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/random.h>
33#include <linux/sched/signal.h>
34#include <linux/export.h>
35#include <linux/swap.h>
36#include <linux/uio.h>
37#include <linux/khugepaged.h>
38#include <linux/hugetlb.h>
39#include <linux/frontswap.h>
40#include <linux/fs_parser.h>
41
42#include <asm/tlbflush.h> /* for arch/microblaze update_mmu_cache() */
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/blkdev.h>
64#include <linux/pagevec.h>
65#include <linux/percpu_counter.h>
66#include <linux/falloc.h>
67#include <linux/splice.h>
68#include <linux/security.h>
69#include <linux/swapops.h>
70#include <linux/mempolicy.h>
71#include <linux/namei.h>
72#include <linux/ctype.h>
73#include <linux/migrate.h>
74#include <linux/highmem.h>
75#include <linux/seq_file.h>
76#include <linux/magic.h>
77#include <linux/syscalls.h>
78#include <linux/fcntl.h>
79#include <uapi/linux/memfd.h>
80#include <linux/userfaultfd_k.h>
81#include <linux/rmap.h>
82#include <linux/uuid.h>
83
84#include <linux/uaccess.h>
85
86#include "internal.h"
87
88#define BLOCKS_PER_PAGE (PAGE_SIZE/512)
89#define VM_ACCT(size) (PAGE_ALIGN(size) >> PAGE_SHIFT)
90
91/* Pretend that each entry is of this size in directory's i_size */
92#define BOGO_DIRENT_SIZE 20
93
94/* Symlink up to this size is kmalloc'ed instead of using a swappable page */
95#define SHORT_SYMLINK_LEN 128
96
97/*
98 * shmem_fallocate communicates with shmem_fault or shmem_writepage via
99 * inode->i_private (with i_mutex making sure that it has only one user at
100 * a time): we would prefer not to enlarge the shmem inode just for that.
101 */
102struct shmem_falloc {
103 wait_queue_head_t *waitq; /* faults into hole wait for punch to end */
104 pgoff_t start; /* start of range currently being fallocated */
105 pgoff_t next; /* the next page offset to be fallocated */
106 pgoff_t nr_falloced; /* how many new pages have been fallocated */
107 pgoff_t nr_unswapped; /* how often writepage refused to swap out */
108};
109
110struct shmem_options {
111 unsigned long long blocks;
112 unsigned long long inodes;
113 struct mempolicy *mpol;
114 kuid_t uid;
115 kgid_t gid;
116 umode_t mode;
117 bool full_inums;
118 int huge;
119 int seen;
120#define SHMEM_SEEN_BLOCKS 1
121#define SHMEM_SEEN_INODES 2
122#define SHMEM_SEEN_HUGE 4
123#define SHMEM_SEEN_INUMS 8
124};
125
126#ifdef CONFIG_TMPFS
127static unsigned long shmem_default_max_blocks(void)
128{
129 return totalram_pages() / 2;
130}
131
132static unsigned long shmem_default_max_inodes(void)
133{
134 unsigned long nr_pages = totalram_pages();
135
136 return min(nr_pages - totalhigh_pages(), nr_pages / 2);
137}
138#endif
139
140static bool shmem_should_replace_page(struct page *page, gfp_t gfp);
141static int shmem_replace_page(struct page **pagep, gfp_t gfp,
142 struct shmem_inode_info *info, pgoff_t index);
143static int shmem_swapin_page(struct inode *inode, pgoff_t index,
144 struct page **pagep, enum sgp_type sgp,
145 gfp_t gfp, struct vm_area_struct *vma,
146 vm_fault_t *fault_type);
147static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
148 struct page **pagep, enum sgp_type sgp,
149 gfp_t gfp, struct vm_area_struct *vma,
150 struct vm_fault *vmf, vm_fault_t *fault_type);
151
152int shmem_getpage(struct inode *inode, pgoff_t index,
153 struct page **pagep, enum sgp_type sgp)
154{
155 return shmem_getpage_gfp(inode, index, pagep, sgp,
156 mapping_gfp_mask(inode->i_mapping), NULL, NULL, NULL);
157}
158
159static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
160{
161 return sb->s_fs_info;
162}
163
164/*
165 * shmem_file_setup pre-accounts the whole fixed size of a VM object,
166 * for shared memory and for shared anonymous (/dev/zero) mappings
167 * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
168 * consistent with the pre-accounting of private mappings ...
169 */
170static inline int shmem_acct_size(unsigned long flags, loff_t size)
171{
172 return (flags & VM_NORESERVE) ?
173 0 : security_vm_enough_memory_mm(current->mm, VM_ACCT(size));
174}
175
176static inline void shmem_unacct_size(unsigned long flags, loff_t size)
177{
178 if (!(flags & VM_NORESERVE))
179 vm_unacct_memory(VM_ACCT(size));
180}
181
182static inline int shmem_reacct_size(unsigned long flags,
183 loff_t oldsize, loff_t newsize)
184{
185 if (!(flags & VM_NORESERVE)) {
186 if (VM_ACCT(newsize) > VM_ACCT(oldsize))
187 return security_vm_enough_memory_mm(current->mm,
188 VM_ACCT(newsize) - VM_ACCT(oldsize));
189 else if (VM_ACCT(newsize) < VM_ACCT(oldsize))
190 vm_unacct_memory(VM_ACCT(oldsize) - VM_ACCT(newsize));
191 }
192 return 0;
193}
194
195/*
196 * ... whereas tmpfs objects are accounted incrementally as
197 * pages are allocated, in order to allow large sparse files.
198 * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
199 * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
200 */
201static inline int shmem_acct_block(unsigned long flags, long pages)
202{
203 if (!(flags & VM_NORESERVE))
204 return 0;
205
206 return security_vm_enough_memory_mm(current->mm,
207 pages * VM_ACCT(PAGE_SIZE));
208}
209
210static inline void shmem_unacct_blocks(unsigned long flags, long pages)
211{
212 if (flags & VM_NORESERVE)
213 vm_unacct_memory(pages * VM_ACCT(PAGE_SIZE));
214}
215
216static inline bool shmem_inode_acct_block(struct inode *inode, long pages)
217{
218 struct shmem_inode_info *info = SHMEM_I(inode);
219 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
220
221 if (shmem_acct_block(info->flags, pages))
222 return false;
223
224 if (sbinfo->max_blocks) {
225 if (percpu_counter_compare(&sbinfo->used_blocks,
226 sbinfo->max_blocks - pages) > 0)
227 goto unacct;
228 percpu_counter_add(&sbinfo->used_blocks, pages);
229 }
230
231 return true;
232
233unacct:
234 shmem_unacct_blocks(info->flags, pages);
235 return false;
236}
237
238static inline void shmem_inode_unacct_blocks(struct inode *inode, long pages)
239{
240 struct shmem_inode_info *info = SHMEM_I(inode);
241 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
242
243 if (sbinfo->max_blocks)
244 percpu_counter_sub(&sbinfo->used_blocks, pages);
245 shmem_unacct_blocks(info->flags, pages);
246}
247
248static const struct super_operations shmem_ops;
249static const struct address_space_operations shmem_aops;
250static const struct file_operations shmem_file_operations;
251static const struct inode_operations shmem_inode_operations;
252static const struct inode_operations shmem_dir_inode_operations;
253static const struct inode_operations shmem_special_inode_operations;
254static const struct vm_operations_struct shmem_vm_ops;
255static struct file_system_type shmem_fs_type;
256
257bool vma_is_shmem(struct vm_area_struct *vma)
258{
259 return vma->vm_ops == &shmem_vm_ops;
260}
261
262static LIST_HEAD(shmem_swaplist);
263static DEFINE_MUTEX(shmem_swaplist_mutex);
264
265/*
266 * shmem_reserve_inode() performs bookkeeping to reserve a shmem inode, and
267 * produces a novel ino for the newly allocated inode.
268 *
269 * It may also be called when making a hard link to permit the space needed by
270 * each dentry. However, in that case, no new inode number is needed since that
271 * internally draws from another pool of inode numbers (currently global
272 * get_next_ino()). This case is indicated by passing NULL as inop.
273 */
274#define SHMEM_INO_BATCH 1024
275static int shmem_reserve_inode(struct super_block *sb, ino_t *inop)
276{
277 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
278 ino_t ino;
279
280 if (!(sb->s_flags & SB_KERNMOUNT)) {
281 spin_lock(&sbinfo->stat_lock);
282 if (sbinfo->max_inodes) {
283 if (!sbinfo->free_inodes) {
284 spin_unlock(&sbinfo->stat_lock);
285 return -ENOSPC;
286 }
287 sbinfo->free_inodes--;
288 }
289 if (inop) {
290 ino = sbinfo->next_ino++;
291 if (unlikely(is_zero_ino(ino)))
292 ino = sbinfo->next_ino++;
293 if (unlikely(!sbinfo->full_inums &&
294 ino > UINT_MAX)) {
295 /*
296 * Emulate get_next_ino uint wraparound for
297 * compatibility
298 */
299 if (IS_ENABLED(CONFIG_64BIT))
300 pr_warn("%s: inode number overflow on device %d, consider using inode64 mount option\n",
301 __func__, MINOR(sb->s_dev));
302 sbinfo->next_ino = 1;
303 ino = sbinfo->next_ino++;
304 }
305 *inop = ino;
306 }
307 spin_unlock(&sbinfo->stat_lock);
308 } else if (inop) {
309 /*
310 * __shmem_file_setup, one of our callers, is lock-free: it
311 * doesn't hold stat_lock in shmem_reserve_inode since
312 * max_inodes is always 0, and is called from potentially
313 * unknown contexts. As such, use a per-cpu batched allocator
314 * which doesn't require the per-sb stat_lock unless we are at
315 * the batch boundary.
316 *
317 * We don't need to worry about inode{32,64} since SB_KERNMOUNT
318 * shmem mounts are not exposed to userspace, so we don't need
319 * to worry about things like glibc compatibility.
320 */
321 ino_t *next_ino;
322 next_ino = per_cpu_ptr(sbinfo->ino_batch, get_cpu());
323 ino = *next_ino;
324 if (unlikely(ino % SHMEM_INO_BATCH == 0)) {
325 spin_lock(&sbinfo->stat_lock);
326 ino = sbinfo->next_ino;
327 sbinfo->next_ino += SHMEM_INO_BATCH;
328 spin_unlock(&sbinfo->stat_lock);
329 if (unlikely(is_zero_ino(ino)))
330 ino++;
331 }
332 *inop = ino;
333 *next_ino = ++ino;
334 put_cpu();
335 }
336
337 return 0;
338}
339
340static void shmem_free_inode(struct super_block *sb)
341{
342 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
343 if (sbinfo->max_inodes) {
344 spin_lock(&sbinfo->stat_lock);
345 sbinfo->free_inodes++;
346 spin_unlock(&sbinfo->stat_lock);
347 }
348}
349
350/**
351 * shmem_recalc_inode - recalculate the block usage of an inode
352 * @inode: inode to recalc
353 *
354 * We have to calculate the free blocks since the mm can drop
355 * undirtied hole pages behind our back.
356 *
357 * But normally info->alloced == inode->i_mapping->nrpages + info->swapped
358 * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
359 *
360 * It has to be called with the spinlock held.
361 */
362static void shmem_recalc_inode(struct inode *inode)
363{
364 struct shmem_inode_info *info = SHMEM_I(inode);
365 long freed;
366
367 freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
368 if (freed > 0) {
369 info->alloced -= freed;
370 inode->i_blocks -= freed * BLOCKS_PER_PAGE;
371 shmem_inode_unacct_blocks(inode, freed);
372 }
373}
374
375bool shmem_charge(struct inode *inode, long pages)
376{
377 struct shmem_inode_info *info = SHMEM_I(inode);
378 unsigned long flags;
379
380 if (!shmem_inode_acct_block(inode, pages))
381 return false;
382
383 /* nrpages adjustment first, then shmem_recalc_inode() when balanced */
384 inode->i_mapping->nrpages += pages;
385
386 spin_lock_irqsave(&info->lock, flags);
387 info->alloced += pages;
388 inode->i_blocks += pages * BLOCKS_PER_PAGE;
389 shmem_recalc_inode(inode);
390 spin_unlock_irqrestore(&info->lock, flags);
391
392 return true;
393}
394
395void shmem_uncharge(struct inode *inode, long pages)
396{
397 struct shmem_inode_info *info = SHMEM_I(inode);
398 unsigned long flags;
399
400 /* nrpages adjustment done by __delete_from_page_cache() or caller */
401
402 spin_lock_irqsave(&info->lock, flags);
403 info->alloced -= pages;
404 inode->i_blocks -= pages * BLOCKS_PER_PAGE;
405 shmem_recalc_inode(inode);
406 spin_unlock_irqrestore(&info->lock, flags);
407
408 shmem_inode_unacct_blocks(inode, pages);
409}
410
411/*
412 * Replace item expected in xarray by a new item, while holding xa_lock.
413 */
414static int shmem_replace_entry(struct address_space *mapping,
415 pgoff_t index, void *expected, void *replacement)
416{
417 XA_STATE(xas, &mapping->i_pages, index);
418 void *item;
419
420 VM_BUG_ON(!expected);
421 VM_BUG_ON(!replacement);
422 item = xas_load(&xas);
423 if (item != expected)
424 return -ENOENT;
425 xas_store(&xas, replacement);
426 return 0;
427}
428
429/*
430 * Sometimes, before we decide whether to proceed or to fail, we must check
431 * that an entry was not already brought back from swap by a racing thread.
432 *
433 * Checking page is not enough: by the time a SwapCache page is locked, it
434 * might be reused, and again be SwapCache, using the same swap as before.
435 */
436static bool shmem_confirm_swap(struct address_space *mapping,
437 pgoff_t index, swp_entry_t swap)
438{
439 return xa_load(&mapping->i_pages, index) == swp_to_radix_entry(swap);
440}
441
442/*
443 * Definitions for "huge tmpfs": tmpfs mounted with the huge= option
444 *
445 * SHMEM_HUGE_NEVER:
446 * disables huge pages for the mount;
447 * SHMEM_HUGE_ALWAYS:
448 * enables huge pages for the mount;
449 * SHMEM_HUGE_WITHIN_SIZE:
450 * only allocate huge pages if the page will be fully within i_size,
451 * also respect fadvise()/madvise() hints;
452 * SHMEM_HUGE_ADVISE:
453 * only allocate huge pages if requested with fadvise()/madvise();
454 */
455
456#define SHMEM_HUGE_NEVER 0
457#define SHMEM_HUGE_ALWAYS 1
458#define SHMEM_HUGE_WITHIN_SIZE 2
459#define SHMEM_HUGE_ADVISE 3
460
461/*
462 * Special values.
463 * Only can be set via /sys/kernel/mm/transparent_hugepage/shmem_enabled:
464 *
465 * SHMEM_HUGE_DENY:
466 * disables huge on shm_mnt and all mounts, for emergency use;
467 * SHMEM_HUGE_FORCE:
468 * enables huge on shm_mnt and all mounts, w/o needing option, for testing;
469 *
470 */
471#define SHMEM_HUGE_DENY (-1)
472#define SHMEM_HUGE_FORCE (-2)
473
474#ifdef CONFIG_TRANSPARENT_HUGEPAGE
475/* ifdef here to avoid bloating shmem.o when not necessary */
476
477static int shmem_huge __read_mostly;
478
479#if defined(CONFIG_SYSFS)
480static int shmem_parse_huge(const char *str)
481{
482 if (!strcmp(str, "never"))
483 return SHMEM_HUGE_NEVER;
484 if (!strcmp(str, "always"))
485 return SHMEM_HUGE_ALWAYS;
486 if (!strcmp(str, "within_size"))
487 return SHMEM_HUGE_WITHIN_SIZE;
488 if (!strcmp(str, "advise"))
489 return SHMEM_HUGE_ADVISE;
490 if (!strcmp(str, "deny"))
491 return SHMEM_HUGE_DENY;
492 if (!strcmp(str, "force"))
493 return SHMEM_HUGE_FORCE;
494 return -EINVAL;
495}
496#endif
497
498#if defined(CONFIG_SYSFS) || defined(CONFIG_TMPFS)
499static const char *shmem_format_huge(int huge)
500{
501 switch (huge) {
502 case SHMEM_HUGE_NEVER:
503 return "never";
504 case SHMEM_HUGE_ALWAYS:
505 return "always";
506 case SHMEM_HUGE_WITHIN_SIZE:
507 return "within_size";
508 case SHMEM_HUGE_ADVISE:
509 return "advise";
510 case SHMEM_HUGE_DENY:
511 return "deny";
512 case SHMEM_HUGE_FORCE:
513 return "force";
514 default:
515 VM_BUG_ON(1);
516 return "bad_val";
517 }
518}
519#endif
520
521static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
522 struct shrink_control *sc, unsigned long nr_to_split)
523{
524 LIST_HEAD(list), *pos, *next;
525 LIST_HEAD(to_remove);
526 struct inode *inode;
527 struct shmem_inode_info *info;
528 struct page *page;
529 unsigned long batch = sc ? sc->nr_to_scan : 128;
530 int removed = 0, split = 0;
531
532 if (list_empty(&sbinfo->shrinklist))
533 return SHRINK_STOP;
534
535 spin_lock(&sbinfo->shrinklist_lock);
536 list_for_each_safe(pos, next, &sbinfo->shrinklist) {
537 info = list_entry(pos, struct shmem_inode_info, shrinklist);
538
539 /* pin the inode */
540 inode = igrab(&info->vfs_inode);
541
542 /* inode is about to be evicted */
543 if (!inode) {
544 list_del_init(&info->shrinklist);
545 removed++;
546 goto next;
547 }
548
549 /* Check if there's anything to gain */
550 if (round_up(inode->i_size, PAGE_SIZE) ==
551 round_up(inode->i_size, HPAGE_PMD_SIZE)) {
552 list_move(&info->shrinklist, &to_remove);
553 removed++;
554 goto next;
555 }
556
557 list_move(&info->shrinklist, &list);
558next:
559 if (!--batch)
560 break;
561 }
562 spin_unlock(&sbinfo->shrinklist_lock);
563
564 list_for_each_safe(pos, next, &to_remove) {
565 info = list_entry(pos, struct shmem_inode_info, shrinklist);
566 inode = &info->vfs_inode;
567 list_del_init(&info->shrinklist);
568 iput(inode);
569 }
570
571 list_for_each_safe(pos, next, &list) {
572 int ret;
573
574 info = list_entry(pos, struct shmem_inode_info, shrinklist);
575 inode = &info->vfs_inode;
576
577 if (nr_to_split && split >= nr_to_split)
578 goto leave;
579
580 page = find_get_page(inode->i_mapping,
581 (inode->i_size & HPAGE_PMD_MASK) >> PAGE_SHIFT);
582 if (!page)
583 goto drop;
584
585 /* No huge page at the end of the file: nothing to split */
586 if (!PageTransHuge(page)) {
587 put_page(page);
588 goto drop;
589 }
590
591 /*
592 * Leave the inode on the list if we failed to lock
593 * the page at this time.
594 *
595 * Waiting for the lock may lead to deadlock in the
596 * reclaim path.
597 */
598 if (!trylock_page(page)) {
599 put_page(page);
600 goto leave;
601 }
602
603 ret = split_huge_page(page);
604 unlock_page(page);
605 put_page(page);
606
607 /* If split failed leave the inode on the list */
608 if (ret)
609 goto leave;
610
611 split++;
612drop:
613 list_del_init(&info->shrinklist);
614 removed++;
615leave:
616 iput(inode);
617 }
618
619 spin_lock(&sbinfo->shrinklist_lock);
620 list_splice_tail(&list, &sbinfo->shrinklist);
621 sbinfo->shrinklist_len -= removed;
622 spin_unlock(&sbinfo->shrinklist_lock);
623
624 return split;
625}
626
627static long shmem_unused_huge_scan(struct super_block *sb,
628 struct shrink_control *sc)
629{
630 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
631
632 if (!READ_ONCE(sbinfo->shrinklist_len))
633 return SHRINK_STOP;
634
635 return shmem_unused_huge_shrink(sbinfo, sc, 0);
636}
637
638static long shmem_unused_huge_count(struct super_block *sb,
639 struct shrink_control *sc)
640{
641 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
642 return READ_ONCE(sbinfo->shrinklist_len);
643}
644#else /* !CONFIG_TRANSPARENT_HUGEPAGE */
645
646#define shmem_huge SHMEM_HUGE_DENY
647
648static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
649 struct shrink_control *sc, unsigned long nr_to_split)
650{
651 return 0;
652}
653#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
654
655static inline bool is_huge_enabled(struct shmem_sb_info *sbinfo)
656{
657 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
658 (shmem_huge == SHMEM_HUGE_FORCE || sbinfo->huge) &&
659 shmem_huge != SHMEM_HUGE_DENY)
660 return true;
661 return false;
662}
663
664/*
665 * Like add_to_page_cache_locked, but error if expected item has gone.
666 */
667static int shmem_add_to_page_cache(struct page *page,
668 struct address_space *mapping,
669 pgoff_t index, void *expected, gfp_t gfp,
670 struct mm_struct *charge_mm)
671{
672 XA_STATE_ORDER(xas, &mapping->i_pages, index, compound_order(page));
673 unsigned long i = 0;
674 unsigned long nr = compound_nr(page);
675 int error;
676
677 VM_BUG_ON_PAGE(PageTail(page), page);
678 VM_BUG_ON_PAGE(index != round_down(index, nr), page);
679 VM_BUG_ON_PAGE(!PageLocked(page), page);
680 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
681 VM_BUG_ON(expected && PageTransHuge(page));
682
683 page_ref_add(page, nr);
684 page->mapping = mapping;
685 page->index = index;
686
687 if (!PageSwapCache(page)) {
688 error = mem_cgroup_charge(page, charge_mm, gfp);
689 if (error) {
690 if (PageTransHuge(page)) {
691 count_vm_event(THP_FILE_FALLBACK);
692 count_vm_event(THP_FILE_FALLBACK_CHARGE);
693 }
694 goto error;
695 }
696 }
697 cgroup_throttle_swaprate(page, gfp);
698
699 do {
700 void *entry;
701 xas_lock_irq(&xas);
702 entry = xas_find_conflict(&xas);
703 if (entry != expected)
704 xas_set_err(&xas, -EEXIST);
705 xas_create_range(&xas);
706 if (xas_error(&xas))
707 goto unlock;
708next:
709 xas_store(&xas, page);
710 if (++i < nr) {
711 xas_next(&xas);
712 goto next;
713 }
714 if (PageTransHuge(page)) {
715 count_vm_event(THP_FILE_ALLOC);
716 __inc_node_page_state(page, NR_SHMEM_THPS);
717 }
718 mapping->nrpages += nr;
719 __mod_lruvec_page_state(page, NR_FILE_PAGES, nr);
720 __mod_lruvec_page_state(page, NR_SHMEM, nr);
721unlock:
722 xas_unlock_irq(&xas);
723 } while (xas_nomem(&xas, gfp));
724
725 if (xas_error(&xas)) {
726 error = xas_error(&xas);
727 goto error;
728 }
729
730 return 0;
731error:
732 page->mapping = NULL;
733 page_ref_sub(page, nr);
734 return error;
735}
736
737/*
738 * Like delete_from_page_cache, but substitutes swap for page.
739 */
740static void shmem_delete_from_page_cache(struct page *page, void *radswap)
741{
742 struct address_space *mapping = page->mapping;
743 int error;
744
745 VM_BUG_ON_PAGE(PageCompound(page), page);
746
747 xa_lock_irq(&mapping->i_pages);
748 error = shmem_replace_entry(mapping, page->index, page, radswap);
749 page->mapping = NULL;
750 mapping->nrpages--;
751 __dec_lruvec_page_state(page, NR_FILE_PAGES);
752 __dec_lruvec_page_state(page, NR_SHMEM);
753 xa_unlock_irq(&mapping->i_pages);
754 put_page(page);
755 BUG_ON(error);
756}
757
758/*
759 * Remove swap entry from page cache, free the swap and its page cache.
760 */
761static int shmem_free_swap(struct address_space *mapping,
762 pgoff_t index, void *radswap)
763{
764 void *old;
765
766 old = xa_cmpxchg_irq(&mapping->i_pages, index, radswap, NULL, 0);
767 if (old != radswap)
768 return -ENOENT;
769 free_swap_and_cache(radix_to_swp_entry(radswap));
770 return 0;
771}
772
773/*
774 * Determine (in bytes) how many of the shmem object's pages mapped by the
775 * given offsets are swapped out.
