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/pagemap.h>
29#include <linux/file.h>
30#include <linux/mm.h>
31#include <linux/export.h>
32#include <linux/swap.h>
33
34static struct vfsmount *shm_mnt;
35
36#ifdef CONFIG_SHMEM
37/*
38 * This virtual memory filesystem is heavily based on the ramfs. It
39 * extends ramfs by the ability to use swap and honor resource limits
40 * which makes it a completely usable filesystem.
41 */
42
43#include <linux/xattr.h>
44#include <linux/exportfs.h>
45#include <linux/posix_acl.h>
46#include <linux/generic_acl.h>
47#include <linux/mman.h>
48#include <linux/string.h>
49#include <linux/slab.h>
50#include <linux/backing-dev.h>
51#include <linux/shmem_fs.h>
52#include <linux/writeback.h>
53#include <linux/blkdev.h>
54#include <linux/pagevec.h>
55#include <linux/percpu_counter.h>
56#include <linux/falloc.h>
57#include <linux/splice.h>
58#include <linux/security.h>
59#include <linux/swapops.h>
60#include <linux/mempolicy.h>
61#include <linux/namei.h>
62#include <linux/ctype.h>
63#include <linux/migrate.h>
64#include <linux/highmem.h>
65#include <linux/seq_file.h>
66#include <linux/magic.h>
67
68#include <asm/uaccess.h>
69#include <asm/pgtable.h>
70
71#define BLOCKS_PER_PAGE (PAGE_CACHE_SIZE/512)
72#define VM_ACCT(size) (PAGE_CACHE_ALIGN(size) >> PAGE_SHIFT)
73
74/* Pretend that each entry is of this size in directory's i_size */
75#define BOGO_DIRENT_SIZE 20
76
77/* Symlink up to this size is kmalloc'ed instead of using a swappable page */
78#define SHORT_SYMLINK_LEN 128
79
80struct shmem_xattr {
81 struct list_head list; /* anchored by shmem_inode_info->xattr_list */
82 char *name; /* xattr name */
83 size_t size;
84 char value[0];
85};
86
87/*
88 * shmem_fallocate and shmem_writepage communicate via inode->i_private
89 * (with i_mutex making sure that it has only one user at a time):
90 * we would prefer not to enlarge the shmem inode just for that.
91 */
92struct shmem_falloc {
93 pgoff_t start; /* start of range currently being fallocated */
94 pgoff_t next; /* the next page offset to be fallocated */
95 pgoff_t nr_falloced; /* how many new pages have been fallocated */
96 pgoff_t nr_unswapped; /* how often writepage refused to swap out */
97};
98
99/* Flag allocation requirements to shmem_getpage */
100enum sgp_type {
101 SGP_READ, /* don't exceed i_size, don't allocate page */
102 SGP_CACHE, /* don't exceed i_size, may allocate page */
103 SGP_DIRTY, /* like SGP_CACHE, but set new page dirty */
104 SGP_WRITE, /* may exceed i_size, may allocate !Uptodate page */
105 SGP_FALLOC, /* like SGP_WRITE, but make existing page Uptodate */
106};
107
108#ifdef CONFIG_TMPFS
109static unsigned long shmem_default_max_blocks(void)
110{
111 return totalram_pages / 2;
112}
113
114static unsigned long shmem_default_max_inodes(void)
115{
116 return min(totalram_pages - totalhigh_pages, totalram_pages / 2);
117}
118#endif
119
120static bool shmem_should_replace_page(struct page *page, gfp_t gfp);
121static int shmem_replace_page(struct page **pagep, gfp_t gfp,
122 struct shmem_inode_info *info, pgoff_t index);
123static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
124 struct page **pagep, enum sgp_type sgp, gfp_t gfp, int *fault_type);
125
126static inline int shmem_getpage(struct inode *inode, pgoff_t index,
127 struct page **pagep, enum sgp_type sgp, int *fault_type)
128{
129 return shmem_getpage_gfp(inode, index, pagep, sgp,
130 mapping_gfp_mask(inode->i_mapping), fault_type);
131}
132
133static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
134{
135 return sb->s_fs_info;
136}
137
138/*
139 * shmem_file_setup pre-accounts the whole fixed size of a VM object,
140 * for shared memory and for shared anonymous (/dev/zero) mappings
141 * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
142 * consistent with the pre-accounting of private mappings ...
143 */
144static inline int shmem_acct_size(unsigned long flags, loff_t size)
145{
146 return (flags & VM_NORESERVE) ?
147 0 : security_vm_enough_memory_mm(current->mm, VM_ACCT(size));
148}
149
150static inline void shmem_unacct_size(unsigned long flags, loff_t size)
151{
152 if (!(flags & VM_NORESERVE))
153 vm_unacct_memory(VM_ACCT(size));
154}
155
156/*
157 * ... whereas tmpfs objects are accounted incrementally as
158 * pages are allocated, in order to allow huge sparse files.
159 * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
160 * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
161 */
162static inline int shmem_acct_block(unsigned long flags)
163{
164 return (flags & VM_NORESERVE) ?
165 security_vm_enough_memory_mm(current->mm, VM_ACCT(PAGE_CACHE_SIZE)) : 0;
166}
167
168static inline void shmem_unacct_blocks(unsigned long flags, long pages)
169{
170 if (flags & VM_NORESERVE)
171 vm_unacct_memory(pages * VM_ACCT(PAGE_CACHE_SIZE));
172}
173
174static const struct super_operations shmem_ops;
175static const struct address_space_operations shmem_aops;
176static const struct file_operations shmem_file_operations;
177static const struct inode_operations shmem_inode_operations;
178static const struct inode_operations shmem_dir_inode_operations;
179static const struct inode_operations shmem_special_inode_operations;
180static const struct vm_operations_struct shmem_vm_ops;
181
182static struct backing_dev_info shmem_backing_dev_info __read_mostly = {
183 .ra_pages = 0, /* No readahead */
184 .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_SWAP_BACKED,
185};
186
187static LIST_HEAD(shmem_swaplist);
188static DEFINE_MUTEX(shmem_swaplist_mutex);
189
190static int shmem_reserve_inode(struct super_block *sb)
191{
192 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
193 if (sbinfo->max_inodes) {
194 spin_lock(&sbinfo->stat_lock);
195 if (!sbinfo->free_inodes) {
196 spin_unlock(&sbinfo->stat_lock);
197 return -ENOSPC;
198 }
199 sbinfo->free_inodes--;
200 spin_unlock(&sbinfo->stat_lock);
201 }
202 return 0;
203}
204
205static void shmem_free_inode(struct super_block *sb)
206{
207 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
208 if (sbinfo->max_inodes) {
209 spin_lock(&sbinfo->stat_lock);
210 sbinfo->free_inodes++;
211 spin_unlock(&sbinfo->stat_lock);
212 }
213}
214
215/**
216 * shmem_recalc_inode - recalculate the block usage of an inode
217 * @inode: inode to recalc
218 *
219 * We have to calculate the free blocks since the mm can drop
220 * undirtied hole pages behind our back.
221 *
222 * But normally info->alloced == inode->i_mapping->nrpages + info->swapped
223 * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
224 *
225 * It has to be called with the spinlock held.
226 */
227static void shmem_recalc_inode(struct inode *inode)
228{
229 struct shmem_inode_info *info = SHMEM_I(inode);
230 long freed;
231
232 freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
233 if (freed > 0) {
234 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
235 if (sbinfo->max_blocks)
236 percpu_counter_add(&sbinfo->used_blocks, -freed);
237 info->alloced -= freed;
238 inode->i_blocks -= freed * BLOCKS_PER_PAGE;
239 shmem_unacct_blocks(info->flags, freed);
240 }
241}
242
243/*
244 * Replace item expected in radix tree by a new item, while holding tree lock.
245 */
246static int shmem_radix_tree_replace(struct address_space *mapping,
247 pgoff_t index, void *expected, void *replacement)
248{
249 void **pslot;
250 void *item = NULL;
251
252 VM_BUG_ON(!expected);
253 pslot = radix_tree_lookup_slot(&mapping->page_tree, index);
254 if (pslot)
255 item = radix_tree_deref_slot_protected(pslot,
256 &mapping->tree_lock);
257 if (item != expected)
258 return -ENOENT;
259 if (replacement)
260 radix_tree_replace_slot(pslot, replacement);
261 else
262 radix_tree_delete(&mapping->page_tree, index);
263 return 0;
264}
265
266/*
267 * Sometimes, before we decide whether to proceed or to fail, we must check
268 * that an entry was not already brought back from swap by a racing thread.
269 *
270 * Checking page is not enough: by the time a SwapCache page is locked, it
271 * might be reused, and again be SwapCache, using the same swap as before.
272 */
273static bool shmem_confirm_swap(struct address_space *mapping,
274 pgoff_t index, swp_entry_t swap)
275{
276 void *item;
277
278 rcu_read_lock();
279 item = radix_tree_lookup(&mapping->page_tree, index);
280 rcu_read_unlock();
281 return item == swp_to_radix_entry(swap);
282}
283
284/*
285 * Like add_to_page_cache_locked, but error if expected item has gone.
286 */
287static int shmem_add_to_page_cache(struct page *page,
288 struct address_space *mapping,
289 pgoff_t index, gfp_t gfp, void *expected)
290{
291 int error;
292
293 VM_BUG_ON(!PageLocked(page));
294 VM_BUG_ON(!PageSwapBacked(page));
295
296 page_cache_get(page);
297 page->mapping = mapping;
298 page->index = index;
299
300 spin_lock_irq(&mapping->tree_lock);
301 if (!expected)
302 error = radix_tree_insert(&mapping->page_tree, index, page);
303 else
304 error = shmem_radix_tree_replace(mapping, index, expected,
305 page);
306 if (!error) {
307 mapping->nrpages++;
308 __inc_zone_page_state(page, NR_FILE_PAGES);
309 __inc_zone_page_state(page, NR_SHMEM);
310 spin_unlock_irq(&mapping->tree_lock);
311 } else {
312 page->mapping = NULL;
313 spin_unlock_irq(&mapping->tree_lock);
314 page_cache_release(page);
315 }
316 return error;
317}
318
319/*
320 * Like delete_from_page_cache, but substitutes swap for page.
321 */
322static void shmem_delete_from_page_cache(struct page *page, void *radswap)
323{
324 struct address_space *mapping = page->mapping;
325 int error;
326
327 spin_lock_irq(&mapping->tree_lock);
328 error = shmem_radix_tree_replace(mapping, page->index, page, radswap);
329 page->mapping = NULL;
330 mapping->nrpages--;
331 __dec_zone_page_state(page, NR_FILE_PAGES);
332 __dec_zone_page_state(page, NR_SHMEM);
333 spin_unlock_irq(&mapping->tree_lock);
334 page_cache_release(page);
335 BUG_ON(error);
336}
337
338/*
339 * Like find_get_pages, but collecting swap entries as well as pages.
340 */
341static unsigned shmem_find_get_pages_and_swap(struct address_space *mapping,
342 pgoff_t start, unsigned int nr_pages,
343 struct page **pages, pgoff_t *indices)
344{
345 unsigned int i;
346 unsigned int ret;
347 unsigned int nr_found;
348
349 rcu_read_lock();
350restart:
351 nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
352 (void ***)pages, indices, start, nr_pages);
353 ret = 0;
354 for (i = 0; i < nr_found; i++) {
355 struct page *page;
356repeat:
357 page = radix_tree_deref_slot((void **)pages[i]);
358 if (unlikely(!page))
359 continue;
360 if (radix_tree_exception(page)) {
361 if (radix_tree_deref_retry(page))
362 goto restart;
363 /*
364 * Otherwise, we must be storing a swap entry
365 * here as an exceptional entry: so return it
366 * without attempting to raise page count.
367 */
368 goto export;
369 }
370 if (!page_cache_get_speculative(page))
371 goto repeat;
372
373 /* Has the page moved? */
374 if (unlikely(page != *((void **)pages[i]))) {
375 page_cache_release(page);
376 goto repeat;
377 }
378export:
379 indices[ret] = indices[i];
380 pages[ret] = page;
381 ret++;
382 }
383 if (unlikely(!ret && nr_found))
384 goto restart;
385 rcu_read_unlock();
386 return ret;
387}
388
389/*
390 * Remove swap entry from radix tree, free the swap and its page cache.
