | S_IRWXUGO, 0, flags);
3396 if (!inode)
3397 goto put_memory;
3398
3399 inode->i_flags |= i_flags;
3400 d_instantiate(path.dentry, inode);
3401 inode->i_size = size;
3402 clear_nlink(inode); /* It is unlinked */
3403 res = ERR_PTR(ramfs_nommu_expand_for_mapping(inode, size));
3404 if (IS_ERR(res))
3405 goto put_path;
3406
3407 res = alloc_file(&path, FMODE_WRITE | FMODE_READ,
3408 &shmem_file_operations);
3409 if (IS_ERR(res))
3410 goto put_path;
3411
3412 return res;
3413
3414put_memory:
3415 shmem_unacct_size(flags, size);
3416put_path:
3417 path_put(&path);
3418 return res;
3419}
3420
3421/**
3422 * shmem_kernel_file_setup - get an unlinked file living in tmpfs which must be
3423 * kernel internal. There will be NO LSM permission checks against the
3424 * underlying inode. So users of this interface must do LSM checks at a
3425 * higher layer. The users are the big_key and shm implementations. LSM
3426 * checks are provided at the key or shm level rather than the inode.
3427 * @name: name for dentry (to be seen in /proc/<pid>/maps
3428 * @size: size to be set for the file
3429 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
3430 */
3431struct file *shmem_kernel_file_setup(const char *name, loff_t size, unsigned long flags)
3432{
3433 return __shmem_file_setup(name, size, flags, S_PRIVATE);
3434}
3435
3436/**
3437 * shmem_file_setup - get an unlinked file living in tmpfs
3438 * @name: name for dentry (to be seen in /proc/<pid>/maps
3439 * @size: size to be set for the file
3440 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
3441 */
3442struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
3443{
3444 return __shmem_file_setup(name, size, flags, 0);
3445}
3446EXPORT_SYMBOL_GPL(shmem_file_setup);
3447
3448/**
3449 * shmem_zero_setup - setup a shared anonymous mapping
3450 * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
3451 */
3452int shmem_zero_setup(struct vm_area_struct *vma)
3453{
3454 struct file *file;
3455 loff_t size = vma->vm_end - vma->vm_start;
3456
3457 /*
3458 * Cloning a new file under mmap_sem leads to a lock ordering conflict
3459 * between XFS directory reading and selinux: since this file is only
3460 * accessible to the user through its mapping, use S_PRIVATE flag to
3461 * bypass file security, in the same way as shmem_kernel_file_setup().
3462 */
3463 file = __shmem_file_setup("dev/zero", size, vma->vm_flags, S_PRIVATE);
3464 if (IS_ERR(file))
3465 return PTR_ERR(file);
3466
3467 if (vma->vm_file)
3468 fput(vma->vm_file);
3469 vma->vm_file = file;
3470 vma->vm_ops = &shmem_vm_ops;
3471 return 0;
3472}
3473
3474/**
3475 * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
3476 * @mapping: the page's address_space
3477 * @index: the page index
3478 * @gfp: the page allocator flags to use if allocating
3479 *
3480 * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
3481 * with any new page allocations done using the specified allocation flags.
3482 * But read_cache_page_gfp() uses the ->readpage() method: which does not
3483 * suit tmpfs, since it may have pages in swapcache, and needs to find those
3484 * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
3485 *
3486 * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
3487 * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
3488 */
3489struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
3490 pgoff_t index, gfp_t gfp)
3491{
3492#ifdef CONFIG_SHMEM
3493 struct inode *inode = mapping->host;
3494 struct page *page;
3495 int error;
3496
3497 BUG_ON(mapping->a_ops != &shmem_aops);
3498 error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE, gfp, NULL);
3499 if (error)
3500 page = ERR_PTR(error);
3501 else
3502 unlock_page(page);
3503 return page;
3504#else
3505 /*
3506 * The tiny !SHMEM case uses ramfs without swap
3507 */
3508 return read_cache_page_gfp(mapping, index, gfp);
3509#endif
3510}
3511EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);