776 *
777 * This is safe to call without i_mutex or the i_pages lock thanks to RCU,
778 * as long as the inode doesn't go away and racy results are not a problem.
779 */
780unsigned long shmem_partial_swap_usage(struct address_space *mapping,
781 pgoff_t start, pgoff_t end)
782{
783 XA_STATE(xas, &mapping->i_pages, start);
784 struct page *page;
785 unsigned long swapped = 0;
786
787 rcu_read_lock();
788 xas_for_each(&xas, page, end - 1) {
789 if (xas_retry(&xas, page))
790 continue;
791 if (xa_is_value(page))
792 swapped++;
793
794 if (need_resched()) {
795 xas_pause(&xas);
796 cond_resched_rcu();
797 }
798 }
799
800 rcu_read_unlock();
801
802 return swapped << PAGE_SHIFT;
803}
804
805/*
806 * Determine (in bytes) how many of the shmem object's pages mapped by the
807 * given vma is swapped out.
808 *
809 * This is safe to call without i_mutex or the i_pages lock thanks to RCU,
810 * as long as the inode doesn't go away and racy results are not a problem.
811 */
812unsigned long shmem_swap_usage(struct vm_area_struct *vma)
813{
814 struct inode *inode = file_inode(vma->vm_file);
815 struct shmem_inode_info *info = SHMEM_I(inode);
816 struct address_space *mapping = inode->i_mapping;
817 unsigned long swapped;
818
819 /* Be careful as we don't hold info->lock */
820 swapped = READ_ONCE(info->swapped);
821
822 /*
823 * The easier cases are when the shmem object has nothing in swap, or
824 * the vma maps it whole. Then we can simply use the stats that we
825 * already track.
826 */
827 if (!swapped)
828 return 0;
829
830 if (!vma->vm_pgoff && vma->vm_end - vma->vm_start >= inode->i_size)
831 return swapped << PAGE_SHIFT;
832
833 /* Here comes the more involved part */
834 return shmem_partial_swap_usage(mapping,
835 linear_page_index(vma, vma->vm_start),
836 linear_page_index(vma, vma->vm_end));
837}
838
839/*
840 * SysV IPC SHM_UNLOCK restore Unevictable pages to their evictable lists.
841 */
842void shmem_unlock_mapping(struct address_space *mapping)
843{
844 struct pagevec pvec;
845 pgoff_t indices[PAGEVEC_SIZE];
846 pgoff_t index = 0;
847
848 pagevec_init(&pvec);
849 /*
850 * Minor point, but we might as well stop if someone else SHM_LOCKs it.
851 */
852 while (!mapping_unevictable(mapping)) {
853 /*
854 * Avoid pagevec_lookup(): find_get_pages() returns 0 as if it
855 * has finished, if it hits a row of PAGEVEC_SIZE swap entries.
856 */
857 pvec.nr = find_get_entries(mapping, index,
858 PAGEVEC_SIZE, pvec.pages, indices);
859 if (!pvec.nr)
860 break;
861 index = indices[pvec.nr - 1] + 1;
862 pagevec_remove_exceptionals(&pvec);
863 check_move_unevictable_pages(&pvec);
864 pagevec_release(&pvec);
865 cond_resched();
866 }
867}
868
869/*
870 * Check whether a hole-punch or truncation needs to split a huge page,
871 * returning true if no split was required, or the split has been successful.
872 *
873 * Eviction (or truncation to 0 size) should never need to split a huge page;
874 * but in rare cases might do so, if shmem_undo_range() failed to trylock on
875 * head, and then succeeded to trylock on tail.
876 *
877 * A split can only succeed when there are no additional references on the
878 * huge page: so the split below relies upon find_get_entries() having stopped
879 * when it found a subpage of the huge page, without getting further references.
880 */
881static bool shmem_punch_compound(struct page *page, pgoff_t start, pgoff_t end)
882{
883 if (!PageTransCompound(page))
884 return true;
885
886 /* Just proceed to delete a huge page wholly within the range punched */
887 if (PageHead(page) &&
888 page->index >= start && page->index + HPAGE_PMD_NR <= end)
889 return true;
890
891 /* Try to split huge page, so we can truly punch the hole or truncate */
892 return split_huge_page(page) >= 0;
893}
894
895/*
896 * Remove range of pages and swap entries from page cache, and free them.
897 * If !unfalloc, truncate or punch hole; if unfalloc, undo failed fallocate.
898 */
899static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
900 bool unfalloc)
901{
902 struct address_space *mapping = inode->i_mapping;
903 struct shmem_inode_info *info = SHMEM_I(inode);
904 pgoff_t start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
905 pgoff_t end = (lend + 1) >> PAGE_SHIFT;
906 unsigned int partial_start = lstart & (PAGE_SIZE - 1);
907 unsigned int partial_end = (lend + 1) & (PAGE_SIZE - 1);
908 struct pagevec pvec;
909 pgoff_t indices[PAGEVEC_SIZE];
910 long nr_swaps_freed = 0;
911 pgoff_t index;
912 int i;
913
914 if (lend == -1)
915 end = -1; /* unsigned, so actually very big */
916
917 pagevec_init(&pvec);
918 index = start;
919 while (index < end) {
920 pvec.nr = find_get_entries(mapping, index,
921 min(end - index, (pgoff_t)PAGEVEC_SIZE),
922 pvec.pages, indices);
923 if (!pvec.nr)
924 break;
925 for (i = 0; i < pagevec_count(&pvec); i++) {
926 struct page *page = pvec.pages[i];
927
928 index = indices[i];
929 if (index >= end)
930 break;
931
932 if (xa_is_value(page)) {
933 if (unfalloc)
934 continue;
935 nr_swaps_freed += !shmem_free_swap(mapping,
936 index, page);
937 continue;
938 }
939
940 VM_BUG_ON_PAGE(page_to_pgoff(page) != index, page);
941
942 if (!trylock_page(page))
943 continue;
944
945 if ((!unfalloc || !PageUptodate(page)) &&
946 page_mapping(page) == mapping) {
947 VM_BUG_ON_PAGE(PageWriteback(page), page);
948 if (shmem_punch_compound(page, start, end))
949 truncate_inode_page(mapping, page);
950 }
951 unlock_page(page);
952 }
953 pagevec_remove_exceptionals(&pvec);
954 pagevec_release(&pvec);
955 cond_resched();
956 index++;
957 }
958
959 if (partial_start) {
960 struct page *page = NULL;
961 shmem_getpage(inode, start - 1, &page, SGP_READ);
962 if (page) {
963 unsigned int top = PAGE_SIZE;
964 if (start > end) {
965 top = partial_end;
966 partial_end = 0;
967 }
968 zero_user_segment(page, partial_start, top);
969 set_page_dirty(page);
970 unlock_page(page);
971 put_page(page);
972 }
973 }
974 if (partial_end) {
975 struct page *page = NULL;
976 shmem_getpage(inode, end, &page, SGP_READ);
977 if (page) {
978 zero_user_segment(page, 0, partial_end);
979 set_page_dirty(page);
980 unlock_page(page);
981 put_page(page);
982 }
983 }
984 if (start >= end)
985 return;
986
987 index = start;
988 while (index < end) {
989 cond_resched();
990
991 pvec.nr = find_get_entries(mapping, index,
992 min(end - index, (pgoff_t)PAGEVEC_SIZE),
993 pvec.pages, indices);
994 if (!pvec.nr) {
995 /* If all gone or hole-punch or unfalloc, we're done */
996 if (index == start || end != -1)
997 break;
998 /* But if truncating, restart to make sure all gone */
999 index = start;
1000 continue;
1001 }
1002 for (i = 0; i < pagevec_count(&pvec); i++) {
1003 struct page *page = pvec.pages[i];
1004
1005 index = indices[i];
1006 if (index >= end)
1007 break;
1008
1009 if (xa_is_value(page)) {
1010 if (unfalloc)
1011 continue;
1012 if (shmem_free_swap(mapping, index, page)) {
1013 /* Swap was replaced by page: retry */
1014 index--;
1015 break;
1016 }
1017 nr_swaps_freed++;
1018 continue;
1019 }
1020
1021 lock_page(page);
1022
1023 if (!unfalloc || !PageUptodate(page)) {
1024 if (page_mapping(page) != mapping) {
1025 /* Page was replaced by swap: retry */
1026 unlock_page(page);
1027 index--;
1028 break;
1029 }
1030 VM_BUG_ON_PAGE(PageWriteback(page), page);
1031 if (shmem_punch_compound(page, start, end))
1032 truncate_inode_page(mapping, page);
1033 else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
1034 /* Wipe the page and don't get stuck */
1035 clear_highpage(page);
1036 flush_dcache_page(page);
1037 set_page_dirty(page);
1038 if (index <
1039 round_up(start, HPAGE_PMD_NR))
1040 start = index + 1;
1041 }
1042 }
1043 unlock_page(page);
1044 }
1045 pagevec_remove_exceptionals(&pvec);
1046 pagevec_release(&pvec);
1047 index++;
1048 }
1049
1050 spin_lock_irq(&info->lock);
1051 info->swapped -= nr_swaps_freed;
1052 shmem_recalc_inode(inode);
1053 spin_unlock_irq(&info->lock);
1054}
1055
1056void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
1057{
1058 shmem_undo_range(inode, lstart, lend, false);
1059 inode->i_ctime = inode->i_mtime = current_time(inode);
1060}
1061EXPORT_SYMBOL_GPL(shmem_truncate_range);
1062
1063static int shmem_getattr(const struct path *path, struct kstat *stat,
1064 u32 request_mask, unsigned int query_flags)
1065{
1066 struct inode *inode = path->dentry->d_inode;
1067 struct shmem_inode_info *info = SHMEM_I(inode);
1068 struct shmem_sb_info *sb_info = SHMEM_SB(inode->i_sb);
1069
1070 if (info->alloced - info->swapped != inode->i_mapping->nrpages) {
1071 spin_lock_irq(&info->lock);
1072 shmem_recalc_inode(inode);
1073 spin_unlock_irq(&info->lock);
1074 }
1075 generic_fillattr(inode, stat);
1076
1077 if (is_huge_enabled(sb_info))
1078 stat->blksize = HPAGE_PMD_SIZE;
1079
1080 return 0;
1081}
1082
1083static int shmem_setattr(struct dentry *dentry, struct iattr *attr)
1084{
1085 struct inode *inode = d_inode(dentry);
1086 struct shmem_inode_info *info = SHMEM_I(inode);
1087 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1088 int error;
1089
1090 error = setattr_prepare(dentry, attr);
1091 if (error)
1092 return error;
1093
1094 if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
1095 loff_t oldsize = inode->i_size;
1096 loff_t newsize = attr->ia_size;
1097
1098 /* protected by i_mutex */
1099 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
1100 (newsize > oldsize && (info->seals & F_SEAL_GROW)))
1101 return -EPERM;
1102
1103 if (newsize != oldsize) {
1104 error = shmem_reacct_size(SHMEM_I(inode)->flags,
1105 oldsize, newsize);
1106 if (error)
1107 return error;
1108 i_size_write(inode, newsize);
1109 inode->i_ctime = inode->i_mtime = current_time(inode);
1110 }
1111 if (newsize <= oldsize) {
1112 loff_t holebegin = round_up(newsize, PAGE_SIZE);
1113 if (oldsize > holebegin)
1114 unmap_mapping_range(inode->i_mapping,
1115 holebegin, 0, 1);
1116 if (info->alloced)
1117 shmem_truncate_range(inode,
1118 newsize, (loff_t)-1);
1119 /* unmap again to remove racily COWed private pages */
1120 if (oldsize > holebegin)
1121 unmap_mapping_range(inode->i_mapping,
1122 holebegin, 0, 1);
1123
1124 /*
1125 * Part of the huge page can be beyond i_size: subject
1126 * to shrink under memory pressure.
1127 */
1128 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
1129 spin_lock(&sbinfo->shrinklist_lock);
1130 /*
1131 * _careful to defend against unlocked access to
1132 * ->shrink_list in shmem_unused_huge_shrink()
1133 */
1134 if (list_empty_careful(&info->shrinklist)) {
1135 list_add_tail(&info->shrinklist,
1136 &sbinfo->shrinklist);
1137 sbinfo->shrinklist_len++;
1138 }
1139 spin_unlock(&sbinfo->shrinklist_lock);
1140 }
1141 }
1142 }
1143
1144 setattr_copy(inode, attr);
1145 if (attr->ia_valid & ATTR_MODE)
1146 error = posix_acl_chmod(inode, inode->i_mode);
1147 return error;
1148}
1149
1150static void shmem_evict_inode(struct inode *inode)
1151{
1152 struct shmem_inode_info *info = SHMEM_I(inode);
1153 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1154
1155 if (inode->i_mapping->a_ops == &shmem_aops) {
1156 shmem_unacct_size(info->flags, inode->i_size);
1157 inode->i_size = 0;
1158 shmem_truncate_range(inode, 0, (loff_t)-1);
1159 if (!list_empty(&info->shrinklist)) {
1160 spin_lock(&sbinfo->shrinklist_lock);
1161 if (!list_empty(&info->shrinklist)) {
1162 list_del_init(&info->shrinklist);
1163 sbinfo->shrinklist_len--;
1164 }
1165 spin_unlock(&sbinfo->shrinklist_lock);
1166 }
1167 while (!list_empty(&info->swaplist)) {
1168 /* Wait while shmem_unuse() is scanning this inode... */
1169 wait_var_event(&info->stop_eviction,
1170 !atomic_read(&info->stop_eviction));
1171 mutex_lock(&shmem_swaplist_mutex);
1172 /* ...but beware of the race if we peeked too early */
1173 if (!atomic_read(&info->stop_eviction))
1174 list_del_init(&info->swaplist);
1175 mutex_unlock(&shmem_swaplist_mutex);
1176 }
1177 }
1178
1179 simple_xattrs_free(&info->xattrs);
1180 WARN_ON(inode->i_blocks);
1181 shmem_free_inode(inode->i_sb);
1182 clear_inode(inode);
1183}
1184
1185extern struct swap_info_struct *swap_info[];
1186
1187static int shmem_find_swap_entries(struct address_space *mapping,
1188 pgoff_t start, unsigned int nr_entries,
1189 struct page **entries, pgoff_t *indices,
1190 unsigned int type, bool frontswap)
1191{
1192 XA_STATE(xas, &mapping->i_pages, start);
1193 struct page *page;
1194 swp_entry_t entry;
1195 unsigned int ret = 0;
1196
1197 if (!nr_entries)
1198 return 0;
1199
1200 rcu_read_lock();
1201 xas_for_each(&xas, page, ULONG_MAX) {
1202 if (xas_retry(&xas, page))
1203 continue;
1204
1205 if (!xa_is_value(page))
1206 continue;
1207
1208 entry = radix_to_swp_entry(page);
1209 if (swp_type(entry) != type)
1210 continue;
1211 if (frontswap &&
1212 !frontswap_test(swap_info[type], swp_offset(entry)))
1213 continue;
1214
1215 indices[ret] = xas.xa_index;
1216 entries[ret] = page;
1217
1218 if (need_resched()) {
1219 xas_pause(&xas);
1220 cond_resched_rcu();
1221 }
1222 if (++ret == nr_entries)
1223 break;
1224 }
1225 rcu_read_unlock();
1226
1227 return ret;
1228}
1229
1230/*
1231 * Move the swapped pages for an inode to page cache. Returns the count
1232 * of pages swapped in, or the error in case of failure.
1233 */
1234static int shmem_unuse_swap_entries(struct inode *inode, struct pagevec pvec,
1235 pgoff_t *indices)
1236{
1237 int i = 0;
1238 int ret = 0;
1239 int error = 0;
1240 struct address_space *mapping = inode->i_mapping;
1241
1242 for (i = 0; i < pvec.nr; i++) {
1243 struct page *page = pvec.pages[i];
1244
1245 if (!xa_is_value(page))
1246 continue;
1247 error = shmem_swapin_page(inode, indices[i],
1248 &page, SGP_CACHE,
1249 mapping_gfp_mask(mapping),
1250 NULL, NULL);
1251 if (error == 0) {
1252 unlock_page(page);
1253 put_page(page);
1254 ret++;
1255 }
1256 if (error == -ENOMEM)
1257 break;
1258 error = 0;
1259 }
1260 return error ? error : ret;
1261}
1262
1263/*
1264 * If swap found in inode, free it and move page from swapcache to filecache.
1265 */
1266static int shmem_unuse_inode(struct inode *inode, unsigned int type,
1267 bool frontswap, unsigned long *fs_pages_to_unuse)
1268{
1269 struct address_space *mapping = inode->i_mapping;
1270 pgoff_t start = 0;
1271 struct pagevec pvec;
1272 pgoff_t indices[PAGEVEC_SIZE];
1273 bool frontswap_partial = (frontswap && *fs_pages_to_unuse > 0);
1274 int ret = 0;
1275
1276 pagevec_init(&pvec);
1277 do {
1278 unsigned int nr_entries = PAGEVEC_SIZE;
1279
1280 if (frontswap_partial && *fs_pages_to_unuse < PAGEVEC_SIZE)
1281 nr_entries = *fs_pages_to_unuse;
1282
1283 pvec.nr = shmem_find_swap_entries(mapping, start, nr_entries,
1284 pvec.pages, indices,
1285 type, frontswap);
1286 if (pvec.nr == 0) {
1287 ret = 0;
1288 break;
1289 }
1290
1291 ret = shmem_unuse_swap_entries(inode, pvec, indices);
1292 if (ret < 0)
1293 break;
1294
1295 if (frontswap_partial) {
1296 *fs_pages_to_unuse -= ret;
1297 if (*fs_pages_to_unuse == 0) {
1298 ret = FRONTSWAP_PAGES_UNUSED;
1299 break;
1300 }
1301 }
1302
1303 start = indices[pvec.nr - 1];
1304 } while (true);
1305
1306 return ret;
1307}
1308
1309/*
1310 * Read all the shared memory data that resides in the swap
1311 * device 'type' back into memory, so the swap device can be
1312 * unused.
1313 */
1314int shmem_unuse(unsigned int type, bool frontswap,
1315 unsigned long *fs_pages_to_unuse)
1316{
1317 struct shmem_inode_info *info, *next;
1318 int error = 0;
1319
1320 if (list_empty(&shmem_swaplist))
1321 return 0;
1322
1323 mutex_lock(&shmem_swaplist_mutex);
1324 list_for_each_entry_safe(info, next, &shmem_swaplist, swaplist) {
1325 if (!info->swapped) {
1326 list_del_init(&info->swaplist);
1327 continue;
1328 }
1329 /*
1330 * Drop the swaplist mutex while searching the inode for swap;
1331 * but before doing so, make sure shmem_evict_inode() will not
1332 * remove placeholder inode from swaplist, nor let it be freed
1333 * (igrab() would protect from unlink, but not from unmount).
1334 */
1335 atomic_inc(&info->stop_eviction);
1336 mutex_unlock(&shmem_swaplist_mutex);
1337
1338 error = shmem_unuse_inode(&info->vfs_inode, type, frontswap,
1339 fs_pages_to_unuse);
1340 cond_resched();
1341
1342 mutex_lock(&shmem_swaplist_mutex);
1343 next = list_next_entry(info, swaplist);
1344 if (!info->swapped)
1345 list_del_init(&info->swaplist);
1346 if (atomic_dec_and_test(&info->stop_eviction))
1347 wake_up_var(&info->stop_eviction);
1348 if (error)
1349 break;
1350 }
1351 mutex_unlock(&shmem_swaplist_mutex);
1352
1353 return error;
1354}
1355
1356/*
1357 * Move the page from the page cache to the swap cache.
1358 */
1359static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1360{
1361 struct shmem_inode_info *info;
1362 struct address_space *mapping;
1363 struct inode *inode;
1364 swp_entry_t swap;
1365 pgoff_t index;
1366
1367 VM_BUG_ON_PAGE(PageCompound(page), page);
1368 BUG_ON(!PageLocked(page));
1369 mapping = page->mapping;
1370 index = page->index;
1371 inode = mapping->host;
1372 info = SHMEM_I(inode);
1373 if (info->flags & VM_LOCKED)
1374 goto redirty;
1375 if (!total_swap_pages)
1376 goto redirty;
1377
1378 /*
1379 * Our capabilities prevent regular writeback or sync from ever calling
1380 * shmem_writepage; but a stacking filesystem might use ->writepage of
1381 * its underlying filesystem, in which case tmpfs should write out to
1382 * swap only in response to memory pressure, and not for the writeback
1383 * threads or sync.
1384 */
1385 if (!wbc->for_reclaim) {
1386 WARN_ON_ONCE(1); /* Still happens? Tell us about it! */
1387 goto redirty;
1388 }
1389
1390 /*
1391 * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
1392 * value into swapfile.c, the only way we can correctly account for a
1393 * fallocated page arriving here is now to initialize it and write it.
1394 *
1395 * That's okay for a page already fallocated earlier, but if we have
1396 * not yet completed the fallocation, then (a) we want to keep track
1397 * of this page in case we have to undo it, and (b) it may not be a
1398 * good idea to continue anyway, once we're pushing into swap. So
1399 * reactivate the page, and let shmem_fallocate() quit when too many.
1400 */
1401 if (!PageUptodate(page)) {
1402 if (inode->i_private) {
1403 struct shmem_falloc *shmem_falloc;
1404 spin_lock(&inode->i_lock);
1405 shmem_falloc = inode->i_private;
1406 if (shmem_falloc &&
1407 !shmem_falloc->waitq &&
1408 index >= shmem_falloc->start &&
1409 index < shmem_falloc->next)
1410 shmem_falloc->nr_unswapped++;
1411 else
1412 shmem_falloc = NULL;
1413 spin_unlock(&inode->i_lock);
1414 if (shmem_falloc)
1415 goto redirty;
1416 }
1417 clear_highpage(page);
1418 flush_dcache_page(page);
1419 SetPageUptodate(page);
1420 }
1421
1422 swap = get_swap_page(page);
1423 if (!swap.val)
1424 goto redirty;
1425
1426 /*
1427 * Add inode to shmem_unuse()'s list of swapped-out inodes,
1428 * if it's not already there. Do it now before the page is
1429 * moved to swap cache, when its pagelock no longer protects
1430 * the inode from eviction. But don't unlock the mutex until
1431 * we've incremented swapped, because shmem_unuse_inode() will
1432 * prune a !swapped inode from the swaplist under this mutex.
