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/mm.h>
32#include <linux/sched/signal.h>
33#include <linux/export.h>
34#include <linux/swap.h>
35#include <linux/uio.h>
36#include <linux/khugepaged.h>
37#include <linux/hugetlb.h>
38
39#include <asm/tlbflush.h> /* for arch/microblaze update_mmu_cache() */
40
41static struct vfsmount *shm_mnt;
42
43#ifdef CONFIG_SHMEM
44/*
45 * This virtual memory filesystem is heavily based on the ramfs. It
46 * extends ramfs by the ability to use swap and honor resource limits
47 * which makes it a completely usable filesystem.
48 */
49
50#include <linux/xattr.h>
51#include <linux/exportfs.h>
52#include <linux/posix_acl.h>
53#include <linux/posix_acl_xattr.h>
54#include <linux/mman.h>
55#include <linux/string.h>
56#include <linux/slab.h>
57#include <linux/backing-dev.h>
58#include <linux/shmem_fs.h>
59#include <linux/writeback.h>
60#include <linux/blkdev.h>
61#include <linux/pagevec.h>
62#include <linux/percpu_counter.h>
63#include <linux/falloc.h>
64#include <linux/splice.h>
65#include <linux/security.h>
66#include <linux/swapops.h>
67#include <linux/mempolicy.h>
68#include <linux/namei.h>
69#include <linux/ctype.h>
70#include <linux/migrate.h>
71#include <linux/highmem.h>
72#include <linux/seq_file.h>
73#include <linux/magic.h>
74#include <linux/syscalls.h>
75#include <linux/fcntl.h>
76#include <uapi/linux/memfd.h>
77#include <linux/userfaultfd_k.h>
78#include <linux/rmap.h>
79#include <linux/uuid.h>
80
81#include <linux/uaccess.h>
82#include <asm/pgtable.h>
83
84#include "internal.h"
85
86#define BLOCKS_PER_PAGE (PAGE_SIZE/512)
87#define VM_ACCT(size) (PAGE_ALIGN(size) >> PAGE_SHIFT)
88
89/* Pretend that each entry is of this size in directory's i_size */
90#define BOGO_DIRENT_SIZE 20
91
92/* Symlink up to this size is kmalloc'ed instead of using a swappable page */
93#define SHORT_SYMLINK_LEN 128
94
95/*
96 * shmem_fallocate communicates with shmem_fault or shmem_writepage via
97 * inode->i_private (with i_mutex making sure that it has only one user at
98 * a time): we would prefer not to enlarge the shmem inode just for that.
99 */
100struct shmem_falloc {
101 wait_queue_head_t *waitq; /* faults into hole wait for punch to end */
102 pgoff_t start; /* start of range currently being fallocated */
103 pgoff_t next; /* the next page offset to be fallocated */
104 pgoff_t nr_falloced; /* how many new pages have been fallocated */
105 pgoff_t nr_unswapped; /* how often writepage refused to swap out */
106};
107
108#ifdef CONFIG_TMPFS
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,
125 gfp_t gfp, struct vm_area_struct *vma,
126 struct vm_fault *vmf, int *fault_type);
127
128int shmem_getpage(struct inode *inode, pgoff_t index,
129 struct page **pagep, enum sgp_type sgp)
130{
131 return shmem_getpage_gfp(inode, index, pagep, sgp,
132 mapping_gfp_mask(inode->i_mapping), NULL, NULL, NULL);
133}
134
135static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
136{
137 return sb->s_fs_info;
138}
139
140/*
141 * shmem_file_setup pre-accounts the whole fixed size of a VM object,
142 * for shared memory and for shared anonymous (/dev/zero) mappings
143 * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
144 * consistent with the pre-accounting of private mappings ...
145 */
146static inline int shmem_acct_size(unsigned long flags, loff_t size)
147{
148 return (flags & VM_NORESERVE) ?
149 0 : security_vm_enough_memory_mm(current->mm, VM_ACCT(size));
150}
151
152static inline void shmem_unacct_size(unsigned long flags, loff_t size)
153{
154 if (!(flags & VM_NORESERVE))
155 vm_unacct_memory(VM_ACCT(size));
156}
157
158static inline int shmem_reacct_size(unsigned long flags,
159 loff_t oldsize, loff_t newsize)
160{
161 if (!(flags & VM_NORESERVE)) {
162 if (VM_ACCT(newsize) > VM_ACCT(oldsize))
163 return security_vm_enough_memory_mm(current->mm,
164 VM_ACCT(newsize) - VM_ACCT(oldsize));
165 else if (VM_ACCT(newsize) < VM_ACCT(oldsize))
166 vm_unacct_memory(VM_ACCT(oldsize) - VM_ACCT(newsize));
167 }
168 return 0;
169}
170
171/*
172 * ... whereas tmpfs objects are accounted incrementally as
173 * pages are allocated, in order to allow large sparse files.
174 * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
175 * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
176 */
177static inline int shmem_acct_block(unsigned long flags, long pages)
178{
179 if (!(flags & VM_NORESERVE))
180 return 0;
181
182 return security_vm_enough_memory_mm(current->mm,
183 pages * VM_ACCT(PAGE_SIZE));
184}
185
186static inline void shmem_unacct_blocks(unsigned long flags, long pages)
187{
188 if (flags & VM_NORESERVE)
189 vm_unacct_memory(pages * VM_ACCT(PAGE_SIZE));
190}
191
192static inline bool shmem_inode_acct_block(struct inode *inode, long pages)
193{
194 struct shmem_inode_info *info = SHMEM_I(inode);
195 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
196
197 if (shmem_acct_block(info->flags, pages))
198 return false;
199
200 if (sbinfo->max_blocks) {
201 if (percpu_counter_compare(&sbinfo->used_blocks,
202 sbinfo->max_blocks - pages) > 0)
203 goto unacct;
204 percpu_counter_add(&sbinfo->used_blocks, pages);
205 }
206
207 return true;
208
209unacct:
210 shmem_unacct_blocks(info->flags, pages);
211 return false;
212}
213
214static inline void shmem_inode_unacct_blocks(struct inode *inode, long pages)
215{
216 struct shmem_inode_info *info = SHMEM_I(inode);
217 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
218
219 if (sbinfo->max_blocks)
220 percpu_counter_sub(&sbinfo->used_blocks, pages);
221 shmem_unacct_blocks(info->flags, pages);
222}
223
224static const struct super_operations shmem_ops;
225static const struct address_space_operations shmem_aops;
226static const struct file_operations shmem_file_operations;
227static const struct inode_operations shmem_inode_operations;
228static const struct inode_operations shmem_dir_inode_operations;
229static const struct inode_operations shmem_special_inode_operations;
230static const struct vm_operations_struct shmem_vm_ops;
231static struct file_system_type shmem_fs_type;
232
233bool vma_is_shmem(struct vm_area_struct *vma)
234{
235 return vma->vm_ops == &shmem_vm_ops;
236}
237
238static LIST_HEAD(shmem_swaplist);
239static DEFINE_MUTEX(shmem_swaplist_mutex);
240
241static int shmem_reserve_inode(struct super_block *sb)
242{
243 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
244 if (sbinfo->max_inodes) {
245 spin_lock(&sbinfo->stat_lock);
246 if (!sbinfo->free_inodes) {
247 spin_unlock(&sbinfo->stat_lock);
248 return -ENOSPC;
249 }
250 sbinfo->free_inodes--;
251 spin_unlock(&sbinfo->stat_lock);
252 }
253 return 0;
254}
255
256static void shmem_free_inode(struct super_block *sb)
257{
258 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
259 if (sbinfo->max_inodes) {
260 spin_lock(&sbinfo->stat_lock);
261 sbinfo->free_inodes++;
262 spin_unlock(&sbinfo->stat_lock);
263 }
264}
265
266/**
267 * shmem_recalc_inode - recalculate the block usage of an inode
268 * @inode: inode to recalc
269 *
270 * We have to calculate the free blocks since the mm can drop
271 * undirtied hole pages behind our back.
272 *
273 * But normally info->alloced == inode->i_mapping->nrpages + info->swapped
274 * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
275 *
276 * It has to be called with the spinlock held.
277 */
278static void shmem_recalc_inode(struct inode *inode)
279{
280 struct shmem_inode_info *info = SHMEM_I(inode);
281 long freed;
282
283 freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
284 if (freed > 0) {
285 info->alloced -= freed;
286 inode->i_blocks -= freed * BLOCKS_PER_PAGE;
287 shmem_inode_unacct_blocks(inode, freed);
288 }
289}
290
291bool shmem_charge(struct inode *inode, long pages)
292{
293 struct shmem_inode_info *info = SHMEM_I(inode);
294 unsigned long flags;
295
296 if (!shmem_inode_acct_block(inode, pages))
297 return false;
298
299 spin_lock_irqsave(&info->lock, flags);
300 info->alloced += pages;
301 inode->i_blocks += pages * BLOCKS_PER_PAGE;
302 shmem_recalc_inode(inode);
303 spin_unlock_irqrestore(&info->lock, flags);
304 inode->i_mapping->nrpages += pages;
305
306 return true;
307}
308
309void shmem_uncharge(struct inode *inode, long pages)
310{
311 struct shmem_inode_info *info = SHMEM_I(inode);
312 unsigned long flags;
313
314 spin_lock_irqsave(&info->lock, flags);
315 info->alloced -= pages;
316 inode->i_blocks -= pages * BLOCKS_PER_PAGE;
317 shmem_recalc_inode(inode);
318 spin_unlock_irqrestore(&info->lock, flags);
319
320 shmem_inode_unacct_blocks(inode, pages);
321}
322
323/*
324 * Replace item expected in radix tree by a new item, while holding tree lock.
325 */
326static int shmem_radix_tree_replace(struct address_space *mapping,
327 pgoff_t index, void *expected, void *replacement)
328{
329 struct radix_tree_node *node;
330 void **pslot;
331 void *item;
332
333 VM_BUG_ON(!expected);
334 VM_BUG_ON(!replacement);
335 item = __radix_tree_lookup(&mapping->i_pages, index, &node, &pslot);
336 if (!item)
337 return -ENOENT;
338 if (item != expected)
339 return -ENOENT;
340 __radix_tree_replace(&mapping->i_pages, node, pslot,
341 replacement, NULL);
342 return 0;
343}
344
345/*
346 * Sometimes, before we decide whether to proceed or to fail, we must check
347 * that an entry was not already brought back from swap by a racing thread.
348 *
349 * Checking page is not enough: by the time a SwapCache page is locked, it
350 * might be reused, and again be SwapCache, using the same swap as before.
351 */
352static bool shmem_confirm_swap(struct address_space *mapping,
353 pgoff_t index, swp_entry_t swap)
354{
355 void *item;
356
357 rcu_read_lock();
358 item = radix_tree_lookup(&mapping->i_pages, index);
359 rcu_read_unlock();
360 return item == swp_to_radix_entry(swap);
361}
362
363/*
364 * Definitions for "huge tmpfs": tmpfs mounted with the huge= option
365 *
366 * SHMEM_HUGE_NEVER:
367 * disables huge pages for the mount;
368 * SHMEM_HUGE_ALWAYS:
369 * enables huge pages for the mount;
370 * SHMEM_HUGE_WITHIN_SIZE:
371 * only allocate huge pages if the page will be fully within i_size,
372 * also respect fadvise()/madvise() hints;
373 * SHMEM_HUGE_ADVISE:
374 * only allocate huge pages if requested with fadvise()/madvise();
375 */
376
377#define SHMEM_HUGE_NEVER 0
378#define SHMEM_HUGE_ALWAYS 1
379#define SHMEM_HUGE_WITHIN_SIZE 2
380#define SHMEM_HUGE_ADVISE 3
381
382/*
383 * Special values.
384 * Only can be set via /sys/kernel/mm/transparent_hugepage/shmem_enabled:
385 *
386 * SHMEM_HUGE_DENY:
387 * disables huge on shm_mnt and all mounts, for emergency use;
388 * SHMEM_HUGE_FORCE:
389 * enables huge on shm_mnt and all mounts, w/o needing option, for testing;
390 *
391 */
392#define SHMEM_HUGE_DENY (-1)
393#define SHMEM_HUGE_FORCE (-2)
394
395#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
396/* ifdef here to avoid bloating shmem.o when not necessary */
397
398int shmem_huge __read_mostly;
399
400#if defined(CONFIG_SYSFS) || defined(CONFIG_TMPFS)
401static int shmem_parse_huge(const char *str)
402{
403 if (!strcmp(str, "never"))
404 return SHMEM_HUGE_NEVER;
405 if (!strcmp(str, "always"))
406 return SHMEM_HUGE_ALWAYS;
407 if (!strcmp(str, "within_size"))
408 return SHMEM_HUGE_WITHIN_SIZE;
409 if (!strcmp(str, "advise"))
410 return SHMEM_HUGE_ADVISE;
411 if (!strcmp(str, "deny"))
412 return SHMEM_HUGE_DENY;
413 if (!strcmp(str, "force"))
414 return SHMEM_HUGE_FORCE;
415 return -EINVAL;
416}
417
418static const char *shmem_format_huge(int huge)
419{
420 switch (huge) {
421 case SHMEM_HUGE_NEVER:
422 return "never";
423 case SHMEM_HUGE_ALWAYS:
424 return "always";
425 case SHMEM_HUGE_WITHIN_SIZE:
426 return "within_size";
427 case SHMEM_HUGE_ADVISE:
428 return "advise";
429 case SHMEM_HUGE_DENY:
430 return "deny";
431 case SHMEM_HUGE_FORCE:
432 return "force";
433 default:
434 VM_BUG_ON(1);
435 return "bad_val";
436 }
437}
438#endif
439
440static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
441 struct shrink_control *sc, unsigned long nr_to_split)
442{
443 LIST_HEAD(list), *pos, *next;
444 LIST_HEAD(to_remove);
445 struct inode *inode;
446 struct shmem_inode_info *info;
447 struct page *page;
448 unsigned long batch = sc ? sc->nr_to_scan : 128;
449 int removed = 0, split = 0;
450
451 if (list_empty(&sbinfo->shrinklist))
452 return SHRINK_STOP;
453
454 spin_lock(&sbinfo->shrinklist_lock);
455 list_for_each_safe(pos, next, &sbinfo->shrinklist) {
456 info = list_entry(pos, struct shmem_inode_info, shrinklist);
457
458 /* pin the inode */
459 inode = igrab(&info->vfs_inode);
460
461 /* inode is about to be evicted */
462 if (!inode) {
463 list_del_init(&info->shrinklist);
464 removed++;
465 goto next;
466 }
467
468 /* Check if there's anything to gain */
469 if (round_up(inode->i_size, PAGE_SIZE) ==
470 round_up(inode->i_size, HPAGE_PMD_SIZE)) {
471 list_move(&info->shrinklist, &to_remove);
472 removed++;
473 goto next;
474 }
475
476 list_move(&info->shrinklist, &list);
477next:
478 if (!--batch)
479 break;
480 }
481 spin_unlock(&sbinfo->shrinklist_lock);
482
483 list_for_each_safe(pos, next, &to_remove) {
484 info = list_entry(pos, struct shmem_inode_info, shrinklist);
485 inode = &info->vfs_inode;
486 list_del_init(&info->shrinklist);
487 iput(inode);
488 }
489
490 list_for_each_safe(pos, next, &list) {
491 int ret;
492
493 info = list_entry(pos, struct shmem_inode_info, shrinklist);
494 inode = &info->vfs_inode;
495
496 if (nr_to_split && split >= nr_to_split)
497 goto leave;
498
499 page = find_get_page(inode->i_mapping,
500 (inode->i_size & HPAGE_PMD_MASK) >> PAGE_SHIFT);
501 if (!page)
502 goto drop;
503
504 /* No huge page at the end of the file: nothing to split */
505 if (!PageTransHuge(page)) {
506 put_page(page);
507 goto drop;
508 }
509
510 /*
511 * Leave the inode on the list if we failed to lock
512 * the page at this time.
513 *
514 * Waiting for the lock may lead to deadlock in the
515 * reclaim path.
516 */
517 if (!trylock_page(page)) {
518 put_page(page);
519 goto leave;
520 }
521
522 ret = split_huge_page(page);
523 unlock_page(page);
524 put_page(page);
525
526 /* If split failed leave the inode on the list */
527 if (ret)
528 goto leave;
529
530 split++;
531drop:
532 list_del_init(&info->shrinklist);
533 removed++;
534leave:
535 iput(inode);
536 }
537
538 spin_lock(&sbinfo->shrinklist_lock);
539 list_splice_tail(&list, &sbinfo->shrinklist);
540 sbinfo->shrinklist_len -= removed;
541 spin_unlock(&sbinfo->shrinklist_lock);
542
543 return split;
544}
545
546static long shmem_unused_huge_scan(struct super_block *sb,
547 struct shrink_control *sc)
548{
549 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
550
551 if (!READ_ONCE(sbinfo->shrinklist_len))
552 return SHRINK_STOP;
553
554 return shmem_unused_huge_shrink(sbinfo, sc, 0);
555}
556
557static long shmem_unused_huge_count(struct super_block *sb,
558 struct shrink_control *sc)
559{
560 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
561 return READ_ONCE(sbinfo->shrinklist_len);
562}
563#else /* !CONFIG_TRANSPARENT_HUGE_PAGECACHE */
564
565#define shmem_huge SHMEM_HUGE_DENY
566
567static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
568 struct shrink_control *sc, unsigned long nr_to_split)
569{
570 return 0;
571}
572#endif /* CONFIG_TRANSPARENT_HUGE_PAGECACHE */
573
574/*
575 * Like add_to_page_cache_locked, but error if expected item has gone.
576 */
577static int shmem_add_to_page_cache(struct page *page,
578 struct address_space *mapping,
579 pgoff_t index, void *expected)
580{
581 int error, nr = hpage_nr_pages(page);
582
583 VM_BUG_ON_PAGE(PageTail(page), page);
584 VM_BUG_ON_PAGE(index != round_down(index, nr), page);
585 VM_BUG_ON_PAGE(!PageLocked(page), page);
586 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
587 VM_BUG_ON(expected && PageTransHuge(page));
588
589 page_ref_add(page, nr);
590 page->mapping = mapping;
591 page->index = index;
592
593 xa_lock_irq(&mapping->i_pages);
594 if (PageTransHuge(page)) {
595 void __rcu **results;
596 pgoff_t idx;
597 int i;
598
599 error = 0;
600 if (radix_tree_gang_lookup_slot(&mapping->i_pages,
601 &results, &idx, index, 1) &&
602 idx < index + HPAGE_PMD_NR) {
603 error = -EEXIST;
604 }
605
606 if (!error) {
607 for (i = 0; i < HPAGE_PMD_NR; i++) {
608 error = radix_tree_insert(&mapping->i_pages,
609 index + i, page + i);
610 VM_BUG_ON(error);
611 }
612 count_vm_event(THP_FILE_ALLOC);
613 }
614 } else if (!expected) {
615 error = radix_tree_insert(&mapping->i_pages, index, page);
616 } else {
617 error = shmem_radix_tree_replace(mapping, index, expected,
618 page);
619 }
620
621 if (!error) {
622 mapping->nrpages += nr;
623 if (PageTransHuge(page))
624 __inc_node_page_state(page, NR_SHMEM_THPS);
625 __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, nr);
626 __mod_node_page_state(page_pgdat(page), NR_SHMEM, nr);
627 xa_unlock_irq(&mapping->i_pages);
628 } else {
629 page->mapping = NULL;
630 xa_unlock_irq(&mapping->i_pages);
631 page_ref_sub(page, nr);
632 }
633 return error;
634}
635
636/*
637 * Like delete_from_page_cache, but substitutes swap for page.