391 */
392static int shmem_free_swap(struct address_space *mapping,
393 pgoff_t index, void *radswap)
394{
395 int error;
396
397 spin_lock_irq(&mapping->tree_lock);
398 error = shmem_radix_tree_replace(mapping, index, radswap, NULL);
399 spin_unlock_irq(&mapping->tree_lock);
400 if (!error)
401 free_swap_and_cache(radix_to_swp_entry(radswap));
402 return error;
403}
404
405/*
406 * Pagevec may contain swap entries, so shuffle up pages before releasing.
407 */
408static void shmem_deswap_pagevec(struct pagevec *pvec)
409{
410 int i, j;
411
412 for (i = 0, j = 0; i < pagevec_count(pvec); i++) {
413 struct page *page = pvec->pages[i];
414 if (!radix_tree_exceptional_entry(page))
415 pvec->pages[j++] = page;
416 }
417 pvec->nr = j;
418}
419
420/*
421 * SysV IPC SHM_UNLOCK restore Unevictable pages to their evictable lists.
422 */
423void shmem_unlock_mapping(struct address_space *mapping)
424{
425 struct pagevec pvec;
426 pgoff_t indices[PAGEVEC_SIZE];
427 pgoff_t index = 0;
428
429 pagevec_init(&pvec, 0);
430 /*
431 * Minor point, but we might as well stop if someone else SHM_LOCKs it.
432 */
433 while (!mapping_unevictable(mapping)) {
434 /*
435 * Avoid pagevec_lookup(): find_get_pages() returns 0 as if it
436 * has finished, if it hits a row of PAGEVEC_SIZE swap entries.
437 */
438 pvec.nr = shmem_find_get_pages_and_swap(mapping, index,
439 PAGEVEC_SIZE, pvec.pages, indices);
440 if (!pvec.nr)
441 break;
442 index = indices[pvec.nr - 1] + 1;
443 shmem_deswap_pagevec(&pvec);
444 check_move_unevictable_pages(pvec.pages, pvec.nr);
445 pagevec_release(&pvec);
446 cond_resched();
447 }
448}
449
450/*
451 * Remove range of pages and swap entries from radix tree, and free them.
452 * If !unfalloc, truncate or punch hole; if unfalloc, undo failed fallocate.
453 */
454static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
455 bool unfalloc)
456{
457 struct address_space *mapping = inode->i_mapping;
458 struct shmem_inode_info *info = SHMEM_I(inode);
459 pgoff_t start = (lstart + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
460 pgoff_t end = (lend + 1) >> PAGE_CACHE_SHIFT;
461 unsigned int partial_start = lstart & (PAGE_CACHE_SIZE - 1);
462 unsigned int partial_end = (lend + 1) & (PAGE_CACHE_SIZE - 1);
463 struct pagevec pvec;
464 pgoff_t indices[PAGEVEC_SIZE];
465 long nr_swaps_freed = 0;
466 pgoff_t index;
467 int i;
468
469 if (lend == -1)
470 end = -1; /* unsigned, so actually very big */
471
472 pagevec_init(&pvec, 0);
473 index = start;
474 while (index < end) {
475 pvec.nr = shmem_find_get_pages_and_swap(mapping, index,
476 min(end - index, (pgoff_t)PAGEVEC_SIZE),
477 pvec.pages, indices);
478 if (!pvec.nr)
479 break;
480 mem_cgroup_uncharge_start();
481 for (i = 0; i < pagevec_count(&pvec); i++) {
482 struct page *page = pvec.pages[i];
483
484 index = indices[i];
485 if (index >= end)
486 break;
487
488 if (radix_tree_exceptional_entry(page)) {
489 if (unfalloc)
490 continue;
491 nr_swaps_freed += !shmem_free_swap(mapping,
492 index, page);
493 continue;
494 }
495
496 if (!trylock_page(page))
497 continue;
498 if (!unfalloc || !PageUptodate(page)) {
499 if (page->mapping == mapping) {
500 VM_BUG_ON(PageWriteback(page));
501 truncate_inode_page(mapping, page);
502 }
503 }
504 unlock_page(page);
505 }
506 shmem_deswap_pagevec(&pvec);
507 pagevec_release(&pvec);
508 mem_cgroup_uncharge_end();
509 cond_resched();
510 index++;
511 }
512
513 if (partial_start) {
514 struct page *page = NULL;
515 shmem_getpage(inode, start - 1, &page, SGP_READ, NULL);
516 if (page) {
517 unsigned int top = PAGE_CACHE_SIZE;
518 if (start > end) {
519 top = partial_end;
520 partial_end = 0;
521 }
522 zero_user_segment(page, partial_start, top);
523 set_page_dirty(page);
524 unlock_page(page);
525 page_cache_release(page);
526 }
527 }
528 if (partial_end) {
529 struct page *page = NULL;
530 shmem_getpage(inode, end, &page, SGP_READ, NULL);
531 if (page) {
532 zero_user_segment(page, 0, partial_end);
533 set_page_dirty(page);
534 unlock_page(page);
535 page_cache_release(page);
536 }
537 }
538 if (start >= end)
539 return;
540
541 index = start;
542 for ( ; ; ) {
543 cond_resched();
544 pvec.nr = shmem_find_get_pages_and_swap(mapping, index,
545 min(end - index, (pgoff_t)PAGEVEC_SIZE),
546 pvec.pages, indices);
547 if (!pvec.nr) {
548 if (index == start || unfalloc)
549 break;
550 index = start;
551 continue;
552 }
553 if ((index == start || unfalloc) && indices[0] >= end) {
554 shmem_deswap_pagevec(&pvec);
555 pagevec_release(&pvec);
556 break;
557 }
558 mem_cgroup_uncharge_start();
559 for (i = 0; i < pagevec_count(&pvec); i++) {
560 struct page *page = pvec.pages[i];
561
562 index = indices[i];
563 if (index >= end)
564 break;
565
566 if (radix_tree_exceptional_entry(page)) {
567 if (unfalloc)
568 continue;
569 nr_swaps_freed += !shmem_free_swap(mapping,
570 index, page);
571 continue;
572 }
573
574 lock_page(page);
575 if (!unfalloc || !PageUptodate(page)) {
576 if (page->mapping == mapping) {
577 VM_BUG_ON(PageWriteback(page));
578 truncate_inode_page(mapping, page);
579 }
580 }
581 unlock_page(page);
582 }
583 shmem_deswap_pagevec(&pvec);
584 pagevec_release(&pvec);
585 mem_cgroup_uncharge_end();
586 index++;
587 }
588
589 spin_lock(&info->lock);
590 info->swapped -= nr_swaps_freed;
591 shmem_recalc_inode(inode);
592 spin_unlock(&info->lock);
593}
594
595void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
596{
597 shmem_undo_range(inode, lstart, lend, false);
598 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
599}
600EXPORT_SYMBOL_GPL(shmem_truncate_range);
601
602static int shmem_setattr(struct dentry *dentry, struct iattr *attr)
603{
604 struct inode *inode = dentry->d_inode;
605 int error;
606
607 error = inode_change_ok(inode, attr);
608 if (error)
609 return error;
610
611 if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
612 loff_t oldsize = inode->i_size;
613 loff_t newsize = attr->ia_size;
614
615 if (newsize != oldsize) {
616 i_size_write(inode, newsize);
617 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
618 }
619 if (newsize < oldsize) {
620 loff_t holebegin = round_up(newsize, PAGE_SIZE);
621 unmap_mapping_range(inode->i_mapping, holebegin, 0, 1);
622 shmem_truncate_range(inode, newsize, (loff_t)-1);
623 /* unmap again to remove racily COWed private pages */
624 unmap_mapping_range(inode->i_mapping, holebegin, 0, 1);
625 }
626 }
627
628 setattr_copy(inode, attr);
629#ifdef CONFIG_TMPFS_POSIX_ACL
630 if (attr->ia_valid & ATTR_MODE)
631 error = generic_acl_chmod(inode);
632#endif
633 return error;
634}
635
636static void shmem_evict_inode(struct inode *inode)
637{
638 struct shmem_inode_info *info = SHMEM_I(inode);
639 struct shmem_xattr *xattr, *nxattr;
640
641 if (inode->i_mapping->a_ops == &shmem_aops) {
642 shmem_unacct_size(info->flags, inode->i_size);
643 inode->i_size = 0;
644 shmem_truncate_range(inode, 0, (loff_t)-1);
645 if (!list_empty(&info->swaplist)) {
646 mutex_lock(&shmem_swaplist_mutex);
647 list_del_init(&info->swaplist);
648 mutex_unlock(&shmem_swaplist_mutex);
649 }
650 } else
651 kfree(info->symlink);
652
653 list_for_each_entry_safe(xattr, nxattr, &info->xattr_list, list) {
654 kfree(xattr->name);
655 kfree(xattr);
656 }
657 BUG_ON(inode->i_blocks);
658 shmem_free_inode(inode->i_sb);
659 clear_inode(inode);
660}
661
662/*
663 * If swap found in inode, free it and move page from swapcache to filecache.
664 */
665static int shmem_unuse_inode(struct shmem_inode_info *info,
666 swp_entry_t swap, struct page **pagep)
667{
668 struct address_space *mapping = info->vfs_inode.i_mapping;
669 void *radswap;
670 pgoff_t index;
671 gfp_t gfp;
672 int error = 0;
673
674 radswap = swp_to_radix_entry(swap);
675 index = radix_tree_locate_item(&mapping->page_tree, radswap);
676 if (index == -1)
677 return 0;
678
679 /*
680 * Move _head_ to start search for next from here.
681 * But be careful: shmem_evict_inode checks list_empty without taking
682 * mutex, and there's an instant in list_move_tail when info->swaplist
683 * would appear empty, if it were the only one on shmem_swaplist.
684 */
685 if (shmem_swaplist.next != &info->swaplist)
686 list_move_tail(&shmem_swaplist, &info->swaplist);
687
688 gfp = mapping_gfp_mask(mapping);
689 if (shmem_should_replace_page(*pagep, gfp)) {
690 mutex_unlock(&shmem_swaplist_mutex);
691 error = shmem_replace_page(pagep, gfp, info, index);
692 mutex_lock(&shmem_swaplist_mutex);
693 /*
694 * We needed to drop mutex to make that restrictive page
695 * allocation, but the inode might have been freed while we
696 * dropped it: although a racing shmem_evict_inode() cannot
697 * complete without emptying the radix_tree, our page lock
698 * on this swapcache page is not enough to prevent that -
699 * free_swap_and_cache() of our swap entry will only
700 * trylock_page(), removing swap from radix_tree whatever.
701 *
702 * We must not proceed to shmem_add_to_page_cache() if the
703 * inode has been freed, but of course we cannot rely on
704 * inode or mapping or info to check that. However, we can
705 * safely check if our swap entry is still in use (and here
706 * it can't have got reused for another page): if it's still
707 * in use, then the inode cannot have been freed yet, and we
708 * can safely proceed (if it's no longer in use, that tells
709 * nothing about the inode, but we don't need to unuse swap).
710 */
711 if (!page_swapcount(*pagep))
712 error = -ENOENT;
713 }
714
715 /*
716 * We rely on shmem_swaplist_mutex, not only to protect the swaplist,
717 * but also to hold up shmem_evict_inode(): so inode cannot be freed
718 * beneath us (pagelock doesn't help until the page is in pagecache).
719 */
720 if (!error)
721 error = shmem_add_to_page_cache(*pagep, mapping, index,
722 GFP_NOWAIT, radswap);
723 if (error != -ENOMEM) {
724 /*
725 * Truncation and eviction use free_swap_and_cache(), which
726 * only does trylock page: if we raced, best clean up here.
727 */
728 delete_from_swap_cache(*pagep);
729 set_page_dirty(*pagep);
730 if (!error) {
731 spin_lock(&info->lock);
732 info->swapped--;
733 spin_unlock(&info->lock);
734 swap_free(swap);
735 }
736 error = 1; /* not an error, but entry was found */
737 }
738 return error;
739}
740
741/*
742 * Search through swapped inodes to find and replace swap by page.
743 */
744int shmem_unuse(swp_entry_t swap, struct page *page)
745{
746 struct list_head *this, *next;
747 struct shmem_inode_info *info;
748 int found = 0;
749 int error = 0;
750
751 /*
752 * There's a faint possibility that swap page was replaced before
753 * caller locked it: caller will come back later with the right page.