1433 */
1434 mutex_lock(&shmem_swaplist_mutex);
1435 if (list_empty(&info->swaplist))
1436 list_add(&info->swaplist, &shmem_swaplist);
1437
1438 if (add_to_swap_cache(page, swap,
1439 __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN,
1440 NULL) == 0) {
1441 spin_lock_irq(&info->lock);
1442 shmem_recalc_inode(inode);
1443 info->swapped++;
1444 spin_unlock_irq(&info->lock);
1445
1446 swap_shmem_alloc(swap);
1447 shmem_delete_from_page_cache(page, swp_to_radix_entry(swap));
1448
1449 mutex_unlock(&shmem_swaplist_mutex);
1450 BUG_ON(page_mapped(page));
1451 swap_writepage(page, wbc);
1452 return 0;
1453 }
1454
1455 mutex_unlock(&shmem_swaplist_mutex);
1456 put_swap_page(page, swap);
1457redirty:
1458 set_page_dirty(page);
1459 if (wbc->for_reclaim)
1460 return AOP_WRITEPAGE_ACTIVATE; /* Return with page locked */
1461 unlock_page(page);
1462 return 0;
1463}
1464
1465#if defined(CONFIG_NUMA) && defined(CONFIG_TMPFS)
1466static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1467{
1468 char buffer[64];
1469
1470 if (!mpol || mpol->mode == MPOL_DEFAULT)
1471 return; /* show nothing */
1472
1473 mpol_to_str(buffer, sizeof(buffer), mpol);
1474
1475 seq_printf(seq, ",mpol=%s", buffer);
1476}
1477
1478static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1479{
1480 struct mempolicy *mpol = NULL;
1481 if (sbinfo->mpol) {
1482 spin_lock(&sbinfo->stat_lock); /* prevent replace/use races */
1483 mpol = sbinfo->mpol;
1484 mpol_get(mpol);
1485 spin_unlock(&sbinfo->stat_lock);
1486 }
1487 return mpol;
1488}
1489#else /* !CONFIG_NUMA || !CONFIG_TMPFS */
1490static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1491{
1492}
1493static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1494{
1495 return NULL;
1496}
1497#endif /* CONFIG_NUMA && CONFIG_TMPFS */
1498#ifndef CONFIG_NUMA
1499#define vm_policy vm_private_data
1500#endif
1501
1502static void shmem_pseudo_vma_init(struct vm_area_struct *vma,
1503 struct shmem_inode_info *info, pgoff_t index)
1504{
1505 /* Create a pseudo vma that just contains the policy */
1506 vma_init(vma, NULL);
1507 /* Bias interleave by inode number to distribute better across nodes */
1508 vma->vm_pgoff = index + info->vfs_inode.i_ino;
1509 vma->vm_policy = mpol_shared_policy_lookup(&info->policy, index);
1510}
1511
1512static void shmem_pseudo_vma_destroy(struct vm_area_struct *vma)
1513{
1514 /* Drop reference taken by mpol_shared_policy_lookup() */
1515 mpol_cond_put(vma->vm_policy);
1516}
1517
1518static struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
1519 struct shmem_inode_info *info, pgoff_t index)
1520{
1521 struct vm_area_struct pvma;
1522 struct page *page;
1523 struct vm_fault vmf;
1524
1525 shmem_pseudo_vma_init(&pvma, info, index);
1526 vmf.vma = &pvma;
1527 vmf.address = 0;
1528 page = swap_cluster_readahead(swap, gfp, &vmf);
1529 shmem_pseudo_vma_destroy(&pvma);
1530
1531 return page;
1532}
1533
1534static struct page *shmem_alloc_hugepage(gfp_t gfp,
1535 struct shmem_inode_info *info, pgoff_t index)
1536{
1537 struct vm_area_struct pvma;
1538 struct address_space *mapping = info->vfs_inode.i_mapping;
1539 pgoff_t hindex;
1540 struct page *page;
1541
1542 hindex = round_down(index, HPAGE_PMD_NR);
1543 if (xa_find(&mapping->i_pages, &hindex, hindex + HPAGE_PMD_NR - 1,
1544 XA_PRESENT))
1545 return NULL;
1546
1547 shmem_pseudo_vma_init(&pvma, info, hindex);
1548 page = alloc_pages_vma(gfp | __GFP_COMP | __GFP_NORETRY | __GFP_NOWARN,
1549 HPAGE_PMD_ORDER, &pvma, 0, numa_node_id(), true);
1550 shmem_pseudo_vma_destroy(&pvma);
1551 if (page)
1552 prep_transhuge_page(page);
1553 else
1554 count_vm_event(THP_FILE_FALLBACK);
1555 return page;
1556}
1557
1558static struct page *shmem_alloc_page(gfp_t gfp,
1559 struct shmem_inode_info *info, pgoff_t index)
1560{
1561 struct vm_area_struct pvma;
1562 struct page *page;
1563
1564 shmem_pseudo_vma_init(&pvma, info, index);
1565 page = alloc_page_vma(gfp, &pvma, 0);
1566 shmem_pseudo_vma_destroy(&pvma);
1567
1568 return page;
1569}
1570
1571static struct page *shmem_alloc_and_acct_page(gfp_t gfp,
1572 struct inode *inode,
1573 pgoff_t index, bool huge)
1574{
1575 struct shmem_inode_info *info = SHMEM_I(inode);
1576 struct page *page;
1577 int nr;
1578 int err = -ENOSPC;
1579
1580 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1581 huge = false;
1582 nr = huge ? HPAGE_PMD_NR : 1;
1583
1584 if (!shmem_inode_acct_block(inode, nr))
1585 goto failed;
1586
1587 if (huge)
1588 page = shmem_alloc_hugepage(gfp, info, index);
1589 else
1590 page = shmem_alloc_page(gfp, info, index);
1591 if (page) {
1592 __SetPageLocked(page);
1593 __SetPageSwapBacked(page);
1594 return page;
1595 }
1596
1597 err = -ENOMEM;
1598 shmem_inode_unacct_blocks(inode, nr);
1599failed:
1600 return ERR_PTR(err);
1601}
1602
1603/*
1604 * When a page is moved from swapcache to shmem filecache (either by the
1605 * usual swapin of shmem_getpage_gfp(), or by the less common swapoff of
1606 * shmem_unuse_inode()), it may have been read in earlier from swap, in
1607 * ignorance of the mapping it belongs to. If that mapping has special
1608 * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
1609 * we may need to copy to a suitable page before moving to filecache.
1610 *
1611 * In a future release, this may well be extended to respect cpuset and
1612 * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
1613 * but for now it is a simple matter of zone.
1614 */
1615static bool shmem_should_replace_page(struct page *page, gfp_t gfp)
1616{
1617 return page_zonenum(page) > gfp_zone(gfp);
1618}
1619
1620static int shmem_replace_page(struct page **pagep, gfp_t gfp,
1621 struct shmem_inode_info *info, pgoff_t index)
1622{
1623 struct page *oldpage, *newpage;
1624 struct address_space *swap_mapping;
1625 swp_entry_t entry;
1626 pgoff_t swap_index;
1627 int error;
1628
1629 oldpage = *pagep;
1630 entry.val = page_private(oldpage);
1631 swap_index = swp_offset(entry);
1632 swap_mapping = page_mapping(oldpage);
1633
1634 /*
1635 * We have arrived here because our zones are constrained, so don't
1636 * limit chance of success by further cpuset and node constraints.
1637 */
1638 gfp &= ~GFP_CONSTRAINT_MASK;
1639 newpage = shmem_alloc_page(gfp, info, index);
1640 if (!newpage)
1641 return -ENOMEM;
1642
1643 get_page(newpage);
1644 copy_highpage(newpage, oldpage);
1645 flush_dcache_page(newpage);
1646
1647 __SetPageLocked(newpage);
1648 __SetPageSwapBacked(newpage);
1649 SetPageUptodate(newpage);
1650 set_page_private(newpage, entry.val);
1651 SetPageSwapCache(newpage);
1652
1653 /*
1654 * Our caller will very soon move newpage out of swapcache, but it's
1655 * a nice clean interface for us to replace oldpage by newpage there.
1656 */
1657 xa_lock_irq(&swap_mapping->i_pages);
1658 error = shmem_replace_entry(swap_mapping, swap_index, oldpage, newpage);
1659 if (!error) {
1660 mem_cgroup_migrate(oldpage, newpage);
1661 __inc_lruvec_page_state(newpage, NR_FILE_PAGES);
1662 __dec_lruvec_page_state(oldpage, NR_FILE_PAGES);
1663 }
1664 xa_unlock_irq(&swap_mapping->i_pages);
1665
1666 if (unlikely(error)) {
1667 /*
1668 * Is this possible? I think not, now that our callers check
1669 * both PageSwapCache and page_private after getting page lock;
1670 * but be defensive. Reverse old to newpage for clear and free.
1671 */
1672 oldpage = newpage;
1673 } else {
1674 lru_cache_add(newpage);
1675 *pagep = newpage;
1676 }
1677
1678 ClearPageSwapCache(oldpage);
1679 set_page_private(oldpage, 0);
1680
1681 unlock_page(oldpage);
1682 put_page(oldpage);
1683 put_page(oldpage);
1684 return error;
1685}
1686
1687/*
1688 * Swap in the page pointed to by *pagep.
1689 * Caller has to make sure that *pagep contains a valid swapped page.
1690 * Returns 0 and the page in pagep if success. On failure, returns the
1691 * error code and NULL in *pagep.
1692 */
1693static int shmem_swapin_page(struct inode *inode, pgoff_t index,
1694 struct page **pagep, enum sgp_type sgp,
1695 gfp_t gfp, struct vm_area_struct *vma,
1696 vm_fault_t *fault_type)
1697{
1698 struct address_space *mapping = inode->i_mapping;
1699 struct shmem_inode_info *info = SHMEM_I(inode);
1700 struct mm_struct *charge_mm = vma ? vma->vm_mm : current->mm;
1701 struct page *page;
1702 swp_entry_t swap;
1703 int error;
1704
1705 VM_BUG_ON(!*pagep || !xa_is_value(*pagep));
1706 swap = radix_to_swp_entry(*pagep);
1707 *pagep = NULL;
1708
1709 /* Look it up and read it in.. */
1710 page = lookup_swap_cache(swap, NULL, 0);
1711 if (!page) {
1712 /* Or update major stats only when swapin succeeds?? */
1713 if (fault_type) {
1714 *fault_type |= VM_FAULT_MAJOR;
1715 count_vm_event(PGMAJFAULT);
1716 count_memcg_event_mm(charge_mm, PGMAJFAULT);
1717 }
1718 /* Here we actually start the io */
1719 page = shmem_swapin(swap, gfp, info, index);
1720 if (!page) {
1721 error = -ENOMEM;
1722 goto failed;
1723 }
1724 }
1725
1726 /* We have to do this with page locked to prevent races */
1727 lock_page(page);
1728 if (!PageSwapCache(page) || page_private(page) != swap.val ||
1729 !shmem_confirm_swap(mapping, index, swap)) {
1730 error = -EEXIST;
1731 goto unlock;
1732 }
1733 if (!PageUptodate(page)) {
1734 error = -EIO;
1735 goto failed;
1736 }
1737 wait_on_page_writeback(page);
1738
1739 if (shmem_should_replace_page(page, gfp)) {
1740 error = shmem_replace_page(&page, gfp, info, index);
1741 if (error)
1742 goto failed;
1743 }
1744
1745 error = shmem_add_to_page_cache(page, mapping, index,
1746 swp_to_radix_entry(swap), gfp,
1747 charge_mm);
1748 if (error)
1749 goto failed;
1750
1751 spin_lock_irq(&info->lock);
1752 info->swapped--;
1753 shmem_recalc_inode(inode);
1754 spin_unlock_irq(&info->lock);
1755
1756 if (sgp == SGP_WRITE)
1757 mark_page_accessed(page);
1758
1759 delete_from_swap_cache(page);
1760 set_page_dirty(page);
1761 swap_free(swap);
1762
1763 *pagep = page;
1764 return 0;
1765failed:
1766 if (!shmem_confirm_swap(mapping, index, swap))
1767 error = -EEXIST;
1768unlock:
1769 if (page) {
1770 unlock_page(page);
1771 put_page(page);
1772 }
1773
1774 return error;
1775}
1776
1777/*
1778 * shmem_getpage_gfp - find page in cache, or get from swap, or allocate
1779 *
1780 * If we allocate a new one we do not mark it dirty. That's up to the
1781 * vm. If we swap it in we mark it dirty since we also free the swap
1782 * entry since a page cannot live in both the swap and page cache.
1783 *
1784 * vmf and fault_type are only supplied by shmem_fault:
1785 * otherwise they are NULL.
1786 */
1787static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
1788 struct page **pagep, enum sgp_type sgp, gfp_t gfp,
1789 struct vm_area_struct *vma, struct vm_fault *vmf,
1790 vm_fault_t *fault_type)
1791{
1792 struct address_space *mapping = inode->i_mapping;
1793 struct shmem_inode_info *info = SHMEM_I(inode);
1794 struct shmem_sb_info *sbinfo;
1795 struct mm_struct *charge_mm;
1796 struct page *page;
1797 enum sgp_type sgp_huge = sgp;
1798 pgoff_t hindex = index;
1799 int error;
1800 int once = 0;
1801 int alloced = 0;
1802
1803 if (index > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
1804 return -EFBIG;
1805 if (sgp == SGP_NOHUGE || sgp == SGP_HUGE)
1806 sgp = SGP_CACHE;
1807repeat:
1808 if (sgp <= SGP_CACHE &&
1809 ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1810 return -EINVAL;
1811 }
1812
1813 sbinfo = SHMEM_SB(inode->i_sb);
1814 charge_mm = vma ? vma->vm_mm : current->mm;
1815
1816 page = find_lock_entry(mapping, index);
1817 if (xa_is_value(page)) {
1818 error = shmem_swapin_page(inode, index, &page,
1819 sgp, gfp, vma, fault_type);
1820 if (error == -EEXIST)
1821 goto repeat;
1822
1823 *pagep = page;
1824 return error;
1825 }
1826
1827 if (page && sgp == SGP_WRITE)
1828 mark_page_accessed(page);
1829
1830 /* fallocated page? */
1831 if (page && !PageUptodate(page)) {
1832 if (sgp != SGP_READ)
1833 goto clear;
1834 unlock_page(page);
1835 put_page(page);
1836 page = NULL;
1837 }
1838 if (page || sgp == SGP_READ) {
1839 *pagep = page;
1840 return 0;
1841 }
1842
1843 /*
1844 * Fast cache lookup did not find it:
1845 * bring it back from swap or allocate.
1846 */
1847
1848 if (vma && userfaultfd_missing(vma)) {
1849 *fault_type = handle_userfault(vmf, VM_UFFD_MISSING);
1850 return 0;
1851 }
1852
1853 /* shmem_symlink() */
1854 if (mapping->a_ops != &shmem_aops)
1855 goto alloc_nohuge;
1856 if (shmem_huge == SHMEM_HUGE_DENY || sgp_huge == SGP_NOHUGE)
1857 goto alloc_nohuge;
1858 if (shmem_huge == SHMEM_HUGE_FORCE)
1859 goto alloc_huge;
1860 switch (sbinfo->huge) {
1861 case SHMEM_HUGE_NEVER:
1862 goto alloc_nohuge;
1863 case SHMEM_HUGE_WITHIN_SIZE: {
1864 loff_t i_size;
1865 pgoff_t off;
1866
1867 off = round_up(index, HPAGE_PMD_NR);
1868 i_size = round_up(i_size_read(inode), PAGE_SIZE);
1869 if (i_size >= HPAGE_PMD_SIZE &&
1870 i_size >> PAGE_SHIFT >= off)
1871 goto alloc_huge;
1872
1873 fallthrough;
1874 }
1875 case SHMEM_HUGE_ADVISE:
1876 if (sgp_huge == SGP_HUGE)
1877 goto alloc_huge;
1878 /* TODO: implement fadvise() hints */
1879 goto alloc_nohuge;
1880 }
1881
1882alloc_huge:
1883 page = shmem_alloc_and_acct_page(gfp, inode, index, true);
1884 if (IS_ERR(page)) {
1885alloc_nohuge:
1886 page = shmem_alloc_and_acct_page(gfp, inode,
1887 index, false);
1888 }
1889 if (IS_ERR(page)) {
1890 int retry = 5;
1891
1892 error = PTR_ERR(page);
1893 page = NULL;
1894 if (error != -ENOSPC)
1895 goto unlock;
1896 /*
1897 * Try to reclaim some space by splitting a huge page
1898 * beyond i_size on the filesystem.
1899 */
1900 while (retry--) {
1901 int ret;
1902
1903 ret = shmem_unused_huge_shrink(sbinfo, NULL, 1);
1904 if (ret == SHRINK_STOP)
1905 break;
1906 if (ret)
1907 goto alloc_nohuge;
1908 }
1909 goto unlock;
1910 }
1911
1912 if (PageTransHuge(page))
1913 hindex = round_down(index, HPAGE_PMD_NR);
1914 else
1915 hindex = index;
1916
1917 if (sgp == SGP_WRITE)
1918 __SetPageReferenced(page);
1919
1920 error = shmem_add_to_page_cache(page, mapping, hindex,
1921 NULL, gfp & GFP_RECLAIM_MASK,
1922 charge_mm);
1923 if (error)
1924 goto unacct;
1925 lru_cache_add(page);
1926
1927 spin_lock_irq(&info->lock);
1928 info->alloced += compound_nr(page);
1929 inode->i_blocks += BLOCKS_PER_PAGE << compound_order(page);
1930 shmem_recalc_inode(inode);
1931 spin_unlock_irq(&info->lock);
1932 alloced = true;
1933
1934 if (PageTransHuge(page) &&
1935 DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE) <
1936 hindex + HPAGE_PMD_NR - 1) {
1937 /*
1938 * Part of the huge page is beyond i_size: subject
1939 * to shrink under memory pressure.
1940 */
1941 spin_lock(&sbinfo->shrinklist_lock);
1942 /*
1943 * _careful to defend against unlocked access to
1944 * ->shrink_list in shmem_unused_huge_shrink()
1945 */
1946 if (list_empty_careful(&info->shrinklist)) {
1947 list_add_tail(&info->shrinklist,
1948 &sbinfo->shrinklist);
1949 sbinfo->shrinklist_len++;
1950 }
1951 spin_unlock(&sbinfo->shrinklist_lock);
1952 }
1953
1954 /*
1955 * Let SGP_FALLOC use the SGP_WRITE optimization on a new page.
1956 */
1957 if (sgp == SGP_FALLOC)
1958 sgp = SGP_WRITE;
1959clear:
1960 /*
1961 * Let SGP_WRITE caller clear ends if write does not fill page;
1962 * but SGP_FALLOC on a page fallocated earlier must initialize
1963 * it now, lest undo on failure cancel our earlier guarantee.
1964 */
1965 if (sgp != SGP_WRITE && !PageUptodate(page)) {
1966 struct page *head = compound_head(page);
1967 int i;
1968
1969 for (i = 0; i < compound_nr(head); i++) {
1970 clear_highpage(head + i);
1971 flush_dcache_page(head + i);
1972 }
1973 SetPageUptodate(head);
1974 }
1975
1976 /* Perhaps the file has been truncated since we checked */
1977 if (sgp <= SGP_CACHE &&
1978 ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1979 if (alloced) {
1980 ClearPageDirty(page);
1981 delete_from_page_cache(page);
1982 spin_lock_irq(&info->lock);
1983 shmem_recalc_inode(inode);
1984 spin_unlock_irq(&info->lock);
1985 }
1986 error = -EINVAL;
1987 goto unlock;
1988 }
1989 *pagep = page + index - hindex;
1990 return 0;
1991
1992 /*
1993 * Error recovery.
1994 */
1995unacct:
1996 shmem_inode_unacct_blocks(inode, compound_nr(page));
1997
1998 if (PageTransHuge(page)) {
1999 unlock_page(page);
2000 put_page(page);
2001 goto alloc_nohuge;
2002 }
2003unlock:
2004 if (page) {
2005 unlock_page(page);
2006 put_page(page);
2007 }
2008 if (error == -ENOSPC && !once++) {
2009 spin_lock_irq(&info->lock);
2010 shmem_recalc_inode(inode);
2011 spin_unlock_irq(&info->lock);
2012 goto repeat;
2013 }
2014 if (error == -EEXIST)
2015 goto repeat;
2016 return error;
2017}
2018
2019/*
2020 * This is like autoremove_wake_function, but it removes the wait queue
2021 * entry unconditionally - even if something else had already woken the
2022 * target.
2023 */
2024static int synchronous_wake_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
2025{
2026 int ret = default_wake_function(wait, mode, sync, key);
2027 list_del_init(&wait->entry);
2028 return ret;
2029}
2030
2031static vm_fault_t shmem_fault(struct vm_fault *vmf)
2032{
2033 struct vm_area_struct *vma = vmf->vma;
2034 struct inode *inode = file_inode(vma->vm_file);
2035 gfp_t gfp = mapping_gfp_mask(inode->i_mapping);
2036 enum sgp_type sgp;
2037 int err;
2038 vm_fault_t ret = VM_FAULT_LOCKED;
2039
2040 /*
2041 * Trinity finds that probing a hole which tmpfs is punching can
2042 * prevent the hole-punch from ever completing: which in turn
2043 * locks writers out with its hold on i_mutex. So refrain from
2044 * faulting pages into the hole while it's being punched. Although
2045 * shmem_undo_range() does remove the additions, it may be unable to
2046 * keep up, as each new page needs its own unmap_mapping_range() call,
2047 * and the i_mmap tree grows ever slower to scan if new vmas are added.
2048 *
2049 * It does not matter if we sometimes reach this check just before the
2050 * hole-punch begins, so that one fault then races with the punch:
2051 * we just need to make racing faults a rare case.
2052 *
2053 * The implementation below would be much simpler if we just used a
2054 * standard mutex or completion: but we cannot take i_mutex in fault,
2055 * and bloating every shmem inode for this unlikely case would be sad.
2056 */
2057 if (unlikely(inode->i_private)) {
2058 struct shmem_falloc *shmem_falloc;
2059
2060 spin_lock(&inode->i_lock);
2061 shmem_falloc = inode->i_private;
2062 if (shmem_falloc &&
2063 shmem_falloc->waitq &&
2064 vmf->pgoff >= shmem_falloc->start &&
2065 vmf->pgoff < shmem_falloc->next) {
2066 struct file *fpin;
2067 wait_queue_head_t *shmem_falloc_waitq;
2068 DEFINE_WAIT_FUNC(shmem_fault_wait, synchronous_wake_function);
2069
2070 ret = VM_FAULT_NOPAGE;
2071 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2072 if (fpin)
2073 ret = VM_FAULT_RETRY;
2074
2075 shmem_falloc_waitq = shmem_falloc->waitq;
2076 prepare_to_wait(shmem_falloc_waitq, &shmem_fault_wait,
2077 TASK_UNINTERRUPTIBLE);
2078 spin_unlock(&inode->i_lock);
2079 schedule();
2080
2081 /*
2082 * shmem_falloc_waitq points into the shmem_fallocate()
2083 * stack of the hole-punching task: shmem_falloc_waitq
2084 * is usually invalid by the time we reach here, but
2085 * finish_wait() does not dereference it in that case;
2086 * though i_lock needed lest racing with wake_up_all().
2087 */
2088 spin_lock(&inode->i_lock);
2089 finish_wait(shmem_falloc_waitq, &shmem_fault_wait);
2090 spin_unlock(&inode->i_lock);
2091
2092 if (fpin)
2093 fput(fpin);
2094 return ret;
2095 }
2096 spin_unlock(&inode->i_lock);
2097 }
2098
2099 sgp = SGP_CACHE;
2100
2101 if ((vma->vm_flags & VM_NOHUGEPAGE) ||
2102 test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
2103 sgp = SGP_NOHUGE;
2104 else if (vma->vm_flags & VM_HUGEPAGE)
2105 sgp = SGP_HUGE;
2106
2107 err = shmem_getpage_gfp(inode, vmf->pgoff, &vmf->page, sgp,
2108 gfp, vma, vmf, &ret);
2109 if (err)
2110 return vmf_error(err);
2111 return ret;
2112}
2113
2114unsigned long shmem_get_unmapped_area(struct file *file,
2115 unsigned long uaddr, unsigned long len,
2116 unsigned long pgoff, unsigned long flags)
2117{
2118 unsigned long (*get_area)(struct file *,
2119 unsigned long, unsigned long, unsigned long, unsigned long);
2120 unsigned long addr;
2121 unsigned long offset;
2122 unsigned long inflated_len;
2123 unsigned long inflated_addr;
2124 unsigned long inflated_offset;
2125
2126 if (len > TASK_SIZE)
2127 return -ENOMEM;
2128
2129 get_area = current->mm->get_unmapped_area;
2130 addr = get_area(file, uaddr, len, pgoff, flags);
2131
2132 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
2133 return addr;
2134 if (IS_ERR_VALUE(addr))
2135 return addr;
2136 if (addr & ~PAGE_MASK)
2137 return addr;
2138 if (addr > TASK_SIZE - len)
2139 return addr;
2140
2141 if (shmem_huge == SHMEM_HUGE_DENY)
2142 return addr;
2143 if (len < HPAGE_PMD_SIZE)
2144 return addr;
2145 if (flags & MAP_FIXED)
2146 return addr;
2147 /*
2148 * Our priority is to support MAP_SHARED mapped hugely;
2149 * and support MAP_PRIVATE mapped hugely too, until it is COWed.
2150 * But if caller specified an address hint and we allocated area there
2151 * successfully, respect that as before.
2152 */
2153 if (uaddr == addr)
2154 return addr;
2155
2156 if (shmem_huge != SHMEM_HUGE_FORCE) {
2157 struct super_block *sb;
2158
2159 if (file) {
2160 VM_BUG_ON(file->f_op != &shmem_file_operations);
2161 sb = file_inode(file)->i_sb;
2162 } else {
2163 /*
2164 * Called directly from mm/mmap.c, or drivers/char/mem.c
2165 * for "/dev/zero", to create a shared anonymous object.