638 */
639static void shmem_delete_from_page_cache(struct page *page, void *radswap)
640{
641 struct address_space *mapping = page->mapping;
642 int error;
643
644 VM_BUG_ON_PAGE(PageCompound(page), page);
645
646 xa_lock_irq(&mapping->i_pages);
647 error = shmem_radix_tree_replace(mapping, page->index, page, radswap);
648 page->mapping = NULL;
649 mapping->nrpages--;
650 __dec_node_page_state(page, NR_FILE_PAGES);
651 __dec_node_page_state(page, NR_SHMEM);
652 xa_unlock_irq(&mapping->i_pages);
653 put_page(page);
654 BUG_ON(error);
655}
656
657/*
658 * Remove swap entry from radix tree, free the swap and its page cache.
659 */
660static int shmem_free_swap(struct address_space *mapping,
661 pgoff_t index, void *radswap)
662{
663 void *old;
664
665 xa_lock_irq(&mapping->i_pages);
666 old = radix_tree_delete_item(&mapping->i_pages, index, radswap);
667 xa_unlock_irq(&mapping->i_pages);
668 if (old != radswap)
669 return -ENOENT;
670 free_swap_and_cache(radix_to_swp_entry(radswap));
671 return 0;
672}
673
674/*
675 * Determine (in bytes) how many of the shmem object's pages mapped by the
676 * given offsets are swapped out.
677 *
678 * This is safe to call without i_mutex or the i_pages lock thanks to RCU,
679 * as long as the inode doesn't go away and racy results are not a problem.
680 */
681unsigned long shmem_partial_swap_usage(struct address_space *mapping,
682 pgoff_t start, pgoff_t end)
683{
684 struct radix_tree_iter iter;
685 void **slot;
686 struct page *page;
687 unsigned long swapped = 0;
688
689 rcu_read_lock();
690
691 radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start) {
692 if (iter.index >= end)
693 break;
694
695 page = radix_tree_deref_slot(slot);
696
697 if (radix_tree_deref_retry(page)) {
698 slot = radix_tree_iter_retry(&iter);
699 continue;
700 }
701
702 if (radix_tree_exceptional_entry(page))
703 swapped++;
704
705 if (need_resched()) {
706 slot = radix_tree_iter_resume(slot, &iter);
707 cond_resched_rcu();
708 }
709 }
710
711 rcu_read_unlock();
712
713 return swapped << PAGE_SHIFT;
714}
715
716/*
717 * Determine (in bytes) how many of the shmem object's pages mapped by the
718 * given vma is swapped out.
719 *
720 * This is safe to call without i_mutex or the i_pages lock thanks to RCU,
721 * as long as the inode doesn't go away and racy results are not a problem.
722 */
723unsigned long shmem_swap_usage(struct vm_area_struct *vma)
724{
725 struct inode *inode = file_inode(vma->vm_file);
726 struct shmem_inode_info *info = SHMEM_I(inode);
727 struct address_space *mapping = inode->i_mapping;
728 unsigned long swapped;
729
730 /* Be careful as we don't hold info->lock */
731 swapped = READ_ONCE(info->swapped);
732
733 /*
734 * The easier cases are when the shmem object has nothing in swap, or
735 * the vma maps it whole. Then we can simply use the stats that we
736 * already track.
737 */
738 if (!swapped)
739 return 0;
740
741 if (!vma->vm_pgoff && vma->vm_end - vma->vm_start >= inode->i_size)
742 return swapped << PAGE_SHIFT;
743
744 /* Here comes the more involved part */
745 return shmem_partial_swap_usage(mapping,
746 linear_page_index(vma, vma->vm_start),
747 linear_page_index(vma, vma->vm_end));
748}
749
750/*
751 * SysV IPC SHM_UNLOCK restore Unevictable pages to their evictable lists.
752 */
753void shmem_unlock_mapping(struct address_space *mapping)
754{
755 struct pagevec pvec;
756 pgoff_t indices[PAGEVEC_SIZE];
757 pgoff_t index = 0;
758
759 pagevec_init(&pvec);
760 /*
761 * Minor point, but we might as well stop if someone else SHM_LOCKs it.
762 */
763 while (!mapping_unevictable(mapping)) {
764 /*
765 * Avoid pagevec_lookup(): find_get_pages() returns 0 as if it
766 * has finished, if it hits a row of PAGEVEC_SIZE swap entries.
767 */
768 pvec.nr = find_get_entries(mapping, index,
769 PAGEVEC_SIZE, pvec.pages, indices);
770 if (!pvec.nr)
771 break;
772 index = indices[pvec.nr - 1] + 1;
773 pagevec_remove_exceptionals(&pvec);
774 check_move_unevictable_pages(pvec.pages, pvec.nr);
775 pagevec_release(&pvec);
776 cond_resched();
777 }
778}
779
780/*
781 * Remove range of pages and swap entries from radix tree, and free them.
782 * If !unfalloc, truncate or punch hole; if unfalloc, undo failed fallocate.
783 */
784static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
785 bool unfalloc)
786{
787 struct address_space *mapping = inode->i_mapping;
788 struct shmem_inode_info *info = SHMEM_I(inode);
789 pgoff_t start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
790 pgoff_t end = (lend + 1) >> PAGE_SHIFT;
791 unsigned int partial_start = lstart & (PAGE_SIZE - 1);
792 unsigned int partial_end = (lend + 1) & (PAGE_SIZE - 1);
793 struct pagevec pvec;
794 pgoff_t indices[PAGEVEC_SIZE];
795 long nr_swaps_freed = 0;
796 pgoff_t index;
797 int i;
798
799 if (lend == -1)
800 end = -1; /* unsigned, so actually very big */
801
802 pagevec_init(&pvec);
803 index = start;
804 while (index < end) {
805 pvec.nr = find_get_entries(mapping, index,
806 min(end - index, (pgoff_t)PAGEVEC_SIZE),
807 pvec.pages, indices);
808 if (!pvec.nr)
809 break;
810 for (i = 0; i < pagevec_count(&pvec); i++) {
811 struct page *page = pvec.pages[i];
812
813 index = indices[i];
814 if (index >= end)
815 break;
816
817 if (radix_tree_exceptional_entry(page)) {
818 if (unfalloc)
819 continue;
820 nr_swaps_freed += !shmem_free_swap(mapping,
821 index, page);
822 continue;
823 }
824
825 VM_BUG_ON_PAGE(page_to_pgoff(page) != index, page);
826
827 if (!trylock_page(page))
828 continue;
829
830 if (PageTransTail(page)) {
831 /* Middle of THP: zero out the page */
832 clear_highpage(page);
833 unlock_page(page);
834 continue;
835 } else if (PageTransHuge(page)) {
836 if (index == round_down(end, HPAGE_PMD_NR)) {
837 /*
838 * Range ends in the middle of THP:
839 * zero out the page
840 */
841 clear_highpage(page);
842 unlock_page(page);
843 continue;
844 }
845 index += HPAGE_PMD_NR - 1;
846 i += HPAGE_PMD_NR - 1;
847 }
848
849 if (!unfalloc || !PageUptodate(page)) {
850 VM_BUG_ON_PAGE(PageTail(page), page);
851 if (page_mapping(page) == mapping) {
852 VM_BUG_ON_PAGE(PageWriteback(page), page);
853 truncate_inode_page(mapping, page);
854 }
855 }
856 unlock_page(page);
857 }
858 pagevec_remove_exceptionals(&pvec);
859 pagevec_release(&pvec);
860 cond_resched();
861 index++;
862 }
863
864 if (partial_start) {
865 struct page *page = NULL;
866 shmem_getpage(inode, start - 1, &page, SGP_READ);
867 if (page) {
868 unsigned int top = PAGE_SIZE;
869 if (start > end) {
870 top = partial_end;
871 partial_end = 0;
872 }
873 zero_user_segment(page, partial_start, top);
874 set_page_dirty(page);
875 unlock_page(page);
876 put_page(page);
877 }
878 }
879 if (partial_end) {
880 struct page *page = NULL;
881 shmem_getpage(inode, end, &page, SGP_READ);
882 if (page) {
883 zero_user_segment(page, 0, partial_end);
884 set_page_dirty(page);
885 unlock_page(page);
886 put_page(page);
887 }
888 }
889 if (start >= end)
890 return;
891
892 index = start;
893 while (index < end) {
894 cond_resched();
895
896 pvec.nr = find_get_entries(mapping, index,
897 min(end - index, (pgoff_t)PAGEVEC_SIZE),
898 pvec.pages, indices);
899 if (!pvec.nr) {
900 /* If all gone or hole-punch or unfalloc, we're done */
901 if (index == start || end != -1)
902 break;
903 /* But if truncating, restart to make sure all gone */
904 index = start;
905 continue;
906 }
907 for (i = 0; i < pagevec_count(&pvec); i++) {
908 struct page *page = pvec.pages[i];
909
910 index = indices[i];
911 if (index >= end)
912 break;
913
914 if (radix_tree_exceptional_entry(page)) {
915 if (unfalloc)
916 continue;
917 if (shmem_free_swap(mapping, index, page)) {
918 /* Swap was replaced by page: retry */
919 index--;
920 break;
921 }
922 nr_swaps_freed++;
923 continue;
924 }
925
926 lock_page(page);
927
928 if (PageTransTail(page)) {
929 /* Middle of THP: zero out the page */
930 clear_highpage(page);
931 unlock_page(page);
932 /*
933 * Partial thp truncate due 'start' in middle
934 * of THP: don't need to look on these pages
935 * again on !pvec.nr restart.
936 */
937 if (index != round_down(end, HPAGE_PMD_NR))
938 start++;
939 continue;
940 } else if (PageTransHuge(page)) {
941 if (index == round_down(end, HPAGE_PMD_NR)) {
942 /*
943 * Range ends in the middle of THP:
944 * zero out the page
945 */
946 clear_highpage(page);
947 unlock_page(page);
948 continue;
949 }
950 index += HPAGE_PMD_NR - 1;
951 i += HPAGE_PMD_NR - 1;
952 }
953
954 if (!unfalloc || !PageUptodate(page)) {
955 VM_BUG_ON_PAGE(PageTail(page), page);
956 if (page_mapping(page) == mapping) {
957 VM_BUG_ON_PAGE(PageWriteback(page), page);
958 truncate_inode_page(mapping, page);
959 } else {
960 /* Page was replaced by swap: retry */
961 unlock_page(page);
962 index--;
963 break;
964 }
965 }
966 unlock_page(page);
967 }
968 pagevec_remove_exceptionals(&pvec);
969 pagevec_release(&pvec);
970 index++;
971 }
972
973 spin_lock_irq(&info->lock);
974 info->swapped -= nr_swaps_freed;
975 shmem_recalc_inode(inode);
976 spin_unlock_irq(&info->lock);
977}
978
979void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
980{
981 shmem_undo_range(inode, lstart, lend, false);
982 inode->i_ctime = inode->i_mtime = current_time(inode);
983}
984EXPORT_SYMBOL_GPL(shmem_truncate_range);
985
986static int shmem_getattr(const struct path *path, struct kstat *stat,
987 u32 request_mask, unsigned int query_flags)
988{
989 struct inode *inode = path->dentry->d_inode;
990 struct shmem_inode_info *info = SHMEM_I(inode);
991
992 if (info->alloced - info->swapped != inode->i_mapping->nrpages) {
993 spin_lock_irq(&info->lock);
994 shmem_recalc_inode(inode);
995 spin_unlock_irq(&info->lock);
996 }
997 generic_fillattr(inode, stat);
998 return 0;
999}
1000
1001static int shmem_setattr(struct dentry *dentry, struct iattr *attr)
1002{
1003 struct inode *inode = d_inode(dentry);
1004 struct shmem_inode_info *info = SHMEM_I(inode);
1005 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1006 int error;
1007
1008 error = setattr_prepare(dentry, attr);
1009 if (error)
1010 return error;
1011
1012 if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
1013 loff_t oldsize = inode->i_size;
1014 loff_t newsize = attr->ia_size;
1015
1016 /* protected by i_mutex */
1017 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
1018 (newsize > oldsize && (info->seals & F_SEAL_GROW)))
1019 return -EPERM;
1020
1021 if (newsize != oldsize) {
1022 error = shmem_reacct_size(SHMEM_I(inode)->flags,
1023 oldsize, newsize);
1024 if (error)
1025 return error;
1026 i_size_write(inode, newsize);
1027 inode->i_ctime = inode->i_mtime = current_time(inode);
1028 }
1029 if (newsize <= oldsize) {
1030 loff_t holebegin = round_up(newsize, PAGE_SIZE);
1031 if (oldsize > holebegin)
1032 unmap_mapping_range(inode->i_mapping,
1033 holebegin, 0, 1);
1034 if (info->alloced)
1035 shmem_truncate_range(inode,
1036 newsize, (loff_t)-1);
1037 /* unmap again to remove racily COWed private pages */
1038 if (oldsize > holebegin)
1039 unmap_mapping_range(inode->i_mapping,
1040 holebegin, 0, 1);
1041
1042 /*
1043 * Part of the huge page can be beyond i_size: subject
1044 * to shrink under memory pressure.
1045 */
1046 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
1047 spin_lock(&sbinfo->shrinklist_lock);
1048 /*
1049 * _careful to defend against unlocked access to
1050 * ->shrink_list in shmem_unused_huge_shrink()
1051 */
1052 if (list_empty_careful(&info->shrinklist)) {
1053 list_add_tail(&info->shrinklist,
1054 &sbinfo->shrinklist);
1055 sbinfo->shrinklist_len++;
1056 }
1057 spin_unlock(&sbinfo->shrinklist_lock);
1058 }
1059 }
1060 }
1061
1062 setattr_copy(inode, attr);
1063 if (attr->ia_valid & ATTR_MODE)
1064 error = posix_acl_chmod(inode, inode->i_mode);
1065 return error;
1066}
1067
1068static void shmem_evict_inode(struct inode *inode)
1069{
1070 struct shmem_inode_info *info = SHMEM_I(inode);
1071 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1072
1073 if (inode->i_mapping->a_ops == &shmem_aops) {
1074 shmem_unacct_size(info->flags, inode->i_size);
1075 inode->i_size = 0;
1076 shmem_truncate_range(inode, 0, (loff_t)-1);
1077 if (!list_empty(&info->shrinklist)) {
1078 spin_lock(&sbinfo->shrinklist_lock);
1079 if (!list_empty(&info->shrinklist)) {
1080 list_del_init(&info->shrinklist);
1081 sbinfo->shrinklist_len--;
1082 }
1083 spin_unlock(&sbinfo->shrinklist_lock);
1084 }
1085 if (!list_empty(&info->swaplist)) {
1086 mutex_lock(&shmem_swaplist_mutex);
1087 list_del_init(&info->swaplist);
1088 mutex_unlock(&shmem_swaplist_mutex);
1089 }
1090 }
1091
1092 simple_xattrs_free(&info->xattrs);
1093 WARN_ON(inode->i_blocks);
1094 shmem_free_inode(inode->i_sb);
1095 clear_inode(inode);
1096}
1097
1098static unsigned long find_swap_entry(struct radix_tree_root *root, void *item)
1099{
1100 struct radix_tree_iter iter;
1101 void **slot;
1102 unsigned long found = -1;
1103 unsigned int checked = 0;
1104
1105 rcu_read_lock();
1106 radix_tree_for_each_slot(slot, root, &iter, 0) {
1107 if (*slot == item) {
1108 found = iter.index;
1109 break;
1110 }
1111 checked++;
1112 if ((checked % 4096) != 0)
1113 continue;
1114 slot = radix_tree_iter_resume(slot, &iter);
1115 cond_resched_rcu();
1116 }
1117
1118 rcu_read_unlock();
1119 return found;
1120}
1121
1122/*
1123 * If swap found in inode, free it and move page from swapcache to filecache.
1124 */
1125static int shmem_unuse_inode(struct shmem_inode_info *info,
1126 swp_entry_t swap, struct page **pagep)
1127{
1128 struct address_space *mapping = info->vfs_inode.i_mapping;
1129 void *radswap;
1130 pgoff_t index;
1131 gfp_t gfp;
1132 int error = 0;
1133
1134 radswap = swp_to_radix_entry(swap);
1135 index = find_swap_entry(&mapping->i_pages, radswap);
1136 if (index == -1)
1137 return -EAGAIN; /* tell shmem_unuse we found nothing */
1138
1139 /*
1140 * Move _head_ to start search for next from here.
1141 * But be careful: shmem_evict_inode checks list_empty without taking
1142 * mutex, and there's an instant in list_move_tail when info->swaplist
1143 * would appear empty, if it were the only one on shmem_swaplist.
1144 */
1145 if (shmem_swaplist.next != &info->swaplist)
1146 list_move_tail(&shmem_swaplist, &info->swaplist);
1147
1148 gfp = mapping_gfp_mask(mapping);
1149 if (shmem_should_replace_page(*pagep, gfp)) {
1150 mutex_unlock(&shmem_swaplist_mutex);
1151 error = shmem_replace_page(pagep, gfp, info, index);
1152 mutex_lock(&shmem_swaplist_mutex);
1153 /*
1154 * We needed to drop mutex to make that restrictive page
1155 * allocation, but the inode might have been freed while we
1156 * dropped it: although a racing shmem_evict_inode() cannot
1157 * complete without emptying the radix_tree, our page lock
1158 * on this swapcache page is not enough to prevent that -
1159 * free_swap_and_cache() of our swap entry will only
1160 * trylock_page(), removing swap from radix_tree whatever.
1161 *
1162 * We must not proceed to shmem_add_to_page_cache() if the
1163 * inode has been freed, but of course we cannot rely on
1164 * inode or mapping or info to check that. However, we can
1165 * safely check if our swap entry is still in use (and here
1166 * it can't have got reused for another page): if it's still
1167 * in use, then the inode cannot have been freed yet, and we
1168 * can safely proceed (if it's no longer in use, that tells
1169 * nothing about the inode, but we don't need to unuse swap).