754 */
755 if (unlikely(!PageSwapCache(page) || page_private(page) != swap.val))
756 goto out;
757
758 /*
759 * Charge page using GFP_KERNEL while we can wait, before taking
760 * the shmem_swaplist_mutex which might hold up shmem_writepage().
761 * Charged back to the user (not to caller) when swap account is used.
762 */
763 error = mem_cgroup_cache_charge(page, current->mm, GFP_KERNEL);
764 if (error)
765 goto out;
766 /* No radix_tree_preload: swap entry keeps a place for page in tree */
767
768 mutex_lock(&shmem_swaplist_mutex);
769 list_for_each_safe(this, next, &shmem_swaplist) {
770 info = list_entry(this, struct shmem_inode_info, swaplist);
771 if (info->swapped)
772 found = shmem_unuse_inode(info, swap, &page);
773 else
774 list_del_init(&info->swaplist);
775 cond_resched();
776 if (found)
777 break;
778 }
779 mutex_unlock(&shmem_swaplist_mutex);
780
781 if (found < 0)
782 error = found;
783out:
784 unlock_page(page);
785 page_cache_release(page);
786 return error;
787}
788
789/*
790 * Move the page from the page cache to the swap cache.
791 */
792static int shmem_writepage(struct page *page, struct writeback_control *wbc)
793{
794 struct shmem_inode_info *info;
795 struct address_space *mapping;
796 struct inode *inode;
797 swp_entry_t swap;
798 pgoff_t index;
799
800 BUG_ON(!PageLocked(page));
801 mapping = page->mapping;
802 index = page->index;
803 inode = mapping->host;
804 info = SHMEM_I(inode);
805 if (info->flags & VM_LOCKED)
806 goto redirty;
807 if (!total_swap_pages)
808 goto redirty;
809
810 /*
811 * shmem_backing_dev_info's capabilities prevent regular writeback or
812 * sync from ever calling shmem_writepage; but a stacking filesystem
813 * might use ->writepage of its underlying filesystem, in which case
814 * tmpfs should write out to swap only in response to memory pressure,
815 * and not for the writeback threads or sync.
816 */
817 if (!wbc->for_reclaim) {
818 WARN_ON_ONCE(1); /* Still happens? Tell us about it! */
819 goto redirty;
820 }
821
822 /*
823 * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
824 * value into swapfile.c, the only way we can correctly account for a
825 * fallocated page arriving here is now to initialize it and write it.
826 *
827 * That's okay for a page already fallocated earlier, but if we have
828 * not yet completed the fallocation, then (a) we want to keep track
829 * of this page in case we have to undo it, and (b) it may not be a
830 * good idea to continue anyway, once we're pushing into swap. So
831 * reactivate the page, and let shmem_fallocate() quit when too many.
832 */
833 if (!PageUptodate(page)) {
834 if (inode->i_private) {
835 struct shmem_falloc *shmem_falloc;
836 spin_lock(&inode->i_lock);
837 shmem_falloc = inode->i_private;
838 if (shmem_falloc &&
839 index >= shmem_falloc->start &&
840 index < shmem_falloc->next)
841 shmem_falloc->nr_unswapped++;
842 else
843 shmem_falloc = NULL;
844 spin_unlock(&inode->i_lock);
845 if (shmem_falloc)
846 goto redirty;
847 }
848 clear_highpage(page);
849 flush_dcache_page(page);
850 SetPageUptodate(page);
851 }
852
853 swap = get_swap_page();
854 if (!swap.val)
855 goto redirty;
856
857 /*
858 * Add inode to shmem_unuse()'s list of swapped-out inodes,
859 * if it's not already there. Do it now before the page is
860 * moved to swap cache, when its pagelock no longer protects
861 * the inode from eviction. But don't unlock the mutex until
862 * we've incremented swapped, because shmem_unuse_inode() will
863 * prune a !swapped inode from the swaplist under this mutex.
864 */
865 mutex_lock(&shmem_swaplist_mutex);
866 if (list_empty(&info->swaplist))
867 list_add_tail(&info->swaplist, &shmem_swaplist);
868
869 if (add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
870 swap_shmem_alloc(swap);
871 shmem_delete_from_page_cache(page, swp_to_radix_entry(swap));
872
873 spin_lock(&info->lock);
874 info->swapped++;
875 shmem_recalc_inode(inode);
876 spin_unlock(&info->lock);
877
878 mutex_unlock(&shmem_swaplist_mutex);
879 BUG_ON(page_mapped(page));
880 swap_writepage(page, wbc);
881 return 0;
882 }
883
884 mutex_unlock(&shmem_swaplist_mutex);
885 swapcache_free(swap, NULL);
886redirty:
887 set_page_dirty(page);
888 if (wbc->for_reclaim)
889 return AOP_WRITEPAGE_ACTIVATE; /* Return with page locked */
890 unlock_page(page);
891 return 0;
892}
893
894#ifdef CONFIG_NUMA
895#ifdef CONFIG_TMPFS
896static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
897{
898 char buffer[64];
899
900 if (!mpol || mpol->mode == MPOL_DEFAULT)
901 return; /* show nothing */
902
903 mpol_to_str(buffer, sizeof(buffer), mpol, 1);
904
905 seq_printf(seq, ",mpol=%s", buffer);
906}
907
908static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
909{
910 struct mempolicy *mpol = NULL;
911 if (sbinfo->mpol) {
912 spin_lock(&sbinfo->stat_lock); /* prevent replace/use races */
913 mpol = sbinfo->mpol;
914 mpol_get(mpol);
915 spin_unlock(&sbinfo->stat_lock);
916 }
917 return mpol;
918}
919#endif /* CONFIG_TMPFS */
920
921static struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
922 struct shmem_inode_info *info, pgoff_t index)
923{
924 struct mempolicy mpol, *spol;
925 struct vm_area_struct pvma;
926
927 spol = mpol_cond_copy(&mpol,
928 mpol_shared_policy_lookup(&info->policy, index));
929
930 /* Create a pseudo vma that just contains the policy */
931 pvma.vm_start = 0;
932 pvma.vm_pgoff = index;
933 pvma.vm_ops = NULL;
934 pvma.vm_policy = spol;
935 return swapin_readahead(swap, gfp, &pvma, 0);
936}
937
938static struct page *shmem_alloc_page(gfp_t gfp,
939 struct shmem_inode_info *info, pgoff_t index)
940{
941 struct vm_area_struct pvma;
942
943 /* Create a pseudo vma that just contains the policy */
944 pvma.vm_start = 0;
945 pvma.vm_pgoff = index;
946 pvma.vm_ops = NULL;
947 pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, index);
948
949 /*
950 * alloc_page_vma() will drop the shared policy reference
951 */
952 return alloc_page_vma(gfp, &pvma, 0);
953}
954#else /* !CONFIG_NUMA */
955#ifdef CONFIG_TMPFS
956static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
957{
958}
959#endif /* CONFIG_TMPFS */
960
961static inline struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
962 struct shmem_inode_info *info, pgoff_t index)
963{
964 return swapin_readahead(swap, gfp, NULL, 0);
965}
966
967static inline struct page *shmem_alloc_page(gfp_t gfp,
968 struct shmem_inode_info *info, pgoff_t index)
969{
970 return alloc_page(gfp);
971}
972#endif /* CONFIG_NUMA */
973
974#if !defined(CONFIG_NUMA) || !defined(CONFIG_TMPFS)
975static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
976{
977 return NULL;
978}
979#endif
980
981/*
982 * When a page is moved from swapcache to shmem filecache (either by the
983 * usual swapin of shmem_getpage_gfp(), or by the less common swapoff of
984 * shmem_unuse_inode()), it may have been read in earlier from swap, in
985 * ignorance of the mapping it belongs to. If that mapping has special
986 * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
987 * we may need to copy to a suitable page before moving to filecache.
988 *
989 * In a future release, this may well be extended to respect cpuset and
990 * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
991 * but for now it is a simple matter of zone.
992 */
993static bool shmem_should_replace_page(struct page *page, gfp_t gfp)
994{
995 return page_zonenum(page) > gfp_zone(gfp);
996}
997
998static int shmem_replace_page(struct page **pagep, gfp_t gfp,
999 struct shmem_inode_info *info, pgoff_t index)
1000{
1001 struct page *oldpage, *newpage;
1002 struct address_space *swap_mapping;
1003 pgoff_t swap_index;
1004 int error;
1005
1006 oldpage = *pagep;
1007 swap_index = page_private(oldpage);
1008 swap_mapping = page_mapping(oldpage);
1009
1010 /*
1011 * We have arrived here because our zones are constrained, so don't
1012 * limit chance of success by further cpuset and node constraints.
1013 */
1014 gfp &= ~GFP_CONSTRAINT_MASK;
1015 newpage = shmem_alloc_page(gfp, info, index);
1016 if (!newpage)
1017 return -ENOMEM;
1018
1019 page_cache_get(newpage);
1020 copy_highpage(newpage, oldpage);
1021 flush_dcache_page(newpage);
1022
1023 __set_page_locked(newpage);
1024 SetPageUptodate(newpage);
1025 SetPageSwapBacked(newpage);
1026 set_page_private(newpage, swap_index);
1027 SetPageSwapCache(newpage);
1028
1029 /*
1030 * Our caller will very soon move newpage out of swapcache, but it's
1031 * a nice clean interface for us to replace oldpage by newpage there.
1032 */
1033 spin_lock_irq(&swap_mapping->tree_lock);
1034 error = shmem_radix_tree_replace(swap_mapping, swap_index, oldpage,
1035 newpage);
1036 if (!error) {
1037 __inc_zone_page_state(newpage, NR_FILE_PAGES);
1038 __dec_zone_page_state(oldpage, NR_FILE_PAGES);
1039 }
1040 spin_unlock_irq(&swap_mapping->tree_lock);
1041
1042 if (unlikely(error)) {
1043 /*
1044 * Is this possible? I think not, now that our callers check
1045 * both PageSwapCache and page_private after getting page lock;
1046 * but be defensive. Reverse old to newpage for clear and free.
1047 */
1048 oldpage = newpage;
1049 } else {
1050 mem_cgroup_replace_page_cache(oldpage, newpage);
1051 lru_cache_add_anon(newpage);
1052 *pagep = newpage;
1053 }
1054
1055 ClearPageSwapCache(oldpage);
1056 set_page_private(oldpage, 0);
1057
1058 unlock_page(oldpage);
1059 page_cache_release(oldpage);
1060 page_cache_release(oldpage);
1061 return error;
1062}
1063
1064/*
1065 * shmem_getpage_gfp - find page in cache, or get from swap, or allocate
1066 *
1067 * If we allocate a new one we do not mark it dirty. That's up to the
1068 * vm. If we swap it in we mark it dirty since we also free the swap
1069 * entry since a page cannot live in both the swap and page cache
1070 */
1071static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
1072 struct page **pagep, enum sgp_type sgp, gfp_t gfp, int *fault_type)
1073{
1074 struct address_space *mapping = inode->i_mapping;
1075 struct shmem_inode_info *info;
1076 struct shmem_sb_info *sbinfo;
1077 struct page *page;
1078 swp_entry_t swap;
1079 int error;
1080 int once = 0;
1081 int alloced = 0;
1082
1083 if (index > (MAX_LFS_FILESIZE >> PAGE_CACHE_SHIFT))
1084 return -EFBIG;
1085repeat:
1086 swap.val = 0;
1087 page = find_lock_page(mapping, index);
1088 if (radix_tree_exceptional_entry(page)) {
1089 swap = radix_to_swp_entry(page);
1090 page = NULL;
1091 }
1092
1093 if (sgp != SGP_WRITE && sgp != SGP_FALLOC &&
1094 ((loff_t)index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
1095 error = -EINVAL;
1096 goto failed;
1097 }
1098
1099 /* fallocated page? */
1100 if (page && !PageUptodate(page)) {
1101 if (sgp != SGP_READ)
1102 goto clear;
1103 unlock_page(page);
1104 page_cache_release(page);
1105 page = NULL;
1106 }
1107 if (page || (sgp == SGP_READ && !swap.val)) {
1108 *pagep = page;
1109 return 0;
1110 }
1111
1112 /*
1113 * Fast cache lookup did not find it:
1114 * bring it back from swap or allocate.