2166 */
2167 if (IS_ERR(shm_mnt))
2168 return addr;
2169 sb = shm_mnt->mnt_sb;
2170 }
2171 if (SHMEM_SB(sb)->huge == SHMEM_HUGE_NEVER)
2172 return addr;
2173 }
2174
2175 offset = (pgoff << PAGE_SHIFT) & (HPAGE_PMD_SIZE-1);
2176 if (offset && offset + len < 2 * HPAGE_PMD_SIZE)
2177 return addr;
2178 if ((addr & (HPAGE_PMD_SIZE-1)) == offset)
2179 return addr;
2180
2181 inflated_len = len + HPAGE_PMD_SIZE - PAGE_SIZE;
2182 if (inflated_len > TASK_SIZE)
2183 return addr;
2184 if (inflated_len < len)
2185 return addr;
2186
2187 inflated_addr = get_area(NULL, uaddr, inflated_len, 0, flags);
2188 if (IS_ERR_VALUE(inflated_addr))
2189 return addr;
2190 if (inflated_addr & ~PAGE_MASK)
2191 return addr;
2192
2193 inflated_offset = inflated_addr & (HPAGE_PMD_SIZE-1);
2194 inflated_addr += offset - inflated_offset;
2195 if (inflated_offset > offset)
2196 inflated_addr += HPAGE_PMD_SIZE;
2197
2198 if (inflated_addr > TASK_SIZE - len)
2199 return addr;
2200 return inflated_addr;
2201}
2202
2203#ifdef CONFIG_NUMA
2204static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
2205{
2206 struct inode *inode = file_inode(vma->vm_file);
2207 return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
2208}
2209
2210static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
2211 unsigned long addr)
2212{
2213 struct inode *inode = file_inode(vma->vm_file);
2214 pgoff_t index;
2215
2216 index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2217 return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
2218}
2219#endif
2220
2221int shmem_lock(struct file *file, int lock, struct user_struct *user)
2222{
2223 struct inode *inode = file_inode(file);
2224 struct shmem_inode_info *info = SHMEM_I(inode);
2225 int retval = -ENOMEM;
2226
2227 /*
2228 * What serializes the accesses to info->flags?
2229 * ipc_lock_object() when called from shmctl_do_lock(),
2230 * no serialization needed when called from shm_destroy().
2231 */
2232 if (lock && !(info->flags & VM_LOCKED)) {
2233 if (!user_shm_lock(inode->i_size, user))
2234 goto out_nomem;
2235 info->flags |= VM_LOCKED;
2236 mapping_set_unevictable(file->f_mapping);
2237 }
2238 if (!lock && (info->flags & VM_LOCKED) && user) {
2239 user_shm_unlock(inode->i_size, user);
2240 info->flags &= ~VM_LOCKED;
2241 mapping_clear_unevictable(file->f_mapping);
2242 }
2243 retval = 0;
2244
2245out_nomem:
2246 return retval;
2247}
2248
2249static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
2250{
2251 struct shmem_inode_info *info = SHMEM_I(file_inode(file));
2252
2253 if (info->seals & F_SEAL_FUTURE_WRITE) {
2254 /*
2255 * New PROT_WRITE and MAP_SHARED mmaps are not allowed when
2256 * "future write" seal active.
2257 */
2258 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
2259 return -EPERM;
2260
2261 /*
2262 * Since an F_SEAL_FUTURE_WRITE sealed memfd can be mapped as
2263 * MAP_SHARED and read-only, take care to not allow mprotect to
2264 * revert protections on such mappings. Do this only for shared
2265 * mappings. For private mappings, don't need to mask
2266 * VM_MAYWRITE as we still want them to be COW-writable.
2267 */
2268 if (vma->vm_flags & VM_SHARED)
2269 vma->vm_flags &= ~(VM_MAYWRITE);
2270 }
2271
2272 file_accessed(file);
2273 vma->vm_ops = &shmem_vm_ops;
2274 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
2275 ((vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK) <
2276 (vma->vm_end & HPAGE_PMD_MASK)) {
2277 khugepaged_enter(vma, vma->vm_flags);
2278 }
2279 return 0;
2280}
2281
2282static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
2283 umode_t mode, dev_t dev, unsigned long flags)
2284{
2285 struct inode *inode;
2286 struct shmem_inode_info *info;
2287 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2288 ino_t ino;
2289
2290 if (shmem_reserve_inode(sb, &ino))
2291 return NULL;
2292
2293 inode = new_inode(sb);
2294 if (inode) {
2295 inode->i_ino = ino;
2296 inode_init_owner(inode, dir, mode);
2297 inode->i_blocks = 0;
2298 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
2299 inode->i_generation = prandom_u32();
2300 info = SHMEM_I(inode);
2301 memset(info, 0, (char *)inode - (char *)info);
2302 spin_lock_init(&info->lock);
2303 atomic_set(&info->stop_eviction, 0);
2304 info->seals = F_SEAL_SEAL;
2305 info->flags = flags & VM_NORESERVE;
2306 INIT_LIST_HEAD(&info->shrinklist);
2307 INIT_LIST_HEAD(&info->swaplist);
2308 simple_xattrs_init(&info->xattrs);
2309 cache_no_acl(inode);
2310
2311 switch (mode & S_IFMT) {
2312 default:
2313 inode->i_op = &shmem_special_inode_operations;
2314 init_special_inode(inode, mode, dev);
2315 break;
2316 case S_IFREG:
2317 inode->i_mapping->a_ops = &shmem_aops;
2318 inode->i_op = &shmem_inode_operations;
2319 inode->i_fop = &shmem_file_operations;
2320 mpol_shared_policy_init(&info->policy,
2321 shmem_get_sbmpol(sbinfo));
2322 break;
2323 case S_IFDIR:
2324 inc_nlink(inode);
2325 /* Some things misbehave if size == 0 on a directory */
2326 inode->i_size = 2 * BOGO_DIRENT_SIZE;
2327 inode->i_op = &shmem_dir_inode_operations;
2328 inode->i_fop = &simple_dir_operations;
2329 break;
2330 case S_IFLNK:
2331 /*
2332 * Must not load anything in the rbtree,
2333 * mpol_free_shared_policy will not be called.
2334 */
2335 mpol_shared_policy_init(&info->policy, NULL);
2336 break;
2337 }
2338
2339 lockdep_annotate_inode_mutex_key(inode);
2340 } else
2341 shmem_free_inode(sb);
2342 return inode;
2343}
2344
2345bool shmem_mapping(struct address_space *mapping)
2346{
2347 return mapping->a_ops == &shmem_aops;
2348}
2349
2350static int shmem_mfill_atomic_pte(struct mm_struct *dst_mm,
2351 pmd_t *dst_pmd,
2352 struct vm_area_struct *dst_vma,
2353 unsigned long dst_addr,
2354 unsigned long src_addr,
2355 bool zeropage,
2356 struct page **pagep)
2357{
2358 struct inode *inode = file_inode(dst_vma->vm_file);
2359 struct shmem_inode_info *info = SHMEM_I(inode);
2360 struct address_space *mapping = inode->i_mapping;
2361 gfp_t gfp = mapping_gfp_mask(mapping);
2362 pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
2363 spinlock_t *ptl;
2364 void *page_kaddr;
2365 struct page *page;
2366 pte_t _dst_pte, *dst_pte;
2367 int ret;
2368 pgoff_t offset, max_off;
2369
2370 ret = -ENOMEM;
2371 if (!shmem_inode_acct_block(inode, 1))
2372 goto out;
2373
2374 if (!*pagep) {
2375 page = shmem_alloc_page(gfp, info, pgoff);
2376 if (!page)
2377 goto out_unacct_blocks;
2378
2379 if (!zeropage) { /* mcopy_atomic */
2380 page_kaddr = kmap_atomic(page);
2381 ret = copy_from_user(page_kaddr,
2382 (const void __user *)src_addr,
2383 PAGE_SIZE);
2384 kunmap_atomic(page_kaddr);
2385
2386 /* fallback to copy_from_user outside mmap_lock */
2387 if (unlikely(ret)) {
2388 *pagep = page;
2389 shmem_inode_unacct_blocks(inode, 1);
2390 /* don't free the page */
2391 return -ENOENT;
2392 }
2393 } else { /* mfill_zeropage_atomic */
2394 clear_highpage(page);
2395 }
2396 } else {
2397 page = *pagep;
2398 *pagep = NULL;
2399 }
2400
2401 VM_BUG_ON(PageLocked(page) || PageSwapBacked(page));
2402 __SetPageLocked(page);
2403 __SetPageSwapBacked(page);
2404 __SetPageUptodate(page);
2405
2406 ret = -EFAULT;
2407 offset = linear_page_index(dst_vma, dst_addr);
2408 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2409 if (unlikely(offset >= max_off))
2410 goto out_release;
2411
2412 ret = shmem_add_to_page_cache(page, mapping, pgoff, NULL,
2413 gfp & GFP_RECLAIM_MASK, dst_mm);
2414 if (ret)
2415 goto out_release;
2416
2417 _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
2418 if (dst_vma->vm_flags & VM_WRITE)
2419 _dst_pte = pte_mkwrite(pte_mkdirty(_dst_pte));
2420 else {
2421 /*
2422 * We don't set the pte dirty if the vma has no
2423 * VM_WRITE permission, so mark the page dirty or it
2424 * could be freed from under us. We could do it
2425 * unconditionally before unlock_page(), but doing it
2426 * only if VM_WRITE is not set is faster.
2427 */
2428 set_page_dirty(page);
2429 }
2430
2431 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
2432
2433 ret = -EFAULT;
2434 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2435 if (unlikely(offset >= max_off))
2436 goto out_release_unlock;
2437
2438 ret = -EEXIST;
2439 if (!pte_none(*dst_pte))
2440 goto out_release_unlock;
2441
2442 lru_cache_add(page);
2443
2444 spin_lock_irq(&info->lock);
2445 info->alloced++;
2446 inode->i_blocks += BLOCKS_PER_PAGE;
2447 shmem_recalc_inode(inode);
2448 spin_unlock_irq(&info->lock);
2449
2450 inc_mm_counter(dst_mm, mm_counter_file(page));
2451 page_add_file_rmap(page, false);
2452 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
2453
2454 /* No need to invalidate - it was non-present before */
2455 update_mmu_cache(dst_vma, dst_addr, dst_pte);
2456 pte_unmap_unlock(dst_pte, ptl);
2457 unlock_page(page);
2458 ret = 0;
2459out:
2460 return ret;
2461out_release_unlock:
2462 pte_unmap_unlock(dst_pte, ptl);
2463 ClearPageDirty(page);
2464 delete_from_page_cache(page);
2465out_release:
2466 unlock_page(page);
2467 put_page(page);
2468out_unacct_blocks:
2469 shmem_inode_unacct_blocks(inode, 1);
2470 goto out;
2471}
2472
2473int shmem_mcopy_atomic_pte(struct mm_struct *dst_mm,
2474 pmd_t *dst_pmd,
2475 struct vm_area_struct *dst_vma,
2476 unsigned long dst_addr,
2477 unsigned long src_addr,
2478 struct page **pagep)
2479{
2480 return shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
2481 dst_addr, src_addr, false, pagep);
2482}
2483
2484int shmem_mfill_zeropage_pte(struct mm_struct *dst_mm,
2485 pmd_t *dst_pmd,
2486 struct vm_area_struct *dst_vma,
2487 unsigned long dst_addr)
2488{
2489 struct page *page = NULL;
2490
2491 return shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
2492 dst_addr, 0, true, &page);
2493}
2494
2495#ifdef CONFIG_TMPFS
2496static const struct inode_operations shmem_symlink_inode_operations;
2497static const struct inode_operations shmem_short_symlink_operations;
2498
2499#ifdef CONFIG_TMPFS_XATTR
2500static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
2501#else
2502#define shmem_initxattrs NULL
2503#endif
2504
2505static int
2506shmem_write_begin(struct file *file, struct address_space *mapping,
2507 loff_t pos, unsigned len, unsigned flags,
2508 struct page **pagep, void **fsdata)
2509{
2510 struct inode *inode = mapping->host;
2511 struct shmem_inode_info *info = SHMEM_I(inode);
2512 pgoff_t index = pos >> PAGE_SHIFT;
2513
2514 /* i_mutex is held by caller */
2515 if (unlikely(info->seals & (F_SEAL_GROW |
2516 F_SEAL_WRITE | F_SEAL_FUTURE_WRITE))) {
2517 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE))
2518 return -EPERM;
2519 if ((info->seals & F_SEAL_GROW) && pos + len > inode->i_size)
2520 return -EPERM;
2521 }
2522
2523 return shmem_getpage(inode, index, pagep, SGP_WRITE);
2524}
2525
2526static int
2527shmem_write_end(struct file *file, struct address_space *mapping,
2528 loff_t pos, unsigned len, unsigned copied,
2529 struct page *page, void *fsdata)
2530{
2531 struct inode *inode = mapping->host;
2532
2533 if (pos + copied > inode->i_size)
2534 i_size_write(inode, pos + copied);
2535
2536 if (!PageUptodate(page)) {
2537 struct page *head = compound_head(page);
2538 if (PageTransCompound(page)) {
2539 int i;
2540
2541 for (i = 0; i < HPAGE_PMD_NR; i++) {
2542 if (head + i == page)
2543 continue;
2544 clear_highpage(head + i);
2545 flush_dcache_page(head + i);
2546 }
2547 }
2548 if (copied < PAGE_SIZE) {
2549 unsigned from = pos & (PAGE_SIZE - 1);
2550 zero_user_segments(page, 0, from,
2551 from + copied, PAGE_SIZE);
2552 }
2553 SetPageUptodate(head);
2554 }
2555 set_page_dirty(page);
2556 unlock_page(page);
2557 put_page(page);
2558
2559 return copied;
2560}
2561
2562static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
2563{
2564 struct file *file = iocb->ki_filp;
2565 struct inode *inode = file_inode(file);
2566 struct address_space *mapping = inode->i_mapping;
2567 pgoff_t index;
2568 unsigned long offset;
2569 enum sgp_type sgp = SGP_READ;
2570 int error = 0;
2571 ssize_t retval = 0;
2572 loff_t *ppos = &iocb->ki_pos;
2573
2574 /*
2575 * Might this read be for a stacking filesystem? Then when reading
2576 * holes of a sparse file, we actually need to allocate those pages,
2577 * and even mark them dirty, so it cannot exceed the max_blocks limit.
2578 */
2579 if (!iter_is_iovec(to))
2580 sgp = SGP_CACHE;
2581
2582 index = *ppos >> PAGE_SHIFT;
2583 offset = *ppos & ~PAGE_MASK;
2584
2585 for (;;) {
2586 struct page *page = NULL;
2587 pgoff_t end_index;
2588 unsigned long nr, ret;
2589 loff_t i_size = i_size_read(inode);
2590
2591 end_index = i_size >> PAGE_SHIFT;
2592 if (index > end_index)
2593 break;
2594 if (index == end_index) {
2595 nr = i_size & ~PAGE_MASK;
2596 if (nr <= offset)
2597 break;
2598 }
2599
2600 error = shmem_getpage(inode, index, &page, sgp);
2601 if (error) {
2602 if (error == -EINVAL)
2603 error = 0;
2604 break;
2605 }
2606 if (page) {
2607 if (sgp == SGP_CACHE)
2608 set_page_dirty(page);
2609 unlock_page(page);
2610 }
2611
2612 /*
2613 * We must evaluate after, since reads (unlike writes)
2614 * are called without i_mutex protection against truncate
2615 */
2616 nr = PAGE_SIZE;
2617 i_size = i_size_read(inode);
2618 end_index = i_size >> PAGE_SHIFT;
2619 if (index == end_index) {
2620 nr = i_size & ~PAGE_MASK;
2621 if (nr <= offset) {
2622 if (page)
2623 put_page(page);
2624 break;
2625 }
2626 }
2627 nr -= offset;
2628
2629 if (page) {
2630 /*
2631 * If users can be writing to this page using arbitrary
2632 * virtual addresses, take care about potential aliasing
2633 * before reading the page on the kernel side.
2634 */
2635 if (mapping_writably_mapped(mapping))
2636 flush_dcache_page(page);
2637 /*
2638 * Mark the page accessed if we read the beginning.
2639 */
2640 if (!offset)
2641 mark_page_accessed(page);
2642 } else {
2643 page = ZERO_PAGE(0);
2644 get_page(page);
2645 }
2646
2647 /*
2648 * Ok, we have the page, and it's up-to-date, so
2649 * now we can copy it to user space...
2650 */
2651 ret = copy_page_to_iter(page, offset, nr, to);
2652 retval += ret;
2653 offset += ret;
2654 index += offset >> PAGE_SHIFT;
2655 offset &= ~PAGE_MASK;
2656
2657 put_page(page);
2658 if (!iov_iter_count(to))
2659 break;
2660 if (ret < nr) {
2661 error = -EFAULT;
2662 break;
2663 }
2664 cond_resched();
2665 }
2666
2667 *ppos = ((loff_t) index << PAGE_SHIFT) + offset;
2668 file_accessed(file);
2669 return retval ? retval : error;
2670}
2671
2672/*
2673 * llseek SEEK_DATA or SEEK_HOLE through the page cache.
2674 */
2675static pgoff_t shmem_seek_hole_data(struct address_space *mapping,
2676 pgoff_t index, pgoff_t end, int whence)
2677{
2678 struct page *page;
2679 struct pagevec pvec;
2680 pgoff_t indices[PAGEVEC_SIZE];
2681 bool done = false;
2682 int i;
2683
2684 pagevec_init(&pvec);
2685 pvec.nr = 1; /* start small: we may be there already */
2686 while (!done) {
2687 pvec.nr = find_get_entries(mapping, index,
2688 pvec.nr, pvec.pages, indices);
2689 if (!pvec.nr) {
2690 if (whence == SEEK_DATA)
2691 index = end;
2692 break;
2693 }
2694 for (i = 0; i < pvec.nr; i++, index++) {
2695 if (index < indices[i]) {
2696 if (whence == SEEK_HOLE) {
2697 done = true;
2698 break;
2699 }
2700 index = indices[i];
2701 }
2702 page = pvec.pages[i];
2703 if (page && !xa_is_value(page)) {
2704 if (!PageUptodate(page))
2705 page = NULL;
2706 }
2707 if (index >= end ||
2708 (page && whence == SEEK_DATA) ||
2709 (!page && whence == SEEK_HOLE)) {
2710 done = true;
2711 break;
2712 }
2713 }
2714 pagevec_remove_exceptionals(&pvec);
2715 pagevec_release(&pvec);
2716 pvec.nr = PAGEVEC_SIZE;
2717 cond_resched();
2718 }
2719 return index;
2720}
2721
2722static loff_t shmem_file_llseek(struct file *file, loff_t offset, int whence)
2723{
2724 struct address_space *mapping = file->f_mapping;
2725 struct inode *inode = mapping->host;
2726 pgoff_t start, end;
2727 loff_t new_offset;
2728
2729 if (whence != SEEK_DATA && whence != SEEK_HOLE)
2730 return generic_file_llseek_size(file, offset, whence,
2731 MAX_LFS_FILESIZE, i_size_read(inode));
2732 inode_lock(inode);
2733 /* We're holding i_mutex so we can access i_size directly */
2734
2735 if (offset < 0 || offset >= inode->i_size)
2736 offset = -ENXIO;
2737 else {
2738 start = offset >> PAGE_SHIFT;
2739 end = (inode->i_size + PAGE_SIZE - 1) >> PAGE_SHIFT;
2740 new_offset = shmem_seek_hole_data(mapping, start, end, whence);
2741 new_offset <<= PAGE_SHIFT;
2742 if (new_offset > offset) {
2743 if (new_offset < inode->i_size)
2744 offset = new_offset;
2745 else if (whence == SEEK_DATA)
2746 offset = -ENXIO;
2747 else
2748 offset = inode->i_size;
2749 }
2750 }
2751
2752 if (offset >= 0)
2753 offset = vfs_setpos(file, offset, MAX_LFS_FILESIZE);
2754 inode_unlock(inode);
2755 return offset;
2756}
2757
2758static long shmem_fallocate(struct file *file, int mode, loff_t offset,
2759 loff_t len)
2760{
2761 struct inode *inode = file_inode(file);
2762 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
2763 struct shmem_inode_info *info = SHMEM_I(inode);
2764 struct shmem_falloc shmem_falloc;
2765 pgoff_t start, index, end;
2766 int error;
2767
2768 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2769 return -EOPNOTSUPP;
2770
2771 inode_lock(inode);
2772
2773 if (mode & FALLOC_FL_PUNCH_HOLE) {
2774 struct address_space *mapping = file->f_mapping;
2775 loff_t unmap_start = round_up(offset, PAGE_SIZE);
2776 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
2777 DECLARE_WAIT_QUEUE_HEAD_ONSTACK(shmem_falloc_waitq);
2778
2779 /* protected by i_mutex */
2780 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
2781 error = -EPERM;
2782 goto out;
2783 }
2784
2785 shmem_falloc.waitq = &shmem_falloc_waitq;
2786 shmem_falloc.start = (u64)unmap_start >> PAGE_SHIFT;
2787 shmem_falloc.next = (unmap_end + 1) >> PAGE_SHIFT;
2788 spin_lock(&inode->i_lock);
2789 inode->i_private = &shmem_falloc;
2790 spin_unlock(&inode->i_lock);
2791
2792 if ((u64)unmap_end > (u64)unmap_start)
2793 unmap_mapping_range(mapping, unmap_start,
2794 1 + unmap_end - unmap_start, 0);
2795 shmem_truncate_range(inode, offset, offset + len - 1);
2796 /* No need to unmap again: hole-punching leaves COWed pages */
2797
2798 spin_lock(&inode->i_lock);
2799 inode->i_private = NULL;
2800 wake_up_all(&shmem_falloc_waitq);
2801 WARN_ON_ONCE(!list_empty(&shmem_falloc_waitq.head));
2802 spin_unlock(&inode->i_lock);
2803 error = 0;
2804 goto out;
2805 }
2806
2807 /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
2808 error = inode_newsize_ok(inode, offset + len);
2809 if (error)
2810 goto out;
2811
2812 if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
2813 error = -EPERM;
2814 goto out;
2815 }
2816
2817 start = offset >> PAGE_SHIFT;
2818 end = (offset + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
2819 /* Try to avoid a swapstorm if len is impossible to satisfy */
2820 if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
2821 error = -ENOSPC;
2822 goto out;
2823 }
2824
2825 shmem_falloc.waitq = NULL;
2826 shmem_falloc.start = start;
2827 shmem_falloc.next = start;
2828 shmem_falloc.nr_falloced = 0;
2829 shmem_falloc.nr_unswapped = 0;
2830 spin_lock(&inode->i_lock);
2831 inode->i_private = &shmem_falloc;
2832 spin_unlock(&inode->i_lock);
2833
2834 for (index = start; index < end; index++) {
2835 struct page *page;
2836
2837 /*
2838 * Good, the fallocate(2) manpage permits EINTR: we may have
2839 * been interrupted because we are using up too much memory.
2840 */
2841 if (signal_pending(current))
2842 error = -EINTR;
2843 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
2844 error = -ENOMEM;
2845 else
2846 error = shmem_getpage(inode, index, &page, SGP_FALLOC);
2847 if (error) {
2848 /* Remove the !PageUptodate pages we added */
2849 if (index > start) {
2850 shmem_undo_range(inode,
2851 (loff_t)start << PAGE_SHIFT,
2852 ((loff_t)index << PAGE_SHIFT) - 1, true);
2853 }
2854 goto undone;
2855 }
2856
2857 /*
2858 * Inform shmem_writepage() how far we have reached.
2859 * No need for lock or barrier: we have the page lock.
2860 */
2861 shmem_falloc.next++;
2862 if (!PageUptodate(page))
2863 shmem_falloc.nr_falloced++;
2864
2865 /*
2866 * If !PageUptodate, leave it that way so that freeable pages
2867 * can be recognized if we need to rollback on error later.
2868 * But set_page_dirty so that memory pressure will swap rather
2869 * than free the pages we are allocating (and SGP_CACHE pages
2870 * might still be clean: we now need to mark those dirty too).
2871 */
2872 set_page_dirty(page);
2873 unlock_page(page);
2874 put_page(page);
2875 cond_resched();
2876 }
2877
2878 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
2879 i_size_write(inode, offset + len);
2880 inode->i_ctime = current_time(inode);
2881undone:
2882 spin_lock(&inode->i_lock);
2883 inode->i_private = NULL;
2884 spin_unlock(&inode->i_lock);
2885out:
2886 inode_unlock(inode);
2887 return error;
2888}
2889
2890static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
2891{
2892 struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
2893
2894 buf->f_type = TMPFS_MAGIC;
2895 buf->f_bsize = PAGE_SIZE;
2896 buf->f_namelen = NAME_MAX;
2897 if (sbinfo->max_blocks) {
2898 buf->f_blocks = sbinfo->max_blocks;
2899 buf->f_bavail =
2900 buf->f_bfree = sbinfo->max_blocks -
2901 percpu_counter_sum(&sbinfo->used_blocks);
2902 }
2903 if (sbinfo->max_inodes) {
2904 buf->f_files = sbinfo->max_inodes;
2905 buf->f_ffree = sbinfo->free_inodes;
2906 }
2907 /* else leave those fields 0 like simple_statfs */
2908 return 0;
2909}
2910
2911/*
2912 * File creation. Allocate an inode, and we're done..