1170 */
1171 if (!page_swapcount(*pagep))
1172 error = -ENOENT;
1173 }
1174
1175 /*
1176 * We rely on shmem_swaplist_mutex, not only to protect the swaplist,
1177 * but also to hold up shmem_evict_inode(): so inode cannot be freed
1178 * beneath us (pagelock doesn't help until the page is in pagecache).
1179 */
1180 if (!error)
1181 error = shmem_add_to_page_cache(*pagep, mapping, index,
1182 radswap);
1183 if (error != -ENOMEM) {
1184 /*
1185 * Truncation and eviction use free_swap_and_cache(), which
1186 * only does trylock page: if we raced, best clean up here.
1187 */
1188 delete_from_swap_cache(*pagep);
1189 set_page_dirty(*pagep);
1190 if (!error) {
1191 spin_lock_irq(&info->lock);
1192 info->swapped--;
1193 spin_unlock_irq(&info->lock);
1194 swap_free(swap);
1195 }
1196 }
1197 return error;
1198}
1199
1200/*
1201 * Search through swapped inodes to find and replace swap by page.
1202 */
1203int shmem_unuse(swp_entry_t swap, struct page *page)
1204{
1205 struct list_head *this, *next;
1206 struct shmem_inode_info *info;
1207 struct mem_cgroup *memcg;
1208 int error = 0;
1209
1210 /*
1211 * There's a faint possibility that swap page was replaced before
1212 * caller locked it: caller will come back later with the right page.
1213 */
1214 if (unlikely(!PageSwapCache(page) || page_private(page) != swap.val))
1215 goto out;
1216
1217 /*
1218 * Charge page using GFP_KERNEL while we can wait, before taking
1219 * the shmem_swaplist_mutex which might hold up shmem_writepage().
1220 * Charged back to the user (not to caller) when swap account is used.
1221 */
1222 error = mem_cgroup_try_charge(page, current->mm, GFP_KERNEL, &memcg,
1223 false);
1224 if (error)
1225 goto out;
1226 /* No radix_tree_preload: swap entry keeps a place for page in tree */
1227 error = -EAGAIN;
1228
1229 mutex_lock(&shmem_swaplist_mutex);
1230 list_for_each_safe(this, next, &shmem_swaplist) {
1231 info = list_entry(this, struct shmem_inode_info, swaplist);
1232 if (info->swapped)
1233 error = shmem_unuse_inode(info, swap, &page);
1234 else
1235 list_del_init(&info->swaplist);
1236 cond_resched();
1237 if (error != -EAGAIN)
1238 break;
1239 /* found nothing in this: move on to search the next */
1240 }
1241 mutex_unlock(&shmem_swaplist_mutex);
1242
1243 if (error) {
1244 if (error != -ENOMEM)
1245 error = 0;
1246 mem_cgroup_cancel_charge(page, memcg, false);
1247 } else
1248 mem_cgroup_commit_charge(page, memcg, true, false);
1249out:
1250 unlock_page(page);
1251 put_page(page);
1252 return error;
1253}
1254
1255/*
1256 * Move the page from the page cache to the swap cache.
1257 */
1258static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1259{
1260 struct shmem_inode_info *info;
1261 struct address_space *mapping;
1262 struct inode *inode;
1263 swp_entry_t swap;
1264 pgoff_t index;
1265
1266 VM_BUG_ON_PAGE(PageCompound(page), page);
1267 BUG_ON(!PageLocked(page));
1268 mapping = page->mapping;
1269 index = page->index;
1270 inode = mapping->host;
1271 info = SHMEM_I(inode);
1272 if (info->flags & VM_LOCKED)
1273 goto redirty;
1274 if (!total_swap_pages)
1275 goto redirty;
1276
1277 /*
1278 * Our capabilities prevent regular writeback or sync from ever calling
1279 * shmem_writepage; but a stacking filesystem might use ->writepage of
1280 * its underlying filesystem, in which case tmpfs should write out to
1281 * swap only in response to memory pressure, and not for the writeback
1282 * threads or sync.
1283 */
1284 if (!wbc->for_reclaim) {
1285 WARN_ON_ONCE(1); /* Still happens? Tell us about it! */
1286 goto redirty;
1287 }
1288
1289 /*
1290 * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
1291 * value into swapfile.c, the only way we can correctly account for a
1292 * fallocated page arriving here is now to initialize it and write it.
1293 *
1294 * That's okay for a page already fallocated earlier, but if we have
1295 * not yet completed the fallocation, then (a) we want to keep track
1296 * of this page in case we have to undo it, and (b) it may not be a
1297 * good idea to continue anyway, once we're pushing into swap. So
1298 * reactivate the page, and let shmem_fallocate() quit when too many.
1299 */
1300 if (!PageUptodate(page)) {
1301 if (inode->i_private) {
1302 struct shmem_falloc *shmem_falloc;
1303 spin_lock(&inode->i_lock);
1304 shmem_falloc = inode->i_private;
1305 if (shmem_falloc &&
1306 !shmem_falloc->waitq &&
1307 index >= shmem_falloc->start &&
1308 index < shmem_falloc->next)
1309 shmem_falloc->nr_unswapped++;
1310 else
1311 shmem_falloc = NULL;
1312 spin_unlock(&inode->i_lock);
1313 if (shmem_falloc)
1314 goto redirty;
1315 }
1316 clear_highpage(page);
1317 flush_dcache_page(page);
1318 SetPageUptodate(page);
1319 }
1320
1321 swap = get_swap_page(page);
1322 if (!swap.val)
1323 goto redirty;
1324
1325 if (mem_cgroup_try_charge_swap(page, swap))
1326 goto free_swap;
1327
1328 /*
1329 * Add inode to shmem_unuse()'s list of swapped-out inodes,
1330 * if it's not already there. Do it now before the page is
1331 * moved to swap cache, when its pagelock no longer protects
1332 * the inode from eviction. But don't unlock the mutex until
1333 * we've incremented swapped, because shmem_unuse_inode() will
1334 * prune a !swapped inode from the swaplist under this mutex.
1335 */
1336 mutex_lock(&shmem_swaplist_mutex);
1337 if (list_empty(&info->swaplist))
1338 list_add_tail(&info->swaplist, &shmem_swaplist);
1339
1340 if (add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
1341 spin_lock_irq(&info->lock);
1342 shmem_recalc_inode(inode);
1343 info->swapped++;
1344 spin_unlock_irq(&info->lock);
1345
1346 swap_shmem_alloc(swap);
1347 shmem_delete_from_page_cache(page, swp_to_radix_entry(swap));
1348
1349 mutex_unlock(&shmem_swaplist_mutex);
1350 BUG_ON(page_mapped(page));
1351 swap_writepage(page, wbc);
1352 return 0;
1353 }
1354
1355 mutex_unlock(&shmem_swaplist_mutex);
1356free_swap:
1357 put_swap_page(page, swap);
1358redirty:
1359 set_page_dirty(page);
1360 if (wbc->for_reclaim)
1361 return AOP_WRITEPAGE_ACTIVATE; /* Return with page locked */
1362 unlock_page(page);
1363 return 0;
1364}
1365
1366#if defined(CONFIG_NUMA) && defined(CONFIG_TMPFS)
1367static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1368{
1369 char buffer[64];
1370
1371 if (!mpol || mpol->mode == MPOL_DEFAULT)
1372 return; /* show nothing */
1373
1374 mpol_to_str(buffer, sizeof(buffer), mpol);
1375
1376 seq_printf(seq, ",mpol=%s", buffer);
1377}
1378
1379static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1380{
1381 struct mempolicy *mpol = NULL;
1382 if (sbinfo->mpol) {
1383 spin_lock(&sbinfo->stat_lock); /* prevent replace/use races */
1384 mpol = sbinfo->mpol;
1385 mpol_get(mpol);
1386 spin_unlock(&sbinfo->stat_lock);
1387 }
1388 return mpol;
1389}
1390#else /* !CONFIG_NUMA || !CONFIG_TMPFS */
1391static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1392{
1393}
1394static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1395{
1396 return NULL;
1397}
1398#endif /* CONFIG_NUMA && CONFIG_TMPFS */
1399#ifndef CONFIG_NUMA
1400#define vm_policy vm_private_data
1401#endif
1402
1403static void shmem_pseudo_vma_init(struct vm_area_struct *vma,
1404 struct shmem_inode_info *info, pgoff_t index)
1405{
1406 /* Create a pseudo vma that just contains the policy */
1407 vma->vm_start = 0;
1408 /* Bias interleave by inode number to distribute better across nodes */
1409 vma->vm_pgoff = index + info->vfs_inode.i_ino;
1410 vma->vm_ops = NULL;
1411 vma->vm_policy = mpol_shared_policy_lookup(&info->policy, index);
1412}
1413
1414static void shmem_pseudo_vma_destroy(struct vm_area_struct *vma)
1415{
1416 /* Drop reference taken by mpol_shared_policy_lookup() */
1417 mpol_cond_put(vma->vm_policy);
1418}
1419
1420static struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
1421 struct shmem_inode_info *info, pgoff_t index)
1422{
1423 struct vm_area_struct pvma;
1424 struct page *page;
1425 struct vm_fault vmf;
1426
1427 shmem_pseudo_vma_init(&pvma, info, index);
1428 vmf.vma = &pvma;
1429 vmf.address = 0;
1430 page = swap_cluster_readahead(swap, gfp, &vmf);
1431 shmem_pseudo_vma_destroy(&pvma);
1432
1433 return page;
1434}
1435
1436static struct page *shmem_alloc_hugepage(gfp_t gfp,
1437 struct shmem_inode_info *info, pgoff_t index)
1438{
1439 struct vm_area_struct pvma;
1440 struct inode *inode = &info->vfs_inode;
1441 struct address_space *mapping = inode->i_mapping;
1442 pgoff_t idx, hindex;
1443 void __rcu **results;
1444 struct page *page;
1445
1446 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
1447 return NULL;
1448
1449 hindex = round_down(index, HPAGE_PMD_NR);
1450 rcu_read_lock();
1451 if (radix_tree_gang_lookup_slot(&mapping->i_pages, &results, &idx,
1452 hindex, 1) && idx < hindex + HPAGE_PMD_NR) {
1453 rcu_read_unlock();
1454 return NULL;
1455 }
1456 rcu_read_unlock();
1457
1458 shmem_pseudo_vma_init(&pvma, info, hindex);
1459 page = alloc_pages_vma(gfp | __GFP_COMP | __GFP_NORETRY | __GFP_NOWARN,
1460 HPAGE_PMD_ORDER, &pvma, 0, numa_node_id(), true);
1461 shmem_pseudo_vma_destroy(&pvma);
1462 if (page)
1463 prep_transhuge_page(page);
1464 return page;
1465}
1466
1467static struct page *shmem_alloc_page(gfp_t gfp,
1468 struct shmem_inode_info *info, pgoff_t index)
1469{
1470 struct vm_area_struct pvma;
1471 struct page *page;
1472
1473 shmem_pseudo_vma_init(&pvma, info, index);
1474 page = alloc_page_vma(gfp, &pvma, 0);
1475 shmem_pseudo_vma_destroy(&pvma);
1476
1477 return page;
1478}
1479
1480static struct page *shmem_alloc_and_acct_page(gfp_t gfp,
1481 struct inode *inode,
1482 pgoff_t index, bool huge)
1483{
1484 struct shmem_inode_info *info = SHMEM_I(inode);
1485 struct page *page;
1486 int nr;
1487 int err = -ENOSPC;
1488
1489 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
1490 huge = false;
1491 nr = huge ? HPAGE_PMD_NR : 1;
1492
1493 if (!shmem_inode_acct_block(inode, nr))
1494 goto failed;
1495
1496 if (huge)
1497 page = shmem_alloc_hugepage(gfp, info, index);
1498 else
1499 page = shmem_alloc_page(gfp, info, index);
1500 if (page) {
1501 __SetPageLocked(page);
1502 __SetPageSwapBacked(page);
1503 return page;
1504 }
1505
1506 err = -ENOMEM;
1507 shmem_inode_unacct_blocks(inode, nr);
1508failed:
1509 return ERR_PTR(err);
1510}
1511
1512/*
1513 * When a page is moved from swapcache to shmem filecache (either by the
1514 * usual swapin of shmem_getpage_gfp(), or by the less common swapoff of
1515 * shmem_unuse_inode()), it may have been read in earlier from swap, in
1516 * ignorance of the mapping it belongs to. If that mapping has special
1517 * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
1518 * we may need to copy to a suitable page before moving to filecache.
1519 *
1520 * In a future release, this may well be extended to respect cpuset and
1521 * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
1522 * but for now it is a simple matter of zone.
1523 */
1524static bool shmem_should_replace_page(struct page *page, gfp_t gfp)
1525{
1526 return page_zonenum(page) > gfp_zone(gfp);
1527}
1528
1529static int shmem_replace_page(struct page **pagep, gfp_t gfp,
1530 struct shmem_inode_info *info, pgoff_t index)
1531{
1532 struct page *oldpage, *newpage;
1533 struct address_space *swap_mapping;
1534 pgoff_t swap_index;
1535 int error;
1536
1537 oldpage = *pagep;
1538 swap_index = page_private(oldpage);
1539 swap_mapping = page_mapping(oldpage);
1540
1541 /*
1542 * We have arrived here because our zones are constrained, so don't
1543 * limit chance of success by further cpuset and node constraints.
1544 */
1545 gfp &= ~GFP_CONSTRAINT_MASK;
1546 newpage = shmem_alloc_page(gfp, info, index);
1547 if (!newpage)
1548 return -ENOMEM;
1549
1550 get_page(newpage);
1551 copy_highpage(newpage, oldpage);
1552 flush_dcache_page(newpage);
1553
1554 __SetPageLocked(newpage);
1555 __SetPageSwapBacked(newpage);
1556 SetPageUptodate(newpage);
1557 set_page_private(newpage, swap_index);
1558 SetPageSwapCache(newpage);
1559
1560 /*
1561 * Our caller will very soon move newpage out of swapcache, but it's
1562 * a nice clean interface for us to replace oldpage by newpage there.
1563 */
1564 xa_lock_irq(&swap_mapping->i_pages);
1565 error = shmem_radix_tree_replace(swap_mapping, swap_index, oldpage,
1566 newpage);
1567 if (!error) {
1568 __inc_node_page_state(newpage, NR_FILE_PAGES);
1569 __dec_node_page_state(oldpage, NR_FILE_PAGES);
1570 }
1571 xa_unlock_irq(&swap_mapping->i_pages);
1572
1573 if (unlikely(error)) {
1574 /*
1575 * Is this possible? I think not, now that our callers check
1576 * both PageSwapCache and page_private after getting page lock;
1577 * but be defensive. Reverse old to newpage for clear and free.
1578 */
1579 oldpage = newpage;
1580 } else {
1581 mem_cgroup_migrate(oldpage, newpage);
1582 lru_cache_add_anon(newpage);
1583 *pagep = newpage;
1584 }
1585
1586 ClearPageSwapCache(oldpage);
1587 set_page_private(oldpage, 0);
1588
1589 unlock_page(oldpage);
1590 put_page(oldpage);
1591 put_page(oldpage);
1592 return error;
1593}
1594
1595/*
1596 * shmem_getpage_gfp - find page in cache, or get from swap, or allocate
1597 *
1598 * If we allocate a new one we do not mark it dirty. That's up to the
1599 * vm. If we swap it in we mark it dirty since we also free the swap
1600 * entry since a page cannot live in both the swap and page cache.
1601 *
1602 * fault_mm and fault_type are only supplied by shmem_fault:
1603 * otherwise they are NULL.
1604 */
1605static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
1606 struct page **pagep, enum sgp_type sgp, gfp_t gfp,
1607 struct vm_area_struct *vma, struct vm_fault *vmf, int *fault_type)
1608{
1609 struct address_space *mapping = inode->i_mapping;
1610 struct shmem_inode_info *info = SHMEM_I(inode);
1611 struct shmem_sb_info *sbinfo;
1612 struct mm_struct *charge_mm;
1613 struct mem_cgroup *memcg;
1614 struct page *page;
1615 swp_entry_t swap;
1616 enum sgp_type sgp_huge = sgp;
1617 pgoff_t hindex = index;
1618 int error;
1619 int once = 0;
1620 int alloced = 0;
1621
1622 if (index > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
1623 return -EFBIG;
1624 if (sgp == SGP_NOHUGE || sgp == SGP_HUGE)
1625 sgp = SGP_CACHE;
1626repeat:
1627 swap.val = 0;
1628 page = find_lock_entry(mapping, index);
1629 if (radix_tree_exceptional_entry(page)) {
1630 swap = radix_to_swp_entry(page);
1631 page = NULL;
1632 }
1633
1634 if (sgp <= SGP_CACHE &&
1635 ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1636 error = -EINVAL;
1637 goto unlock;
1638 }
1639
1640 if (page && sgp == SGP_WRITE)
1641 mark_page_accessed(page);
1642
1643 /* fallocated page? */
1644 if (page && !PageUptodate(page)) {
1645 if (sgp != SGP_READ)
1646 goto clear;
1647 unlock_page(page);
1648 put_page(page);
1649 page = NULL;
1650 }
1651 if (page || (sgp == SGP_READ && !swap.val)) {
1652 *pagep = page;
1653 return 0;
1654 }
1655
1656 /*
1657 * Fast cache lookup did not find it:
1658 * bring it back from swap or allocate.
1659 */
1660 sbinfo = SHMEM_SB(inode->i_sb);
1661 charge_mm = vma ? vma->vm_mm : current->mm;
1662
1663 if (swap.val) {
1664 /* Look it up and read it in.. */
1665 page = lookup_swap_cache(swap, NULL, 0);
1666 if (!page) {
1667 /* Or update major stats only when swapin succeeds?? */
1668 if (fault_type) {
1669 *fault_type |= VM_FAULT_MAJOR;
1670 count_vm_event(PGMAJFAULT);
1671 count_memcg_event_mm(charge_mm, PGMAJFAULT);
1672 }
1673 /* Here we actually start the io */
1674 page = shmem_swapin(swap, gfp, info, index);
1675 if (!page) {
1676 error = -ENOMEM;
1677 goto failed;
1678 }
1679 }
1680
1681 /* We have to do this with page locked to prevent races */
1682 lock_page(page);
1683 if (!PageSwapCache(page) || page_private(page) != swap.val ||
1684 !shmem_confirm_swap(mapping, index, swap)) {
1685 error = -EEXIST; /* try again */
1686 goto unlock;
1687 }
1688 if (!PageUptodate(page)) {
1689 error = -EIO;
1690 goto failed;
1691 }
1692 wait_on_page_writeback(page);
1693
1694 if (shmem_should_replace_page(page, gfp)) {
1695 error = shmem_replace_page(&page, gfp, info, index);
1696 if (error)
1697 goto failed;
1698 }
1699
1700 error = mem_cgroup_try_charge(page, charge_mm, gfp, &memcg,
1701 false);
1702 if (!error) {
1703 error = shmem_add_to_page_cache(page, mapping, index,
1704 swp_to_radix_entry(swap));
1705 /*
1706 * We already confirmed swap under page lock, and make
1707 * no memory allocation here, so usually no possibility
1708 * of error; but free_swap_and_cache() only trylocks a
1709 * page, so it is just possible that the entry has been
1710 * truncated or holepunched since swap was confirmed.