1115 */
1116 info = SHMEM_I(inode);
1117 sbinfo = SHMEM_SB(inode->i_sb);
1118
1119 if (swap.val) {
1120 /* Look it up and read it in.. */
1121 page = lookup_swap_cache(swap);
1122 if (!page) {
1123 /* here we actually do the io */
1124 if (fault_type)
1125 *fault_type |= VM_FAULT_MAJOR;
1126 page = shmem_swapin(swap, gfp, info, index);
1127 if (!page) {
1128 error = -ENOMEM;
1129 goto failed;
1130 }
1131 }
1132
1133 /* We have to do this with page locked to prevent races */
1134 lock_page(page);
1135 if (!PageSwapCache(page) || page_private(page) != swap.val ||
1136 !shmem_confirm_swap(mapping, index, swap)) {
1137 error = -EEXIST; /* try again */
1138 goto unlock;
1139 }
1140 if (!PageUptodate(page)) {
1141 error = -EIO;
1142 goto failed;
1143 }
1144 wait_on_page_writeback(page);
1145
1146 if (shmem_should_replace_page(page, gfp)) {
1147 error = shmem_replace_page(&page, gfp, info, index);
1148 if (error)
1149 goto failed;
1150 }
1151
1152 error = mem_cgroup_cache_charge(page, current->mm,
1153 gfp & GFP_RECLAIM_MASK);
1154 if (!error) {
1155 error = shmem_add_to_page_cache(page, mapping, index,
1156 gfp, swp_to_radix_entry(swap));
1157 /* We already confirmed swap, and make no allocation */
1158 VM_BUG_ON(error);
1159 }
1160 if (error)
1161 goto failed;
1162
1163 spin_lock(&info->lock);
1164 info->swapped--;
1165 shmem_recalc_inode(inode);
1166 spin_unlock(&info->lock);
1167
1168 delete_from_swap_cache(page);
1169 set_page_dirty(page);
1170 swap_free(swap);
1171
1172 } else {
1173 if (shmem_acct_block(info->flags)) {
1174 error = -ENOSPC;
1175 goto failed;
1176 }
1177 if (sbinfo->max_blocks) {
1178 if (percpu_counter_compare(&sbinfo->used_blocks,
1179 sbinfo->max_blocks) >= 0) {
1180 error = -ENOSPC;
1181 goto unacct;
1182 }
1183 percpu_counter_inc(&sbinfo->used_blocks);
1184 }
1185
1186 page = shmem_alloc_page(gfp, info, index);
1187 if (!page) {
1188 error = -ENOMEM;
1189 goto decused;
1190 }
1191
1192 SetPageSwapBacked(page);
1193 __set_page_locked(page);
1194 error = mem_cgroup_cache_charge(page, current->mm,
1195 gfp & GFP_RECLAIM_MASK);
1196 if (error)
1197 goto decused;
1198 error = radix_tree_preload(gfp & GFP_RECLAIM_MASK);
1199 if (!error) {
1200 error = shmem_add_to_page_cache(page, mapping, index,
1201 gfp, NULL);
1202 radix_tree_preload_end();
1203 }
1204 if (error) {
1205 mem_cgroup_uncharge_cache_page(page);
1206 goto decused;
1207 }
1208 lru_cache_add_anon(page);
1209
1210 spin_lock(&info->lock);
1211 info->alloced++;
1212 inode->i_blocks += BLOCKS_PER_PAGE;
1213 shmem_recalc_inode(inode);
1214 spin_unlock(&info->lock);
1215 alloced = true;
1216
1217 /*
1218 * Let SGP_FALLOC use the SGP_WRITE optimization on a new page.
1219 */
1220 if (sgp == SGP_FALLOC)
1221 sgp = SGP_WRITE;
1222clear:
1223 /*
1224 * Let SGP_WRITE caller clear ends if write does not fill page;
1225 * but SGP_FALLOC on a page fallocated earlier must initialize
1226 * it now, lest undo on failure cancel our earlier guarantee.
1227 */
1228 if (sgp != SGP_WRITE) {
1229 clear_highpage(page);
1230 flush_dcache_page(page);
1231 SetPageUptodate(page);
1232 }
1233 if (sgp == SGP_DIRTY)
1234 set_page_dirty(page);
1235 }
1236
1237 /* Perhaps the file has been truncated since we checked */
1238 if (sgp != SGP_WRITE && sgp != SGP_FALLOC &&
1239 ((loff_t)index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
1240 error = -EINVAL;
1241 if (alloced)
1242 goto trunc;
1243 else
1244 goto failed;
1245 }
1246 *pagep = page;
1247 return 0;
1248
1249 /*
1250 * Error recovery.
1251 */
1252trunc:
1253 info = SHMEM_I(inode);
1254 ClearPageDirty(page);
1255 delete_from_page_cache(page);
1256 spin_lock(&info->lock);
1257 info->alloced--;
1258 inode->i_blocks -= BLOCKS_PER_PAGE;
1259 spin_unlock(&info->lock);
1260decused:
1261 sbinfo = SHMEM_SB(inode->i_sb);
1262 if (sbinfo->max_blocks)
1263 percpu_counter_add(&sbinfo->used_blocks, -1);
1264unacct:
1265 shmem_unacct_blocks(info->flags, 1);
1266failed:
1267 if (swap.val && error != -EINVAL &&
1268 !shmem_confirm_swap(mapping, index, swap))
1269 error = -EEXIST;
1270unlock:
1271 if (page) {
1272 unlock_page(page);
1273 page_cache_release(page);
1274 }
1275 if (error == -ENOSPC && !once++) {
1276 info = SHMEM_I(inode);
1277 spin_lock(&info->lock);
1278 shmem_recalc_inode(inode);
1279 spin_unlock(&info->lock);
1280 goto repeat;
1281 }
1282 if (error == -EEXIST) /* from above or from radix_tree_insert */
1283 goto repeat;
1284 return error;
1285}
1286
1287static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1288{
1289 struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1290 int error;
1291 int ret = VM_FAULT_LOCKED;
1292
1293 error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_CACHE, &ret);
1294 if (error)
1295 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1296
1297 if (ret & VM_FAULT_MAJOR) {
1298 count_vm_event(PGMAJFAULT);
1299 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1300 }
1301 return ret;
1302}
1303
1304#ifdef CONFIG_NUMA
1305static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
1306{
1307 struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1308 return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
1309}
1310
1311static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
1312 unsigned long addr)
1313{
1314 struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1315 pgoff_t index;
1316
1317 index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1318 return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
1319}
1320#endif
1321
1322int shmem_lock(struct file *file, int lock, struct user_struct *user)
1323{
1324 struct inode *inode = file->f_path.dentry->d_inode;
1325 struct shmem_inode_info *info = SHMEM_I(inode);
1326 int retval = -ENOMEM;
1327
1328 spin_lock(&info->lock);
1329 if (lock && !(info->flags & VM_LOCKED)) {
1330 if (!user_shm_lock(inode->i_size, user))
1331 goto out_nomem;
1332 info->flags |= VM_LOCKED;
1333 mapping_set_unevictable(file->f_mapping);
1334 }
1335 if (!lock && (info->flags & VM_LOCKED) && user) {
1336 user_shm_unlock(inode->i_size, user);
1337 info->flags &= ~VM_LOCKED;
1338 mapping_clear_unevictable(file->f_mapping);
1339 }
1340 retval = 0;
1341
1342out_nomem:
1343 spin_unlock(&info->lock);
1344 return retval;
1345}
1346
1347static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1348{
1349 file_accessed(file);
1350 vma->vm_ops = &shmem_vm_ops;
1351 vma->vm_flags |= VM_CAN_NONLINEAR;
1352 return 0;
1353}
1354
1355static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
1356 umode_t mode, dev_t dev, unsigned long flags)
1357{
1358 struct inode *inode;
1359 struct shmem_inode_info *info;
1360 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1361
1362 if (shmem_reserve_inode(sb))
1363 return NULL;
1364
1365 inode = new_inode(sb);
1366 if (inode) {
1367 inode->i_ino = get_next_ino();
1368 inode_init_owner(inode, dir, mode);
1369 inode->i_blocks = 0;
1370 inode->i_mapping->backing_dev_info = &shmem_backing_dev_info;
1371 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1372 inode->i_generation = get_seconds();
1373 info = SHMEM_I(inode);
1374 memset(info, 0, (char *)inode - (char *)info);
1375 spin_lock_init(&info->lock);
1376 info->flags = flags & VM_NORESERVE;
1377 INIT_LIST_HEAD(&info->swaplist);
1378 INIT_LIST_HEAD(&info->xattr_list);
1379 cache_no_acl(inode);
1380
1381 switch (mode & S_IFMT) {
1382 default:
1383 inode->i_op = &shmem_special_inode_operations;
1384 init_special_inode(inode, mode, dev);
1385 break;
1386 case S_IFREG:
1387 inode->i_mapping->a_ops = &shmem_aops;
1388 inode->i_op = &shmem_inode_operations;
1389 inode->i_fop = &shmem_file_operations;
1390 mpol_shared_policy_init(&info->policy,
1391 shmem_get_sbmpol(sbinfo));
1392 break;
1393 case S_IFDIR:
1394 inc_nlink(inode);
1395 /* Some things misbehave if size == 0 on a directory */
1396 inode->i_size = 2 * BOGO_DIRENT_SIZE;
1397 inode->i_op = &shmem_dir_inode_operations;
1398 inode->i_fop = &simple_dir_operations;
1399 break;
1400 case S_IFLNK:
1401 /*
1402 * Must not load anything in the rbtree,
1403 * mpol_free_shared_policy will not be called.
1404 */
1405 mpol_shared_policy_init(&info->policy, NULL);
1406 break;
1407 }
1408 } else
1409 shmem_free_inode(sb);
1410 return inode;
1411}
1412
1413#ifdef CONFIG_TMPFS
1414static const struct inode_operations shmem_symlink_inode_operations;
1415static const struct inode_operations shmem_short_symlink_operations;
1416
1417#ifdef CONFIG_TMPFS_XATTR
1418static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
1419#else
1420#define shmem_initxattrs NULL
1421#endif
1422
1423static int
1424shmem_write_begin(struct file *file, struct address_space *mapping,
1425 loff_t pos, unsigned len, unsigned flags,
1426 struct page **pagep, void **fsdata)
1427{
1428 struct inode *inode = mapping->host;
1429 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1430 return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1431}
1432
1433static int
1434shmem_write_end(struct file *file, struct address_space *mapping,
1435 loff_t pos, unsigned len, unsigned copied,
1436 struct page *page, void *fsdata)
1437{
1438 struct inode *inode = mapping->host;
1439
1440 if (pos + copied > inode->i_size)
1441 i_size_write(inode, pos + copied);
1442
1443 if (!PageUptodate(page)) {
1444 if (copied < PAGE_CACHE_SIZE) {
1445 unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1446 zero_user_segments(page, 0, from,
1447 from + copied, PAGE_CACHE_SIZE);
1448 }
1449 SetPageUptodate(page);
1450 }
1451 set_page_dirty(page);
1452 unlock_page(page);
1453 page_cache_release(page);
1454
1455 return copied;
1456}
1457
1458static void do_shmem_file_read(struct file *filp, loff_t *ppos, read_descriptor_t *desc, read_actor_t actor)
1459{
1460 struct inode *inode = filp->f_path.dentry->d_inode;
1461 struct address_space *mapping = inode->i_mapping;
1462 pgoff_t index;
1463 unsigned long offset;
1464 enum sgp_type sgp = SGP_READ;
1465
1466 /*
1467 * Might this read be for a stacking filesystem? Then when reading
1468 * holes of a sparse file, we actually need to allocate those pages,
1469 * and even mark them dirty, so it cannot exceed the max_blocks limit.
1470 */
1471 if (segment_eq(get_fs(), KERNEL_DS))
1472 sgp = SGP_DIRTY;
1473
1474 index = *ppos >> PAGE_CACHE_SHIFT;
1475 offset = *ppos & ~PAGE_CACHE_MASK;
1476
1477 for (;;) {
1478 struct page *page = NULL;
1479 pgoff_t end_index;
1480 unsigned long nr, ret;
1481 loff_t i_size = i_size_read(inode);
1482
1483 end_index = i_size >> PAGE_CACHE_SHIFT;
1484 if (index > end_index)
1485 break;
1486 if (index == end_index) {
1487 nr = i_size & ~PAGE_CACHE_MASK;
1488 if (nr <= offset)
1489 break;
1490 }
1491
1492 desc->error = shmem_getpage(inode, index, &page, sgp, NULL);
1493 if (desc->error) {
1494 if (desc->error == -EINVAL)
1495 desc->error = 0;
1496 break;
1497 }
1498 if (page)
1499 unlock_page(page);
1500
1501 /*
1502 * We must evaluate after, since reads (unlike writes)
1503 * are called without i_mutex protection against truncate
1504 */
1505 nr = PAGE_CACHE_SIZE;
1506 i_size = i_size_read(inode);
1507 end_index = i_size >> PAGE_CACHE_SHIFT;
1508 if (index == end_index) {
1509 nr = i_size & ~PAGE_CACHE_MASK;
1510 if (nr <= offset) {
1511 if (page)
1512 page_cache_release(page);
1513 break;
1514 }
1515 }
1516 nr -= offset;
1517
1518 if (page) {
1519 /*
1520 * If users can be writing to this page using arbitrary
1521 * virtual addresses, take care about potential aliasing
1522 * before reading the page on the kernel side.