2913 */
2914static int
2915shmem_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
2916{
2917 struct inode *inode;
2918 int error = -ENOSPC;
2919
2920 inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
2921 if (inode) {
2922 error = simple_acl_create(dir, inode);
2923 if (error)
2924 goto out_iput;
2925 error = security_inode_init_security(inode, dir,
2926 &dentry->d_name,
2927 shmem_initxattrs, NULL);
2928 if (error && error != -EOPNOTSUPP)
2929 goto out_iput;
2930
2931 error = 0;
2932 dir->i_size += BOGO_DIRENT_SIZE;
2933 dir->i_ctime = dir->i_mtime = current_time(dir);
2934 d_instantiate(dentry, inode);
2935 dget(dentry); /* Extra count - pin the dentry in core */
2936 }
2937 return error;
2938out_iput:
2939 iput(inode);
2940 return error;
2941}
2942
2943static int
2944shmem_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
2945{
2946 struct inode *inode;
2947 int error = -ENOSPC;
2948
2949 inode = shmem_get_inode(dir->i_sb, dir, mode, 0, VM_NORESERVE);
2950 if (inode) {
2951 error = security_inode_init_security(inode, dir,
2952 NULL,
2953 shmem_initxattrs, NULL);
2954 if (error && error != -EOPNOTSUPP)
2955 goto out_iput;
2956 error = simple_acl_create(dir, inode);
2957 if (error)
2958 goto out_iput;
2959 d_tmpfile(dentry, inode);
2960 }
2961 return error;
2962out_iput:
2963 iput(inode);
2964 return error;
2965}
2966
2967static int shmem_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
2968{
2969 int error;
2970
2971 if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
2972 return error;
2973 inc_nlink(dir);
2974 return 0;
2975}
2976
2977static int shmem_create(struct inode *dir, struct dentry *dentry, umode_t mode,
2978 bool excl)
2979{
2980 return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
2981}
2982
2983/*
2984 * Link a file..
2985 */
2986static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
2987{
2988 struct inode *inode = d_inode(old_dentry);
2989 int ret = 0;
2990
2991 /*
2992 * No ordinary (disk based) filesystem counts links as inodes;
2993 * but each new link needs a new dentry, pinning lowmem, and
2994 * tmpfs dentries cannot be pruned until they are unlinked.
2995 * But if an O_TMPFILE file is linked into the tmpfs, the
2996 * first link must skip that, to get the accounting right.
2997 */
2998 if (inode->i_nlink) {
2999 ret = shmem_reserve_inode(inode->i_sb, NULL);
3000 if (ret)
3001 goto out;
3002 }
3003
3004 dir->i_size += BOGO_DIRENT_SIZE;
3005 inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
3006 inc_nlink(inode);
3007 ihold(inode); /* New dentry reference */
3008 dget(dentry); /* Extra pinning count for the created dentry */
3009 d_instantiate(dentry, inode);
3010out:
3011 return ret;
3012}
3013
3014static int shmem_unlink(struct inode *dir, struct dentry *dentry)
3015{
3016 struct inode *inode = d_inode(dentry);
3017
3018 if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
3019 shmem_free_inode(inode->i_sb);
3020
3021 dir->i_size -= BOGO_DIRENT_SIZE;
3022 inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
3023 drop_nlink(inode);
3024 dput(dentry); /* Undo the count from "create" - this does all the work */
3025 return 0;
3026}
3027
3028static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
3029{
3030 if (!simple_empty(dentry))
3031 return -ENOTEMPTY;
3032
3033 drop_nlink(d_inode(dentry));
3034 drop_nlink(dir);
3035 return shmem_unlink(dir, dentry);
3036}
3037
3038static int shmem_exchange(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
3039{
3040 bool old_is_dir = d_is_dir(old_dentry);
3041 bool new_is_dir = d_is_dir(new_dentry);
3042
3043 if (old_dir != new_dir && old_is_dir != new_is_dir) {
3044 if (old_is_dir) {
3045 drop_nlink(old_dir);
3046 inc_nlink(new_dir);
3047 } else {
3048 drop_nlink(new_dir);
3049 inc_nlink(old_dir);
3050 }
3051 }
3052 old_dir->i_ctime = old_dir->i_mtime =
3053 new_dir->i_ctime = new_dir->i_mtime =
3054 d_inode(old_dentry)->i_ctime =
3055 d_inode(new_dentry)->i_ctime = current_time(old_dir);
3056
3057 return 0;
3058}
3059
3060static int shmem_whiteout(struct inode *old_dir, struct dentry *old_dentry)
3061{
3062 struct dentry *whiteout;
3063 int error;
3064
3065 whiteout = d_alloc(old_dentry->d_parent, &old_dentry->d_name);
3066 if (!whiteout)
3067 return -ENOMEM;
3068
3069 error = shmem_mknod(old_dir, whiteout,
3070 S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
3071 dput(whiteout);
3072 if (error)
3073 return error;
3074
3075 /*
3076 * Cheat and hash the whiteout while the old dentry is still in
3077 * place, instead of playing games with FS_RENAME_DOES_D_MOVE.
3078 *
3079 * d_lookup() will consistently find one of them at this point,
3080 * not sure which one, but that isn't even important.
3081 */
3082 d_rehash(whiteout);
3083 return 0;
3084}
3085
3086/*
3087 * The VFS layer already does all the dentry stuff for rename,
3088 * we just have to decrement the usage count for the target if
3089 * it exists so that the VFS layer correctly free's it when it
3090 * gets overwritten.
3091 */
3092static int shmem_rename2(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry, unsigned int flags)
3093{
3094 struct inode *inode = d_inode(old_dentry);
3095 int they_are_dirs = S_ISDIR(inode->i_mode);
3096
3097 if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
3098 return -EINVAL;
3099
3100 if (flags & RENAME_EXCHANGE)
3101 return shmem_exchange(old_dir, old_dentry, new_dir, new_dentry);
3102
3103 if (!simple_empty(new_dentry))
3104 return -ENOTEMPTY;
3105
3106 if (flags & RENAME_WHITEOUT) {
3107 int error;
3108
3109 error = shmem_whiteout(old_dir, old_dentry);
3110 if (error)
3111 return error;
3112 }
3113
3114 if (d_really_is_positive(new_dentry)) {
3115 (void) shmem_unlink(new_dir, new_dentry);
3116 if (they_are_dirs) {
3117 drop_nlink(d_inode(new_dentry));
3118 drop_nlink(old_dir);
3119 }
3120 } else if (they_are_dirs) {
3121 drop_nlink(old_dir);
3122 inc_nlink(new_dir);
3123 }
3124
3125 old_dir->i_size -= BOGO_DIRENT_SIZE;
3126 new_dir->i_size += BOGO_DIRENT_SIZE;
3127 old_dir->i_ctime = old_dir->i_mtime =
3128 new_dir->i_ctime = new_dir->i_mtime =
3129 inode->i_ctime = current_time(old_dir);
3130 return 0;
3131}
3132
3133static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
3134{
3135 int error;
3136 int len;
3137 struct inode *inode;
3138 struct page *page;
3139
3140 len = strlen(symname) + 1;
3141 if (len > PAGE_SIZE)
3142 return -ENAMETOOLONG;
3143
3144 inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK | 0777, 0,
3145 VM_NORESERVE);
3146 if (!inode)
3147 return -ENOSPC;
3148
3149 error = security_inode_init_security(inode, dir, &dentry->d_name,
3150 shmem_initxattrs, NULL);
3151 if (error && error != -EOPNOTSUPP) {
3152 iput(inode);
3153 return error;
3154 }
3155
3156 inode->i_size = len-1;
3157 if (len <= SHORT_SYMLINK_LEN) {
3158 inode->i_link = kmemdup(symname, len, GFP_KERNEL);
3159 if (!inode->i_link) {
3160 iput(inode);
3161 return -ENOMEM;
3162 }
3163 inode->i_op = &shmem_short_symlink_operations;
3164 } else {
3165 inode_nohighmem(inode);
3166 error = shmem_getpage(inode, 0, &page, SGP_WRITE);
3167 if (error) {
3168 iput(inode);
3169 return error;
3170 }
3171 inode->i_mapping->a_ops = &shmem_aops;
3172 inode->i_op = &shmem_symlink_inode_operations;
3173 memcpy(page_address(page), symname, len);
3174 SetPageUptodate(page);
3175 set_page_dirty(page);
3176 unlock_page(page);
3177 put_page(page);
3178 }
3179 dir->i_size += BOGO_DIRENT_SIZE;
3180 dir->i_ctime = dir->i_mtime = current_time(dir);
3181 d_instantiate(dentry, inode);
3182 dget(dentry);
3183 return 0;
3184}
3185
3186static void shmem_put_link(void *arg)
3187{
3188 mark_page_accessed(arg);
3189 put_page(arg);
3190}
3191
3192static const char *shmem_get_link(struct dentry *dentry,
3193 struct inode *inode,
3194 struct delayed_call *done)
3195{
3196 struct page *page = NULL;
3197 int error;
3198 if (!dentry) {
3199 page = find_get_page(inode->i_mapping, 0);
3200 if (!page)
3201 return ERR_PTR(-ECHILD);
3202 if (!PageUptodate(page)) {
3203 put_page(page);
3204 return ERR_PTR(-ECHILD);
3205 }
3206 } else {
3207 error = shmem_getpage(inode, 0, &page, SGP_READ);
3208 if (error)
3209 return ERR_PTR(error);
3210 unlock_page(page);
3211 }
3212 set_delayed_call(done, shmem_put_link, page);
3213 return page_address(page);
3214}
3215
3216#ifdef CONFIG_TMPFS_XATTR
3217/*
3218 * Superblocks without xattr inode operations may get some security.* xattr
3219 * support from the LSM "for free". As soon as we have any other xattrs
3220 * like ACLs, we also need to implement the security.* handlers at
3221 * filesystem level, though.
3222 */
3223
3224/*
3225 * Callback for security_inode_init_security() for acquiring xattrs.
3226 */
3227static int shmem_initxattrs(struct inode *inode,
3228 const struct xattr *xattr_array,
3229 void *fs_info)
3230{
3231 struct shmem_inode_info *info = SHMEM_I(inode);
3232 const struct xattr *xattr;
3233 struct simple_xattr *new_xattr;
3234 size_t len;
3235
3236 for (xattr = xattr_array; xattr->name != NULL; xattr++) {
3237 new_xattr = simple_xattr_alloc(xattr->value, xattr->value_len);
3238 if (!new_xattr)
3239 return -ENOMEM;
3240
3241 len = strlen(xattr->name) + 1;
3242 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
3243 GFP_KERNEL);
3244 if (!new_xattr->name) {
3245 kvfree(new_xattr);
3246 return -ENOMEM;
3247 }
3248
3249 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
3250 XATTR_SECURITY_PREFIX_LEN);
3251 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
3252 xattr->name, len);
3253
3254 simple_xattr_list_add(&info->xattrs, new_xattr);
3255 }
3256
3257 return 0;
3258}
3259
3260static int shmem_xattr_handler_get(const struct xattr_handler *handler,
3261 struct dentry *unused, struct inode *inode,
3262 const char *name, void *buffer, size_t size)
3263{
3264 struct shmem_inode_info *info = SHMEM_I(inode);
3265
3266 name = xattr_full_name(handler, name);
3267 return simple_xattr_get(&info->xattrs, name, buffer, size);
3268}
3269
3270static int shmem_xattr_handler_set(const struct xattr_handler *handler,
3271 struct dentry *unused, struct inode *inode,
3272 const char *name, const void *value,
3273 size_t size, int flags)
3274{
3275 struct shmem_inode_info *info = SHMEM_I(inode);
3276
3277 name = xattr_full_name(handler, name);
3278 return simple_xattr_set(&info->xattrs, name, value, size, flags, NULL);
3279}
3280
3281static const struct xattr_handler shmem_security_xattr_handler = {
3282 .prefix = XATTR_SECURITY_PREFIX,
3283 .get = shmem_xattr_handler_get,
3284 .set = shmem_xattr_handler_set,
3285};
3286
3287static const struct xattr_handler shmem_trusted_xattr_handler = {
3288 .prefix = XATTR_TRUSTED_PREFIX,
3289 .get = shmem_xattr_handler_get,
3290 .set = shmem_xattr_handler_set,
3291};
3292
3293static const struct xattr_handler *shmem_xattr_handlers[] = {
3294#ifdef CONFIG_TMPFS_POSIX_ACL
3295 &posix_acl_access_xattr_handler,
3296 &posix_acl_default_xattr_handler,
3297#endif
3298 &shmem_security_xattr_handler,
3299 &shmem_trusted_xattr_handler,
3300 NULL
3301};
3302
3303static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
3304{
3305 struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
3306 return simple_xattr_list(d_inode(dentry), &info->xattrs, buffer, size);
3307}
3308#endif /* CONFIG_TMPFS_XATTR */
3309
3310static const struct inode_operations shmem_short_symlink_operations = {
3311 .get_link = simple_get_link,
3312#ifdef CONFIG_TMPFS_XATTR
3313 .listxattr = shmem_listxattr,
3314#endif
3315};
3316
3317static const struct inode_operations shmem_symlink_inode_operations = {
3318 .get_link = shmem_get_link,
3319#ifdef CONFIG_TMPFS_XATTR
3320 .listxattr = shmem_listxattr,
3321#endif
3322};
3323
3324static struct dentry *shmem_get_parent(struct dentry *child)
3325{
3326 return ERR_PTR(-ESTALE);
3327}
3328
3329static int shmem_match(struct inode *ino, void *vfh)
3330{
3331 __u32 *fh = vfh;
3332 __u64 inum = fh[2];
3333 inum = (inum << 32) | fh[1];
3334 return ino->i_ino == inum && fh[0] == ino->i_generation;
3335}
3336
3337/* Find any alias of inode, but prefer a hashed alias */
3338static struct dentry *shmem_find_alias(struct inode *inode)
3339{
3340 struct dentry *alias = d_find_alias(inode);
3341
3342 return alias ?: d_find_any_alias(inode);
3343}
3344
3345
3346static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
3347 struct fid *fid, int fh_len, int fh_type)
3348{
3349 struct inode *inode;
3350 struct dentry *dentry = NULL;
3351 u64 inum;
3352
3353 if (fh_len < 3)
3354 return NULL;
3355
3356 inum = fid->raw[2];
3357 inum = (inum << 32) | fid->raw[1];
3358
3359 inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
3360 shmem_match, fid->raw);
3361 if (inode) {
3362 dentry = shmem_find_alias(inode);
3363 iput(inode);
3364 }
3365
3366 return dentry;
3367}
3368
3369static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
3370 struct inode *parent)
3371{
3372 if (*len < 3) {
3373 *len = 3;
3374 return FILEID_INVALID;
3375 }
3376
3377 if (inode_unhashed(inode)) {
3378 /* Unfortunately insert_inode_hash is not idempotent,
3379 * so as we hash inodes here rather than at creation
3380 * time, we need a lock to ensure we only try
3381 * to do it once
3382 */
3383 static DEFINE_SPINLOCK(lock);
3384 spin_lock(&lock);
3385 if (inode_unhashed(inode))
3386 __insert_inode_hash(inode,
3387 inode->i_ino + inode->i_generation);
3388 spin_unlock(&lock);
3389 }
3390
3391 fh[0] = inode->i_generation;
3392 fh[1] = inode->i_ino;
3393 fh[2] = ((__u64)inode->i_ino) >> 32;
3394
3395 *len = 3;
3396 return 1;
3397}
3398
3399static const struct export_operations shmem_export_ops = {
3400 .get_parent = shmem_get_parent,
3401 .encode_fh = shmem_encode_fh,
3402 .fh_to_dentry = shmem_fh_to_dentry,
3403};
3404
3405enum shmem_param {
3406 Opt_gid,
3407 Opt_huge,
3408 Opt_mode,
3409 Opt_mpol,
3410 Opt_nr_blocks,
3411 Opt_nr_inodes,
3412 Opt_size,
3413 Opt_uid,
3414 Opt_inode32,
3415 Opt_inode64,
3416};
3417
3418static const struct constant_table shmem_param_enums_huge[] = {
3419 {"never", SHMEM_HUGE_NEVER },
3420 {"always", SHMEM_HUGE_ALWAYS },
3421 {"within_size", SHMEM_HUGE_WITHIN_SIZE },
3422 {"advise", SHMEM_HUGE_ADVISE },
3423 {}
3424};
3425
3426const struct fs_parameter_spec shmem_fs_parameters[] = {
3427 fsparam_u32 ("gid", Opt_gid),
3428 fsparam_enum ("huge", Opt_huge, shmem_param_enums_huge),
3429 fsparam_u32oct("mode", Opt_mode),
3430 fsparam_string("mpol", Opt_mpol),
3431 fsparam_string("nr_blocks", Opt_nr_blocks),
3432 fsparam_string("nr_inodes", Opt_nr_inodes),
3433 fsparam_string("size", Opt_size),
3434 fsparam_u32 ("uid", Opt_uid),
3435 fsparam_flag ("inode32", Opt_inode32),
3436 fsparam_flag ("inode64", Opt_inode64),
3437 {}
3438};
3439
3440static int shmem_parse_one(struct fs_context *fc, struct fs_parameter *param)
3441{
3442 struct shmem_options *ctx = fc->fs_private;
3443 struct fs_parse_result result;
3444 unsigned long long size;
3445 char *rest;
3446 int opt;
3447
3448 opt = fs_parse(fc, shmem_fs_parameters, param, &result);
3449 if (opt < 0)
3450 return opt;
3451
3452 switch (opt) {
3453 case Opt_size:
3454 size = memparse(param->string, &rest);
3455 if (*rest == '%') {
3456 size <<= PAGE_SHIFT;
3457 size *= totalram_pages();
3458 do_div(size, 100);
3459 rest++;
3460 }
3461 if (*rest)
3462 goto bad_value;
3463 ctx->blocks = DIV_ROUND_UP(size, PAGE_SIZE);
3464 ctx->seen |= SHMEM_SEEN_BLOCKS;
3465 break;
3466 case Opt_nr_blocks:
3467 ctx->blocks = memparse(param->string, &rest);
3468 if (*rest)
3469 goto bad_value;
3470 ctx->seen |= SHMEM_SEEN_BLOCKS;
3471 break;
3472 case Opt_nr_inodes:
3473 ctx->inodes = memparse(param->string, &rest);
3474 if (*rest)
3475 goto bad_value;
3476 ctx->seen |= SHMEM_SEEN_INODES;
3477 break;
3478 case Opt_mode:
3479 ctx->mode = result.uint_32 & 07777;
3480 break;
3481 case Opt_uid:
3482 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
3483 if (!uid_valid(ctx->uid))
3484 goto bad_value;
3485 break;
3486 case Opt_gid:
3487 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
3488 if (!gid_valid(ctx->gid))
3489 goto bad_value;
3490 break;
3491 case Opt_huge:
3492 ctx->huge = result.uint_32;
3493 if (ctx->huge != SHMEM_HUGE_NEVER &&
3494 !(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
3495 has_transparent_hugepage()))
3496 goto unsupported_parameter;
3497 ctx->seen |= SHMEM_SEEN_HUGE;
3498 break;
3499 case Opt_mpol:
3500 if (IS_ENABLED(CONFIG_NUMA)) {
3501 mpol_put(ctx->mpol);
3502 ctx->mpol = NULL;
3503 if (mpol_parse_str(param->string, &ctx->mpol))
3504 goto bad_value;
3505 break;
3506 }
3507 goto unsupported_parameter;
3508 case Opt_inode32:
3509 ctx->full_inums = false;
3510 ctx->seen |= SHMEM_SEEN_INUMS;
3511 break;
3512 case Opt_inode64:
3513 if (sizeof(ino_t) < 8) {
3514 return invalfc(fc,
3515 "Cannot use inode64 with <64bit inums in kernel\n");
3516 }
3517 ctx->full_inums = true;
3518 ctx->seen |= SHMEM_SEEN_INUMS;
3519 break;
3520 }
3521 return 0;
3522
3523unsupported_parameter:
3524 return invalfc(fc, "Unsupported parameter '%s'", param->key);
3525bad_value:
3526 return invalfc(fc, "Bad value for '%s'", param->key);
3527}
3528
3529static int shmem_parse_options(struct fs_context *fc, void *data)
3530{
3531 char *options = data;
3532
3533 if (options) {
3534 int err = security_sb_eat_lsm_opts(options, &fc->security);
3535 if (err)
3536 return err;
3537 }
3538
3539 while (options != NULL) {
3540 char *this_char = options;
3541 for (;;) {
3542 /*
3543 * NUL-terminate this option: unfortunately,
3544 * mount options form a comma-separated list,
3545 * but mpol's nodelist may also contain commas.
3546 */
3547 options = strchr(options, ',');
3548 if (options == NULL)
3549 break;
3550 options++;
3551 if (!isdigit(*options)) {
3552 options[-1] = '\0';
3553 break;
3554 }
3555 }
3556 if (*this_char) {
3557 char *value = strchr(this_char,'=');
3558 size_t len = 0;
3559 int err;
3560
3561 if (value) {
3562 *value++ = '\0';
3563 len = strlen(value);
3564 }
3565 err = vfs_parse_fs_string(fc, this_char, value, len);
3566 if (err < 0)
3567 return err;
3568 }
3569 }
3570 return 0;
3571}
3572
3573/*
3574 * Reconfigure a shmem filesystem.
3575 *
3576 * Note that we disallow change from limited->unlimited blocks/inodes while any
3577 * are in use; but we must separately disallow unlimited->limited, because in
3578 * that case we have no record of how much is already in use.
3579 */
3580static int shmem_reconfigure(struct fs_context *fc)
3581{
3582 struct shmem_options *ctx = fc->fs_private;
3583 struct shmem_sb_info *sbinfo = SHMEM_SB(fc->root->d_sb);
3584 unsigned long inodes;
3585 const char *err;
3586
3587 spin_lock(&sbinfo->stat_lock);
3588 inodes = sbinfo->max_inodes - sbinfo->free_inodes;
3589 if ((ctx->seen & SHMEM_SEEN_BLOCKS) && ctx->blocks) {
3590 if (!sbinfo->max_blocks) {
3591 err = "Cannot retroactively limit size";
3592 goto out;
3593 }
3594 if (percpu_counter_compare(&sbinfo->used_blocks,
3595 ctx->blocks) > 0) {
3596 err = "Too small a size for current use";
3597 goto out;
3598 }
3599 }
3600 if ((ctx->seen & SHMEM_SEEN_INODES) && ctx->inodes) {
3601 if (!sbinfo->max_inodes) {
3602 err = "Cannot retroactively limit inodes";
3603 goto out;
3604 }
3605 if (ctx->inodes < inodes) {
3606 err = "Too few inodes for current use";
3607 goto out;
3608 }
3609 }
3610
3611 if ((ctx->seen & SHMEM_SEEN_INUMS) && !ctx->full_inums &&
3612 sbinfo->next_ino > UINT_MAX) {
3613 err = "Current inum too high to switch to 32-bit inums";
3614 goto out;
3615 }
3616
3617 if (ctx->seen & SHMEM_SEEN_HUGE)
3618 sbinfo->huge = ctx->huge;
3619 if (ctx->seen & SHMEM_SEEN_INUMS)
3620 sbinfo->full_inums = ctx->full_inums;
3621 if (ctx->seen & SHMEM_SEEN_BLOCKS)
3622 sbinfo->max_blocks = ctx->blocks;
3623 if (ctx->seen & SHMEM_SEEN_INODES) {
3624 sbinfo->max_inodes = ctx->inodes;
3625 sbinfo->free_inodes = ctx->inodes - inodes;
3626 }
3627
3628 /*
3629 * Preserve previous mempolicy unless mpol remount option was specified.