1711 * shmem_undo_range() will have done some of the
1712 * unaccounting, now delete_from_swap_cache() will do
1713 * the rest.
1714 * Reset swap.val? No, leave it so "failed" goes back to
1715 * "repeat": reading a hole and writing should succeed.
1716 */
1717 if (error) {
1718 mem_cgroup_cancel_charge(page, memcg, false);
1719 delete_from_swap_cache(page);
1720 }
1721 }
1722 if (error)
1723 goto failed;
1724
1725 mem_cgroup_commit_charge(page, memcg, true, false);
1726
1727 spin_lock_irq(&info->lock);
1728 info->swapped--;
1729 shmem_recalc_inode(inode);
1730 spin_unlock_irq(&info->lock);
1731
1732 if (sgp == SGP_WRITE)
1733 mark_page_accessed(page);
1734
1735 delete_from_swap_cache(page);
1736 set_page_dirty(page);
1737 swap_free(swap);
1738
1739 } else {
1740 if (vma && userfaultfd_missing(vma)) {
1741 *fault_type = handle_userfault(vmf, VM_UFFD_MISSING);
1742 return 0;
1743 }
1744
1745 /* shmem_symlink() */
1746 if (mapping->a_ops != &shmem_aops)
1747 goto alloc_nohuge;
1748 if (shmem_huge == SHMEM_HUGE_DENY || sgp_huge == SGP_NOHUGE)
1749 goto alloc_nohuge;
1750 if (shmem_huge == SHMEM_HUGE_FORCE)
1751 goto alloc_huge;
1752 switch (sbinfo->huge) {
1753 loff_t i_size;
1754 pgoff_t off;
1755 case SHMEM_HUGE_NEVER:
1756 goto alloc_nohuge;
1757 case SHMEM_HUGE_WITHIN_SIZE:
1758 off = round_up(index, HPAGE_PMD_NR);
1759 i_size = round_up(i_size_read(inode), PAGE_SIZE);
1760 if (i_size >= HPAGE_PMD_SIZE &&
1761 i_size >> PAGE_SHIFT >= off)
1762 goto alloc_huge;
1763 /* fallthrough */
1764 case SHMEM_HUGE_ADVISE:
1765 if (sgp_huge == SGP_HUGE)
1766 goto alloc_huge;
1767 /* TODO: implement fadvise() hints */
1768 goto alloc_nohuge;
1769 }
1770
1771alloc_huge:
1772 page = shmem_alloc_and_acct_page(gfp, inode, index, true);
1773 if (IS_ERR(page)) {
1774alloc_nohuge: page = shmem_alloc_and_acct_page(gfp, inode,
1775 index, false);
1776 }
1777 if (IS_ERR(page)) {
1778 int retry = 5;
1779 error = PTR_ERR(page);
1780 page = NULL;
1781 if (error != -ENOSPC)
1782 goto failed;
1783 /*
1784 * Try to reclaim some spece by splitting a huge page
1785 * beyond i_size on the filesystem.
1786 */
1787 while (retry--) {
1788 int ret;
1789 ret = shmem_unused_huge_shrink(sbinfo, NULL, 1);
1790 if (ret == SHRINK_STOP)
1791 break;
1792 if (ret)
1793 goto alloc_nohuge;
1794 }
1795 goto failed;
1796 }
1797
1798 if (PageTransHuge(page))
1799 hindex = round_down(index, HPAGE_PMD_NR);
1800 else
1801 hindex = index;
1802
1803 if (sgp == SGP_WRITE)
1804 __SetPageReferenced(page);
1805
1806 error = mem_cgroup_try_charge(page, charge_mm, gfp, &memcg,
1807 PageTransHuge(page));
1808 if (error)
1809 goto unacct;
1810 error = radix_tree_maybe_preload_order(gfp & GFP_RECLAIM_MASK,
1811 compound_order(page));
1812 if (!error) {
1813 error = shmem_add_to_page_cache(page, mapping, hindex,
1814 NULL);
1815 radix_tree_preload_end();
1816 }
1817 if (error) {
1818 mem_cgroup_cancel_charge(page, memcg,
1819 PageTransHuge(page));
1820 goto unacct;
1821 }
1822 mem_cgroup_commit_charge(page, memcg, false,
1823 PageTransHuge(page));
1824 lru_cache_add_anon(page);
1825
1826 spin_lock_irq(&info->lock);
1827 info->alloced += 1 << compound_order(page);
1828 inode->i_blocks += BLOCKS_PER_PAGE << compound_order(page);
1829 shmem_recalc_inode(inode);
1830 spin_unlock_irq(&info->lock);
1831 alloced = true;
1832
1833 if (PageTransHuge(page) &&
1834 DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE) <
1835 hindex + HPAGE_PMD_NR - 1) {
1836 /*
1837 * Part of the huge page is beyond i_size: subject
1838 * to shrink under memory pressure.
1839 */
1840 spin_lock(&sbinfo->shrinklist_lock);
1841 /*
1842 * _careful to defend against unlocked access to
1843 * ->shrink_list in shmem_unused_huge_shrink()
1844 */
1845 if (list_empty_careful(&info->shrinklist)) {
1846 list_add_tail(&info->shrinklist,
1847 &sbinfo->shrinklist);
1848 sbinfo->shrinklist_len++;
1849 }
1850 spin_unlock(&sbinfo->shrinklist_lock);
1851 }
1852
1853 /*
1854 * Let SGP_FALLOC use the SGP_WRITE optimization on a new page.
1855 */
1856 if (sgp == SGP_FALLOC)
1857 sgp = SGP_WRITE;
1858clear:
1859 /*
1860 * Let SGP_WRITE caller clear ends if write does not fill page;
1861 * but SGP_FALLOC on a page fallocated earlier must initialize
1862 * it now, lest undo on failure cancel our earlier guarantee.
1863 */
1864 if (sgp != SGP_WRITE && !PageUptodate(page)) {
1865 struct page *head = compound_head(page);
1866 int i;
1867
1868 for (i = 0; i < (1 << compound_order(head)); i++) {
1869 clear_highpage(head + i);
1870 flush_dcache_page(head + i);
1871 }
1872 SetPageUptodate(head);
1873 }
1874 }
1875
1876 /* Perhaps the file has been truncated since we checked */
1877 if (sgp <= SGP_CACHE &&
1878 ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1879 if (alloced) {
1880 ClearPageDirty(page);
1881 delete_from_page_cache(page);
1882 spin_lock_irq(&info->lock);
1883 shmem_recalc_inode(inode);
1884 spin_unlock_irq(&info->lock);
1885 }
1886 error = -EINVAL;
1887 goto unlock;
1888 }
1889 *pagep = page + index - hindex;
1890 return 0;
1891
1892 /*
1893 * Error recovery.
1894 */
1895unacct:
1896 shmem_inode_unacct_blocks(inode, 1 << compound_order(page));
1897
1898 if (PageTransHuge(page)) {
1899 unlock_page(page);
1900 put_page(page);
1901 goto alloc_nohuge;
1902 }
1903failed:
1904 if (swap.val && !shmem_confirm_swap(mapping, index, swap))
1905 error = -EEXIST;
1906unlock:
1907 if (page) {
1908 unlock_page(page);
1909 put_page(page);
1910 }
1911 if (error == -ENOSPC && !once++) {
1912 spin_lock_irq(&info->lock);
1913 shmem_recalc_inode(inode);
1914 spin_unlock_irq(&info->lock);
1915 goto repeat;
1916 }
1917 if (error == -EEXIST) /* from above or from radix_tree_insert */
1918 goto repeat;
1919 return error;
1920}
1921
1922/*
1923 * This is like autoremove_wake_function, but it removes the wait queue
1924 * entry unconditionally - even if something else had already woken the
1925 * target.
1926 */
1927static int synchronous_wake_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
1928{
1929 int ret = default_wake_function(wait, mode, sync, key);
1930 list_del_init(&wait->entry);
1931 return ret;
1932}
1933
1934static int shmem_fault(struct vm_fault *vmf)
1935{
1936 struct vm_area_struct *vma = vmf->vma;
1937 struct inode *inode = file_inode(vma->vm_file);
1938 gfp_t gfp = mapping_gfp_mask(inode->i_mapping);
1939 enum sgp_type sgp;
1940 int error;
1941 int ret = VM_FAULT_LOCKED;
1942
1943 /*
1944 * Trinity finds that probing a hole which tmpfs is punching can
1945 * prevent the hole-punch from ever completing: which in turn
1946 * locks writers out with its hold on i_mutex. So refrain from
1947 * faulting pages into the hole while it's being punched. Although
1948 * shmem_undo_range() does remove the additions, it may be unable to
1949 * keep up, as each new page needs its own unmap_mapping_range() call,
1950 * and the i_mmap tree grows ever slower to scan if new vmas are added.
1951 *
1952 * It does not matter if we sometimes reach this check just before the
1953 * hole-punch begins, so that one fault then races with the punch:
1954 * we just need to make racing faults a rare case.
1955 *
1956 * The implementation below would be much simpler if we just used a
1957 * standard mutex or completion: but we cannot take i_mutex in fault,
1958 * and bloating every shmem inode for this unlikely case would be sad.
1959 */
1960 if (unlikely(inode->i_private)) {
1961 struct shmem_falloc *shmem_falloc;
1962
1963 spin_lock(&inode->i_lock);
1964 shmem_falloc = inode->i_private;
1965 if (shmem_falloc &&
1966 shmem_falloc->waitq &&
1967 vmf->pgoff >= shmem_falloc->start &&
1968 vmf->pgoff < shmem_falloc->next) {
1969 wait_queue_head_t *shmem_falloc_waitq;
1970 DEFINE_WAIT_FUNC(shmem_fault_wait, synchronous_wake_function);
1971
1972 ret = VM_FAULT_NOPAGE;
1973 if ((vmf->flags & FAULT_FLAG_ALLOW_RETRY) &&
1974 !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
1975 /* It's polite to up mmap_sem if we can */
1976 up_read(&vma->vm_mm->mmap_sem);
1977 ret = VM_FAULT_RETRY;
1978 }
1979
1980 shmem_falloc_waitq = shmem_falloc->waitq;
1981 prepare_to_wait(shmem_falloc_waitq, &shmem_fault_wait,
1982 TASK_UNINTERRUPTIBLE);
1983 spin_unlock(&inode->i_lock);
1984 schedule();
1985
1986 /*
1987 * shmem_falloc_waitq points into the shmem_fallocate()
1988 * stack of the hole-punching task: shmem_falloc_waitq
1989 * is usually invalid by the time we reach here, but
1990 * finish_wait() does not dereference it in that case;
1991 * though i_lock needed lest racing with wake_up_all().
1992 */
1993 spin_lock(&inode->i_lock);
1994 finish_wait(shmem_falloc_waitq, &shmem_fault_wait);
1995 spin_unlock(&inode->i_lock);
1996 return ret;
1997 }
1998 spin_unlock(&inode->i_lock);
1999 }
2000
2001 sgp = SGP_CACHE;
2002
2003 if ((vma->vm_flags & VM_NOHUGEPAGE) ||
2004 test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
2005 sgp = SGP_NOHUGE;
2006 else if (vma->vm_flags & VM_HUGEPAGE)
2007 sgp = SGP_HUGE;
2008
2009 error = shmem_getpage_gfp(inode, vmf->pgoff, &vmf->page, sgp,
2010 gfp, vma, vmf, &ret);
2011 if (error)
2012 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
2013 return ret;
2014}
2015
2016unsigned long shmem_get_unmapped_area(struct file *file,
2017 unsigned long uaddr, unsigned long len,
2018 unsigned long pgoff, unsigned long flags)
2019{
2020 unsigned long (*get_area)(struct file *,
2021 unsigned long, unsigned long, unsigned long, unsigned long);
2022 unsigned long addr;
2023 unsigned long offset;
2024 unsigned long inflated_len;
2025 unsigned long inflated_addr;
2026 unsigned long inflated_offset;
2027
2028 if (len > TASK_SIZE)
2029 return -ENOMEM;
2030
2031 get_area = current->mm->get_unmapped_area;
2032 addr = get_area(file, uaddr, len, pgoff, flags);
2033
2034 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
2035 return addr;
2036 if (IS_ERR_VALUE(addr))
2037 return addr;
2038 if (addr & ~PAGE_MASK)
2039 return addr;
2040 if (addr > TASK_SIZE - len)
2041 return addr;
2042
2043 if (shmem_huge == SHMEM_HUGE_DENY)
2044 return addr;
2045 if (len < HPAGE_PMD_SIZE)
2046 return addr;
2047 if (flags & MAP_FIXED)
2048 return addr;
2049 /*
2050 * Our priority is to support MAP_SHARED mapped hugely;
2051 * and support MAP_PRIVATE mapped hugely too, until it is COWed.
2052 * But if caller specified an address hint, respect that as before.
2053 */
2054 if (uaddr)
2055 return addr;
2056
2057 if (shmem_huge != SHMEM_HUGE_FORCE) {
2058 struct super_block *sb;
2059
2060 if (file) {
2061 VM_BUG_ON(file->f_op != &shmem_file_operations);
2062 sb = file_inode(file)->i_sb;
2063 } else {
2064 /*
2065 * Called directly from mm/mmap.c, or drivers/char/mem.c
2066 * for "/dev/zero", to create a shared anonymous object.
2067 */
2068 if (IS_ERR(shm_mnt))
2069 return addr;
2070 sb = shm_mnt->mnt_sb;
2071 }
2072 if (SHMEM_SB(sb)->huge == SHMEM_HUGE_NEVER)
2073 return addr;
2074 }
2075
2076 offset = (pgoff << PAGE_SHIFT) & (HPAGE_PMD_SIZE-1);
2077 if (offset && offset + len < 2 * HPAGE_PMD_SIZE)
2078 return addr;
2079 if ((addr & (HPAGE_PMD_SIZE-1)) == offset)
2080 return addr;
2081
2082 inflated_len = len + HPAGE_PMD_SIZE - PAGE_SIZE;
2083 if (inflated_len > TASK_SIZE)
2084 return addr;
2085 if (inflated_len < len)
2086 return addr;
2087
2088 inflated_addr = get_area(NULL, 0, inflated_len, 0, flags);
2089 if (IS_ERR_VALUE(inflated_addr))
2090 return addr;
2091 if (inflated_addr & ~PAGE_MASK)
2092 return addr;
2093
2094 inflated_offset = inflated_addr & (HPAGE_PMD_SIZE-1);
2095 inflated_addr += offset - inflated_offset;
2096 if (inflated_offset > offset)
2097 inflated_addr += HPAGE_PMD_SIZE;
2098
2099 if (inflated_addr > TASK_SIZE - len)
2100 return addr;
2101 return inflated_addr;
2102}
2103
2104#ifdef CONFIG_NUMA
2105static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
2106{
2107 struct inode *inode = file_inode(vma->vm_file);
2108 return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
2109}
2110
2111static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
2112 unsigned long addr)
2113{
2114 struct inode *inode = file_inode(vma->vm_file);
2115 pgoff_t index;
2116
2117 index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2118 return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
2119}
2120#endif
2121
2122int shmem_lock(struct file *file, int lock, struct user_struct *user)
2123{
2124 struct inode *inode = file_inode(file);
2125 struct shmem_inode_info *info = SHMEM_I(inode);
2126 int retval = -ENOMEM;
2127
2128 spin_lock_irq(&info->lock);
2129 if (lock && !(info->flags & VM_LOCKED)) {
2130 if (!user_shm_lock(inode->i_size, user))
2131 goto out_nomem;
2132 info->flags |= VM_LOCKED;
2133 mapping_set_unevictable(file->f_mapping);
2134 }
2135 if (!lock && (info->flags & VM_LOCKED) && user) {
2136 user_shm_unlock(inode->i_size, user);
2137 info->flags &= ~VM_LOCKED;
2138 mapping_clear_unevictable(file->f_mapping);
2139 }
2140 retval = 0;
2141
2142out_nomem:
2143 spin_unlock_irq(&info->lock);
2144 return retval;
2145}
2146
2147static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
2148{
2149 file_accessed(file);
2150 vma->vm_ops = &shmem_vm_ops;
2151 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE) &&
2152 ((vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK) <
2153 (vma->vm_end & HPAGE_PMD_MASK)) {
2154 khugepaged_enter(vma, vma->vm_flags);
2155 }
2156 return 0;
2157}
2158
2159static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
2160 umode_t mode, dev_t dev, unsigned long flags)
2161{
2162 struct inode *inode;
2163 struct shmem_inode_info *info;
2164 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2165
2166 if (shmem_reserve_inode(sb))
2167 return NULL;
2168
2169 inode = new_inode(sb);
2170 if (inode) {
2171 inode->i_ino = get_next_ino();
2172 inode_init_owner(inode, dir, mode);
2173 inode->i_blocks = 0;
2174 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
2175 inode->i_generation = get_seconds();
2176 info = SHMEM_I(inode);
2177 memset(info, 0, (char *)inode - (char *)info);
2178 spin_lock_init(&info->lock);
2179 info->seals = F_SEAL_SEAL;
2180 info->flags = flags & VM_NORESERVE;
2181 INIT_LIST_HEAD(&info->shrinklist);
2182 INIT_LIST_HEAD(&info->swaplist);
2183 simple_xattrs_init(&info->xattrs);
2184 cache_no_acl(inode);
2185
2186 switch (mode & S_IFMT) {
2187 default:
2188 inode->i_op = &shmem_special_inode_operations;
2189 init_special_inode(inode, mode, dev);
2190 break;
2191 case S_IFREG:
2192 inode->i_mapping->a_ops = &shmem_aops;
2193 inode->i_op = &shmem_inode_operations;
2194 inode->i_fop = &shmem_file_operations;
2195 mpol_shared_policy_init(&info->policy,
2196 shmem_get_sbmpol(sbinfo));
2197 break;
2198 case S_IFDIR:
2199 inc_nlink(inode);
2200 /* Some things misbehave if size == 0 on a directory */
2201 inode->i_size = 2 * BOGO_DIRENT_SIZE;
2202 inode->i_op = &shmem_dir_inode_operations;
2203 inode->i_fop = &simple_dir_operations;
2204 break;
2205 case S_IFLNK:
2206 /*
2207 * Must not load anything in the rbtree,
2208 * mpol_free_shared_policy will not be called.