1523 */
1524 if (mapping_writably_mapped(mapping))
1525 flush_dcache_page(page);
1526 /*
1527 * Mark the page accessed if we read the beginning.
1528 */
1529 if (!offset)
1530 mark_page_accessed(page);
1531 } else {
1532 page = ZERO_PAGE(0);
1533 page_cache_get(page);
1534 }
1535
1536 /*
1537 * Ok, we have the page, and it's up-to-date, so
1538 * now we can copy it to user space...
1539 *
1540 * The actor routine returns how many bytes were actually used..
1541 * NOTE! This may not be the same as how much of a user buffer
1542 * we filled up (we may be padding etc), so we can only update
1543 * "pos" here (the actor routine has to update the user buffer
1544 * pointers and the remaining count).
1545 */
1546 ret = actor(desc, page, offset, nr);
1547 offset += ret;
1548 index += offset >> PAGE_CACHE_SHIFT;
1549 offset &= ~PAGE_CACHE_MASK;
1550
1551 page_cache_release(page);
1552 if (ret != nr || !desc->count)
1553 break;
1554
1555 cond_resched();
1556 }
1557
1558 *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1559 file_accessed(filp);
1560}
1561
1562static ssize_t shmem_file_aio_read(struct kiocb *iocb,
1563 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
1564{
1565 struct file *filp = iocb->ki_filp;
1566 ssize_t retval;
1567 unsigned long seg;
1568 size_t count;
1569 loff_t *ppos = &iocb->ki_pos;
1570
1571 retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1572 if (retval)
1573 return retval;
1574
1575 for (seg = 0; seg < nr_segs; seg++) {
1576 read_descriptor_t desc;
1577
1578 desc.written = 0;
1579 desc.arg.buf = iov[seg].iov_base;
1580 desc.count = iov[seg].iov_len;
1581 if (desc.count == 0)
1582 continue;
1583 desc.error = 0;
1584 do_shmem_file_read(filp, ppos, &desc, file_read_actor);
1585 retval += desc.written;
1586 if (desc.error) {
1587 retval = retval ?: desc.error;
1588 break;
1589 }
1590 if (desc.count > 0)
1591 break;
1592 }
1593 return retval;
1594}
1595
1596static ssize_t shmem_file_splice_read(struct file *in, loff_t *ppos,
1597 struct pipe_inode_info *pipe, size_t len,
1598 unsigned int flags)
1599{
1600 struct address_space *mapping = in->f_mapping;
1601 struct inode *inode = mapping->host;
1602 unsigned int loff, nr_pages, req_pages;
1603 struct page *pages[PIPE_DEF_BUFFERS];
1604 struct partial_page partial[PIPE_DEF_BUFFERS];
1605 struct page *page;
1606 pgoff_t index, end_index;
1607 loff_t isize, left;
1608 int error, page_nr;
1609 struct splice_pipe_desc spd = {
1610 .pages = pages,
1611 .partial = partial,
1612 .nr_pages_max = PIPE_DEF_BUFFERS,
1613 .flags = flags,
1614 .ops = &page_cache_pipe_buf_ops,
1615 .spd_release = spd_release_page,
1616 };
1617
1618 isize = i_size_read(inode);
1619 if (unlikely(*ppos >= isize))
1620 return 0;
1621
1622 left = isize - *ppos;
1623 if (unlikely(left < len))
1624 len = left;
1625
1626 if (splice_grow_spd(pipe, &spd))
1627 return -ENOMEM;
1628
1629 index = *ppos >> PAGE_CACHE_SHIFT;
1630 loff = *ppos & ~PAGE_CACHE_MASK;
1631 req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1632 nr_pages = min(req_pages, pipe->buffers);
1633
1634 spd.nr_pages = find_get_pages_contig(mapping, index,
1635 nr_pages, spd.pages);
1636 index += spd.nr_pages;
1637 error = 0;
1638
1639 while (spd.nr_pages < nr_pages) {
1640 error = shmem_getpage(inode, index, &page, SGP_CACHE, NULL);
1641 if (error)
1642 break;
1643 unlock_page(page);
1644 spd.pages[spd.nr_pages++] = page;
1645 index++;
1646 }
1647
1648 index = *ppos >> PAGE_CACHE_SHIFT;
1649 nr_pages = spd.nr_pages;
1650 spd.nr_pages = 0;
1651
1652 for (page_nr = 0; page_nr < nr_pages; page_nr++) {
1653 unsigned int this_len;
1654
1655 if (!len)
1656 break;
1657
1658 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
1659 page = spd.pages[page_nr];
1660
1661 if (!PageUptodate(page) || page->mapping != mapping) {
1662 error = shmem_getpage(inode, index, &page,
1663 SGP_CACHE, NULL);
1664 if (error)
1665 break;
1666 unlock_page(page);
1667 page_cache_release(spd.pages[page_nr]);
1668 spd.pages[page_nr] = page;
1669 }
1670
1671 isize = i_size_read(inode);
1672 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1673 if (unlikely(!isize || index > end_index))
1674 break;
1675
1676 if (end_index == index) {
1677 unsigned int plen;
1678
1679 plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1680 if (plen <= loff)
1681 break;
1682
1683 this_len = min(this_len, plen - loff);
1684 len = this_len;
1685 }
1686
1687 spd.partial[page_nr].offset = loff;
1688 spd.partial[page_nr].len = this_len;
1689 len -= this_len;
1690 loff = 0;
1691 spd.nr_pages++;
1692 index++;
1693 }
1694
1695 while (page_nr < nr_pages)
1696 page_cache_release(spd.pages[page_nr++]);
1697
1698 if (spd.nr_pages)
1699 error = splice_to_pipe(pipe, &spd);
1700
1701 splice_shrink_spd(&spd);
1702
1703 if (error > 0) {
1704 *ppos += error;
1705 file_accessed(in);
1706 }
1707 return error;
1708}
1709
1710static long shmem_fallocate(struct file *file, int mode, loff_t offset,
1711 loff_t len)
1712{
1713 struct inode *inode = file->f_path.dentry->d_inode;
1714 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1715 struct shmem_falloc shmem_falloc;
1716 pgoff_t start, index, end;
1717 int error;
1718
1719 mutex_lock(&inode->i_mutex);
1720
1721 if (mode & FALLOC_FL_PUNCH_HOLE) {
1722 struct address_space *mapping = file->f_mapping;
1723 loff_t unmap_start = round_up(offset, PAGE_SIZE);
1724 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
1725
1726 if ((u64)unmap_end > (u64)unmap_start)
1727 unmap_mapping_range(mapping, unmap_start,
1728 1 + unmap_end - unmap_start, 0);
1729 shmem_truncate_range(inode, offset, offset + len - 1);
1730 /* No need to unmap again: hole-punching leaves COWed pages */
1731 error = 0;
1732 goto out;
1733 }
1734
1735 /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
1736 error = inode_newsize_ok(inode, offset + len);
1737 if (error)
1738 goto out;
1739
1740 start = offset >> PAGE_CACHE_SHIFT;
1741 end = (offset + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1742 /* Try to avoid a swapstorm if len is impossible to satisfy */
1743 if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
1744 error = -ENOSPC;
1745 goto out;
1746 }
1747
1748 shmem_falloc.start = start;
1749 shmem_falloc.next = start;
1750 shmem_falloc.nr_falloced = 0;
1751 shmem_falloc.nr_unswapped = 0;
1752 spin_lock(&inode->i_lock);
1753 inode->i_private = &shmem_falloc;
1754 spin_unlock(&inode->i_lock);
1755
1756 for (index = start; index < end; index++) {
1757 struct page *page;
1758
1759 /*
1760 * Good, the fallocate(2) manpage permits EINTR: we may have
1761 * been interrupted because we are using up too much memory.
1762 */
1763 if (signal_pending(current))
1764 error = -EINTR;
1765 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
1766 error = -ENOMEM;
1767 else
1768 error = shmem_getpage(inode, index, &page, SGP_FALLOC,
1769 NULL);
1770 if (error) {
1771 /* Remove the !PageUptodate pages we added */
1772 shmem_undo_range(inode,
1773 (loff_t)start << PAGE_CACHE_SHIFT,
1774 (loff_t)index << PAGE_CACHE_SHIFT, true);
1775 goto undone;
1776 }
1777
1778 /*
1779 * Inform shmem_writepage() how far we have reached.
1780 * No need for lock or barrier: we have the page lock.
1781 */
1782 shmem_falloc.next++;
1783 if (!PageUptodate(page))
1784 shmem_falloc.nr_falloced++;
1785
1786 /*
1787 * If !PageUptodate, leave it that way so that freeable pages
1788 * can be recognized if we need to rollback on error later.
1789 * But set_page_dirty so that memory pressure will swap rather
1790 * than free the pages we are allocating (and SGP_CACHE pages
1791 * might still be clean: we now need to mark those dirty too).
1792 */
1793 set_page_dirty(page);
1794 unlock_page(page);
1795 page_cache_release(page);
1796 cond_resched();
1797 }
1798
1799 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
1800 i_size_write(inode, offset + len);
1801 inode->i_ctime = CURRENT_TIME;
1802undone:
1803 spin_lock(&inode->i_lock);
1804 inode->i_private = NULL;
1805 spin_unlock(&inode->i_lock);
1806out:
1807 mutex_unlock(&inode->i_mutex);
1808 return error;
1809}
1810
1811static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
1812{
1813 struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
1814
1815 buf->f_type = TMPFS_MAGIC;
1816 buf->f_bsize = PAGE_CACHE_SIZE;
1817 buf->f_namelen = NAME_MAX;
1818 if (sbinfo->max_blocks) {
1819 buf->f_blocks = sbinfo->max_blocks;
1820 buf->f_bavail =
1821 buf->f_bfree = sbinfo->max_blocks -
1822 percpu_counter_sum(&sbinfo->used_blocks);
1823 }
1824 if (sbinfo->max_inodes) {
1825 buf->f_files = sbinfo->max_inodes;
1826 buf->f_ffree = sbinfo->free_inodes;
1827 }
1828 /* else leave those fields 0 like simple_statfs */
1829 return 0;
1830}
1831
1832/*
1833 * File creation. Allocate an inode, and we're done..
1834 */
1835static int
1836shmem_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
1837{
1838 struct inode *inode;
1839 int error = -ENOSPC;
1840
1841 inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
1842 if (inode) {
1843 error = security_inode_init_security(inode, dir,
1844 &dentry->d_name,
1845 shmem_initxattrs, NULL);
1846 if (error) {
1847 if (error != -EOPNOTSUPP) {
1848 iput(inode);
1849 return error;
1850 }
1851 }
1852#ifdef CONFIG_TMPFS_POSIX_ACL
1853 error = generic_acl_init(inode, dir);
1854 if (error) {
1855 iput(inode);
1856 return error;
1857 }
1858#else
1859 error = 0;
1860#endif
1861 dir->i_size += BOGO_DIRENT_SIZE;
1862 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1863 d_instantiate(dentry, inode);
1864 dget(dentry); /* Extra count - pin the dentry in core */
1865 }
1866 return error;
1867}
1868
1869static int shmem_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
1870{
1871 int error;
1872
1873 if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
1874 return error;
1875 inc_nlink(dir);
1876 return 0;
1877}
1878
1879static int shmem_create(struct inode *dir, struct dentry *dentry, umode_t mode,
1880 struct nameidata *nd)
1881{
1882 return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
1883}
1884
1885/*
1886 * Link a file..