3630 */
3631 if (ctx->mpol) {
3632 mpol_put(sbinfo->mpol);
3633 sbinfo->mpol = ctx->mpol; /* transfers initial ref */
3634 ctx->mpol = NULL;
3635 }
3636 spin_unlock(&sbinfo->stat_lock);
3637 return 0;
3638out:
3639 spin_unlock(&sbinfo->stat_lock);
3640 return invalfc(fc, "%s", err);
3641}
3642
3643static int shmem_show_options(struct seq_file *seq, struct dentry *root)
3644{
3645 struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
3646
3647 if (sbinfo->max_blocks != shmem_default_max_blocks())
3648 seq_printf(seq, ",size=%luk",
3649 sbinfo->max_blocks << (PAGE_SHIFT - 10));
3650 if (sbinfo->max_inodes != shmem_default_max_inodes())
3651 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
3652 if (sbinfo->mode != (0777 | S_ISVTX))
3653 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
3654 if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
3655 seq_printf(seq, ",uid=%u",
3656 from_kuid_munged(&init_user_ns, sbinfo->uid));
3657 if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
3658 seq_printf(seq, ",gid=%u",
3659 from_kgid_munged(&init_user_ns, sbinfo->gid));
3660
3661 /*
3662 * Showing inode{64,32} might be useful even if it's the system default,
3663 * since then people don't have to resort to checking both here and
3664 * /proc/config.gz to confirm 64-bit inums were successfully applied
3665 * (which may not even exist if IKCONFIG_PROC isn't enabled).
3666 *
3667 * We hide it when inode64 isn't the default and we are using 32-bit
3668 * inodes, since that probably just means the feature isn't even under
3669 * consideration.
3670 *
3671 * As such:
3672 *
3673 * +-----------------+-----------------+
3674 * | TMPFS_INODE64=y | TMPFS_INODE64=n |
3675 * +------------------+-----------------+-----------------+
3676 * | full_inums=true | show | show |
3677 * | full_inums=false | show | hide |
3678 * +------------------+-----------------+-----------------+
3679 *
3680 */
3681 if (IS_ENABLED(CONFIG_TMPFS_INODE64) || sbinfo->full_inums)
3682 seq_printf(seq, ",inode%d", (sbinfo->full_inums ? 64 : 32));
3683#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3684 /* Rightly or wrongly, show huge mount option unmasked by shmem_huge */
3685 if (sbinfo->huge)
3686 seq_printf(seq, ",huge=%s", shmem_format_huge(sbinfo->huge));
3687#endif
3688 shmem_show_mpol(seq, sbinfo->mpol);
3689 return 0;
3690}
3691
3692#endif /* CONFIG_TMPFS */
3693
3694static void shmem_put_super(struct super_block *sb)
3695{
3696 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
3697
3698 free_percpu(sbinfo->ino_batch);
3699 percpu_counter_destroy(&sbinfo->used_blocks);
3700 mpol_put(sbinfo->mpol);
3701 kfree(sbinfo);
3702 sb->s_fs_info = NULL;
3703}
3704
3705static int shmem_fill_super(struct super_block *sb, struct fs_context *fc)
3706{
3707 struct shmem_options *ctx = fc->fs_private;
3708 struct inode *inode;
3709 struct shmem_sb_info *sbinfo;
3710 int err = -ENOMEM;
3711
3712 /* Round up to L1_CACHE_BYTES to resist false sharing */
3713 sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
3714 L1_CACHE_BYTES), GFP_KERNEL);
3715 if (!sbinfo)
3716 return -ENOMEM;
3717
3718 sb->s_fs_info = sbinfo;
3719
3720#ifdef CONFIG_TMPFS
3721 /*
3722 * Per default we only allow half of the physical ram per
3723 * tmpfs instance, limiting inodes to one per page of lowmem;
3724 * but the internal instance is left unlimited.
3725 */
3726 if (!(sb->s_flags & SB_KERNMOUNT)) {
3727 if (!(ctx->seen & SHMEM_SEEN_BLOCKS))
3728 ctx->blocks = shmem_default_max_blocks();
3729 if (!(ctx->seen & SHMEM_SEEN_INODES))
3730 ctx->inodes = shmem_default_max_inodes();
3731 if (!(ctx->seen & SHMEM_SEEN_INUMS))
3732 ctx->full_inums = IS_ENABLED(CONFIG_TMPFS_INODE64);
3733 } else {
3734 sb->s_flags |= SB_NOUSER;
3735 }
3736 sb->s_export_op = &shmem_export_ops;
3737 sb->s_flags |= SB_NOSEC;
3738#else
3739 sb->s_flags |= SB_NOUSER;
3740#endif
3741 sbinfo->max_blocks = ctx->blocks;
3742 sbinfo->free_inodes = sbinfo->max_inodes = ctx->inodes;
3743 if (sb->s_flags & SB_KERNMOUNT) {
3744 sbinfo->ino_batch = alloc_percpu(ino_t);
3745 if (!sbinfo->ino_batch)
3746 goto failed;
3747 }
3748 sbinfo->uid = ctx->uid;
3749 sbinfo->gid = ctx->gid;
3750 sbinfo->full_inums = ctx->full_inums;
3751 sbinfo->mode = ctx->mode;
3752 sbinfo->huge = ctx->huge;
3753 sbinfo->mpol = ctx->mpol;
3754 ctx->mpol = NULL;
3755
3756 spin_lock_init(&sbinfo->stat_lock);
3757 if (percpu_counter_init(&sbinfo->used_blocks, 0, GFP_KERNEL))
3758 goto failed;
3759 spin_lock_init(&sbinfo->shrinklist_lock);
3760 INIT_LIST_HEAD(&sbinfo->shrinklist);
3761
3762 sb->s_maxbytes = MAX_LFS_FILESIZE;
3763 sb->s_blocksize = PAGE_SIZE;
3764 sb->s_blocksize_bits = PAGE_SHIFT;
3765 sb->s_magic = TMPFS_MAGIC;
3766 sb->s_op = &shmem_ops;
3767 sb->s_time_gran = 1;
3768#ifdef CONFIG_TMPFS_XATTR
3769 sb->s_xattr = shmem_xattr_handlers;
3770#endif
3771#ifdef CONFIG_TMPFS_POSIX_ACL
3772 sb->s_flags |= SB_POSIXACL;
3773#endif
3774 uuid_gen(&sb->s_uuid);
3775
3776 inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
3777 if (!inode)
3778 goto failed;
3779 inode->i_uid = sbinfo->uid;
3780 inode->i_gid = sbinfo->gid;
3781 sb->s_root = d_make_root(inode);
3782 if (!sb->s_root)
3783 goto failed;
3784 return 0;
3785
3786failed:
3787 shmem_put_super(sb);
3788 return err;
3789}
3790
3791static int shmem_get_tree(struct fs_context *fc)
3792{
3793 return get_tree_nodev(fc, shmem_fill_super);
3794}
3795
3796static void shmem_free_fc(struct fs_context *fc)
3797{
3798 struct shmem_options *ctx = fc->fs_private;
3799
3800 if (ctx) {
3801 mpol_put(ctx->mpol);
3802 kfree(ctx);
3803 }
3804}
3805
3806static const struct fs_context_operations shmem_fs_context_ops = {
3807 .free = shmem_free_fc,
3808 .get_tree = shmem_get_tree,
3809#ifdef CONFIG_TMPFS
3810 .parse_monolithic = shmem_parse_options,
3811 .parse_param = shmem_parse_one,
3812 .reconfigure = shmem_reconfigure,
3813#endif
3814};
3815
3816static struct kmem_cache *shmem_inode_cachep;
3817
3818static struct inode *shmem_alloc_inode(struct super_block *sb)
3819{
3820 struct shmem_inode_info *info;
3821 info = kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
3822 if (!info)
3823 return NULL;
3824 return &info->vfs_inode;
3825}
3826
3827static void shmem_free_in_core_inode(struct inode *inode)
3828{
3829 if (S_ISLNK(inode->i_mode))
3830 kfree(inode->i_link);
3831 kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
3832}
3833
3834static void shmem_destroy_inode(struct inode *inode)
3835{
3836 if (S_ISREG(inode->i_mode))
3837 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
3838}
3839
3840static void shmem_init_inode(void *foo)
3841{
3842 struct shmem_inode_info *info = foo;
3843 inode_init_once(&info->vfs_inode);
3844}
3845
3846static void shmem_init_inodecache(void)
3847{
3848 shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
3849 sizeof(struct shmem_inode_info),
3850 0, SLAB_PANIC|SLAB_ACCOUNT, shmem_init_inode);
3851}
3852
3853static void shmem_destroy_inodecache(void)
3854{
3855 kmem_cache_destroy(shmem_inode_cachep);
3856}
3857
3858static const struct address_space_operations shmem_aops = {
3859 .writepage = shmem_writepage,
3860 .set_page_dirty = __set_page_dirty_no_writeback,
3861#ifdef CONFIG_TMPFS
3862 .write_begin = shmem_write_begin,
3863 .write_end = shmem_write_end,
3864#endif
3865#ifdef CONFIG_MIGRATION
3866 .migratepage = migrate_page,
3867#endif
3868 .error_remove_page = generic_error_remove_page,
3869};
3870
3871static const struct file_operations shmem_file_operations = {
3872 .mmap = shmem_mmap,
3873 .get_unmapped_area = shmem_get_unmapped_area,
3874#ifdef CONFIG_TMPFS
3875 .llseek = shmem_file_llseek,
3876 .read_iter = shmem_file_read_iter,
3877 .write_iter = generic_file_write_iter,
3878 .fsync = noop_fsync,
3879 .splice_read = generic_file_splice_read,
3880 .splice_write = iter_file_splice_write,
3881 .fallocate = shmem_fallocate,
3882#endif
3883};
3884
3885static const struct inode_operations shmem_inode_operations = {
3886 .getattr = shmem_getattr,
3887 .setattr = shmem_setattr,
3888#ifdef CONFIG_TMPFS_XATTR
3889 .listxattr = shmem_listxattr,
3890 .set_acl = simple_set_acl,
3891#endif
3892};
3893
3894static const struct inode_operations shmem_dir_inode_operations = {
3895#ifdef CONFIG_TMPFS
3896 .create = shmem_create,
3897 .lookup = simple_lookup,
3898 .link = shmem_link,
3899 .unlink = shmem_unlink,
3900 .symlink = shmem_symlink,
3901 .mkdir = shmem_mkdir,
3902 .rmdir = shmem_rmdir,
3903 .mknod = shmem_mknod,
3904 .rename = shmem_rename2,
3905 .tmpfile = shmem_tmpfile,
3906#endif
3907#ifdef CONFIG_TMPFS_XATTR
3908 .listxattr = shmem_listxattr,
3909#endif
3910#ifdef CONFIG_TMPFS_POSIX_ACL
3911 .setattr = shmem_setattr,
3912 .set_acl = simple_set_acl,
3913#endif
3914};
3915
3916static const struct inode_operations shmem_special_inode_operations = {
3917#ifdef CONFIG_TMPFS_XATTR
3918 .listxattr = shmem_listxattr,
3919#endif
3920#ifdef CONFIG_TMPFS_POSIX_ACL
3921 .setattr = shmem_setattr,
3922 .set_acl = simple_set_acl,
3923#endif
3924};
3925
3926static const struct super_operations shmem_ops = {
3927 .alloc_inode = shmem_alloc_inode,
3928 .free_inode = shmem_free_in_core_inode,
3929 .destroy_inode = shmem_destroy_inode,
3930#ifdef CONFIG_TMPFS
3931 .statfs = shmem_statfs,
3932 .show_options = shmem_show_options,
3933#endif
3934 .evict_inode = shmem_evict_inode,
3935 .drop_inode = generic_delete_inode,
3936 .put_super = shmem_put_super,
3937#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3938 .nr_cached_objects = shmem_unused_huge_count,
3939 .free_cached_objects = shmem_unused_huge_scan,
3940#endif
3941};
3942
3943static const struct vm_operations_struct shmem_vm_ops = {
3944 .fault = shmem_fault,
3945 .map_pages = filemap_map_pages,
3946#ifdef CONFIG_NUMA
3947 .set_policy = shmem_set_policy,
3948 .get_policy = shmem_get_policy,
3949#endif
3950};
3951
3952int shmem_init_fs_context(struct fs_context *fc)
3953{
3954 struct shmem_options *ctx;
3955
3956 ctx = kzalloc(sizeof(struct shmem_options), GFP_KERNEL);
3957 if (!ctx)
3958 return -ENOMEM;
3959
3960 ctx->mode = 0777 | S_ISVTX;
3961 ctx->uid = current_fsuid();
3962 ctx->gid = current_fsgid();
3963
3964 fc->fs_private = ctx;
3965 fc->ops = &shmem_fs_context_ops;
3966 return 0;
3967}
3968
3969static struct file_system_type shmem_fs_type = {
3970 .owner = THIS_MODULE,
3971 .name = "tmpfs",
3972 .init_fs_context = shmem_init_fs_context,
3973#ifdef CONFIG_TMPFS
3974 .parameters = shmem_fs_parameters,
3975#endif
3976 .kill_sb = kill_litter_super,
3977 .fs_flags = FS_USERNS_MOUNT,
3978};
3979
3980int __init shmem_init(void)
3981{
3982 int error;
3983
3984 shmem_init_inodecache();
3985
3986 error = register_filesystem(&shmem_fs_type);
3987 if (error) {
3988 pr_err("Could not register tmpfs\n");
3989 goto out2;
3990 }
3991
3992 shm_mnt = kern_mount(&shmem_fs_type);
3993 if (IS_ERR(shm_mnt)) {
3994 error = PTR_ERR(shm_mnt);
3995 pr_err("Could not kern_mount tmpfs\n");
3996 goto out1;
3997 }
3998
3999#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4000 if (has_transparent_hugepage() && shmem_huge > SHMEM_HUGE_DENY)
4001 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
4002 else
4003 shmem_huge = 0; /* just in case it was patched */
4004#endif
4005 return 0;
4006
4007out1:
4008 unregister_filesystem(&shmem_fs_type);
4009out2:
4010 shmem_destroy_inodecache();
4011 shm_mnt = ERR_PTR(error);
4012 return error;
4013}
4014
4015#if defined(CONFIG_TRANSPARENT_HUGEPAGE) && defined(CONFIG_SYSFS)
4016static ssize_t shmem_enabled_show(struct kobject *kobj,
4017 struct kobj_attribute *attr, char *buf)
4018{
4019 static const int values[] = {
4020 SHMEM_HUGE_ALWAYS,
4021 SHMEM_HUGE_WITHIN_SIZE,
4022 SHMEM_HUGE_ADVISE,
4023 SHMEM_HUGE_NEVER,
4024 SHMEM_HUGE_DENY,
4025 SHMEM_HUGE_FORCE,
4026 };
4027 int i, count;
4028
4029 for (i = 0, count = 0; i < ARRAY_SIZE(values); i++) {
4030 const char *fmt = shmem_huge == values[i] ? "[%s] " : "%s ";
4031
4032 count += sprintf(buf + count, fmt,
4033 shmem_format_huge(values[i]));
4034 }
4035 buf[count - 1] = '\n';
4036 return count;
4037}
4038
4039static ssize_t shmem_enabled_store(struct kobject *kobj,
4040 struct kobj_attribute *attr, const char *buf, size_t count)
4041{
4042 char tmp[16];
4043 int huge;
4044
4045 if (count + 1 > sizeof(tmp))
4046 return -EINVAL;
4047 memcpy(tmp, buf, count);
4048 tmp[count] = '\0';
4049 if (count && tmp[count - 1] == '\n')
4050 tmp[count - 1] = '\0';
4051
4052 huge = shmem_parse_huge(tmp);
4053 if (huge == -EINVAL)
4054 return -EINVAL;
4055 if (!has_transparent_hugepage() &&
4056 huge != SHMEM_HUGE_NEVER && huge != SHMEM_HUGE_DENY)
4057 return -EINVAL;
4058
4059 shmem_huge = huge;
4060 if (shmem_huge > SHMEM_HUGE_DENY)
4061 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
4062 return count;
4063}
4064
4065struct kobj_attribute shmem_enabled_attr =
4066 __ATTR(shmem_enabled, 0644, shmem_enabled_show, shmem_enabled_store);
4067#endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_SYSFS */
4068
4069#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4070bool shmem_huge_enabled(struct vm_area_struct *vma)
4071{
4072 struct inode *inode = file_inode(vma->vm_file);
4073 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
4074 loff_t i_size;
4075 pgoff_t off;
4076
4077 if ((vma->vm_flags & VM_NOHUGEPAGE) ||
4078 test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
4079 return false;
4080 if (shmem_huge == SHMEM_HUGE_FORCE)
4081 return true;
4082 if (shmem_huge == SHMEM_HUGE_DENY)
4083 return false;
4084 switch (sbinfo->huge) {
4085 case SHMEM_HUGE_NEVER:
4086 return false;
4087 case SHMEM_HUGE_ALWAYS:
4088 return true;
4089 case SHMEM_HUGE_WITHIN_SIZE:
4090 off = round_up(vma->vm_pgoff, HPAGE_PMD_NR);
4091 i_size = round_up(i_size_read(inode), PAGE_SIZE);
4092 if (i_size >= HPAGE_PMD_SIZE &&
4093 i_size >> PAGE_SHIFT >= off)
4094 return true;
4095 fallthrough;
4096 case SHMEM_HUGE_ADVISE:
4097 /* TODO: implement fadvise() hints */
4098 return (vma->vm_flags & VM_HUGEPAGE);
4099 default:
4100 VM_BUG_ON(1);
4101 return false;
4102 }
4103}
4104#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4105
4106#else /* !CONFIG_SHMEM */
4107
4108/*
4109 * tiny-shmem: simple shmemfs and tmpfs using ramfs code
4110 *
4111 * This is intended for small system where the benefits of the full
4112 * shmem code (swap-backed and resource-limited) are outweighed by
4113 * their complexity. On systems without swap this code should be
4114 * effectively equivalent, but much lighter weight.
4115 */
4116
4117static struct file_system_type shmem_fs_type = {
4118 .name = "tmpfs",
4119 .init_fs_context = ramfs_init_fs_context,
4120 .parameters = ramfs_fs_parameters,
4121 .kill_sb = kill_litter_super,
4122 .fs_flags = FS_USERNS_MOUNT,
4123};
4124
4125int __init shmem_init(void)
4126{
4127 BUG_ON(register_filesystem(&shmem_fs_type) != 0);
4128
4129 shm_mnt = kern_mount(&shmem_fs_type);
4130 BUG_ON(IS_ERR(shm_mnt));
4131
4132 return 0;
4133}
4134
4135int shmem_unuse(unsigned int type, bool frontswap,
4136 unsigned long *fs_pages_to_unuse)
4137{
4138 return 0;
4139}
4140
4141int shmem_lock(struct file *file, int lock, struct user_struct *user)
4142{
4143 return 0;
4144}
4145
4146void shmem_unlock_mapping(struct address_space *mapping)
4147{
4148}
4149
4150#ifdef CONFIG_MMU
4151unsigned long shmem_get_unmapped_area(struct file *file,
4152 unsigned long addr, unsigned long len,
4153 unsigned long pgoff, unsigned long flags)
4154{
4155 return current->mm->get_unmapped_area(file, addr, len, pgoff, flags);
4156}
4157#endif
4158
4159void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
4160{
4161 truncate_inode_pages_range(inode->i_mapping, lstart, lend);
4162}
4163EXPORT_SYMBOL_GPL(shmem_truncate_range);
4164
4165#define shmem_vm_ops generic_file_vm_ops
4166#define shmem_file_operations ramfs_file_operations
4167#define shmem_get_inode(sb, dir, mode, dev, flags) ramfs_get_inode(sb, dir, mode, dev)
4168#define shmem_acct_size(flags, size) 0
4169#define shmem_unacct_size(flags, size) do {} while (0)
4170
4171#endif /* CONFIG_SHMEM */
4172
4173/* common code */
4174
4175static struct file *__shmem_file_setup(struct vfsmount *mnt, const char *name, loff_t size,
4176 unsigned long flags, unsigned int i_flags)
4177{
4178 struct inode *inode;
4179 struct file *res;
4180
4181 if (IS_ERR(mnt))
4182 return ERR_CAST(mnt);
4183
4184 if (size < 0 || size > MAX_LFS_FILESIZE)
4185 return ERR_PTR(-EINVAL);
4186
4187 if (shmem_acct_size(flags, size))
4188 return ERR_PTR(-ENOMEM);
4189
4190 inode = shmem_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0,
4191 flags);
4192 if (unlikely(!inode)) {
4193 shmem_unacct_size(flags, size);
4194 return ERR_PTR(-ENOSPC);
4195 }
4196 inode->i_flags |= i_flags;
4197 inode->i_size = size;
4198 clear_nlink(inode); /* It is unlinked */
4199 res = ERR_PTR(ramfs_nommu_expand_for_mapping(inode, size));
4200 if (!IS_ERR(res))
4201 res = alloc_file_pseudo(inode, mnt, name, O_RDWR,
4202 &shmem_file_operations);
4203 if (IS_ERR(res))
4204 iput(inode);
4205 return res;
4206}
4207
4208/**
4209 * shmem_kernel_file_setup - get an unlinked file living in tmpfs which must be
4210 * kernel internal. There will be NO LSM permission checks against the
4211 * underlying inode. So users of this interface must do LSM checks at a
4212 * higher layer. The users are the big_key and shm implementations. LSM
4213 * checks are provided at the key or shm level rather than the inode.
4214 * @name: name for dentry (to be seen in /proc/<pid>/maps
4215 * @size: size to be set for the file
4216 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4217 */
4218struct file *shmem_kernel_file_setup(const char *name, loff_t size, unsigned long flags)
4219{
4220 return __shmem_file_setup(shm_mnt, name, size, flags, S_PRIVATE);
4221}
4222
4223/**
4224 * shmem_file_setup - get an unlinked file living in tmpfs
4225 * @name: name for dentry (to be seen in /proc/<pid>/maps
4226 * @size: size to be set for the file
4227 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4228 */
4229struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
4230{
4231 return __shmem_file_setup(shm_mnt, name, size, flags, 0);
4232}
4233EXPORT_SYMBOL_GPL(shmem_file_setup);
4234
4235/**
4236 * shmem_file_setup_with_mnt - get an unlinked file living in tmpfs
4237 * @mnt: the tmpfs mount where the file will be created
4238 * @name: name for dentry (to be seen in /proc/<pid>/maps
4239 * @size: size to be set for the file
4240 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4241 */
4242struct file *shmem_file_setup_with_mnt(struct vfsmount *mnt, const char *name,
4243 loff_t size, unsigned long flags)
4244{
4245 return __shmem_file_setup(mnt, name, size, flags, 0);
4246}
4247EXPORT_SYMBOL_GPL(shmem_file_setup_with_mnt);
4248
4249/**
4250 * shmem_zero_setup - setup a shared anonymous mapping
4251 * @vma: the vma to be mmapped is prepared by do_mmap
4252 */
4253int shmem_zero_setup(struct vm_area_struct *vma)
4254{
4255 struct file *file;
4256 loff_t size = vma->vm_end - vma->vm_start;
4257
4258 /*
4259 * Cloning a new file under mmap_lock leads to a lock ordering conflict
4260 * between XFS directory reading and selinux: since this file is only
4261 * accessible to the user through its mapping, use S_PRIVATE flag to
4262 * bypass file security, in the same way as shmem_kernel_file_setup().
4263 */
4264 file = shmem_kernel_file_setup("dev/zero", size, vma->vm_flags);
4265 if (IS_ERR(file))
4266 return PTR_ERR(file);
4267
4268 if (vma->vm_file)
4269 fput(vma->vm_file);
4270 vma->vm_file = file;
4271 vma->vm_ops = &shmem_vm_ops;
4272
4273 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
4274 ((vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK) <
4275 (vma->vm_end & HPAGE_PMD_MASK)) {
4276 khugepaged_enter(vma, vma->vm_flags);
4277 }
4278
4279 return 0;
4280}
4281
4282/**
4283 * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
4284 * @mapping: the page's address_space
4285 * @index: the page index
4286 * @gfp: the page allocator flags to use if allocating
4287 *
4288 * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
4289 * with any new page allocations done using the specified allocation flags.