2209 */
2210 mpol_shared_policy_init(&info->policy, NULL);
2211 break;
2212 }
2213 } else
2214 shmem_free_inode(sb);
2215 return inode;
2216}
2217
2218bool shmem_mapping(struct address_space *mapping)
2219{
2220 return mapping->a_ops == &shmem_aops;
2221}
2222
2223static int shmem_mfill_atomic_pte(struct mm_struct *dst_mm,
2224 pmd_t *dst_pmd,
2225 struct vm_area_struct *dst_vma,
2226 unsigned long dst_addr,
2227 unsigned long src_addr,
2228 bool zeropage,
2229 struct page **pagep)
2230{
2231 struct inode *inode = file_inode(dst_vma->vm_file);
2232 struct shmem_inode_info *info = SHMEM_I(inode);
2233 struct address_space *mapping = inode->i_mapping;
2234 gfp_t gfp = mapping_gfp_mask(mapping);
2235 pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
2236 struct mem_cgroup *memcg;
2237 spinlock_t *ptl;
2238 void *page_kaddr;
2239 struct page *page;
2240 pte_t _dst_pte, *dst_pte;
2241 int ret;
2242
2243 ret = -ENOMEM;
2244 if (!shmem_inode_acct_block(inode, 1))
2245 goto out;
2246
2247 if (!*pagep) {
2248 page = shmem_alloc_page(gfp, info, pgoff);
2249 if (!page)
2250 goto out_unacct_blocks;
2251
2252 if (!zeropage) { /* mcopy_atomic */
2253 page_kaddr = kmap_atomic(page);
2254 ret = copy_from_user(page_kaddr,
2255 (const void __user *)src_addr,
2256 PAGE_SIZE);
2257 kunmap_atomic(page_kaddr);
2258
2259 /* fallback to copy_from_user outside mmap_sem */
2260 if (unlikely(ret)) {
2261 *pagep = page;
2262 shmem_inode_unacct_blocks(inode, 1);
2263 /* don't free the page */
2264 return -EFAULT;
2265 }
2266 } else { /* mfill_zeropage_atomic */
2267 clear_highpage(page);
2268 }
2269 } else {
2270 page = *pagep;
2271 *pagep = NULL;
2272 }
2273
2274 VM_BUG_ON(PageLocked(page) || PageSwapBacked(page));
2275 __SetPageLocked(page);
2276 __SetPageSwapBacked(page);
2277 __SetPageUptodate(page);
2278
2279 ret = mem_cgroup_try_charge(page, dst_mm, gfp, &memcg, false);
2280 if (ret)
2281 goto out_release;
2282
2283 ret = radix_tree_maybe_preload(gfp & GFP_RECLAIM_MASK);
2284 if (!ret) {
2285 ret = shmem_add_to_page_cache(page, mapping, pgoff, NULL);
2286 radix_tree_preload_end();
2287 }
2288 if (ret)
2289 goto out_release_uncharge;
2290
2291 mem_cgroup_commit_charge(page, memcg, false, false);
2292
2293 _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
2294 if (dst_vma->vm_flags & VM_WRITE)
2295 _dst_pte = pte_mkwrite(pte_mkdirty(_dst_pte));
2296
2297 ret = -EEXIST;
2298 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
2299 if (!pte_none(*dst_pte))
2300 goto out_release_uncharge_unlock;
2301
2302 lru_cache_add_anon(page);
2303
2304 spin_lock(&info->lock);
2305 info->alloced++;
2306 inode->i_blocks += BLOCKS_PER_PAGE;
2307 shmem_recalc_inode(inode);
2308 spin_unlock(&info->lock);
2309
2310 inc_mm_counter(dst_mm, mm_counter_file(page));
2311 page_add_file_rmap(page, false);
2312 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
2313
2314 /* No need to invalidate - it was non-present before */
2315 update_mmu_cache(dst_vma, dst_addr, dst_pte);
2316 unlock_page(page);
2317 pte_unmap_unlock(dst_pte, ptl);
2318 ret = 0;
2319out:
2320 return ret;
2321out_release_uncharge_unlock:
2322 pte_unmap_unlock(dst_pte, ptl);
2323out_release_uncharge:
2324 mem_cgroup_cancel_charge(page, memcg, false);
2325out_release:
2326 unlock_page(page);
2327 put_page(page);
2328out_unacct_blocks:
2329 shmem_inode_unacct_blocks(inode, 1);
2330 goto out;
2331}
2332
2333int shmem_mcopy_atomic_pte(struct mm_struct *dst_mm,
2334 pmd_t *dst_pmd,
2335 struct vm_area_struct *dst_vma,
2336 unsigned long dst_addr,
2337 unsigned long src_addr,
2338 struct page **pagep)
2339{
2340 return shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
2341 dst_addr, src_addr, false, pagep);
2342}
2343
2344int shmem_mfill_zeropage_pte(struct mm_struct *dst_mm,
2345 pmd_t *dst_pmd,
2346 struct vm_area_struct *dst_vma,
2347 unsigned long dst_addr)
2348{
2349 struct page *page = NULL;
2350
2351 return shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
2352 dst_addr, 0, true, &page);
2353}
2354
2355#ifdef CONFIG_TMPFS
2356static const struct inode_operations shmem_symlink_inode_operations;
2357static const struct inode_operations shmem_short_symlink_operations;
2358
2359#ifdef CONFIG_TMPFS_XATTR
2360static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
2361#else
2362#define shmem_initxattrs NULL
2363#endif
2364
2365static int
2366shmem_write_begin(struct file *file, struct address_space *mapping,
2367 loff_t pos, unsigned len, unsigned flags,
2368 struct page **pagep, void **fsdata)
2369{
2370 struct inode *inode = mapping->host;
2371 struct shmem_inode_info *info = SHMEM_I(inode);
2372 pgoff_t index = pos >> PAGE_SHIFT;
2373
2374 /* i_mutex is held by caller */
2375 if (unlikely(info->seals & (F_SEAL_WRITE | F_SEAL_GROW))) {
2376 if (info->seals & F_SEAL_WRITE)
2377 return -EPERM;
2378 if ((info->seals & F_SEAL_GROW) && pos + len > inode->i_size)
2379 return -EPERM;
2380 }
2381
2382 return shmem_getpage(inode, index, pagep, SGP_WRITE);
2383}
2384
2385static int
2386shmem_write_end(struct file *file, struct address_space *mapping,
2387 loff_t pos, unsigned len, unsigned copied,
2388 struct page *page, void *fsdata)
2389{
2390 struct inode *inode = mapping->host;
2391
2392 if (pos + copied > inode->i_size)
2393 i_size_write(inode, pos + copied);
2394
2395 if (!PageUptodate(page)) {
2396 struct page *head = compound_head(page);
2397 if (PageTransCompound(page)) {
2398 int i;
2399
2400 for (i = 0; i < HPAGE_PMD_NR; i++) {
2401 if (head + i == page)
2402 continue;
2403 clear_highpage(head + i);
2404 flush_dcache_page(head + i);
2405 }
2406 }
2407 if (copied < PAGE_SIZE) {
2408 unsigned from = pos & (PAGE_SIZE - 1);
2409 zero_user_segments(page, 0, from,
2410 from + copied, PAGE_SIZE);
2411 }
2412 SetPageUptodate(head);
2413 }
2414 set_page_dirty(page);
2415 unlock_page(page);
2416 put_page(page);
2417
2418 return copied;
2419}
2420
2421static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
2422{
2423 struct file *file = iocb->ki_filp;
2424 struct inode *inode = file_inode(file);
2425 struct address_space *mapping = inode->i_mapping;
2426 pgoff_t index;
2427 unsigned long offset;
2428 enum sgp_type sgp = SGP_READ;
2429 int error = 0;
2430 ssize_t retval = 0;
2431 loff_t *ppos = &iocb->ki_pos;
2432
2433 /*
2434 * Might this read be for a stacking filesystem? Then when reading
2435 * holes of a sparse file, we actually need to allocate those pages,
2436 * and even mark them dirty, so it cannot exceed the max_blocks limit.
2437 */
2438 if (!iter_is_iovec(to))
2439 sgp = SGP_CACHE;
2440
2441 index = *ppos >> PAGE_SHIFT;
2442 offset = *ppos & ~PAGE_MASK;
2443
2444 for (;;) {
2445 struct page *page = NULL;
2446 pgoff_t end_index;
2447 unsigned long nr, ret;
2448 loff_t i_size = i_size_read(inode);
2449
2450 end_index = i_size >> PAGE_SHIFT;
2451 if (index > end_index)
2452 break;
2453 if (index == end_index) {
2454 nr = i_size & ~PAGE_MASK;
2455 if (nr <= offset)
2456 break;
2457 }
2458
2459 error = shmem_getpage(inode, index, &page, sgp);
2460 if (error) {
2461 if (error == -EINVAL)
2462 error = 0;
2463 break;
2464 }
2465 if (page) {
2466 if (sgp == SGP_CACHE)
2467 set_page_dirty(page);
2468 unlock_page(page);
2469 }
2470
2471 /*
2472 * We must evaluate after, since reads (unlike writes)
2473 * are called without i_mutex protection against truncate
2474 */
2475 nr = PAGE_SIZE;
2476 i_size = i_size_read(inode);
2477 end_index = i_size >> PAGE_SHIFT;
2478 if (index == end_index) {
2479 nr = i_size & ~PAGE_MASK;
2480 if (nr <= offset) {
2481 if (page)
2482 put_page(page);
2483 break;
2484 }
2485 }
2486 nr -= offset;
2487
2488 if (page) {
2489 /*
2490 * If users can be writing to this page using arbitrary
2491 * virtual addresses, take care about potential aliasing
2492 * before reading the page on the kernel side.
2493 */
2494 if (mapping_writably_mapped(mapping))
2495 flush_dcache_page(page);
2496 /*
2497 * Mark the page accessed if we read the beginning.
2498 */
2499 if (!offset)
2500 mark_page_accessed(page);
2501 } else {
2502 page = ZERO_PAGE(0);
2503 get_page(page);
2504 }
2505
2506 /*
2507 * Ok, we have the page, and it's up-to-date, so
2508 * now we can copy it to user space...
2509 */
2510 ret = copy_page_to_iter(page, offset, nr, to);
2511 retval += ret;
2512 offset += ret;
2513 index += offset >> PAGE_SHIFT;
2514 offset &= ~PAGE_MASK;
2515
2516 put_page(page);
2517 if (!iov_iter_count(to))
2518 break;
2519 if (ret < nr) {
2520 error = -EFAULT;
2521 break;
2522 }
2523 cond_resched();
2524 }
2525
2526 *ppos = ((loff_t) index << PAGE_SHIFT) + offset;
2527 file_accessed(file);
2528 return retval ? retval : error;
2529}
2530
2531/*
2532 * llseek SEEK_DATA or SEEK_HOLE through the radix_tree.
2533 */
2534static pgoff_t shmem_seek_hole_data(struct address_space *mapping,
2535 pgoff_t index, pgoff_t end, int whence)
2536{
2537 struct page *page;
2538 struct pagevec pvec;
2539 pgoff_t indices[PAGEVEC_SIZE];
2540 bool done = false;
2541 int i;
2542
2543 pagevec_init(&pvec);
2544 pvec.nr = 1; /* start small: we may be there already */
2545 while (!done) {
2546 pvec.nr = find_get_entries(mapping, index,
2547 pvec.nr, pvec.pages, indices);
2548 if (!pvec.nr) {
2549 if (whence == SEEK_DATA)
2550 index = end;
2551 break;
2552 }
2553 for (i = 0; i < pvec.nr; i++, index++) {
2554 if (index < indices[i]) {
2555 if (whence == SEEK_HOLE) {
2556 done = true;
2557 break;
2558 }
2559 index = indices[i];
2560 }
2561 page = pvec.pages[i];
2562 if (page && !radix_tree_exceptional_entry(page)) {
2563 if (!PageUptodate(page))
2564 page = NULL;
2565 }
2566 if (index >= end ||
2567 (page && whence == SEEK_DATA) ||
2568 (!page && whence == SEEK_HOLE)) {
2569 done = true;
2570 break;
2571 }
2572 }
2573 pagevec_remove_exceptionals(&pvec);
2574 pagevec_release(&pvec);
2575 pvec.nr = PAGEVEC_SIZE;
2576 cond_resched();
2577 }
2578 return index;
2579}
2580
2581static loff_t shmem_file_llseek(struct file *file, loff_t offset, int whence)
2582{
2583 struct address_space *mapping = file->f_mapping;
2584 struct inode *inode = mapping->host;
2585 pgoff_t start, end;
2586 loff_t new_offset;
2587
2588 if (whence != SEEK_DATA && whence != SEEK_HOLE)
2589 return generic_file_llseek_size(file, offset, whence,
2590 MAX_LFS_FILESIZE, i_size_read(inode));
2591 inode_lock(inode);
2592 /* We're holding i_mutex so we can access i_size directly */
2593
2594 if (offset < 0)
2595 offset = -EINVAL;
2596 else if (offset >= inode->i_size)
2597 offset = -ENXIO;
2598 else {
2599 start = offset >> PAGE_SHIFT;
2600 end = (inode->i_size + PAGE_SIZE - 1) >> PAGE_SHIFT;
2601 new_offset = shmem_seek_hole_data(mapping, start, end, whence);
2602 new_offset <<= PAGE_SHIFT;
2603 if (new_offset > offset) {
2604 if (new_offset < inode->i_size)
2605 offset = new_offset;
2606 else if (whence == SEEK_DATA)
2607 offset = -ENXIO;
2608 else
2609 offset = inode->i_size;
2610 }
2611 }
2612
2613 if (offset >= 0)
2614 offset = vfs_setpos(file, offset, MAX_LFS_FILESIZE);
2615 inode_unlock(inode);
2616 return offset;
2617}
2618
2619/*
2620 * We need a tag: a new tag would expand every radix_tree_node by 8 bytes,
2621 * so reuse a tag which we firmly believe is never set or cleared on shmem.
2622 */
2623#define SHMEM_TAG_PINNED PAGECACHE_TAG_TOWRITE
2624#define LAST_SCAN 4 /* about 150ms max */
2625
2626static void shmem_tag_pins(struct address_space *mapping)
2627{
2628 struct radix_tree_iter iter;
2629 void **slot;
2630 pgoff_t start;
2631 struct page *page;
2632
2633 lru_add_drain();
2634 start = 0;
2635 rcu_read_lock();
2636
2637 radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start) {
2638 page = radix_tree_deref_slot(slot);
2639 if (!page || radix_tree_exception(page)) {
2640 if (radix_tree_deref_retry(page)) {
2641 slot = radix_tree_iter_retry(&iter);
2642 continue;
2643 }
2644 } else if (page_count(page) - page_mapcount(page) > 1) {
2645 xa_lock_irq(&mapping->i_pages);
2646 radix_tree_tag_set(&mapping->i_pages, iter.index,
2647 SHMEM_TAG_PINNED);
2648 xa_unlock_irq(&mapping->i_pages);
2649 }
2650
2651 if (need_resched()) {
2652 slot = radix_tree_iter_resume(slot, &iter);
2653 cond_resched_rcu();
2654 }
2655 }
2656 rcu_read_unlock();
2657}
2658
2659/*
2660 * Setting SEAL_WRITE requires us to verify there's no pending writer. However,
2661 * via get_user_pages(), drivers might have some pending I/O without any active
2662 * user-space mappings (eg., direct-IO, AIO). Therefore, we look at all pages
2663 * and see whether it has an elevated ref-count. If so, we tag them and wait for
2664 * them to be dropped.
2665 * The caller must guarantee that no new user will acquire writable references
2666 * to those pages to avoid races.
2667 */
2668static int shmem_wait_for_pins(struct address_space *mapping)
2669{
2670 struct radix_tree_iter iter;
2671 void **slot;
2672 pgoff_t start;
2673 struct page *page;
2674 int error, scan;
2675
2676 shmem_tag_pins(mapping);
2677
2678 error = 0;
2679 for (scan = 0; scan <= LAST_SCAN; scan++) {
2680 if (!radix_tree_tagged(&mapping->i_pages, SHMEM_TAG_PINNED))
2681 break;
2682
2683 if (!scan)
2684 lru_add_drain_all();
2685 else if (schedule_timeout_killable((HZ << scan) / 200))
2686 scan = LAST_SCAN;
2687
2688 start = 0;
2689 rcu_read_lock();
2690 radix_tree_for_each_tagged(slot, &mapping->i_pages, &iter,
2691 start, SHMEM_TAG_PINNED) {
2692
2693 page = radix_tree_deref_slot(slot);
2694 if (radix_tree_exception(page)) {
2695 if (radix_tree_deref_retry(page)) {
2696 slot = radix_tree_iter_retry(&iter);
2697 continue;
2698 }
2699
2700 page = NULL;
2701 }
2702
2703 if (page &&
2704 page_count(page) - page_mapcount(page) != 1) {
2705 if (scan < LAST_SCAN)
2706 goto continue_resched;
2707
2708 /*
2709 * On the last scan, we clean up all those tags
2710 * we inserted; but make a note that we still
2711 * found pages pinned.
2712 */
2713 error = -EBUSY;
2714 }
2715
2716 xa_lock_irq(&mapping->i_pages);
2717 radix_tree_tag_clear(&mapping->i_pages,
2718 iter.index, SHMEM_TAG_PINNED);
2719 xa_unlock_irq(&mapping->i_pages);
2720continue_resched:
2721 if (need_resched()) {
2722 slot = radix_tree_iter_resume(slot, &iter);
2723 cond_resched_rcu();
2724 }
2725 }
2726 rcu_read_unlock();
2727 }
2728
2729 return error;
2730}
2731
2732static unsigned int *memfd_file_seals_ptr(struct file *file)
2733{
2734 if (file->f_op == &shmem_file_operations)
2735 return &SHMEM_I(file_inode(file))->seals;
2736
2737#ifdef CONFIG_HUGETLBFS
2738 if (file->f_op == &hugetlbfs_file_operations)
2739 return &HUGETLBFS_I(file_inode(file))->seals;
2740#endif
2741
2742 return NULL;
2743}
2744
2745#define F_ALL_SEALS (F_SEAL_SEAL | \
2746 F_SEAL_SHRINK | \
2747 F_SEAL_GROW | \
2748 F_SEAL_WRITE)
2749
2750static int memfd_add_seals(struct file *file, unsigned int seals)
2751{
2752 struct inode *inode = file_inode(file);
2753 unsigned int *file_seals;
2754 int error;
2755
2756 /*
2757 * SEALING
2758 * Sealing allows multiple parties to share a shmem-file but restrict
2759 * access to a specific subset of file operations. Seals can only be
2760 * added, but never removed. This way, mutually untrusted parties can
2761 * share common memory regions with a well-defined policy. A malicious
2762 * peer can thus never perform unwanted operations on a shared object.