1887 */
1888static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
1889{
1890 struct inode *inode = old_dentry->d_inode;
1891 int ret;
1892
1893 /*
1894 * No ordinary (disk based) filesystem counts links as inodes;
1895 * but each new link needs a new dentry, pinning lowmem, and
1896 * tmpfs dentries cannot be pruned until they are unlinked.
1897 */
1898 ret = shmem_reserve_inode(inode->i_sb);
1899 if (ret)
1900 goto out;
1901
1902 dir->i_size += BOGO_DIRENT_SIZE;
1903 inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1904 inc_nlink(inode);
1905 ihold(inode); /* New dentry reference */
1906 dget(dentry); /* Extra pinning count for the created dentry */
1907 d_instantiate(dentry, inode);
1908out:
1909 return ret;
1910}
1911
1912static int shmem_unlink(struct inode *dir, struct dentry *dentry)
1913{
1914 struct inode *inode = dentry->d_inode;
1915
1916 if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
1917 shmem_free_inode(inode->i_sb);
1918
1919 dir->i_size -= BOGO_DIRENT_SIZE;
1920 inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1921 drop_nlink(inode);
1922 dput(dentry); /* Undo the count from "create" - this does all the work */
1923 return 0;
1924}
1925
1926static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
1927{
1928 if (!simple_empty(dentry))
1929 return -ENOTEMPTY;
1930
1931 drop_nlink(dentry->d_inode);
1932 drop_nlink(dir);
1933 return shmem_unlink(dir, dentry);
1934}
1935
1936/*
1937 * The VFS layer already does all the dentry stuff for rename,
1938 * we just have to decrement the usage count for the target if
1939 * it exists so that the VFS layer correctly free's it when it
1940 * gets overwritten.
1941 */
1942static int shmem_rename(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
1943{
1944 struct inode *inode = old_dentry->d_inode;
1945 int they_are_dirs = S_ISDIR(inode->i_mode);
1946
1947 if (!simple_empty(new_dentry))
1948 return -ENOTEMPTY;
1949
1950 if (new_dentry->d_inode) {
1951 (void) shmem_unlink(new_dir, new_dentry);
1952 if (they_are_dirs)
1953 drop_nlink(old_dir);
1954 } else if (they_are_dirs) {
1955 drop_nlink(old_dir);
1956 inc_nlink(new_dir);
1957 }
1958
1959 old_dir->i_size -= BOGO_DIRENT_SIZE;
1960 new_dir->i_size += BOGO_DIRENT_SIZE;
1961 old_dir->i_ctime = old_dir->i_mtime =
1962 new_dir->i_ctime = new_dir->i_mtime =
1963 inode->i_ctime = CURRENT_TIME;
1964 return 0;
1965}
1966
1967static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
1968{
1969 int error;
1970 int len;
1971 struct inode *inode;
1972 struct page *page;
1973 char *kaddr;
1974 struct shmem_inode_info *info;
1975
1976 len = strlen(symname) + 1;
1977 if (len > PAGE_CACHE_SIZE)
1978 return -ENAMETOOLONG;
1979
1980 inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0, VM_NORESERVE);
1981 if (!inode)
1982 return -ENOSPC;
1983
1984 error = security_inode_init_security(inode, dir, &dentry->d_name,
1985 shmem_initxattrs, NULL);
1986 if (error) {
1987 if (error != -EOPNOTSUPP) {
1988 iput(inode);
1989 return error;
1990 }
1991 error = 0;
1992 }
1993
1994 info = SHMEM_I(inode);
1995 inode->i_size = len-1;
1996 if (len <= SHORT_SYMLINK_LEN) {
1997 info->symlink = kmemdup(symname, len, GFP_KERNEL);
1998 if (!info->symlink) {
1999 iput(inode);
2000 return -ENOMEM;
2001 }
2002 inode->i_op = &shmem_short_symlink_operations;
2003 } else {
2004 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
2005 if (error) {
2006 iput(inode);
2007 return error;
2008 }
2009 inode->i_mapping->a_ops = &shmem_aops;
2010 inode->i_op = &shmem_symlink_inode_operations;
2011 kaddr = kmap_atomic(page);
2012 memcpy(kaddr, symname, len);
2013 kunmap_atomic(kaddr);
2014 SetPageUptodate(page);
2015 set_page_dirty(page);
2016 unlock_page(page);
2017 page_cache_release(page);
2018 }
2019 dir->i_size += BOGO_DIRENT_SIZE;
2020 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2021 d_instantiate(dentry, inode);
2022 dget(dentry);
2023 return 0;
2024}
2025
2026static void *shmem_follow_short_symlink(struct dentry *dentry, struct nameidata *nd)
2027{
2028 nd_set_link(nd, SHMEM_I(dentry->d_inode)->symlink);
2029 return NULL;
2030}
2031
2032static void *shmem_follow_link(struct dentry *dentry, struct nameidata *nd)
2033{
2034 struct page *page = NULL;
2035 int error = shmem_getpage(dentry->d_inode, 0, &page, SGP_READ, NULL);
2036 nd_set_link(nd, error ? ERR_PTR(error) : kmap(page));
2037 if (page)
2038 unlock_page(page);
2039 return page;
2040}
2041
2042static void shmem_put_link(struct dentry *dentry, struct nameidata *nd, void *cookie)
2043{
2044 if (!IS_ERR(nd_get_link(nd))) {
2045 struct page *page = cookie;
2046 kunmap(page);
2047 mark_page_accessed(page);
2048 page_cache_release(page);
2049 }
2050}
2051
2052#ifdef CONFIG_TMPFS_XATTR
2053/*
2054 * Superblocks without xattr inode operations may get some security.* xattr
2055 * support from the LSM "for free". As soon as we have any other xattrs
2056 * like ACLs, we also need to implement the security.* handlers at
2057 * filesystem level, though.
2058 */
2059
2060/*
2061 * Allocate new xattr and copy in the value; but leave the name to callers.
2062 */
2063static struct shmem_xattr *shmem_xattr_alloc(const void *value, size_t size)
2064{
2065 struct shmem_xattr *new_xattr;
2066 size_t len;
2067
2068 /* wrap around? */
2069 len = sizeof(*new_xattr) + size;
2070 if (len <= sizeof(*new_xattr))
2071 return NULL;
2072
2073 new_xattr = kmalloc(len, GFP_KERNEL);
2074 if (!new_xattr)
2075 return NULL;
2076
2077 new_xattr->size = size;
2078 memcpy(new_xattr->value, value, size);
2079 return new_xattr;
2080}
2081
2082/*
2083 * Callback for security_inode_init_security() for acquiring xattrs.
2084 */
2085static int shmem_initxattrs(struct inode *inode,
2086 const struct xattr *xattr_array,
2087 void *fs_info)
2088{
2089 struct shmem_inode_info *info = SHMEM_I(inode);
2090 const struct xattr *xattr;
2091 struct shmem_xattr *new_xattr;
2092 size_t len;
2093
2094 for (xattr = xattr_array; xattr->name != NULL; xattr++) {
2095 new_xattr = shmem_xattr_alloc(xattr->value, xattr->value_len);
2096 if (!new_xattr)
2097 return -ENOMEM;
2098
2099 len = strlen(xattr->name) + 1;
2100 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
2101 GFP_KERNEL);
2102 if (!new_xattr->name) {
2103 kfree(new_xattr);
2104 return -ENOMEM;
2105 }
2106
2107 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
2108 XATTR_SECURITY_PREFIX_LEN);
2109 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
2110 xattr->name, len);
2111
2112 spin_lock(&info->lock);
2113 list_add(&new_xattr->list, &info->xattr_list);
2114 spin_unlock(&info->lock);
2115 }
2116
2117 return 0;
2118}
2119
2120static int shmem_xattr_get(struct dentry *dentry, const char *name,
2121 void *buffer, size_t size)
2122{
2123 struct shmem_inode_info *info;
2124 struct shmem_xattr *xattr;
2125 int ret = -ENODATA;
2126
2127 info = SHMEM_I(dentry->d_inode);
2128
2129 spin_lock(&info->lock);
2130 list_for_each_entry(xattr, &info->xattr_list, list) {
2131 if (strcmp(name, xattr->name))
2132 continue;
2133
2134 ret = xattr->size;
2135 if (buffer) {
2136 if (size < xattr->size)
2137 ret = -ERANGE;
2138 else
2139 memcpy(buffer, xattr->value, xattr->size);
2140 }
2141 break;
2142 }
2143 spin_unlock(&info->lock);
2144 return ret;
2145}
2146
2147static int shmem_xattr_set(struct inode *inode, const char *name,
2148 const void *value, size_t size, int flags)
2149{
2150 struct shmem_inode_info *info = SHMEM_I(inode);
2151 struct shmem_xattr *xattr;
2152 struct shmem_xattr *new_xattr = NULL;
2153 int err = 0;
2154
2155 /* value == NULL means remove */
2156 if (value) {
2157 new_xattr = shmem_xattr_alloc(value, size);
2158 if (!new_xattr)
2159 return -ENOMEM;
2160
2161 new_xattr->name = kstrdup(name, GFP_KERNEL);
2162 if (!new_xattr->name) {
2163 kfree(new_xattr);
2164 return -ENOMEM;
2165 }
2166 }
2167
2168 spin_lock(&info->lock);
2169 list_for_each_entry(xattr, &info->xattr_list, list) {
2170 if (!strcmp(name, xattr->name)) {
2171 if (flags & XATTR_CREATE) {
2172 xattr = new_xattr;
2173 err = -EEXIST;
2174 } else if (new_xattr) {
2175 list_replace(&xattr->list, &new_xattr->list);
2176 } else {
2177 list_del(&xattr->list);
2178 }
2179 goto out;
2180 }
2181 }
2182 if (flags & XATTR_REPLACE) {
2183 xattr = new_xattr;
2184 err = -ENODATA;
2185 } else {
2186 list_add(&new_xattr->list, &info->xattr_list);
2187 xattr = NULL;
2188 }
2189out:
2190 spin_unlock(&info->lock);
2191 if (xattr)
2192 kfree(xattr->name);
2193 kfree(xattr);
2194 return err;
2195}
2196
2197static const struct xattr_handler *shmem_xattr_handlers[] = {
2198#ifdef CONFIG_TMPFS_POSIX_ACL
2199 &generic_acl_access_handler,
2200 &generic_acl_default_handler,
2201#endif
2202 NULL
2203};
2204
2205static int shmem_xattr_validate(const char *name)
2206{
2207 struct { const char *prefix; size_t len; } arr[] = {
2208 { XATTR_SECURITY_PREFIX, XATTR_SECURITY_PREFIX_LEN },
2209 { XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN }
2210 };
2211 int i;
2212
2213 for (i = 0; i < ARRAY_SIZE(arr); i++) {
2214 size_t preflen = arr[i].len;
2215 if (strncmp(name, arr[i].prefix, preflen) == 0) {
2216 if (!name[preflen])
2217 return -EINVAL;
2218 return 0;
2219 }
2220 }
2221 return -EOPNOTSUPP;
2222}
2223
2224static ssize_t shmem_getxattr(struct dentry *dentry, const char *name,
2225 void *buffer, size_t size)
2226{
2227 int err;
2228
2229 /*
2230 * If this is a request for a synthetic attribute in the system.*
2231 * namespace use the generic infrastructure to resolve a handler
2232 * for it via sb->s_xattr.
2233 */
2234 if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2235 return generic_getxattr(dentry, name, buffer, size);
2236
2237 err = shmem_xattr_validate(name);
2238 if (err)
2239 return err;
2240
2241 return shmem_xattr_get(dentry, name, buffer, size);
2242}
2243
2244static int shmem_setxattr(struct dentry *dentry, const char *name,
2245 const void *value, size_t size, int flags)
2246{
2247 int err;
2248
2249 /*
2250 * If this is a request for a synthetic attribute in the system.*
2251 * namespace use the generic infrastructure to resolve a handler
2252 * for it via sb->s_xattr.
2253 */
2254 if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2255 return generic_setxattr(dentry, name, value, size, flags);
2256
2257 err = shmem_xattr_validate(name);
2258 if (err)
2259 return err;
2260
2261 if (size == 0)
2262 value = ""; /* empty EA, do not remove */
2263
2264 return shmem_xattr_set(dentry->d_inode, name, value, size, flags);
2265
2266}
2267
2268static int shmem_removexattr(struct dentry *dentry, const char *name)
2269{
2270 int err;
2271
2272 /*
2273 * If this is a request for a synthetic attribute in the system.*
2274 * namespace use the generic infrastructure to resolve a handler
2275 * for it via sb->s_xattr.