4290 * But read_cache_page_gfp() uses the ->readpage() method: which does not
4291 * suit tmpfs, since it may have pages in swapcache, and needs to find those
4292 * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
4293 *
4294 * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
4295 * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
4296 */
4297struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
4298 pgoff_t index, gfp_t gfp)
4299{
4300#ifdef CONFIG_SHMEM
4301 struct inode *inode = mapping->host;
4302 struct page *page;
4303 int error;
4304
4305 BUG_ON(mapping->a_ops != &shmem_aops);
4306 error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE,
4307 gfp, NULL, NULL, NULL);
4308 if (error)
4309 page = ERR_PTR(error);
4310 else
4311 unlock_page(page);
4312 return page;
4313#else
4314 /*
4315 * The tiny !SHMEM case uses ramfs without swap
4316 */
4317 return read_cache_page_gfp(mapping, index, gfp);
4318#endif
4319}
4320EXPORT_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/shmem_fs.h>
37#include <linux/swap.h>
38#include <linux/uio.h>
39#include <linux/hugetlb.h>
40#include <linux/fs_parser.h>
41#include <linux/swapfile.h>
42#include <linux/iversion.h>
43#include "swap.h"
44
45static struct vfsmount *shm_mnt __ro_after_init;
46
47#ifdef CONFIG_SHMEM
48/*
49 * This virtual memory filesystem is heavily based on the ramfs. It
50 * extends ramfs by the ability to use swap and honor resource limits
51 * which makes it a completely usable filesystem.
52 */
53
54#include <linux/xattr.h>
55#include <linux/exportfs.h>
56#include <linux/posix_acl.h>
57#include <linux/posix_acl_xattr.h>
58#include <linux/mman.h>
59#include <linux/string.h>
60#include <linux/slab.h>
61#include <linux/backing-dev.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/rmap.h>
80#include <linux/uuid.h>
81#include <linux/quotaops.h>
82#include <linux/rcupdate_wait.h>
83
84#include <linux/uaccess.h>
85
86#include "internal.h"
87
88#define BLOCKS_PER_PAGE (PAGE_SIZE/512)
89#define VM_ACCT(size) (PAGE_ALIGN(size) >> PAGE_SHIFT)
90
91/* Pretend that each entry is of this size in directory's i_size */
92#define BOGO_DIRENT_SIZE 20
93
94/* Pretend that one inode + its dentry occupy this much memory */
95#define BOGO_INODE_SIZE 1024
96
97/* Symlink up to this size is kmalloc'ed instead of using a swappable page */
98#define SHORT_SYMLINK_LEN 128
99
100/*
101 * shmem_fallocate communicates with shmem_fault or shmem_writepage via
102 * inode->i_private (with i_rwsem making sure that it has only one user at
103 * a time): we would prefer not to enlarge the shmem inode just for that.
104 */
105struct shmem_falloc {
106 wait_queue_head_t *waitq; /* faults into hole wait for punch to end */
107 pgoff_t start; /* start of range currently being fallocated */
108 pgoff_t next; /* the next page offset to be fallocated */
109 pgoff_t nr_falloced; /* how many new pages have been fallocated */
110 pgoff_t nr_unswapped; /* how often writepage refused to swap out */
111};
112
113struct shmem_options {
114 unsigned long long blocks;
115 unsigned long long inodes;
116 struct mempolicy *mpol;
117 kuid_t uid;
118 kgid_t gid;
119 umode_t mode;
120 bool full_inums;
121 int huge;
122 int seen;
123 bool noswap;
124 unsigned short quota_types;
125 struct shmem_quota_limits qlimits;
126#define SHMEM_SEEN_BLOCKS 1
127#define SHMEM_SEEN_INODES 2
128#define SHMEM_SEEN_HUGE 4
129#define SHMEM_SEEN_INUMS 8
130#define SHMEM_SEEN_NOSWAP 16
131#define SHMEM_SEEN_QUOTA 32
132};
133
134#ifdef CONFIG_TMPFS
135static unsigned long shmem_default_max_blocks(void)
136{
137 return totalram_pages() / 2;
138}
139
140static unsigned long shmem_default_max_inodes(void)
141{
142 unsigned long nr_pages = totalram_pages();
143
144 return min3(nr_pages - totalhigh_pages(), nr_pages / 2,
145 ULONG_MAX / BOGO_INODE_SIZE);
146}
147#endif
148
149static int shmem_swapin_folio(struct inode *inode, pgoff_t index,
150 struct folio **foliop, enum sgp_type sgp, gfp_t gfp,
151 struct mm_struct *fault_mm, vm_fault_t *fault_type);
152
153static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
154{
155 return sb->s_fs_info;
156}
157
158/*
159 * shmem_file_setup pre-accounts the whole fixed size of a VM object,
160 * for shared memory and for shared anonymous (/dev/zero) mappings
161 * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
162 * consistent with the pre-accounting of private mappings ...
163 */
164static inline int shmem_acct_size(unsigned long flags, loff_t size)
165{
166 return (flags & VM_NORESERVE) ?
167 0 : security_vm_enough_memory_mm(current->mm, VM_ACCT(size));
168}
169
170static inline void shmem_unacct_size(unsigned long flags, loff_t size)
171{
172 if (!(flags & VM_NORESERVE))
173 vm_unacct_memory(VM_ACCT(size));
174}
175
176static inline int shmem_reacct_size(unsigned long flags,
177 loff_t oldsize, loff_t newsize)
178{
179 if (!(flags & VM_NORESERVE)) {
180 if (VM_ACCT(newsize) > VM_ACCT(oldsize))
181 return security_vm_enough_memory_mm(current->mm,
182 VM_ACCT(newsize) - VM_ACCT(oldsize));
183 else if (VM_ACCT(newsize) < VM_ACCT(oldsize))
184 vm_unacct_memory(VM_ACCT(oldsize) - VM_ACCT(newsize));
185 }
186 return 0;
187}
188
189/*
190 * ... whereas tmpfs objects are accounted incrementally as
191 * pages are allocated, in order to allow large sparse files.
192 * shmem_get_folio reports shmem_acct_blocks failure as -ENOSPC not -ENOMEM,
193 * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
194 */
195static inline int shmem_acct_blocks(unsigned long flags, long pages)
196{
197 if (!(flags & VM_NORESERVE))
198 return 0;
199
200 return security_vm_enough_memory_mm(current->mm,
201 pages * VM_ACCT(PAGE_SIZE));
202}
203
204static inline void shmem_unacct_blocks(unsigned long flags, long pages)
205{
206 if (flags & VM_NORESERVE)
207 vm_unacct_memory(pages * VM_ACCT(PAGE_SIZE));
208}
209
210static int shmem_inode_acct_blocks(struct inode *inode, long pages)
211{
212 struct shmem_inode_info *info = SHMEM_I(inode);
213 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
214 int err = -ENOSPC;
215
216 if (shmem_acct_blocks(info->flags, pages))
217 return err;
218
219 might_sleep(); /* when quotas */
220 if (sbinfo->max_blocks) {
221 if (!percpu_counter_limited_add(&sbinfo->used_blocks,
222 sbinfo->max_blocks, pages))
223 goto unacct;
224
225 err = dquot_alloc_block_nodirty(inode, pages);
226 if (err) {
227 percpu_counter_sub(&sbinfo->used_blocks, pages);
228 goto unacct;
229 }
230 } else {
231 err = dquot_alloc_block_nodirty(inode, pages);
232 if (err)
233 goto unacct;
234 }
235
236 return 0;
237
238unacct:
239 shmem_unacct_blocks(info->flags, pages);
240 return err;
241}
242
243static void shmem_inode_unacct_blocks(struct inode *inode, long pages)
244{
245 struct shmem_inode_info *info = SHMEM_I(inode);
246 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
247
248 might_sleep(); /* when quotas */
249 dquot_free_block_nodirty(inode, pages);
250
251 if (sbinfo->max_blocks)
252 percpu_counter_sub(&sbinfo->used_blocks, pages);
253 shmem_unacct_blocks(info->flags, pages);
254}
255
256static const struct super_operations shmem_ops;
257const struct address_space_operations shmem_aops;
258static const struct file_operations shmem_file_operations;
259static const struct inode_operations shmem_inode_operations;
260static const struct inode_operations shmem_dir_inode_operations;
261static const struct inode_operations shmem_special_inode_operations;
262static const struct vm_operations_struct shmem_vm_ops;
263static const struct vm_operations_struct shmem_anon_vm_ops;
264static struct file_system_type shmem_fs_type;
265
266bool vma_is_anon_shmem(struct vm_area_struct *vma)
267{
268 return vma->vm_ops == &shmem_anon_vm_ops;
269}
270
271bool vma_is_shmem(struct vm_area_struct *vma)
272{
273 return vma_is_anon_shmem(vma) || vma->vm_ops == &shmem_vm_ops;
274}
275
276static LIST_HEAD(shmem_swaplist);
277static DEFINE_MUTEX(shmem_swaplist_mutex);
278
279#ifdef CONFIG_TMPFS_QUOTA
280
281static int shmem_enable_quotas(struct super_block *sb,
282 unsigned short quota_types)
283{
284 int type, err = 0;
285
286 sb_dqopt(sb)->flags |= DQUOT_QUOTA_SYS_FILE | DQUOT_NOLIST_DIRTY;
287 for (type = 0; type < SHMEM_MAXQUOTAS; type++) {
288 if (!(quota_types & (1 << type)))
289 continue;
290 err = dquot_load_quota_sb(sb, type, QFMT_SHMEM,
291 DQUOT_USAGE_ENABLED |
292 DQUOT_LIMITS_ENABLED);
293 if (err)
294 goto out_err;
295 }
296 return 0;
297
298out_err:
299 pr_warn("tmpfs: failed to enable quota tracking (type=%d, err=%d)\n",
300 type, err);
301 for (type--; type >= 0; type--)
302 dquot_quota_off(sb, type);
303 return err;
304}
305
306static void shmem_disable_quotas(struct super_block *sb)
307{
308 int type;
309
310 for (type = 0; type < SHMEM_MAXQUOTAS; type++)
311 dquot_quota_off(sb, type);
312}
313
314static struct dquot **shmem_get_dquots(struct inode *inode)
315{
316 return SHMEM_I(inode)->i_dquot;
317}
318#endif /* CONFIG_TMPFS_QUOTA */
319
320/*
321 * shmem_reserve_inode() performs bookkeeping to reserve a shmem inode, and
322 * produces a novel ino for the newly allocated inode.
323 *
324 * It may also be called when making a hard link to permit the space needed by
325 * each dentry. However, in that case, no new inode number is needed since that
326 * internally draws from another pool of inode numbers (currently global
327 * get_next_ino()). This case is indicated by passing NULL as inop.
328 */
329#define SHMEM_INO_BATCH 1024
330static int shmem_reserve_inode(struct super_block *sb, ino_t *inop)
331{
332 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
333 ino_t ino;
334
335 if (!(sb->s_flags & SB_KERNMOUNT)) {
336 raw_spin_lock(&sbinfo->stat_lock);
337 if (sbinfo->max_inodes) {
338 if (sbinfo->free_ispace < BOGO_INODE_SIZE) {
339 raw_spin_unlock(&sbinfo->stat_lock);
340 return -ENOSPC;
341 }
342 sbinfo->free_ispace -= BOGO_INODE_SIZE;
343 }
344 if (inop) {
345 ino = sbinfo->next_ino++;
346 if (unlikely(is_zero_ino(ino)))
347 ino = sbinfo->next_ino++;
348 if (unlikely(!sbinfo->full_inums &&
349 ino > UINT_MAX)) {
350 /*
351 * Emulate get_next_ino uint wraparound for
352 * compatibility
353 */
354 if (IS_ENABLED(CONFIG_64BIT))
355 pr_warn("%s: inode number overflow on device %d, consider using inode64 mount option\n",
356 __func__, MINOR(sb->s_dev));
357 sbinfo->next_ino = 1;
358 ino = sbinfo->next_ino++;
359 }
360 *inop = ino;
361 }
362 raw_spin_unlock(&sbinfo->stat_lock);
363 } else if (inop) {
364 /*
365 * __shmem_file_setup, one of our callers, is lock-free: it
366 * doesn't hold stat_lock in shmem_reserve_inode since
367 * max_inodes is always 0, and is called from potentially
368 * unknown contexts. As such, use a per-cpu batched allocator
369 * which doesn't require the per-sb stat_lock unless we are at
370 * the batch boundary.
371 *
372 * We don't need to worry about inode{32,64} since SB_KERNMOUNT
373 * shmem mounts are not exposed to userspace, so we don't need
374 * to worry about things like glibc compatibility.
375 */
376 ino_t *next_ino;
377
378 next_ino = per_cpu_ptr(sbinfo->ino_batch, get_cpu());
379 ino = *next_ino;
380 if (unlikely(ino % SHMEM_INO_BATCH == 0)) {
381 raw_spin_lock(&sbinfo->stat_lock);
382 ino = sbinfo->next_ino;
383 sbinfo->next_ino += SHMEM_INO_BATCH;
384 raw_spin_unlock(&sbinfo->stat_lock);
385 if (unlikely(is_zero_ino(ino)))
386 ino++;
387 }
388 *inop = ino;
389 *next_ino = ++ino;
390 put_cpu();
391 }
392
393 return 0;
394}
395
396static void shmem_free_inode(struct super_block *sb, size_t freed_ispace)
397{
398 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
399 if (sbinfo->max_inodes) {
400 raw_spin_lock(&sbinfo->stat_lock);
401 sbinfo->free_ispace += BOGO_INODE_SIZE + freed_ispace;
402 raw_spin_unlock(&sbinfo->stat_lock);
403 }
404}
405
406/**
407 * shmem_recalc_inode - recalculate the block usage of an inode
408 * @inode: inode to recalc
409 * @alloced: the change in number of pages allocated to inode
410 * @swapped: the change in number of pages swapped from inode
411 *
412 * We have to calculate the free blocks since the mm can drop
413 * undirtied hole pages behind our back.
414 *
415 * But normally info->alloced == inode->i_mapping->nrpages + info->swapped
416 * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
417 */
418static void shmem_recalc_inode(struct inode *inode, long alloced, long swapped)
419{
420 struct shmem_inode_info *info = SHMEM_I(inode);
421 long freed;
422
423 spin_lock(&info->lock);
424 info->alloced += alloced;
425 info->swapped += swapped;
426 freed = info->alloced - info->swapped -
427 READ_ONCE(inode->i_mapping->nrpages);
428 /*
429 * Special case: whereas normally shmem_recalc_inode() is called
430 * after i_mapping->nrpages has already been adjusted (up or down),
431 * shmem_writepage() has to raise swapped before nrpages is lowered -
432 * to stop a racing shmem_recalc_inode() from thinking that a page has
433 * been freed. Compensate here, to avoid the need for a followup call.
434 */
435 if (swapped > 0)
436 freed += swapped;
437 if (freed > 0)
438 info->alloced -= freed;
439 spin_unlock(&info->lock);
440
441 /* The quota case may block */
442 if (freed > 0)
443 shmem_inode_unacct_blocks(inode, freed);
444}
445
446bool shmem_charge(struct inode *inode, long pages)
447{
448 struct address_space *mapping = inode->i_mapping;
449
450 if (shmem_inode_acct_blocks(inode, pages))
451 return false;
452
453 /* nrpages adjustment first, then shmem_recalc_inode() when balanced */
454 xa_lock_irq(&mapping->i_pages);
455 mapping->nrpages += pages;
456 xa_unlock_irq(&mapping->i_pages);
457
458 shmem_recalc_inode(inode, pages, 0);
459 return true;
460}
461
462void shmem_uncharge(struct inode *inode, long pages)
463{
464 /* pages argument is currently unused: keep it to help debugging */
465 /* nrpages adjustment done by __filemap_remove_folio() or caller */
466
467 shmem_recalc_inode(inode, 0, 0);
468}
469
470/*
471 * Replace item expected in xarray by a new item, while holding xa_lock.
472 */
473static int shmem_replace_entry(struct address_space *mapping,
474 pgoff_t index, void *expected, void *replacement)
475{
476 XA_STATE(xas, &mapping->i_pages, index);
477 void *item;
478
479 VM_BUG_ON(!expected);
480 VM_BUG_ON(!replacement);
481 item = xas_load(&xas);
482 if (item != expected)
483 return -ENOENT;
484 xas_store(&xas, replacement);
485 return 0;
486}
487
488/*
489 * Sometimes, before we decide whether to proceed or to fail, we must check
490 * that an entry was not already brought back from swap by a racing thread.
491 *
492 * Checking page is not enough: by the time a SwapCache page is locked, it
493 * might be reused, and again be SwapCache, using the same swap as before.
494 */
495static bool shmem_confirm_swap(struct address_space *mapping,
496 pgoff_t index, swp_entry_t swap)
497{
498 return xa_load(&mapping->i_pages, index) == swp_to_radix_entry(swap);
499}
500
501/*
502 * Definitions for "huge tmpfs": tmpfs mounted with the huge= option
503 *
504 * SHMEM_HUGE_NEVER:
505 * disables huge pages for the mount;
506 * SHMEM_HUGE_ALWAYS:
507 * enables huge pages for the mount;
508 * SHMEM_HUGE_WITHIN_SIZE:
509 * only allocate huge pages if the page will be fully within i_size,
510 * also respect fadvise()/madvise() hints;
511 * SHMEM_HUGE_ADVISE:
512 * only allocate huge pages if requested with fadvise()/madvise();
513 */
514
515#define SHMEM_HUGE_NEVER 0
516#define SHMEM_HUGE_ALWAYS 1
517#define SHMEM_HUGE_WITHIN_SIZE 2
518#define SHMEM_HUGE_ADVISE 3
519
520/*
521 * Special values.
522 * Only can be set via /sys/kernel/mm/transparent_hugepage/shmem_enabled:
523 *
524 * SHMEM_HUGE_DENY:
525 * disables huge on shm_mnt and all mounts, for emergency use;
526 * SHMEM_HUGE_FORCE:
527 * enables huge on shm_mnt and all mounts, w/o needing option, for testing;
528 *
529 */
530#define SHMEM_HUGE_DENY (-1)
531#define SHMEM_HUGE_FORCE (-2)
532
533#ifdef CONFIG_TRANSPARENT_HUGEPAGE
534/* ifdef here to avoid bloating shmem.o when not necessary */
535
536static int shmem_huge __read_mostly = SHMEM_HUGE_NEVER;
537
538bool shmem_is_huge(struct inode *inode, pgoff_t index, bool shmem_huge_force,
539 struct mm_struct *mm, unsigned long vm_flags)
540{
541 loff_t i_size;
542
543 if (!S_ISREG(inode->i_mode))
544 return false;
545 if (mm && ((vm_flags & VM_NOHUGEPAGE) || test_bit(MMF_DISABLE_THP, &mm->flags)))
546 return false;
547 if (shmem_huge == SHMEM_HUGE_DENY)
548 return false;
549 if (shmem_huge_force || shmem_huge == SHMEM_HUGE_FORCE)
550 return true;
551
552 switch (SHMEM_SB(inode->i_sb)->huge) {
553 case SHMEM_HUGE_ALWAYS:
554 return true;
555 case SHMEM_HUGE_WITHIN_SIZE:
556 index = round_up(index + 1, HPAGE_PMD_NR);
557 i_size = round_up(i_size_read(inode), PAGE_SIZE);
558 if (i_size >> PAGE_SHIFT >= index)
559 return true;
560 fallthrough;
561 case SHMEM_HUGE_ADVISE:
562 if (mm && (vm_flags & VM_HUGEPAGE))
563 return true;
564 fallthrough;
565 default:
566 return false;
567 }
568}
569
570#if defined(CONFIG_SYSFS)
571static int shmem_parse_huge(const char *str)
572{
573 if (!strcmp(str, "never"))
574 return SHMEM_HUGE_NEVER;
575 if (!strcmp(str, "always"))
576 return SHMEM_HUGE_ALWAYS;
577 if (!strcmp(str, "within_size"))
578 return SHMEM_HUGE_WITHIN_SIZE;
579 if (!strcmp(str, "advise"))
580 return SHMEM_HUGE_ADVISE;
581 if (!strcmp(str, "deny"))
582 return SHMEM_HUGE_DENY;
583 if (!strcmp(str, "force"))
584 return SHMEM_HUGE_FORCE;
585 return -EINVAL;
586}
587#endif
588
589#if defined(CONFIG_SYSFS) || defined(CONFIG_TMPFS)
590static const char *shmem_format_huge(int huge)
591{
592 switch (huge) {
593 case SHMEM_HUGE_NEVER:
594 return "never";
595 case SHMEM_HUGE_ALWAYS:
596 return "always";
597 case SHMEM_HUGE_WITHIN_SIZE:
598 return "within_size";
599 case SHMEM_HUGE_ADVISE:
600 return "advise";
601 case SHMEM_HUGE_DENY:
602 return "deny";
603 case SHMEM_HUGE_FORCE:
604 return "force";
605 default:
606 VM_BUG_ON(1);
607 return "bad_val";
608 }
609}
610#endif
611
612static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
613 struct shrink_control *sc, unsigned long nr_to_split)
614{
615 LIST_HEAD(list), *pos, *next;
616 LIST_HEAD(to_remove);
617 struct inode *inode;
618 struct shmem_inode_info *info;
619 struct folio *folio;
620 unsigned long batch = sc ? sc->nr_to_scan : 128;
621 int split = 0;
622
623 if (list_empty(&sbinfo->shrinklist))
624 return SHRINK_STOP;
625
626 spin_lock(&sbinfo->shrinklist_lock);
627 list_for_each_safe(pos, next, &sbinfo->shrinklist) {
628 info = list_entry(pos, struct shmem_inode_info, shrinklist);
629
630 /* pin the inode */
631 inode = igrab(&info->vfs_inode);
632
633 /* inode is about to be evicted */
634 if (!inode) {
635 list_del_init(&info->shrinklist);
636 goto next;
637 }
638
639 /* Check if there's anything to gain */
640 if (round_up(inode->i_size, PAGE_SIZE) ==
641 round_up(inode->i_size, HPAGE_PMD_SIZE)) {
642 list_move(&info->shrinklist, &to_remove);
643 goto next;
644 }
645
646 list_move(&info->shrinklist, &list);
647next:
648 sbinfo->shrinklist_len--;
649 if (!--batch)
650 break;
651 }
652 spin_unlock(&sbinfo->shrinklist_lock);
653
654 list_for_each_safe(pos, next, &to_remove) {
655 info = list_entry(pos, struct shmem_inode_info, shrinklist);
656 inode = &info->vfs_inode;
657 list_del_init(&info->shrinklist);
658 iput(inode);
659 }
660
661 list_for_each_safe(pos, next, &list) {
662 int ret;
663 pgoff_t index;
664
665 info = list_entry(pos, struct shmem_inode_info, shrinklist);
666 inode = &info->vfs_inode;
667
668 if (nr_to_split && split >= nr_to_split)
669 goto move_back;
670
671 index = (inode->i_size & HPAGE_PMD_MASK) >> PAGE_SHIFT;
672 folio = filemap_get_folio(inode->i_mapping, index);
673 if (IS_ERR(folio))
674 goto drop;
675
676 /* No huge page at the end of the file: nothing to split */
677 if (!folio_test_large(folio)) {
678 folio_put(folio);
679 goto drop;
680 }
681
682 /*
683 * Move the inode on the list back to shrinklist if we failed
684 * to lock the page at this time.
685 *
686 * Waiting for the lock may lead to deadlock in the
687 * reclaim path.
688 */
689 if (!folio_trylock(folio)) {
690 folio_put(folio);
691 goto move_back;
692 }
693
694 ret = split_folio(folio);
695 folio_unlock(folio);
696 folio_put(folio);
697
698 /* If split failed move the inode on the list back to shrinklist */
699 if (ret)
700 goto move_back;
701
702 split++;
703drop:
704 list_del_init(&info->shrinklist);
705 goto put;
706move_back:
707 /*
708 * Make sure the inode is either on the global list or deleted
709 * from any local list before iput() since it could be deleted
710 * in another thread once we put the inode (then the local list
711 * is corrupted).
712 */
713 spin_lock(&sbinfo->shrinklist_lock);
714 list_move(&info->shrinklist, &sbinfo->shrinklist);
715 sbinfo->shrinklist_len++;
716 spin_unlock(&sbinfo->shrinklist_lock);
717put:
718 iput(inode);
719 }
720
721 return split;
722}
723
724static long shmem_unused_huge_scan(struct super_block *sb,
725 struct shrink_control *sc)
726{
727 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
728
729 if (!READ_ONCE(sbinfo->shrinklist_len))
730 return SHRINK_STOP;
731
732 return shmem_unused_huge_shrink(sbinfo, sc, 0);
733}
734
735static long shmem_unused_huge_count(struct super_block *sb,
736 struct shrink_control *sc)
737{
738 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
739 return READ_ONCE(sbinfo->shrinklist_len);
740}
741#else /* !CONFIG_TRANSPARENT_HUGEPAGE */
742
743#define shmem_huge SHMEM_HUGE_DENY
744
745bool shmem_is_huge(struct inode *inode, pgoff_t index, bool shmem_huge_force,
746 struct mm_struct *mm, unsigned long vm_flags)
747{
748 return false;
749}
750
751static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
752 struct shrink_control *sc, unsigned long nr_to_split)
753{
754 return 0;
755}
756#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
757
758/*
759 * Somewhat like filemap_add_folio, but error if expected item has gone.