2763 *
2764 * Seals are only supported on special shmem-files and always affect
2765 * the whole underlying inode. Once a seal is set, it may prevent some
2766 * kinds of access to the file. Currently, the following seals are
2767 * defined:
2768 * SEAL_SEAL: Prevent further seals from being set on this file
2769 * SEAL_SHRINK: Prevent the file from shrinking
2770 * SEAL_GROW: Prevent the file from growing
2771 * SEAL_WRITE: Prevent write access to the file
2772 *
2773 * As we don't require any trust relationship between two parties, we
2774 * must prevent seals from being removed. Therefore, sealing a file
2775 * only adds a given set of seals to the file, it never touches
2776 * existing seals. Furthermore, the "setting seals"-operation can be
2777 * sealed itself, which basically prevents any further seal from being
2778 * added.
2779 *
2780 * Semantics of sealing are only defined on volatile files. Only
2781 * anonymous shmem files support sealing. More importantly, seals are
2782 * never written to disk. Therefore, there's no plan to support it on
2783 * other file types.
2784 */
2785
2786 if (!(file->f_mode & FMODE_WRITE))
2787 return -EPERM;
2788 if (seals & ~(unsigned int)F_ALL_SEALS)
2789 return -EINVAL;
2790
2791 inode_lock(inode);
2792
2793 file_seals = memfd_file_seals_ptr(file);
2794 if (!file_seals) {
2795 error = -EINVAL;
2796 goto unlock;
2797 }
2798
2799 if (*file_seals & F_SEAL_SEAL) {
2800 error = -EPERM;
2801 goto unlock;
2802 }
2803
2804 if ((seals & F_SEAL_WRITE) && !(*file_seals & F_SEAL_WRITE)) {
2805 error = mapping_deny_writable(file->f_mapping);
2806 if (error)
2807 goto unlock;
2808
2809 error = shmem_wait_for_pins(file->f_mapping);
2810 if (error) {
2811 mapping_allow_writable(file->f_mapping);
2812 goto unlock;
2813 }
2814 }
2815
2816 *file_seals |= seals;
2817 error = 0;
2818
2819unlock:
2820 inode_unlock(inode);
2821 return error;
2822}
2823
2824static int memfd_get_seals(struct file *file)
2825{
2826 unsigned int *seals = memfd_file_seals_ptr(file);
2827
2828 return seals ? *seals : -EINVAL;
2829}
2830
2831long memfd_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
2832{
2833 long error;
2834
2835 switch (cmd) {
2836 case F_ADD_SEALS:
2837 /* disallow upper 32bit */
2838 if (arg > UINT_MAX)
2839 return -EINVAL;
2840
2841 error = memfd_add_seals(file, arg);
2842 break;
2843 case F_GET_SEALS:
2844 error = memfd_get_seals(file);
2845 break;
2846 default:
2847 error = -EINVAL;
2848 break;
2849 }
2850
2851 return error;
2852}
2853
2854static long shmem_fallocate(struct file *file, int mode, loff_t offset,
2855 loff_t len)
2856{
2857 struct inode *inode = file_inode(file);
2858 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
2859 struct shmem_inode_info *info = SHMEM_I(inode);
2860 struct shmem_falloc shmem_falloc;
2861 pgoff_t start, index, end;
2862 int error;
2863
2864 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2865 return -EOPNOTSUPP;
2866
2867 inode_lock(inode);
2868
2869 if (mode & FALLOC_FL_PUNCH_HOLE) {
2870 struct address_space *mapping = file->f_mapping;
2871 loff_t unmap_start = round_up(offset, PAGE_SIZE);
2872 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
2873 DECLARE_WAIT_QUEUE_HEAD_ONSTACK(shmem_falloc_waitq);
2874
2875 /* protected by i_mutex */
2876 if (info->seals & F_SEAL_WRITE) {
2877 error = -EPERM;
2878 goto out;
2879 }
2880
2881 shmem_falloc.waitq = &shmem_falloc_waitq;
2882 shmem_falloc.start = unmap_start >> PAGE_SHIFT;
2883 shmem_falloc.next = (unmap_end + 1) >> PAGE_SHIFT;
2884 spin_lock(&inode->i_lock);
2885 inode->i_private = &shmem_falloc;
2886 spin_unlock(&inode->i_lock);
2887
2888 if ((u64)unmap_end > (u64)unmap_start)
2889 unmap_mapping_range(mapping, unmap_start,
2890 1 + unmap_end - unmap_start, 0);
2891 shmem_truncate_range(inode, offset, offset + len - 1);
2892 /* No need to unmap again: hole-punching leaves COWed pages */
2893
2894 spin_lock(&inode->i_lock);
2895 inode->i_private = NULL;
2896 wake_up_all(&shmem_falloc_waitq);
2897 WARN_ON_ONCE(!list_empty(&shmem_falloc_waitq.head));
2898 spin_unlock(&inode->i_lock);
2899 error = 0;
2900 goto out;
2901 }
2902
2903 /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
2904 error = inode_newsize_ok(inode, offset + len);
2905 if (error)
2906 goto out;
2907
2908 if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
2909 error = -EPERM;
2910 goto out;
2911 }
2912
2913 start = offset >> PAGE_SHIFT;
2914 end = (offset + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
2915 /* Try to avoid a swapstorm if len is impossible to satisfy */
2916 if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
2917 error = -ENOSPC;
2918 goto out;
2919 }
2920
2921 shmem_falloc.waitq = NULL;
2922 shmem_falloc.start = start;
2923 shmem_falloc.next = start;
2924 shmem_falloc.nr_falloced = 0;
2925 shmem_falloc.nr_unswapped = 0;
2926 spin_lock(&inode->i_lock);
2927 inode->i_private = &shmem_falloc;
2928 spin_unlock(&inode->i_lock);
2929
2930 for (index = start; index < end; index++) {
2931 struct page *page;
2932
2933 /*
2934 * Good, the fallocate(2) manpage permits EINTR: we may have
2935 * been interrupted because we are using up too much memory.
2936 */
2937 if (signal_pending(current))
2938 error = -EINTR;
2939 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
2940 error = -ENOMEM;
2941 else
2942 error = shmem_getpage(inode, index, &page, SGP_FALLOC);
2943 if (error) {
2944 /* Remove the !PageUptodate pages we added */
2945 if (index > start) {
2946 shmem_undo_range(inode,
2947 (loff_t)start << PAGE_SHIFT,
2948 ((loff_t)index << PAGE_SHIFT) - 1, true);
2949 }
2950 goto undone;
2951 }
2952
2953 /*
2954 * Inform shmem_writepage() how far we have reached.
2955 * No need for lock or barrier: we have the page lock.
2956 */
2957 shmem_falloc.next++;
2958 if (!PageUptodate(page))
2959 shmem_falloc.nr_falloced++;
2960
2961 /*
2962 * If !PageUptodate, leave it that way so that freeable pages
2963 * can be recognized if we need to rollback on error later.
2964 * But set_page_dirty so that memory pressure will swap rather
2965 * than free the pages we are allocating (and SGP_CACHE pages
2966 * might still be clean: we now need to mark those dirty too).
2967 */
2968 set_page_dirty(page);
2969 unlock_page(page);
2970 put_page(page);
2971 cond_resched();
2972 }
2973
2974 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
2975 i_size_write(inode, offset + len);
2976 inode->i_ctime = current_time(inode);
2977undone:
2978 spin_lock(&inode->i_lock);
2979 inode->i_private = NULL;
2980 spin_unlock(&inode->i_lock);
2981out:
2982 inode_unlock(inode);
2983 return error;
2984}
2985
2986static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
2987{
2988 struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
2989
2990 buf->f_type = TMPFS_MAGIC;
2991 buf->f_bsize = PAGE_SIZE;
2992 buf->f_namelen = NAME_MAX;
2993 if (sbinfo->max_blocks) {
2994 buf->f_blocks = sbinfo->max_blocks;
2995 buf->f_bavail =
2996 buf->f_bfree = sbinfo->max_blocks -
2997 percpu_counter_sum(&sbinfo->used_blocks);
2998 }
2999 if (sbinfo->max_inodes) {
3000 buf->f_files = sbinfo->max_inodes;
3001 buf->f_ffree = sbinfo->free_inodes;
3002 }
3003 /* else leave those fields 0 like simple_statfs */
3004 return 0;
3005}
3006
3007/*
3008 * File creation. Allocate an inode, and we're done..
3009 */
3010static int
3011shmem_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
3012{
3013 struct inode *inode;
3014 int error = -ENOSPC;
3015
3016 inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
3017 if (inode) {
3018 error = simple_acl_create(dir, inode);
3019 if (error)
3020 goto out_iput;
3021 error = security_inode_init_security(inode, dir,
3022 &dentry->d_name,
3023 shmem_initxattrs, NULL);
3024 if (error && error != -EOPNOTSUPP)
3025 goto out_iput;
3026
3027 error = 0;
3028 dir->i_size += BOGO_DIRENT_SIZE;
3029 dir->i_ctime = dir->i_mtime = current_time(dir);
3030 d_instantiate(dentry, inode);
3031 dget(dentry); /* Extra count - pin the dentry in core */
3032 }
3033 return error;
3034out_iput:
3035 iput(inode);
3036 return error;
3037}
3038
3039static int
3040shmem_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
3041{
3042 struct inode *inode;
3043 int error = -ENOSPC;
3044
3045 inode = shmem_get_inode(dir->i_sb, dir, mode, 0, VM_NORESERVE);
3046 if (inode) {
3047 error = security_inode_init_security(inode, dir,
3048 NULL,
3049 shmem_initxattrs, NULL);
3050 if (error && error != -EOPNOTSUPP)
3051 goto out_iput;
3052 error = simple_acl_create(dir, inode);
3053 if (error)
3054 goto out_iput;
3055 d_tmpfile(dentry, inode);
3056 }
3057 return error;
3058out_iput:
3059 iput(inode);
3060 return error;
3061}
3062
3063static int shmem_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
3064{
3065 int error;
3066
3067 if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
3068 return error;
3069 inc_nlink(dir);
3070 return 0;
3071}
3072
3073static int shmem_create(struct inode *dir, struct dentry *dentry, umode_t mode,
3074 bool excl)
3075{
3076 return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
3077}
3078
3079/*
3080 * Link a file..
3081 */
3082static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
3083{
3084 struct inode *inode = d_inode(old_dentry);
3085 int ret;
3086
3087 /*
3088 * No ordinary (disk based) filesystem counts links as inodes;
3089 * but each new link needs a new dentry, pinning lowmem, and
3090 * tmpfs dentries cannot be pruned until they are unlinked.
3091 */
3092 ret = shmem_reserve_inode(inode->i_sb);
3093 if (ret)
3094 goto out;
3095
3096 dir->i_size += BOGO_DIRENT_SIZE;
3097 inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
3098 inc_nlink(inode);
3099 ihold(inode); /* New dentry reference */
3100 dget(dentry); /* Extra pinning count for the created dentry */
3101 d_instantiate(dentry, inode);
3102out:
3103 return ret;
3104}
3105
3106static int shmem_unlink(struct inode *dir, struct dentry *dentry)
3107{
3108 struct inode *inode = d_inode(dentry);
3109
3110 if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
3111 shmem_free_inode(inode->i_sb);
3112
3113 dir->i_size -= BOGO_DIRENT_SIZE;
3114 inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
3115 drop_nlink(inode);
3116 dput(dentry); /* Undo the count from "create" - this does all the work */
3117 return 0;
3118}
3119
3120static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
3121{
3122 if (!simple_empty(dentry))
3123 return -ENOTEMPTY;
3124
3125 drop_nlink(d_inode(dentry));
3126 drop_nlink(dir);
3127 return shmem_unlink(dir, dentry);
3128}
3129
3130static int shmem_exchange(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
3131{
3132 bool old_is_dir = d_is_dir(old_dentry);
3133 bool new_is_dir = d_is_dir(new_dentry);
3134
3135 if (old_dir != new_dir && old_is_dir != new_is_dir) {
3136 if (old_is_dir) {
3137 drop_nlink(old_dir);
3138 inc_nlink(new_dir);
3139 } else {
3140 drop_nlink(new_dir);
3141 inc_nlink(old_dir);
3142 }
3143 }
3144 old_dir->i_ctime = old_dir->i_mtime =
3145 new_dir->i_ctime = new_dir->i_mtime =
3146 d_inode(old_dentry)->i_ctime =
3147 d_inode(new_dentry)->i_ctime = current_time(old_dir);
3148
3149 return 0;
3150}
3151
3152static int shmem_whiteout(struct inode *old_dir, struct dentry *old_dentry)
3153{
3154 struct dentry *whiteout;
3155 int error;
3156
3157 whiteout = d_alloc(old_dentry->d_parent, &old_dentry->d_name);
3158 if (!whiteout)
3159 return -ENOMEM;
3160
3161 error = shmem_mknod(old_dir, whiteout,
3162 S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
3163 dput(whiteout);
3164 if (error)
3165 return error;
3166
3167 /*
3168 * Cheat and hash the whiteout while the old dentry is still in
3169 * place, instead of playing games with FS_RENAME_DOES_D_MOVE.
3170 *
3171 * d_lookup() will consistently find one of them at this point,
3172 * not sure which one, but that isn't even important.
3173 */
3174 d_rehash(whiteout);
3175 return 0;
3176}
3177
3178/*
3179 * The VFS layer already does all the dentry stuff for rename,
3180 * we just have to decrement the usage count for the target if
3181 * it exists so that the VFS layer correctly free's it when it
3182 * gets overwritten.
3183 */
3184static int shmem_rename2(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry, unsigned int flags)
3185{
3186 struct inode *inode = d_inode(old_dentry);
3187 int they_are_dirs = S_ISDIR(inode->i_mode);
3188
3189 if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
3190 return -EINVAL;
3191
3192 if (flags & RENAME_EXCHANGE)
3193 return shmem_exchange(old_dir, old_dentry, new_dir, new_dentry);
3194
3195 if (!simple_empty(new_dentry))
3196 return -ENOTEMPTY;
3197
3198 if (flags & RENAME_WHITEOUT) {
3199 int error;
3200
3201 error = shmem_whiteout(old_dir, old_dentry);
3202 if (error)
3203 return error;
3204 }
3205
3206 if (d_really_is_positive(new_dentry)) {
3207 (void) shmem_unlink(new_dir, new_dentry);
3208 if (they_are_dirs) {
3209 drop_nlink(d_inode(new_dentry));
3210 drop_nlink(old_dir);
3211 }
3212 } else if (they_are_dirs) {
3213 drop_nlink(old_dir);
3214 inc_nlink(new_dir);
3215 }
3216
3217 old_dir->i_size -= BOGO_DIRENT_SIZE;
3218 new_dir->i_size += BOGO_DIRENT_SIZE;
3219 old_dir->i_ctime = old_dir->i_mtime =
3220 new_dir->i_ctime = new_dir->i_mtime =
3221 inode->i_ctime = current_time(old_dir);
3222 return 0;
3223}
3224
3225static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
3226{
3227 int error;
3228 int len;
3229 struct inode *inode;
3230 struct page *page;
3231
3232 len = strlen(symname) + 1;
3233 if (len > PAGE_SIZE)
3234 return -ENAMETOOLONG;
3235
3236 inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0, VM_NORESERVE);
3237 if (!inode)
3238 return -ENOSPC;
3239
3240 error = security_inode_init_security(inode, dir, &dentry->d_name,
3241 shmem_initxattrs, NULL);
3242 if (error) {
3243 if (error != -EOPNOTSUPP) {
3244 iput(inode);
3245 return error;
3246 }
3247 error = 0;
3248 }
3249
3250 inode->i_size = len-1;
3251 if (len <= SHORT_SYMLINK_LEN) {
3252 inode->i_link = kmemdup(symname, len, GFP_KERNEL);
3253 if (!inode->i_link) {
3254 iput(inode);
3255 return -ENOMEM;
3256 }
3257 inode->i_op = &shmem_short_symlink_operations;
3258 } else {
3259 inode_nohighmem(inode);
3260 error = shmem_getpage(inode, 0, &page, SGP_WRITE);
3261 if (error) {
3262 iput(inode);
3263 return error;
3264 }
3265 inode->i_mapping->a_ops = &shmem_aops;
3266 inode->i_op = &shmem_symlink_inode_operations;
3267 memcpy(page_address(page), symname, len);
3268 SetPageUptodate(page);
3269 set_page_dirty(page);
3270 unlock_page(page);
3271 put_page(page);
3272 }
3273 dir->i_size += BOGO_DIRENT_SIZE;
3274 dir->i_ctime = dir->i_mtime = current_time(dir);
3275 d_instantiate(dentry, inode);
3276 dget(dentry);
3277 return 0;
3278}
3279
3280static void shmem_put_link(void *arg)
3281{
3282 mark_page_accessed(arg);
3283 put_page(arg);
3284}
3285
3286static const char *shmem_get_link(struct dentry *dentry,
3287 struct inode *inode,
3288 struct delayed_call *done)
3289{
3290 struct page *page = NULL;
3291 int error;
3292 if (!dentry) {
3293 page = find_get_page(inode->i_mapping, 0);
3294 if (!page)
3295 return ERR_PTR(-ECHILD);
3296 if (!PageUptodate(page)) {
3297 put_page(page);
3298 return ERR_PTR(-ECHILD);
3299 }
3300 } else {
3301 error = shmem_getpage(inode, 0, &page, SGP_READ);
3302 if (error)
3303 return ERR_PTR(error);
3304 unlock_page(page);
3305 }
3306 set_delayed_call(done, shmem_put_link, page);
3307 return page_address(page);
3308}
3309
3310#ifdef CONFIG_TMPFS_XATTR
3311/*
3312 * Superblocks without xattr inode operations may get some security.* xattr
3313 * support from the LSM "for free". As soon as we have any other xattrs
3314 * like ACLs, we also need to implement the security.* handlers at
3315 * filesystem level, though.
3316 */
3317
3318/*
3319 * Callback for security_inode_init_security() for acquiring xattrs.