2276 */
2277 if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2278 return generic_removexattr(dentry, name);
2279
2280 err = shmem_xattr_validate(name);
2281 if (err)
2282 return err;
2283
2284 return shmem_xattr_set(dentry->d_inode, name, NULL, 0, XATTR_REPLACE);
2285}
2286
2287static bool xattr_is_trusted(const char *name)
2288{
2289 return !strncmp(name, XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN);
2290}
2291
2292static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
2293{
2294 bool trusted = capable(CAP_SYS_ADMIN);
2295 struct shmem_xattr *xattr;
2296 struct shmem_inode_info *info;
2297 size_t used = 0;
2298
2299 info = SHMEM_I(dentry->d_inode);
2300
2301 spin_lock(&info->lock);
2302 list_for_each_entry(xattr, &info->xattr_list, list) {
2303 size_t len;
2304
2305 /* skip "trusted." attributes for unprivileged callers */
2306 if (!trusted && xattr_is_trusted(xattr->name))
2307 continue;
2308
2309 len = strlen(xattr->name) + 1;
2310 used += len;
2311 if (buffer) {
2312 if (size < used) {
2313 used = -ERANGE;
2314 break;
2315 }
2316 memcpy(buffer, xattr->name, len);
2317 buffer += len;
2318 }
2319 }
2320 spin_unlock(&info->lock);
2321
2322 return used;
2323}
2324#endif /* CONFIG_TMPFS_XATTR */
2325
2326static const struct inode_operations shmem_short_symlink_operations = {
2327 .readlink = generic_readlink,
2328 .follow_link = shmem_follow_short_symlink,
2329#ifdef CONFIG_TMPFS_XATTR
2330 .setxattr = shmem_setxattr,
2331 .getxattr = shmem_getxattr,
2332 .listxattr = shmem_listxattr,
2333 .removexattr = shmem_removexattr,
2334#endif
2335};
2336
2337static const struct inode_operations shmem_symlink_inode_operations = {
2338 .readlink = generic_readlink,
2339 .follow_link = shmem_follow_link,
2340 .put_link = shmem_put_link,
2341#ifdef CONFIG_TMPFS_XATTR
2342 .setxattr = shmem_setxattr,
2343 .getxattr = shmem_getxattr,
2344 .listxattr = shmem_listxattr,
2345 .removexattr = shmem_removexattr,
2346#endif
2347};
2348
2349static struct dentry *shmem_get_parent(struct dentry *child)
2350{
2351 return ERR_PTR(-ESTALE);
2352}
2353
2354static int shmem_match(struct inode *ino, void *vfh)
2355{
2356 __u32 *fh = vfh;
2357 __u64 inum = fh[2];
2358 inum = (inum << 32) | fh[1];
2359 return ino->i_ino == inum && fh[0] == ino->i_generation;
2360}
2361
2362static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
2363 struct fid *fid, int fh_len, int fh_type)
2364{
2365 struct inode *inode;
2366 struct dentry *dentry = NULL;
2367 u64 inum = fid->raw[2];
2368 inum = (inum << 32) | fid->raw[1];
2369
2370 if (fh_len < 3)
2371 return NULL;
2372
2373 inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
2374 shmem_match, fid->raw);
2375 if (inode) {
2376 dentry = d_find_alias(inode);
2377 iput(inode);
2378 }
2379
2380 return dentry;
2381}
2382
2383static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
2384 struct inode *parent)
2385{
2386 if (*len < 3) {
2387 *len = 3;
2388 return 255;
2389 }
2390
2391 if (inode_unhashed(inode)) {
2392 /* Unfortunately insert_inode_hash is not idempotent,
2393 * so as we hash inodes here rather than at creation
2394 * time, we need a lock to ensure we only try
2395 * to do it once
2396 */
2397 static DEFINE_SPINLOCK(lock);
2398 spin_lock(&lock);
2399 if (inode_unhashed(inode))
2400 __insert_inode_hash(inode,
2401 inode->i_ino + inode->i_generation);
2402 spin_unlock(&lock);
2403 }
2404
2405 fh[0] = inode->i_generation;
2406 fh[1] = inode->i_ino;
2407 fh[2] = ((__u64)inode->i_ino) >> 32;
2408
2409 *len = 3;
2410 return 1;
2411}
2412
2413static const struct export_operations shmem_export_ops = {
2414 .get_parent = shmem_get_parent,
2415 .encode_fh = shmem_encode_fh,
2416 .fh_to_dentry = shmem_fh_to_dentry,
2417};
2418
2419static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
2420 bool remount)
2421{
2422 char *this_char, *value, *rest;
2423 uid_t uid;
2424 gid_t gid;
2425
2426 while (options != NULL) {
2427 this_char = options;
2428 for (;;) {
2429 /*
2430 * NUL-terminate this option: unfortunately,
2431 * mount options form a comma-separated list,
2432 * but mpol's nodelist may also contain commas.
2433 */
2434 options = strchr(options, ',');
2435 if (options == NULL)
2436 break;
2437 options++;
2438 if (!isdigit(*options)) {
2439 options[-1] = '\0';
2440 break;
2441 }
2442 }
2443 if (!*this_char)
2444 continue;
2445 if ((value = strchr(this_char,'=')) != NULL) {
2446 *value++ = 0;
2447 } else {
2448 printk(KERN_ERR
2449 "tmpfs: No value for mount option '%s'\n",
2450 this_char);
2451 return 1;
2452 }
2453
2454 if (!strcmp(this_char,"size")) {
2455 unsigned long long size;
2456 size = memparse(value,&rest);
2457 if (*rest == '%') {
2458 size <<= PAGE_SHIFT;
2459 size *= totalram_pages;
2460 do_div(size, 100);
2461 rest++;
2462 }
2463 if (*rest)
2464 goto bad_val;
2465 sbinfo->max_blocks =
2466 DIV_ROUND_UP(size, PAGE_CACHE_SIZE);
2467 } else if (!strcmp(this_char,"nr_blocks")) {
2468 sbinfo->max_blocks = memparse(value, &rest);
2469 if (*rest)
2470 goto bad_val;
2471 } else if (!strcmp(this_char,"nr_inodes")) {
2472 sbinfo->max_inodes = memparse(value, &rest);
2473 if (*rest)
2474 goto bad_val;
2475 } else if (!strcmp(this_char,"mode")) {
2476 if (remount)
2477 continue;
2478 sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
2479 if (*rest)
2480 goto bad_val;
2481 } else if (!strcmp(this_char,"uid")) {
2482 if (remount)
2483 continue;
2484 uid = simple_strtoul(value, &rest, 0);
2485 if (*rest)
2486 goto bad_val;
2487 sbinfo->uid = make_kuid(current_user_ns(), uid);
2488 if (!uid_valid(sbinfo->uid))
2489 goto bad_val;
2490 } else if (!strcmp(this_char,"gid")) {
2491 if (remount)
2492 continue;
2493 gid = simple_strtoul(value, &rest, 0);
2494 if (*rest)
2495 goto bad_val;
2496 sbinfo->gid = make_kgid(current_user_ns(), gid);
2497 if (!gid_valid(sbinfo->gid))
2498 goto bad_val;
2499 } else if (!strcmp(this_char,"mpol")) {
2500 if (mpol_parse_str(value, &sbinfo->mpol, 1))
2501 goto bad_val;
2502 } else {
2503 printk(KERN_ERR "tmpfs: Bad mount option %s\n",
2504 this_char);
2505 return 1;
2506 }
2507 }
2508 return 0;
2509
2510bad_val:
2511 printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
2512 value, this_char);
2513 return 1;
2514
2515}
2516
2517static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2518{
2519 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2520 struct shmem_sb_info config = *sbinfo;
2521 unsigned long inodes;
2522 int error = -EINVAL;
2523
2524 if (shmem_parse_options(data, &config, true))
2525 return error;
2526
2527 spin_lock(&sbinfo->stat_lock);
2528 inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2529 if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
2530 goto out;
2531 if (config.max_inodes < inodes)
2532 goto out;
2533 /*
2534 * Those tests disallow limited->unlimited while any are in use;
2535 * but we must separately disallow unlimited->limited, because
2536 * in that case we have no record of how much is already in use.
2537 */
2538 if (config.max_blocks && !sbinfo->max_blocks)
2539 goto out;
2540 if (config.max_inodes && !sbinfo->max_inodes)
2541 goto out;
2542
2543 error = 0;
2544 sbinfo->max_blocks = config.max_blocks;
2545 sbinfo->max_inodes = config.max_inodes;
2546 sbinfo->free_inodes = config.max_inodes - inodes;
2547
2548 mpol_put(sbinfo->mpol);
2549 sbinfo->mpol = config.mpol; /* transfers initial ref */
2550out:
2551 spin_unlock(&sbinfo->stat_lock);
2552 return error;
2553}
2554
2555static int shmem_show_options(struct seq_file *seq, struct dentry *root)
2556{
2557 struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
2558
2559 if (sbinfo->max_blocks != shmem_default_max_blocks())
2560 seq_printf(seq, ",size=%luk",
2561 sbinfo->max_blocks << (PAGE_CACHE_SHIFT - 10));
2562 if (sbinfo->max_inodes != shmem_default_max_inodes())
2563 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
2564 if (sbinfo->mode != (S_IRWXUGO | S_ISVTX))
2565 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
2566 if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
2567 seq_printf(seq, ",uid=%u",
2568 from_kuid_munged(&init_user_ns, sbinfo->uid));
2569 if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
2570 seq_printf(seq, ",gid=%u",
2571 from_kgid_munged(&init_user_ns, sbinfo->gid));
2572 shmem_show_mpol(seq, sbinfo->mpol);
2573 return 0;
2574}
2575#endif /* CONFIG_TMPFS */
2576
2577static void shmem_put_super(struct super_block *sb)
2578{
2579 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2580
2581 percpu_counter_destroy(&sbinfo->used_blocks);
2582 kfree(sbinfo);
2583 sb->s_fs_info = NULL;
2584}
2585
2586int shmem_fill_super(struct super_block *sb, void *data, int silent)
2587{
2588 struct inode *inode;
2589 struct shmem_sb_info *sbinfo;
2590 int err = -ENOMEM;
2591
2592 /* Round up to L1_CACHE_BYTES to resist false sharing */
2593 sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
2594 L1_CACHE_BYTES), GFP_KERNEL);
2595 if (!sbinfo)
2596 return -ENOMEM;
2597
2598 sbinfo->mode = S_IRWXUGO | S_ISVTX;
2599 sbinfo->uid = current_fsuid();
2600 sbinfo->gid = current_fsgid();
2601 sb->s_fs_info = sbinfo;
2602
2603#ifdef CONFIG_TMPFS
2604 /*
2605 * Per default we only allow half of the physical ram per
2606 * tmpfs instance, limiting inodes to one per page of lowmem;
2607 * but the internal instance is left unlimited.