3320 */
3321static int shmem_initxattrs(struct inode *inode,
3322 const struct xattr *xattr_array,
3323 void *fs_info)
3324{
3325 struct shmem_inode_info *info = SHMEM_I(inode);
3326 const struct xattr *xattr;
3327 struct simple_xattr *new_xattr;
3328 size_t len;
3329
3330 for (xattr = xattr_array; xattr->name != NULL; xattr++) {
3331 new_xattr = simple_xattr_alloc(xattr->value, xattr->value_len);
3332 if (!new_xattr)
3333 return -ENOMEM;
3334
3335 len = strlen(xattr->name) + 1;
3336 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
3337 GFP_KERNEL);
3338 if (!new_xattr->name) {
3339 kfree(new_xattr);
3340 return -ENOMEM;
3341 }
3342
3343 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
3344 XATTR_SECURITY_PREFIX_LEN);
3345 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
3346 xattr->name, len);
3347
3348 simple_xattr_list_add(&info->xattrs, new_xattr);
3349 }
3350
3351 return 0;
3352}
3353
3354static int shmem_xattr_handler_get(const struct xattr_handler *handler,
3355 struct dentry *unused, struct inode *inode,
3356 const char *name, void *buffer, size_t size)
3357{
3358 struct shmem_inode_info *info = SHMEM_I(inode);
3359
3360 name = xattr_full_name(handler, name);
3361 return simple_xattr_get(&info->xattrs, name, buffer, size);
3362}
3363
3364static int shmem_xattr_handler_set(const struct xattr_handler *handler,
3365 struct dentry *unused, struct inode *inode,
3366 const char *name, const void *value,
3367 size_t size, int flags)
3368{
3369 struct shmem_inode_info *info = SHMEM_I(inode);
3370
3371 name = xattr_full_name(handler, name);
3372 return simple_xattr_set(&info->xattrs, name, value, size, flags);
3373}
3374
3375static const struct xattr_handler shmem_security_xattr_handler = {
3376 .prefix = XATTR_SECURITY_PREFIX,
3377 .get = shmem_xattr_handler_get,
3378 .set = shmem_xattr_handler_set,
3379};
3380
3381static const struct xattr_handler shmem_trusted_xattr_handler = {
3382 .prefix = XATTR_TRUSTED_PREFIX,
3383 .get = shmem_xattr_handler_get,
3384 .set = shmem_xattr_handler_set,
3385};
3386
3387static const struct xattr_handler *shmem_xattr_handlers[] = {
3388#ifdef CONFIG_TMPFS_POSIX_ACL
3389 &posix_acl_access_xattr_handler,
3390 &posix_acl_default_xattr_handler,
3391#endif
3392 &shmem_security_xattr_handler,
3393 &shmem_trusted_xattr_handler,
3394 NULL
3395};
3396
3397static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
3398{
3399 struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
3400 return simple_xattr_list(d_inode(dentry), &info->xattrs, buffer, size);
3401}
3402#endif /* CONFIG_TMPFS_XATTR */
3403
3404static const struct inode_operations shmem_short_symlink_operations = {
3405 .get_link = simple_get_link,
3406#ifdef CONFIG_TMPFS_XATTR
3407 .listxattr = shmem_listxattr,
3408#endif
3409};
3410
3411static const struct inode_operations shmem_symlink_inode_operations = {
3412 .get_link = shmem_get_link,
3413#ifdef CONFIG_TMPFS_XATTR
3414 .listxattr = shmem_listxattr,
3415#endif
3416};
3417
3418static struct dentry *shmem_get_parent(struct dentry *child)
3419{
3420 return ERR_PTR(-ESTALE);
3421}
3422
3423static int shmem_match(struct inode *ino, void *vfh)
3424{
3425 __u32 *fh = vfh;
3426 __u64 inum = fh[2];
3427 inum = (inum << 32) | fh[1];
3428 return ino->i_ino == inum && fh[0] == ino->i_generation;
3429}
3430
3431static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
3432 struct fid *fid, int fh_len, int fh_type)
3433{
3434 struct inode *inode;
3435 struct dentry *dentry = NULL;
3436 u64 inum;
3437
3438 if (fh_len < 3)
3439 return NULL;
3440
3441 inum = fid->raw[2];
3442 inum = (inum << 32) | fid->raw[1];
3443
3444 inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
3445 shmem_match, fid->raw);
3446 if (inode) {
3447 dentry = d_find_alias(inode);
3448 iput(inode);
3449 }
3450
3451 return dentry;
3452}
3453
3454static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
3455 struct inode *parent)
3456{
3457 if (*len < 3) {
3458 *len = 3;
3459 return FILEID_INVALID;
3460 }
3461
3462 if (inode_unhashed(inode)) {
3463 /* Unfortunately insert_inode_hash is not idempotent,
3464 * so as we hash inodes here rather than at creation
3465 * time, we need a lock to ensure we only try
3466 * to do it once
3467 */
3468 static DEFINE_SPINLOCK(lock);
3469 spin_lock(&lock);
3470 if (inode_unhashed(inode))
3471 __insert_inode_hash(inode,
3472 inode->i_ino + inode->i_generation);
3473 spin_unlock(&lock);
3474 }
3475
3476 fh[0] = inode->i_generation;
3477 fh[1] = inode->i_ino;
3478 fh[2] = ((__u64)inode->i_ino) >> 32;
3479
3480 *len = 3;
3481 return 1;
3482}
3483
3484static const struct export_operations shmem_export_ops = {
3485 .get_parent = shmem_get_parent,
3486 .encode_fh = shmem_encode_fh,
3487 .fh_to_dentry = shmem_fh_to_dentry,
3488};
3489
3490static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
3491 bool remount)
3492{
3493 char *this_char, *value, *rest;
3494 struct mempolicy *mpol = NULL;
3495 uid_t uid;
3496 gid_t gid;
3497
3498 while (options != NULL) {
3499 this_char = options;
3500 for (;;) {
3501 /*
3502 * NUL-terminate this option: unfortunately,
3503 * mount options form a comma-separated list,
3504 * but mpol's nodelist may also contain commas.
3505 */
3506 options = strchr(options, ',');
3507 if (options == NULL)
3508 break;
3509 options++;
3510 if (!isdigit(*options)) {
3511 options[-1] = '\0';
3512 break;
3513 }
3514 }
3515 if (!*this_char)
3516 continue;
3517 if ((value = strchr(this_char,'=')) != NULL) {
3518 *value++ = 0;
3519 } else {
3520 pr_err("tmpfs: No value for mount option '%s'\n",
3521 this_char);
3522 goto error;
3523 }
3524
3525 if (!strcmp(this_char,"size")) {
3526 unsigned long long size;
3527 size = memparse(value,&rest);
3528 if (*rest == '%') {
3529 size <<= PAGE_SHIFT;
3530 size *= totalram_pages;
3531 do_div(size, 100);
3532 rest++;
3533 }
3534 if (*rest)
3535 goto bad_val;
3536 sbinfo->max_blocks =
3537 DIV_ROUND_UP(size, PAGE_SIZE);
3538 } else if (!strcmp(this_char,"nr_blocks")) {
3539 sbinfo->max_blocks = memparse(value, &rest);
3540 if (*rest)
3541 goto bad_val;
3542 } else if (!strcmp(this_char,"nr_inodes")) {
3543 sbinfo->max_inodes = memparse(value, &rest);
3544 if (*rest)
3545 goto bad_val;
3546 } else if (!strcmp(this_char,"mode")) {
3547 if (remount)
3548 continue;
3549 sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
3550 if (*rest)
3551 goto bad_val;
3552 } else if (!strcmp(this_char,"uid")) {
3553 if (remount)
3554 continue;
3555 uid = simple_strtoul(value, &rest, 0);
3556 if (*rest)
3557 goto bad_val;
3558 sbinfo->uid = make_kuid(current_user_ns(), uid);
3559 if (!uid_valid(sbinfo->uid))
3560 goto bad_val;
3561 } else if (!strcmp(this_char,"gid")) {
3562 if (remount)
3563 continue;
3564 gid = simple_strtoul(value, &rest, 0);
3565 if (*rest)
3566 goto bad_val;
3567 sbinfo->gid = make_kgid(current_user_ns(), gid);
3568 if (!gid_valid(sbinfo->gid))
3569 goto bad_val;
3570#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3571 } else if (!strcmp(this_char, "huge")) {
3572 int huge;
3573 huge = shmem_parse_huge(value);
3574 if (huge < 0)
3575 goto bad_val;
3576 if (!has_transparent_hugepage() &&
3577 huge != SHMEM_HUGE_NEVER)
3578 goto bad_val;
3579 sbinfo->huge = huge;
3580#endif
3581#ifdef CONFIG_NUMA
3582 } else if (!strcmp(this_char,"mpol")) {
3583 mpol_put(mpol);
3584 mpol = NULL;
3585 if (mpol_parse_str(value, &mpol))
3586 goto bad_val;
3587#endif
3588 } else {
3589 pr_err("tmpfs: Bad mount option %s\n", this_char);
3590 goto error;
3591 }
3592 }
3593 sbinfo->mpol = mpol;
3594 return 0;
3595
3596bad_val:
3597 pr_err("tmpfs: Bad value '%s' for mount option '%s'\n",
3598 value, this_char);
3599error:
3600 mpol_put(mpol);
3601 return 1;
3602
3603}
3604
3605static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
3606{
3607 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
3608 struct shmem_sb_info config = *sbinfo;
3609 unsigned long inodes;
3610 int error = -EINVAL;
3611
3612 config.mpol = NULL;
3613 if (shmem_parse_options(data, &config, true))
3614 return error;
3615
3616 spin_lock(&sbinfo->stat_lock);
3617 inodes = sbinfo->max_inodes - sbinfo->free_inodes;
3618 if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
3619 goto out;
3620 if (config.max_inodes < inodes)
3621 goto out;
3622 /*
3623 * Those tests disallow limited->unlimited while any are in use;
3624 * but we must separately disallow unlimited->limited, because
3625 * in that case we have no record of how much is already in use.
3626 */
3627 if (config.max_blocks && !sbinfo->max_blocks)
3628 goto out;
3629 if (config.max_inodes && !sbinfo->max_inodes)
3630 goto out;
3631
3632 error = 0;
3633 sbinfo->huge = config.huge;
3634 sbinfo->max_blocks = config.max_blocks;
3635 sbinfo->max_inodes = config.max_inodes;
3636 sbinfo->free_inodes = config.max_inodes - inodes;
3637
3638 /*
3639 * Preserve previous mempolicy unless mpol remount option was specified.
3640 */
3641 if (config.mpol) {
3642 mpol_put(sbinfo->mpol);
3643 sbinfo->mpol = config.mpol; /* transfers initial ref */
3644 }
3645out:
3646 spin_unlock(&sbinfo->stat_lock);
3647 return error;
3648}
3649
3650static int shmem_show_options(struct seq_file *seq, struct dentry *root)
3651{
3652 struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
3653
3654 if (sbinfo->max_blocks != shmem_default_max_blocks())
3655 seq_printf(seq, ",size=%luk",
3656 sbinfo->max_blocks << (PAGE_SHIFT - 10));
3657 if (sbinfo->max_inodes != shmem_default_max_inodes())
3658 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
3659 if (sbinfo->mode != (S_IRWXUGO | S_ISVTX))
3660 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
3661 if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
3662 seq_printf(seq, ",uid=%u",
3663 from_kuid_munged(&init_user_ns, sbinfo->uid));
3664 if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
3665 seq_printf(seq, ",gid=%u",
3666 from_kgid_munged(&init_user_ns, sbinfo->gid));
3667#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3668 /* Rightly or wrongly, show huge mount option unmasked by shmem_huge */
3669 if (sbinfo->huge)
3670 seq_printf(seq, ",huge=%s", shmem_format_huge(sbinfo->huge));
3671#endif
3672 shmem_show_mpol(seq, sbinfo->mpol);
3673 return 0;
3674}
3675
3676#define MFD_NAME_PREFIX "memfd:"
3677#define MFD_NAME_PREFIX_LEN (sizeof(MFD_NAME_PREFIX) - 1)
3678#define MFD_NAME_MAX_LEN (NAME_MAX - MFD_NAME_PREFIX_LEN)
3679
3680#define MFD_ALL_FLAGS (MFD_CLOEXEC | MFD_ALLOW_SEALING | MFD_HUGETLB)
3681
3682SYSCALL_DEFINE2(memfd_create,
3683 const char __user *, uname,
3684 unsigned int, flags)
3685{
3686 unsigned int *file_seals;
3687 struct file *file;
3688 int fd, error;
3689 char *name;
3690 long len;
3691
3692 if (!(flags & MFD_HUGETLB)) {
3693 if (flags & ~(unsigned int)MFD_ALL_FLAGS)
3694 return -EINVAL;
3695 } else {
3696 /* Allow huge page size encoding in flags. */
3697 if (flags & ~(unsigned int)(MFD_ALL_FLAGS |
3698 (MFD_HUGE_MASK << MFD_HUGE_SHIFT)))
3699 return -EINVAL;
3700 }
3701
3702 /* length includes terminating zero */
3703 len = strnlen_user(uname, MFD_NAME_MAX_LEN + 1);
3704 if (len <= 0)
3705 return -EFAULT;
3706 if (len > MFD_NAME_MAX_LEN + 1)
3707 return -EINVAL;
3708
3709 name = kmalloc(len + MFD_NAME_PREFIX_LEN, GFP_KERNEL);
3710 if (!name)
3711 return -ENOMEM;
3712
3713 strcpy(name, MFD_NAME_PREFIX);
3714 if (copy_from_user(&name[MFD_NAME_PREFIX_LEN], uname, len)) {
3715 error = -EFAULT;
3716 goto err_name;
3717 }
3718
3719 /* terminating-zero may have changed after strnlen_user() returned */
3720 if (name[len + MFD_NAME_PREFIX_LEN - 1]) {
3721 error = -EFAULT;
3722 goto err_name;
3723 }
3724
3725 fd = get_unused_fd_flags((flags & MFD_CLOEXEC) ? O_CLOEXEC : 0);
3726 if (fd < 0) {
3727 error = fd;
3728 goto err_name;
3729 }
3730
3731 if (flags & MFD_HUGETLB) {
3732 struct user_struct *user = NULL;
3733
3734 file = hugetlb_file_setup(name, 0, VM_NORESERVE, &user,
3735 HUGETLB_ANONHUGE_INODE,
3736 (flags >> MFD_HUGE_SHIFT) &
3737 MFD_HUGE_MASK);
3738 } else
3739 file = shmem_file_setup(name, 0, VM_NORESERVE);
3740 if (IS_ERR(file)) {
3741 error = PTR_ERR(file);
3742 goto err_fd;
3743 }
3744 file->f_mode |= FMODE_LSEEK | FMODE_PREAD | FMODE_PWRITE;
3745 file->f_flags |= O_RDWR | O_LARGEFILE;
3746
3747 if (flags & MFD_ALLOW_SEALING) {
3748 file_seals = memfd_file_seals_ptr(file);
3749 *file_seals &= ~F_SEAL_SEAL;
3750 }
3751
3752 fd_install(fd, file);
3753 kfree(name);
3754 return fd;
3755
3756err_fd:
3757 put_unused_fd(fd);
3758err_name:
3759 kfree(name);
3760 return error;
3761}
3762
3763#endif /* CONFIG_TMPFS */
3764
3765static void shmem_put_super(struct super_block *sb)
3766{
3767 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
3768
3769 percpu_counter_destroy(&sbinfo->used_blocks);
3770 mpol_put(sbinfo->mpol);
3771 kfree(sbinfo);
3772 sb->s_fs_info = NULL;
3773}
3774
3775int shmem_fill_super(struct super_block *sb, void *data, int silent)
3776{
3777 struct inode *inode;
3778 struct shmem_sb_info *sbinfo;
3779 int err = -ENOMEM;
3780
3781 /* Round up to L1_CACHE_BYTES to resist false sharing */
3782 sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
3783 L1_CACHE_BYTES), GFP_KERNEL);
3784 if (!sbinfo)
3785 return -ENOMEM;
3786
3787 sbinfo->mode = S_IRWXUGO | S_ISVTX;
3788 sbinfo->uid = current_fsuid();
3789 sbinfo->gid = current_fsgid();
3790 sb->s_fs_info = sbinfo;
3791
3792#ifdef CONFIG_TMPFS
3793 /*
3794 * Per default we only allow half of the physical ram per
3795 * tmpfs instance, limiting inodes to one per page of lowmem;
3796 * but the internal instance is left unlimited.
3797 */
3798 if (!(sb->s_flags & SB_KERNMOUNT)) {
3799 sbinfo->max_blocks = shmem_default_max_blocks();
3800 sbinfo->max_inodes = shmem_default_max_inodes();
3801 if (shmem_parse_options(data, sbinfo, false)) {
3802 err = -EINVAL;
3803 goto failed;
3804 }
3805 } else {
3806 sb->s_flags |= SB_NOUSER;
3807 }
3808 sb->s_export_op = &shmem_export_ops;
3809 sb->s_flags |= SB_NOSEC;
3810#else
3811 sb->s_flags |= SB_NOUSER;
3812#endif
3813
3814 spin_lock_init(&sbinfo->stat_lock);
3815 if (percpu_counter_init(&sbinfo->used_blocks, 0, GFP_KERNEL))
3816 goto failed;
3817 sbinfo->free_inodes = sbinfo->max_inodes;
3818 spin_lock_init(&sbinfo->shrinklist_lock);
3819 INIT_LIST_HEAD(&sbinfo->shrinklist);
3820
3821 sb->s_maxbytes = MAX_LFS_FILESIZE;
3822 sb->s_blocksize = PAGE_SIZE;
3823 sb->s_blocksize_bits = PAGE_SHIFT;
3824 sb->s_magic = TMPFS_MAGIC;
3825 sb->s_op = &shmem_ops;
3826 sb->s_time_gran = 1;
3827#ifdef CONFIG_TMPFS_XATTR
3828 sb->s_xattr = shmem_xattr_handlers;
3829#endif
3830#ifdef CONFIG_TMPFS_POSIX_ACL
3831 sb->s_flags |= SB_POSIXACL;
3832#endif
3833 uuid_gen(&sb->s_uuid);
3834
3835 inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
3836 if (!inode)
3837 goto failed;
3838 inode->i_uid = sbinfo->uid;
3839 inode->i_gid = sbinfo->gid;
3840 sb->s_root = d_make_root(inode);
3841 if (!sb->s_root)
3842 goto failed;
3843 return 0;
3844
3845failed:
3846 shmem_put_super(sb);
3847 return err;
3848}
3849
3850static struct kmem_cache *shmem_inode_cachep;
3851
3852static struct inode *shmem_alloc_inode(struct super_block *sb)
3853{
3854 struct shmem_inode_info *info;
3855 info = kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
3856 if (!info)
3857 return NULL;
3858 return &info->vfs_inode;
3859}
3860
3861static void shmem_destroy_callback(struct rcu_head *head)
3862{
3863 struct inode *inode = container_of(head, struct inode, i_rcu);
3864 if (S_ISLNK(inode->i_mode))
3865 kfree(inode->i_link);
3866 kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
3867}
3868
3869static void shmem_destroy_inode(struct inode *inode)
3870{
3871 if (S_ISREG(inode->i_mode))
3872 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
3873 call_rcu(&inode->i_rcu, shmem_destroy_callback);
3874}
3875
3876static void shmem_init_inode(void *foo)
3877{
3878 struct shmem_inode_info *info = foo;
3879 inode_init_once(&info->vfs_inode);
3880}
3881
3882static void shmem_init_inodecache(void)
3883{
3884 shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
3885 sizeof(struct shmem_inode_info),
3886 0, SLAB_PANIC|SLAB_ACCOUNT, shmem_init_inode);
3887}
3888
3889static void shmem_destroy_inodecache(void)
3890{
3891 kmem_cache_destroy(shmem_inode_cachep);
3892}
3893
3894static const struct address_space_operations shmem_aops = {
3895 .writepage = shmem_writepage,
3896 .set_page_dirty = __set_page_dirty_no_writeback,
3897#ifdef CONFIG_TMPFS
3898 .write_begin = shmem_write_begin,
3899 .write_end = shmem_write_end,
3900#endif
3901#ifdef CONFIG_MIGRATION
3902 .migratepage = migrate_page,
3903#endif
3904 .error_remove_page = generic_error_remove_page,
3905};
3906
3907static const struct file_operations shmem_file_operations = {
3908 .mmap = shmem_mmap,
3909 .get_unmapped_area = shmem_get_unmapped_area,
3910#ifdef CONFIG_TMPFS
3911 .llseek = shmem_file_llseek,
3912 .read_iter = shmem_file_read_iter,
3913 .write_iter = generic_file_write_iter,
3914 .fsync = noop_fsync,
3915 .splice_read = generic_file_splice_read,
3916 .splice_write = iter_file_splice_write,
3917 .fallocate = shmem_fallocate,
3918#endif
3919};
3920
3921static const struct inode_operations shmem_inode_operations = {
3922 .getattr = shmem_getattr,
3923 .setattr = shmem_setattr,
3924#ifdef CONFIG_TMPFS_XATTR
3925 .listxattr = shmem_listxattr,
3926 .set_acl = simple_set_acl,
3927#endif
3928};
3929
3930static const struct inode_operations shmem_dir_inode_operations = {
3931#ifdef CONFIG_TMPFS
3932 .create = shmem_create,
3933 .lookup = simple_lookup,
3934 .link = shmem_link,
3935 .unlink = shmem_unlink,
3936 .symlink = shmem_symlink,
3937 .mkdir = shmem_mkdir,
3938 .rmdir = shmem_rmdir,
3939 .mknod = shmem_mknod,
3940 .rename = shmem_rename2,
3941 .tmpfile = shmem_tmpfile,
3942#endif
3943#ifdef CONFIG_TMPFS_XATTR
3944 .listxattr = shmem_listxattr,
3945#endif
3946#ifdef CONFIG_TMPFS_POSIX_ACL
3947 .setattr = shmem_setattr,
3948 .set_acl = simple_set_acl,
3949#endif
3950};
3951
3952static const struct inode_operations shmem_special_inode_operations = {
3953#ifdef CONFIG_TMPFS_XATTR
3954 .listxattr = shmem_listxattr,
3955#endif
3956#ifdef CONFIG_TMPFS_POSIX_ACL
3957 .setattr = shmem_setattr,
3958 .set_acl = simple_set_acl,
3959#endif
3960};
3961
3962static const struct super_operations shmem_ops = {
3963 .alloc_inode = shmem_alloc_inode,
3964 .destroy_inode = shmem_destroy_inode,
3965#ifdef CONFIG_TMPFS
3966 .statfs = shmem_statfs,
3967 .remount_fs = shmem_remount_fs,
3968 .show_options = shmem_show_options,
3969#endif
3970 .evict_inode = shmem_evict_inode,
3971 .drop_inode = generic_delete_inode,
3972 .put_super = shmem_put_super,
3973#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3974 .nr_cached_objects = shmem_unused_huge_count,
3975 .free_cached_objects = shmem_unused_huge_scan,
3976#endif
3977};
3978
3979static const struct vm_operations_struct shmem_vm_ops = {
3980 .fault = shmem_fault,
3981 .map_pages = filemap_map_pages,
3982#ifdef CONFIG_NUMA
3983 .set_policy = shmem_set_policy,
3984 .get_policy = shmem_get_policy,
3985#endif
3986};
3987
3988static struct dentry *shmem_mount(struct file_system_type *fs_type,
3989 int flags, const char *dev_name, void *data)
3990{
3991 return mount_nodev(fs_type, flags, data, shmem_fill_super);
3992}
3993
3994static struct file_system_type shmem_fs_type = {
3995 .owner = THIS_MODULE,
3996 .name = "tmpfs",
3997 .mount = shmem_mount,
3998 .kill_sb = kill_litter_super,
3999 .fs_flags = FS_USERNS_MOUNT,
4000};
4001
4002int __init shmem_init(void)
4003{
4004 int error;
4005
4006 /* If rootfs called this, don't re-init */
4007 if (shmem_inode_cachep)
4008 return 0;
4009
4010 shmem_init_inodecache();
4011
4012 error = register_filesystem(&shmem_fs_type);
4013 if (error) {
4014 pr_err("Could not register tmpfs\n");
4015 goto out2;
4016 }
4017
4018 shm_mnt = kern_mount(&shmem_fs_type);
4019 if (IS_ERR(shm_mnt)) {
4020 error = PTR_ERR(shm_mnt);
4021 pr_err("Could not kern_mount tmpfs\n");
4022 goto out1;
4023 }
4024
4025#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
4026 if (has_transparent_hugepage() && shmem_huge > SHMEM_HUGE_DENY)
4027 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
4028 else
4029 shmem_huge = 0; /* just in case it was patched */
4030#endif
4031 return 0;
4032
4033out1:
4034 unregister_filesystem(&shmem_fs_type);
4035out2:
4036 shmem_destroy_inodecache();
4037 shm_mnt = ERR_PTR(error);
4038 return error;
4039}
4040
4041#if defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE) && defined(CONFIG_SYSFS)
4042static ssize_t shmem_enabled_show(struct kobject *kobj,
4043 struct kobj_attribute *attr, char *buf)
4044{
4045 int values[] = {
4046 SHMEM_HUGE_ALWAYS,
4047 SHMEM_HUGE_WITHIN_SIZE,
4048 SHMEM_HUGE_ADVISE,
4049 SHMEM_HUGE_NEVER,
4050 SHMEM_HUGE_DENY,
4051 SHMEM_HUGE_FORCE,
4052 };
4053 int i, count;
4054
4055 for (i = 0, count = 0; i < ARRAY_SIZE(values); i++) {
4056 const char *fmt = shmem_huge == values[i] ? "[%s] " : "%s ";
4057
4058 count += sprintf(buf + count, fmt,
4059 shmem_format_huge(values[i]));
4060 }
4061 buf[count - 1] = '\n';
4062 return count;
4063}
4064
4065static ssize_t shmem_enabled_store(struct kobject *kobj,
4066 struct kobj_attribute *attr, const char *buf, size_t count)
4067{
4068 char tmp[16];
4069 int huge;
4070
4071 if (count + 1 > sizeof(tmp))
4072 return -EINVAL;
4073 memcpy(tmp, buf, count);
4074 tmp[count] = '\0';
4075 if (count && tmp[count - 1] == '\n')
4076 tmp[count - 1] = '\0';
4077
4078 huge = shmem_parse_huge(tmp);
4079 if (huge == -EINVAL)
4080 return -EINVAL;
4081 if (!has_transparent_hugepage() &&
4082 huge != SHMEM_HUGE_NEVER && huge != SHMEM_HUGE_DENY)
4083 return -EINVAL;
4084
4085 shmem_huge = huge;
4086 if (shmem_huge > SHMEM_HUGE_DENY)
4087 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
4088 return count;
4089}
4090
4091struct kobj_attribute shmem_enabled_attr =
4092 __ATTR(shmem_enabled, 0644, shmem_enabled_show, shmem_enabled_store);
4093#endif /* CONFIG_TRANSPARENT_HUGE_PAGECACHE && CONFIG_SYSFS */
4094
4095#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
4096bool shmem_huge_enabled(struct vm_area_struct *vma)
4097{
4098 struct inode *inode = file_inode(vma->vm_file);
4099 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
4100 loff_t i_size;
4101 pgoff_t off;
4102
4103 if (shmem_huge == SHMEM_HUGE_FORCE)
4104 return true;
4105 if (shmem_huge == SHMEM_HUGE_DENY)
4106 return false;
4107 switch (sbinfo->huge) {
4108 case SHMEM_HUGE_NEVER:
4109 return false;
4110 case SHMEM_HUGE_ALWAYS:
4111 return true;
4112 case SHMEM_HUGE_WITHIN_SIZE:
4113 off = round_up(vma->vm_pgoff, HPAGE_PMD_NR);
4114 i_size = round_up(i_size_read(inode), PAGE_SIZE);
4115 if (i_size >= HPAGE_PMD_SIZE &&
4116 i_size >> PAGE_SHIFT >= off)
4117 return true;
4118 /* fall through */
4119 case SHMEM_HUGE_ADVISE:
4120 /* TODO: implement fadvise() hints */
4121 return (vma->vm_flags & VM_HUGEPAGE);
4122 default:
4123 VM_BUG_ON(1);
4124 return false;
4125 }
4126}
4127#endif /* CONFIG_TRANSPARENT_HUGE_PAGECACHE */
4128
4129#else /* !CONFIG_SHMEM */
4130
4131/*
4132 * tiny-shmem: simple shmemfs and tmpfs using ramfs code
4133 *
4134 * This is intended for small system where the benefits of the full
4135 * shmem code (swap-backed and resource-limited) are outweighed by
4136 * their complexity. On systems without swap this code should be
4137 * effectively equivalent, but much lighter weight.
4138 */
4139
4140static struct file_system_type shmem_fs_type = {
4141 .name = "tmpfs",
4142 .mount = ramfs_mount,
4143 .kill_sb = kill_litter_super,
4144 .fs_flags = FS_USERNS_MOUNT,
4145};
4146
4147int __init shmem_init(void)
4148{
4149 BUG_ON(register_filesystem(&shmem_fs_type) != 0);
4150
4151 shm_mnt = kern_mount(&shmem_fs_type);
4152 BUG_ON(IS_ERR(shm_mnt));
4153
4154 return 0;
4155}
4156
4157int shmem_unuse(swp_entry_t swap, struct page *page)
4158{
4159 return 0;
4160}
4161
4162int shmem_lock(struct file *file, int lock, struct user_struct *user)
4163{
4164 return 0;
4165}
4166
4167void shmem_unlock_mapping(struct address_space *mapping)
4168{
4169}
4170
4171#ifdef CONFIG_MMU
4172unsigned long shmem_get_unmapped_area(struct file *file,
4173 unsigned long addr, unsigned long len,
4174 unsigned long pgoff, unsigned long flags)
4175{
4176 return current->mm->get_unmapped_area(file, addr, len, pgoff, flags);
4177}
4178#endif
4179
4180void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
4181{
4182 truncate_inode_pages_range(inode->i_mapping, lstart, lend);
4183}
4184EXPORT_SYMBOL_GPL(shmem_truncate_range);
4185
4186#define shmem_vm_ops generic_file_vm_ops
4187#define shmem_file_operations ramfs_file_operations
4188#define shmem_get_inode(sb, dir, mode, dev, flags) ramfs_get_inode(sb, dir, mode, dev)
4189#define shmem_acct_size(flags, size) 0
4190#define shmem_unacct_size(flags, size) do {} while (0)
4191
4192#endif /* CONFIG_SHMEM */
4193
4194/* common code */
4195
4196static const struct dentry_operations anon_ops = {
4197 .d_dname = simple_dname
4198};
4199
4200static struct file *__shmem_file_setup(struct vfsmount *mnt, const char *name, loff_t size,
4201 unsigned long flags, unsigned int i_flags)
4202{
4203 struct file *res;
4204 struct inode *inode;
4205 struct path path;
4206 struct super_block *sb;
4207 struct qstr this;
4208
4209 if (IS_ERR(mnt))
4210 return ERR_CAST(mnt);
4211
4212 if (size < 0 || size > MAX_LFS_FILESIZE)
4213 return ERR_PTR(-EINVAL);
4214
4215 if (shmem_acct_size(flags, size))
4216 return ERR_PTR(-ENOMEM);
4217
4218 res = ERR_PTR(-ENOMEM);
4219 this.name = name;
4220 this.len = strlen(name);
4221 this.hash = 0; /* will go */
4222 sb = mnt->mnt_sb;
4223 path.mnt = mntget(mnt);
4224 path.dentry = d_alloc_pseudo(sb, &this);
4225 if (!path.dentry)
4226 goto put_memory;
4227 d_set_d_op(path.dentry, &anon_ops);
4228
4229 res = ERR_PTR(-ENOSPC);
4230 inode = shmem_get_inode(sb, NULL, S_IFREG | S_IRWXUGO, 0, flags);
4231 if (!inode)
4232 goto put_memory;
4233
4234 inode->i_flags |= i_flags;
4235 d_instantiate(path.dentry, inode);
4236 inode->i_size = size;
4237 clear_nlink(inode); /* It is unlinked */
4238 res = ERR_PTR(ramfs_nommu_expand_for_mapping(inode, size));
4239 if (IS_ERR(res))
4240 goto put_path;
4241
4242 res = alloc_file(&path, FMODE_WRITE | FMODE_READ,
4243 &shmem_file_operations);
4244 if (IS_ERR(res))
4245 goto put_path;
4246
4247 return res;
4248
4249put_memory:
4250 shmem_unacct_size(flags, size);
4251put_path:
4252 path_put(&path);
4253 return res;
4254}
4255
4256/**
4257 * shmem_kernel_file_setup - get an unlinked file living in tmpfs which must be
4258 * kernel internal. There will be NO LSM permission checks against the
4259 * underlying inode. So users of this interface must do LSM checks at a
4260 * higher layer. The users are the big_key and shm implementations. LSM
4261 * checks are provided at the key or shm level rather than the inode.
4262 * @name: name for dentry (to be seen in /proc/<pid>/maps
4263 * @size: size to be set for the file
4264 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4265 */
4266struct file *shmem_kernel_file_setup(const char *name, loff_t size, unsigned long flags)
4267{
4268 return __shmem_file_setup(shm_mnt, name, size, flags, S_PRIVATE);
4269}
4270
4271/**
4272 * shmem_file_setup - get an unlinked file living in tmpfs
4273 * @name: name for dentry (to be seen in /proc/<pid>/maps
4274 * @size: size to be set for the file
4275 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4276 */
4277struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
4278{
4279 return __shmem_file_setup(shm_mnt, name, size, flags, 0);
4280}
4281EXPORT_SYMBOL_GPL(shmem_file_setup);
4282
4283/**
4284 * shmem_file_setup_with_mnt - get an unlinked file living in tmpfs
4285 * @mnt: the tmpfs mount where the file will be created
4286 * @name: name for dentry (to be seen in /proc/<pid>/maps
4287 * @size: size to be set for the file
4288 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4289 */
4290struct file *shmem_file_setup_with_mnt(struct vfsmount *mnt, const char *name,
4291 loff_t size, unsigned long flags)
4292{
4293 return __shmem_file_setup(mnt, name, size, flags, 0);
4294}
4295EXPORT_SYMBOL_GPL(shmem_file_setup_with_mnt);
4296
4297/**
4298 * shmem_zero_setup - setup a shared anonymous mapping
4299 * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
4300 */
4301int shmem_zero_setup(struct vm_area_struct *vma)
4302{
4303 struct file *file;
4304 loff_t size = vma->vm_end - vma->vm_start;
4305
4306 /*
4307 * Cloning a new file under mmap_sem leads to a lock ordering conflict
4308 * between XFS directory reading and selinux: since this file is only
4309 * accessible to the user through its mapping, use S_PRIVATE flag to
4310 * bypass file security, in the same way as shmem_kernel_file_setup().
4311 */
4312 file = shmem_kernel_file_setup("dev/zero", size, vma->vm_flags);
4313 if (IS_ERR(file))
4314 return PTR_ERR(file);
4315
4316 if (vma->vm_file)
4317 fput(vma->vm_file);
4318 vma->vm_file = file;
4319 vma->vm_ops = &shmem_vm_ops;
4320
4321 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE) &&
4322 ((vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK) <
4323 (vma->vm_end & HPAGE_PMD_MASK)) {
4324 khugepaged_enter(vma, vma->vm_flags);
4325 }
4326
4327 return 0;
4328}
4329
4330/**
4331 * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
4332 * @mapping: the page's address_space
4333 * @index: the page index
4334 * @gfp: the page allocator flags to use if allocating
4335 *
4336 * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
4337 * with any new page allocations done using the specified allocation flags.
4338 * But read_cache_page_gfp() uses the ->readpage() method: which does not
4339 * suit tmpfs, since it may have pages in swapcache, and needs to find those
4340 * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
4341 *
4342 * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
4343 * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
4344 */
4345struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
4346 pgoff_t index, gfp_t gfp)
4347{
4348#ifdef CONFIG_SHMEM
4349 struct inode *inode = mapping->host;
4350 struct page *page;
4351 int error;
4352
4353 BUG_ON(mapping->a_ops != &shmem_aops);
4354 error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE,
4355 gfp, NULL, NULL, NULL);
4356 if (error)
4357 page = ERR_PTR(error);
4358 else
4359 unlock_page(page);
4360 return page;
4361#else
4362 /*
4363 * The tiny !SHMEM case uses ramfs without swap
4364 */
4365 return read_cache_page_gfp(mapping, index, gfp);
4366#endif
4367}
4368EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);
1/*
2 * Resizable virtual memory filesystem for Linux.
3 *
4 * Copyright (C) 2000 Linus Torvalds.
5 * 2000 Transmeta Corp.
6 * 2000-2001 Christoph Rohland
7 * 2000-2001 SAP AG
8 * 2002 Red Hat Inc.
9 * Copyright (C) 2002-2011 Hugh Dickins.
10 * Copyright (C) 2011 Google Inc.
11 * Copyright (C) 2002-2005 VERITAS Software Corporation.
12 * Copyright (C) 2004 Andi Kleen, SuSE Labs
13 *
14 * Extended attribute support for tmpfs:
15 * Copyright (c) 2004, Luke Kenneth Casson Leighton <lkcl@lkcl.net>
16 * Copyright (c) 2004 Red Hat, Inc., James Morris <jmorris@redhat.com>
17 *
18 * tiny-shmem:
19 * Copyright (c) 2004, 2008 Matt Mackall <mpm@selenic.com>
20 *
21 * This file is released under the GPL.
22 */
23
24#include <linux/fs.h>
25#include <linux/init.h>
26#include <linux/vfs.h>
27#include <linux/mount.h>
28#include <linux/ramfs.h>
29#include <linux/pagemap.h>
30#include <linux/file.h>
31#include <linux/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);