2608 */
2609 if (!(sb->s_flags & MS_NOUSER)) {
2610 sbinfo->max_blocks = shmem_default_max_blocks();
2611 sbinfo->max_inodes = shmem_default_max_inodes();
2612 if (shmem_parse_options(data, sbinfo, false)) {
2613 err = -EINVAL;
2614 goto failed;
2615 }
2616 }
2617 sb->s_export_op = &shmem_export_ops;
2618 sb->s_flags |= MS_NOSEC;
2619#else
2620 sb->s_flags |= MS_NOUSER;
2621#endif
2622
2623 spin_lock_init(&sbinfo->stat_lock);
2624 if (percpu_counter_init(&sbinfo->used_blocks, 0))
2625 goto failed;
2626 sbinfo->free_inodes = sbinfo->max_inodes;
2627
2628 sb->s_maxbytes = MAX_LFS_FILESIZE;
2629 sb->s_blocksize = PAGE_CACHE_SIZE;
2630 sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
2631 sb->s_magic = TMPFS_MAGIC;
2632 sb->s_op = &shmem_ops;
2633 sb->s_time_gran = 1;
2634#ifdef CONFIG_TMPFS_XATTR
2635 sb->s_xattr = shmem_xattr_handlers;
2636#endif
2637#ifdef CONFIG_TMPFS_POSIX_ACL
2638 sb->s_flags |= MS_POSIXACL;
2639#endif
2640
2641 inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
2642 if (!inode)
2643 goto failed;
2644 inode->i_uid = sbinfo->uid;
2645 inode->i_gid = sbinfo->gid;
2646 sb->s_root = d_make_root(inode);
2647 if (!sb->s_root)
2648 goto failed;
2649 return 0;
2650
2651failed:
2652 shmem_put_super(sb);
2653 return err;
2654}
2655
2656static struct kmem_cache *shmem_inode_cachep;
2657
2658static struct inode *shmem_alloc_inode(struct super_block *sb)
2659{
2660 struct shmem_inode_info *info;
2661 info = kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
2662 if (!info)
2663 return NULL;
2664 return &info->vfs_inode;
2665}
2666
2667static void shmem_destroy_callback(struct rcu_head *head)
2668{
2669 struct inode *inode = container_of(head, struct inode, i_rcu);
2670 kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
2671}
2672
2673static void shmem_destroy_inode(struct inode *inode)
2674{
2675 if (S_ISREG(inode->i_mode))
2676 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
2677 call_rcu(&inode->i_rcu, shmem_destroy_callback);
2678}
2679
2680static void shmem_init_inode(void *foo)
2681{
2682 struct shmem_inode_info *info = foo;
2683 inode_init_once(&info->vfs_inode);
2684}
2685
2686static int shmem_init_inodecache(void)
2687{
2688 shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
2689 sizeof(struct shmem_inode_info),
2690 0, SLAB_PANIC, shmem_init_inode);
2691 return 0;
2692}
2693
2694static void shmem_destroy_inodecache(void)
2695{
2696 kmem_cache_destroy(shmem_inode_cachep);
2697}
2698
2699static const struct address_space_operations shmem_aops = {
2700 .writepage = shmem_writepage,
2701 .set_page_dirty = __set_page_dirty_no_writeback,
2702#ifdef CONFIG_TMPFS
2703 .write_begin = shmem_write_begin,
2704 .write_end = shmem_write_end,
2705#endif
2706 .migratepage = migrate_page,
2707 .error_remove_page = generic_error_remove_page,
2708};
2709
2710static const struct file_operations shmem_file_operations = {
2711 .mmap = shmem_mmap,
2712#ifdef CONFIG_TMPFS
2713 .llseek = generic_file_llseek,
2714 .read = do_sync_read,
2715 .write = do_sync_write,
2716 .aio_read = shmem_file_aio_read,
2717 .aio_write = generic_file_aio_write,
2718 .fsync = noop_fsync,
2719 .splice_read = shmem_file_splice_read,
2720 .splice_write = generic_file_splice_write,
2721 .fallocate = shmem_fallocate,
2722#endif
2723};
2724
2725static const struct inode_operations shmem_inode_operations = {
2726 .setattr = shmem_setattr,
2727#ifdef CONFIG_TMPFS_XATTR
2728 .setxattr = shmem_setxattr,
2729 .getxattr = shmem_getxattr,
2730 .listxattr = shmem_listxattr,
2731 .removexattr = shmem_removexattr,
2732#endif
2733};
2734
2735static const struct inode_operations shmem_dir_inode_operations = {
2736#ifdef CONFIG_TMPFS
2737 .create = shmem_create,
2738 .lookup = simple_lookup,
2739 .link = shmem_link,
2740 .unlink = shmem_unlink,
2741 .symlink = shmem_symlink,
2742 .mkdir = shmem_mkdir,
2743 .rmdir = shmem_rmdir,
2744 .mknod = shmem_mknod,
2745 .rename = shmem_rename,
2746#endif
2747#ifdef CONFIG_TMPFS_XATTR
2748 .setxattr = shmem_setxattr,
2749 .getxattr = shmem_getxattr,
2750 .listxattr = shmem_listxattr,
2751 .removexattr = shmem_removexattr,
2752#endif
2753#ifdef CONFIG_TMPFS_POSIX_ACL
2754 .setattr = shmem_setattr,
2755#endif
2756};
2757
2758static const struct inode_operations shmem_special_inode_operations = {
2759#ifdef CONFIG_TMPFS_XATTR
2760 .setxattr = shmem_setxattr,
2761 .getxattr = shmem_getxattr,
2762 .listxattr = shmem_listxattr,
2763 .removexattr = shmem_removexattr,
2764#endif
2765#ifdef CONFIG_TMPFS_POSIX_ACL
2766 .setattr = shmem_setattr,
2767#endif
2768};
2769
2770static const struct super_operations shmem_ops = {
2771 .alloc_inode = shmem_alloc_inode,
2772 .destroy_inode = shmem_destroy_inode,
2773#ifdef CONFIG_TMPFS
2774 .statfs = shmem_statfs,
2775 .remount_fs = shmem_remount_fs,
2776 .show_options = shmem_show_options,
2777#endif
2778 .evict_inode = shmem_evict_inode,
2779 .drop_inode = generic_delete_inode,
2780 .put_super = shmem_put_super,
2781};
2782
2783static const struct vm_operations_struct shmem_vm_ops = {
2784 .fault = shmem_fault,
2785#ifdef CONFIG_NUMA
2786 .set_policy = shmem_set_policy,
2787 .get_policy = shmem_get_policy,
2788#endif
2789};
2790
2791static struct dentry *shmem_mount(struct file_system_type *fs_type,
2792 int flags, const char *dev_name, void *data)
2793{
2794 return mount_nodev(fs_type, flags, data, shmem_fill_super);
2795}
2796
2797static struct file_system_type shmem_fs_type = {
2798 .owner = THIS_MODULE,
2799 .name = "tmpfs",
2800 .mount = shmem_mount,
2801 .kill_sb = kill_litter_super,
2802};
2803
2804int __init shmem_init(void)
2805{
2806 int error;
2807
2808 error = bdi_init(&shmem_backing_dev_info);
2809 if (error)
2810 goto out4;
2811
2812 error = shmem_init_inodecache();
2813 if (error)
2814 goto out3;
2815
2816 error = register_filesystem(&shmem_fs_type);
2817 if (error) {
2818 printk(KERN_ERR "Could not register tmpfs\n");
2819 goto out2;
2820 }
2821
2822 shm_mnt = vfs_kern_mount(&shmem_fs_type, MS_NOUSER,
2823 shmem_fs_type.name, NULL);
2824 if (IS_ERR(shm_mnt)) {
2825 error = PTR_ERR(shm_mnt);
2826 printk(KERN_ERR "Could not kern_mount tmpfs\n");
2827 goto out1;
2828 }
2829 return 0;
2830
2831out1:
2832 unregister_filesystem(&shmem_fs_type);
2833out2:
2834 shmem_destroy_inodecache();
2835out3:
2836 bdi_destroy(&shmem_backing_dev_info);
2837out4:
2838 shm_mnt = ERR_PTR(error);
2839 return error;
2840}
2841
2842#else /* !CONFIG_SHMEM */
2843
2844/*
2845 * tiny-shmem: simple shmemfs and tmpfs using ramfs code
2846 *
2847 * This is intended for small system where the benefits of the full
2848 * shmem code (swap-backed and resource-limited) are outweighed by
2849 * their complexity. On systems without swap this code should be
2850 * effectively equivalent, but much lighter weight.
2851 */
2852
2853#include <linux/ramfs.h>
2854
2855static struct file_system_type shmem_fs_type = {
2856 .name = "tmpfs",
2857 .mount = ramfs_mount,
2858 .kill_sb = kill_litter_super,
2859};
2860
2861int __init shmem_init(void)
2862{
2863 BUG_ON(register_filesystem(&shmem_fs_type) != 0);
2864
2865 shm_mnt = kern_mount(&shmem_fs_type);
2866 BUG_ON(IS_ERR(shm_mnt));
2867
2868 return 0;
2869}
2870
2871int shmem_unuse(swp_entry_t swap, struct page *page)
2872{
2873 return 0;
2874}
2875
2876int shmem_lock(struct file *file, int lock, struct user_struct *user)
2877{
2878 return 0;
2879}
2880
2881void shmem_unlock_mapping(struct address_space *mapping)
2882{
2883}
2884
2885void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
2886{
2887 truncate_inode_pages_range(inode->i_mapping, lstart, lend);
2888}
2889EXPORT_SYMBOL_GPL(shmem_truncate_range);
2890
2891#define shmem_vm_ops generic_file_vm_ops
2892#define shmem_file_operations ramfs_file_operations
2893#define shmem_get_inode(sb, dir, mode, dev, flags) ramfs_get_inode(sb, dir, mode, dev)
2894#define shmem_acct_size(flags, size) 0
2895#define shmem_unacct_size(flags, size) do {} while (0)
2896
2897#endif /* CONFIG_SHMEM */
2898
2899/* common code */
2900
2901/**
2902 * shmem_file_setup - get an unlinked file living in tmpfs
2903 * @name: name for dentry (to be seen in /proc/<pid>/maps
2904 * @size: size to be set for the file
2905 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
2906 */
2907struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
2908{
2909 int error;
2910 struct file *file;
2911 struct inode *inode;
2912 struct path path;
2913 struct dentry *root;
2914 struct qstr this;
2915
2916 if (IS_ERR(shm_mnt))
2917 return (void *)shm_mnt;
2918
2919 if (size < 0 || size > MAX_LFS_FILESIZE)
2920 return ERR_PTR(-EINVAL);
2921
2922 if (shmem_acct_size(flags, size))
2923 return ERR_PTR(-ENOMEM);
2924
2925 error = -ENOMEM;
2926 this.name = name;
2927 this.len = strlen(name);
2928 this.hash = 0; /* will go */
2929 root = shm_mnt->mnt_root;
2930 path.dentry = d_alloc(root, &this);
2931 if (!path.dentry)
2932 goto put_memory;
2933 path.mnt = mntget(shm_mnt);
2934
2935 error = -ENOSPC;
2936 inode = shmem_get_inode(root->d_sb, NULL, S_IFREG | S_IRWXUGO, 0, flags);
2937 if (!inode)
2938 goto put_dentry;
2939
2940 d_instantiate(path.dentry, inode);
2941 inode->i_size = size;
2942 clear_nlink(inode); /* It is unlinked */
2943#ifndef CONFIG_MMU
2944 error = ramfs_nommu_expand_for_mapping(inode, size);
2945 if (error)
2946 goto put_dentry;
2947#endif
2948
2949 error = -ENFILE;
2950 file = alloc_file(&path, FMODE_WRITE | FMODE_READ,
2951 &shmem_file_operations);
2952 if (!file)
2953 goto put_dentry;
2954
2955 return file;
2956
2957put_dentry:
2958 path_put(&path);
2959put_memory:
2960 shmem_unacct_size(flags, size);
2961 return ERR_PTR(error);
2962}
2963EXPORT_SYMBOL_GPL(shmem_file_setup);
2964
2965/**
2966 * shmem_zero_setup - setup a shared anonymous mapping
2967 * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
2968 */
2969int shmem_zero_setup(struct vm_area_struct *vma)
2970{
2971 struct file *file;
2972 loff_t size = vma->vm_end - vma->vm_start;
2973
2974 file = shmem_file_setup("dev/zero", size, vma->vm_flags);
2975 if (IS_ERR(file))
2976 return PTR_ERR(file);
2977
2978 if (vma->vm_file)
2979 fput(vma->vm_file);
2980 vma->vm_file = file;
2981 vma->vm_ops = &shmem_vm_ops;
2982 vma->vm_flags |= VM_CAN_NONLINEAR;
2983 return 0;
2984}
2985
2986/**
2987 * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
2988 * @mapping: the page's address_space
2989 * @index: the page index
2990 * @gfp: the page allocator flags to use if allocating
2991 *
2992 * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
2993 * with any new page allocations done using the specified allocation flags.
2994 * But read_cache_page_gfp() uses the ->readpage() method: which does not
2995 * suit tmpfs, since it may have pages in swapcache, and needs to find those
2996 * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
2997 *
2998 * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
2999 * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
3000 */
3001struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
3002 pgoff_t index, gfp_t gfp)
3003{
3004#ifdef CONFIG_SHMEM
3005 struct inode *inode = mapping->host;
3006 struct page *page;
3007 int error;
3008
3009 BUG_ON(mapping->a_ops != &shmem_aops);
3010 error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE, gfp, NULL);
3011 if (error)
3012 page = ERR_PTR(error);
3013 else
3014 unlock_page(page);
3015 return page;
3016#else
3017 /*
3018 * The tiny !SHMEM case uses ramfs without swap
3019 */
3020 return read_cache_page_gfp(mapping, index, gfp);
3021#endif
3022}
3023EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);