760 */
761static int shmem_add_to_page_cache(struct folio *folio,
762 struct address_space *mapping,
763 pgoff_t index, void *expected, gfp_t gfp)
764{
765 XA_STATE_ORDER(xas, &mapping->i_pages, index, folio_order(folio));
766 long nr = folio_nr_pages(folio);
767
768 VM_BUG_ON_FOLIO(index != round_down(index, nr), folio);
769 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
770 VM_BUG_ON_FOLIO(!folio_test_swapbacked(folio), folio);
771 VM_BUG_ON(expected && folio_test_large(folio));
772
773 folio_ref_add(folio, nr);
774 folio->mapping = mapping;
775 folio->index = index;
776
777 gfp &= GFP_RECLAIM_MASK;
778 folio_throttle_swaprate(folio, gfp);
779
780 do {
781 xas_lock_irq(&xas);
782 if (expected != xas_find_conflict(&xas)) {
783 xas_set_err(&xas, -EEXIST);
784 goto unlock;
785 }
786 if (expected && xas_find_conflict(&xas)) {
787 xas_set_err(&xas, -EEXIST);
788 goto unlock;
789 }
790 xas_store(&xas, folio);
791 if (xas_error(&xas))
792 goto unlock;
793 if (folio_test_pmd_mappable(folio))
794 __lruvec_stat_mod_folio(folio, NR_SHMEM_THPS, nr);
795 __lruvec_stat_mod_folio(folio, NR_FILE_PAGES, nr);
796 __lruvec_stat_mod_folio(folio, NR_SHMEM, nr);
797 mapping->nrpages += nr;
798unlock:
799 xas_unlock_irq(&xas);
800 } while (xas_nomem(&xas, gfp));
801
802 if (xas_error(&xas)) {
803 folio->mapping = NULL;
804 folio_ref_sub(folio, nr);
805 return xas_error(&xas);
806 }
807
808 return 0;
809}
810
811/*
812 * Somewhat like filemap_remove_folio, but substitutes swap for @folio.
813 */
814static void shmem_delete_from_page_cache(struct folio *folio, void *radswap)
815{
816 struct address_space *mapping = folio->mapping;
817 long nr = folio_nr_pages(folio);
818 int error;
819
820 xa_lock_irq(&mapping->i_pages);
821 error = shmem_replace_entry(mapping, folio->index, folio, radswap);
822 folio->mapping = NULL;
823 mapping->nrpages -= nr;
824 __lruvec_stat_mod_folio(folio, NR_FILE_PAGES, -nr);
825 __lruvec_stat_mod_folio(folio, NR_SHMEM, -nr);
826 xa_unlock_irq(&mapping->i_pages);
827 folio_put(folio);
828 BUG_ON(error);
829}
830
831/*
832 * Remove swap entry from page cache, free the swap and its page cache.
833 */
834static int shmem_free_swap(struct address_space *mapping,
835 pgoff_t index, void *radswap)
836{
837 void *old;
838
839 old = xa_cmpxchg_irq(&mapping->i_pages, index, radswap, NULL, 0);
840 if (old != radswap)
841 return -ENOENT;
842 free_swap_and_cache(radix_to_swp_entry(radswap));
843 return 0;
844}
845
846/*
847 * Determine (in bytes) how many of the shmem object's pages mapped by the
848 * given offsets are swapped out.
849 *
850 * This is safe to call without i_rwsem or the i_pages lock thanks to RCU,
851 * as long as the inode doesn't go away and racy results are not a problem.
852 */
853unsigned long shmem_partial_swap_usage(struct address_space *mapping,
854 pgoff_t start, pgoff_t end)
855{
856 XA_STATE(xas, &mapping->i_pages, start);
857 struct page *page;
858 unsigned long swapped = 0;
859 unsigned long max = end - 1;
860
861 rcu_read_lock();
862 xas_for_each(&xas, page, max) {
863 if (xas_retry(&xas, page))
864 continue;
865 if (xa_is_value(page))
866 swapped++;
867 if (xas.xa_index == max)
868 break;
869 if (need_resched()) {
870 xas_pause(&xas);
871 cond_resched_rcu();
872 }
873 }
874 rcu_read_unlock();
875
876 return swapped << PAGE_SHIFT;
877}
878
879/*
880 * Determine (in bytes) how many of the shmem object's pages mapped by the
881 * given vma is swapped out.
882 *
883 * This is safe to call without i_rwsem or the i_pages lock thanks to RCU,
884 * as long as the inode doesn't go away and racy results are not a problem.
885 */
886unsigned long shmem_swap_usage(struct vm_area_struct *vma)
887{
888 struct inode *inode = file_inode(vma->vm_file);
889 struct shmem_inode_info *info = SHMEM_I(inode);
890 struct address_space *mapping = inode->i_mapping;
891 unsigned long swapped;
892
893 /* Be careful as we don't hold info->lock */
894 swapped = READ_ONCE(info->swapped);
895
896 /*
897 * The easier cases are when the shmem object has nothing in swap, or
898 * the vma maps it whole. Then we can simply use the stats that we
899 * already track.
900 */
901 if (!swapped)
902 return 0;
903
904 if (!vma->vm_pgoff && vma->vm_end - vma->vm_start >= inode->i_size)
905 return swapped << PAGE_SHIFT;
906
907 /* Here comes the more involved part */
908 return shmem_partial_swap_usage(mapping, vma->vm_pgoff,
909 vma->vm_pgoff + vma_pages(vma));
910}
911
912/*
913 * SysV IPC SHM_UNLOCK restore Unevictable pages to their evictable lists.
914 */
915void shmem_unlock_mapping(struct address_space *mapping)
916{
917 struct folio_batch fbatch;
918 pgoff_t index = 0;
919
920 folio_batch_init(&fbatch);
921 /*
922 * Minor point, but we might as well stop if someone else SHM_LOCKs it.
923 */
924 while (!mapping_unevictable(mapping) &&
925 filemap_get_folios(mapping, &index, ~0UL, &fbatch)) {
926 check_move_unevictable_folios(&fbatch);
927 folio_batch_release(&fbatch);
928 cond_resched();
929 }
930}
931
932static struct folio *shmem_get_partial_folio(struct inode *inode, pgoff_t index)
933{
934 struct folio *folio;
935
936 /*
937 * At first avoid shmem_get_folio(,,,SGP_READ): that fails
938 * beyond i_size, and reports fallocated folios as holes.
939 */
940 folio = filemap_get_entry(inode->i_mapping, index);
941 if (!folio)
942 return folio;
943 if (!xa_is_value(folio)) {
944 folio_lock(folio);
945 if (folio->mapping == inode->i_mapping)
946 return folio;
947 /* The folio has been swapped out */
948 folio_unlock(folio);
949 folio_put(folio);
950 }
951 /*
952 * But read a folio back from swap if any of it is within i_size
953 * (although in some cases this is just a waste of time).
954 */
955 folio = NULL;
956 shmem_get_folio(inode, index, &folio, SGP_READ);
957 return folio;
958}
959
960/*
961 * Remove range of pages and swap entries from page cache, and free them.
962 * If !unfalloc, truncate or punch hole; if unfalloc, undo failed fallocate.
963 */
964static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
965 bool unfalloc)
966{
967 struct address_space *mapping = inode->i_mapping;
968 struct shmem_inode_info *info = SHMEM_I(inode);
969 pgoff_t start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
970 pgoff_t end = (lend + 1) >> PAGE_SHIFT;
971 struct folio_batch fbatch;
972 pgoff_t indices[PAGEVEC_SIZE];
973 struct folio *folio;
974 bool same_folio;
975 long nr_swaps_freed = 0;
976 pgoff_t index;
977 int i;
978
979 if (lend == -1)
980 end = -1; /* unsigned, so actually very big */
981
982 if (info->fallocend > start && info->fallocend <= end && !unfalloc)
983 info->fallocend = start;
984
985 folio_batch_init(&fbatch);
986 index = start;
987 while (index < end && find_lock_entries(mapping, &index, end - 1,
988 &fbatch, indices)) {
989 for (i = 0; i < folio_batch_count(&fbatch); i++) {
990 folio = fbatch.folios[i];
991
992 if (xa_is_value(folio)) {
993 if (unfalloc)
994 continue;
995 nr_swaps_freed += !shmem_free_swap(mapping,
996 indices[i], folio);
997 continue;
998 }
999
1000 if (!unfalloc || !folio_test_uptodate(folio))
1001 truncate_inode_folio(mapping, folio);
1002 folio_unlock(folio);
1003 }
1004 folio_batch_remove_exceptionals(&fbatch);
1005 folio_batch_release(&fbatch);
1006 cond_resched();
1007 }
1008
1009 /*
1010 * When undoing a failed fallocate, we want none of the partial folio
1011 * zeroing and splitting below, but shall want to truncate the whole
1012 * folio when !uptodate indicates that it was added by this fallocate,
1013 * even when [lstart, lend] covers only a part of the folio.
1014 */
1015 if (unfalloc)
1016 goto whole_folios;
1017
1018 same_folio = (lstart >> PAGE_SHIFT) == (lend >> PAGE_SHIFT);
1019 folio = shmem_get_partial_folio(inode, lstart >> PAGE_SHIFT);
1020 if (folio) {
1021 same_folio = lend < folio_pos(folio) + folio_size(folio);
1022 folio_mark_dirty(folio);
1023 if (!truncate_inode_partial_folio(folio, lstart, lend)) {
1024 start = folio_next_index(folio);
1025 if (same_folio)
1026 end = folio->index;
1027 }
1028 folio_unlock(folio);
1029 folio_put(folio);
1030 folio = NULL;
1031 }
1032
1033 if (!same_folio)
1034 folio = shmem_get_partial_folio(inode, lend >> PAGE_SHIFT);
1035 if (folio) {
1036 folio_mark_dirty(folio);
1037 if (!truncate_inode_partial_folio(folio, lstart, lend))
1038 end = folio->index;
1039 folio_unlock(folio);
1040 folio_put(folio);
1041 }
1042
1043whole_folios:
1044
1045 index = start;
1046 while (index < end) {
1047 cond_resched();
1048
1049 if (!find_get_entries(mapping, &index, end - 1, &fbatch,
1050 indices)) {
1051 /* If all gone or hole-punch or unfalloc, we're done */
1052 if (index == start || end != -1)
1053 break;
1054 /* But if truncating, restart to make sure all gone */
1055 index = start;
1056 continue;
1057 }
1058 for (i = 0; i < folio_batch_count(&fbatch); i++) {
1059 folio = fbatch.folios[i];
1060
1061 if (xa_is_value(folio)) {
1062 if (unfalloc)
1063 continue;
1064 if (shmem_free_swap(mapping, indices[i], folio)) {
1065 /* Swap was replaced by page: retry */
1066 index = indices[i];
1067 break;
1068 }
1069 nr_swaps_freed++;
1070 continue;
1071 }
1072
1073 folio_lock(folio);
1074
1075 if (!unfalloc || !folio_test_uptodate(folio)) {
1076 if (folio_mapping(folio) != mapping) {
1077 /* Page was replaced by swap: retry */
1078 folio_unlock(folio);
1079 index = indices[i];
1080 break;
1081 }
1082 VM_BUG_ON_FOLIO(folio_test_writeback(folio),
1083 folio);
1084
1085 if (!folio_test_large(folio)) {
1086 truncate_inode_folio(mapping, folio);
1087 } else if (truncate_inode_partial_folio(folio, lstart, lend)) {
1088 /*
1089 * If we split a page, reset the loop so
1090 * that we pick up the new sub pages.
1091 * Otherwise the THP was entirely
1092 * dropped or the target range was
1093 * zeroed, so just continue the loop as
1094 * is.
1095 */
1096 if (!folio_test_large(folio)) {
1097 folio_unlock(folio);
1098 index = start;
1099 break;
1100 }
1101 }
1102 }
1103 folio_unlock(folio);
1104 }
1105 folio_batch_remove_exceptionals(&fbatch);
1106 folio_batch_release(&fbatch);
1107 }
1108
1109 shmem_recalc_inode(inode, 0, -nr_swaps_freed);
1110}
1111
1112void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
1113{
1114 shmem_undo_range(inode, lstart, lend, false);
1115 inode_set_mtime_to_ts(inode, inode_set_ctime_current(inode));
1116 inode_inc_iversion(inode);
1117}
1118EXPORT_SYMBOL_GPL(shmem_truncate_range);
1119
1120static int shmem_getattr(struct mnt_idmap *idmap,
1121 const struct path *path, struct kstat *stat,
1122 u32 request_mask, unsigned int query_flags)
1123{
1124 struct inode *inode = path->dentry->d_inode;
1125 struct shmem_inode_info *info = SHMEM_I(inode);
1126
1127 if (info->alloced - info->swapped != inode->i_mapping->nrpages)
1128 shmem_recalc_inode(inode, 0, 0);
1129
1130 if (info->fsflags & FS_APPEND_FL)
1131 stat->attributes |= STATX_ATTR_APPEND;
1132 if (info->fsflags & FS_IMMUTABLE_FL)
1133 stat->attributes |= STATX_ATTR_IMMUTABLE;
1134 if (info->fsflags & FS_NODUMP_FL)
1135 stat->attributes |= STATX_ATTR_NODUMP;
1136 stat->attributes_mask |= (STATX_ATTR_APPEND |
1137 STATX_ATTR_IMMUTABLE |
1138 STATX_ATTR_NODUMP);
1139 generic_fillattr(idmap, request_mask, inode, stat);
1140
1141 if (shmem_is_huge(inode, 0, false, NULL, 0))
1142 stat->blksize = HPAGE_PMD_SIZE;
1143
1144 if (request_mask & STATX_BTIME) {
1145 stat->result_mask |= STATX_BTIME;
1146 stat->btime.tv_sec = info->i_crtime.tv_sec;
1147 stat->btime.tv_nsec = info->i_crtime.tv_nsec;
1148 }
1149
1150 return 0;
1151}
1152
1153static int shmem_setattr(struct mnt_idmap *idmap,
1154 struct dentry *dentry, struct iattr *attr)
1155{
1156 struct inode *inode = d_inode(dentry);
1157 struct shmem_inode_info *info = SHMEM_I(inode);
1158 int error;
1159 bool update_mtime = false;
1160 bool update_ctime = true;
1161
1162 error = setattr_prepare(idmap, dentry, attr);
1163 if (error)
1164 return error;
1165
1166 if ((info->seals & F_SEAL_EXEC) && (attr->ia_valid & ATTR_MODE)) {
1167 if ((inode->i_mode ^ attr->ia_mode) & 0111) {
1168 return -EPERM;
1169 }
1170 }
1171
1172 if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
1173 loff_t oldsize = inode->i_size;
1174 loff_t newsize = attr->ia_size;
1175
1176 /* protected by i_rwsem */
1177 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
1178 (newsize > oldsize && (info->seals & F_SEAL_GROW)))
1179 return -EPERM;
1180
1181 if (newsize != oldsize) {
1182 error = shmem_reacct_size(SHMEM_I(inode)->flags,
1183 oldsize, newsize);
1184 if (error)
1185 return error;
1186 i_size_write(inode, newsize);
1187 update_mtime = true;
1188 } else {
1189 update_ctime = false;
1190 }
1191 if (newsize <= oldsize) {
1192 loff_t holebegin = round_up(newsize, PAGE_SIZE);
1193 if (oldsize > holebegin)
1194 unmap_mapping_range(inode->i_mapping,
1195 holebegin, 0, 1);
1196 if (info->alloced)
1197 shmem_truncate_range(inode,
1198 newsize, (loff_t)-1);
1199 /* unmap again to remove racily COWed private pages */
1200 if (oldsize > holebegin)
1201 unmap_mapping_range(inode->i_mapping,
1202 holebegin, 0, 1);
1203 }
1204 }
1205
1206 if (is_quota_modification(idmap, inode, attr)) {
1207 error = dquot_initialize(inode);
1208 if (error)
1209 return error;
1210 }
1211
1212 /* Transfer quota accounting */
1213 if (i_uid_needs_update(idmap, attr, inode) ||
1214 i_gid_needs_update(idmap, attr, inode)) {
1215 error = dquot_transfer(idmap, inode, attr);
1216 if (error)
1217 return error;
1218 }
1219
1220 setattr_copy(idmap, inode, attr);
1221 if (attr->ia_valid & ATTR_MODE)
1222 error = posix_acl_chmod(idmap, dentry, inode->i_mode);
1223 if (!error && update_ctime) {
1224 inode_set_ctime_current(inode);
1225 if (update_mtime)
1226 inode_set_mtime_to_ts(inode, inode_get_ctime(inode));
1227 inode_inc_iversion(inode);
1228 }
1229 return error;
1230}
1231
1232static void shmem_evict_inode(struct inode *inode)
1233{
1234 struct shmem_inode_info *info = SHMEM_I(inode);
1235 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1236 size_t freed = 0;
1237
1238 if (shmem_mapping(inode->i_mapping)) {
1239 shmem_unacct_size(info->flags, inode->i_size);
1240 inode->i_size = 0;
1241 mapping_set_exiting(inode->i_mapping);
1242 shmem_truncate_range(inode, 0, (loff_t)-1);
1243 if (!list_empty(&info->shrinklist)) {
1244 spin_lock(&sbinfo->shrinklist_lock);
1245 if (!list_empty(&info->shrinklist)) {
1246 list_del_init(&info->shrinklist);
1247 sbinfo->shrinklist_len--;
1248 }
1249 spin_unlock(&sbinfo->shrinklist_lock);
1250 }
1251 while (!list_empty(&info->swaplist)) {
1252 /* Wait while shmem_unuse() is scanning this inode... */
1253 wait_var_event(&info->stop_eviction,
1254 !atomic_read(&info->stop_eviction));
1255 mutex_lock(&shmem_swaplist_mutex);
1256 /* ...but beware of the race if we peeked too early */
1257 if (!atomic_read(&info->stop_eviction))
1258 list_del_init(&info->swaplist);
1259 mutex_unlock(&shmem_swaplist_mutex);
1260 }
1261 }
1262
1263 simple_xattrs_free(&info->xattrs, sbinfo->max_inodes ? &freed : NULL);
1264 shmem_free_inode(inode->i_sb, freed);
1265 WARN_ON(inode->i_blocks);
1266 clear_inode(inode);
1267#ifdef CONFIG_TMPFS_QUOTA
1268 dquot_free_inode(inode);
1269 dquot_drop(inode);
1270#endif
1271}
1272
1273static int shmem_find_swap_entries(struct address_space *mapping,
1274 pgoff_t start, struct folio_batch *fbatch,
1275 pgoff_t *indices, unsigned int type)
1276{
1277 XA_STATE(xas, &mapping->i_pages, start);
1278 struct folio *folio;
1279 swp_entry_t entry;
1280
1281 rcu_read_lock();
1282 xas_for_each(&xas, folio, ULONG_MAX) {
1283 if (xas_retry(&xas, folio))
1284 continue;
1285
1286 if (!xa_is_value(folio))
1287 continue;
1288
1289 entry = radix_to_swp_entry(folio);
1290 /*
1291 * swapin error entries can be found in the mapping. But they're
1292 * deliberately ignored here as we've done everything we can do.
1293 */
1294 if (swp_type(entry) != type)
1295 continue;
1296
1297 indices[folio_batch_count(fbatch)] = xas.xa_index;
1298 if (!folio_batch_add(fbatch, folio))
1299 break;
1300
1301 if (need_resched()) {
1302 xas_pause(&xas);
1303 cond_resched_rcu();
1304 }
1305 }
1306 rcu_read_unlock();
1307
1308 return xas.xa_index;
1309}
1310
1311/*
1312 * Move the swapped pages for an inode to page cache. Returns the count
1313 * of pages swapped in, or the error in case of failure.
1314 */
1315static int shmem_unuse_swap_entries(struct inode *inode,
1316 struct folio_batch *fbatch, pgoff_t *indices)
1317{
1318 int i = 0;
1319 int ret = 0;
1320 int error = 0;
1321 struct address_space *mapping = inode->i_mapping;
1322
1323 for (i = 0; i < folio_batch_count(fbatch); i++) {
1324 struct folio *folio = fbatch->folios[i];
1325
1326 if (!xa_is_value(folio))
1327 continue;
1328 error = shmem_swapin_folio(inode, indices[i], &folio, SGP_CACHE,
1329 mapping_gfp_mask(mapping), NULL, NULL);
1330 if (error == 0) {
1331 folio_unlock(folio);
1332 folio_put(folio);
1333 ret++;
1334 }
1335 if (error == -ENOMEM)
1336 break;
1337 error = 0;
1338 }
1339 return error ? error : ret;
1340}
1341
1342/*
1343 * If swap found in inode, free it and move page from swapcache to filecache.
1344 */
1345static int shmem_unuse_inode(struct inode *inode, unsigned int type)
1346{
1347 struct address_space *mapping = inode->i_mapping;
1348 pgoff_t start = 0;
1349 struct folio_batch fbatch;
1350 pgoff_t indices[PAGEVEC_SIZE];
1351 int ret = 0;
1352
1353 do {
1354 folio_batch_init(&fbatch);
1355 shmem_find_swap_entries(mapping, start, &fbatch, indices, type);
1356 if (folio_batch_count(&fbatch) == 0) {
1357 ret = 0;
1358 break;
1359 }
1360
1361 ret = shmem_unuse_swap_entries(inode, &fbatch, indices);
1362 if (ret < 0)
1363 break;
1364
1365 start = indices[folio_batch_count(&fbatch) - 1];
1366 } while (true);
1367
1368 return ret;
1369}
1370
1371/*
1372 * Read all the shared memory data that resides in the swap
1373 * device 'type' back into memory, so the swap device can be
1374 * unused.
1375 */
1376int shmem_unuse(unsigned int type)
1377{
1378 struct shmem_inode_info *info, *next;
1379 int error = 0;
1380
1381 if (list_empty(&shmem_swaplist))
1382 return 0;
1383
1384 mutex_lock(&shmem_swaplist_mutex);
1385 list_for_each_entry_safe(info, next, &shmem_swaplist, swaplist) {
1386 if (!info->swapped) {
1387 list_del_init(&info->swaplist);
1388 continue;
1389 }
1390 /*
1391 * Drop the swaplist mutex while searching the inode for swap;
1392 * but before doing so, make sure shmem_evict_inode() will not
1393 * remove placeholder inode from swaplist, nor let it be freed
1394 * (igrab() would protect from unlink, but not from unmount).
1395 */
1396 atomic_inc(&info->stop_eviction);
1397 mutex_unlock(&shmem_swaplist_mutex);
1398
1399 error = shmem_unuse_inode(&info->vfs_inode, type);
1400 cond_resched();
1401
1402 mutex_lock(&shmem_swaplist_mutex);
1403 next = list_next_entry(info, swaplist);
1404 if (!info->swapped)
1405 list_del_init(&info->swaplist);
1406 if (atomic_dec_and_test(&info->stop_eviction))
1407 wake_up_var(&info->stop_eviction);
1408 if (error)
1409 break;
1410 }
1411 mutex_unlock(&shmem_swaplist_mutex);
1412
1413 return error;
1414}
1415
1416/*
1417 * Move the page from the page cache to the swap cache.
1418 */
1419static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1420{
1421 struct folio *folio = page_folio(page);
1422 struct address_space *mapping = folio->mapping;
1423 struct inode *inode = mapping->host;
1424 struct shmem_inode_info *info = SHMEM_I(inode);
1425 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1426 swp_entry_t swap;
1427 pgoff_t index;
1428
1429 /*
1430 * Our capabilities prevent regular writeback or sync from ever calling
1431 * shmem_writepage; but a stacking filesystem might use ->writepage of
1432 * its underlying filesystem, in which case tmpfs should write out to
1433 * swap only in response to memory pressure, and not for the writeback
1434 * threads or sync.
1435 */
1436 if (WARN_ON_ONCE(!wbc->for_reclaim))
1437 goto redirty;
1438
1439 if (WARN_ON_ONCE((info->flags & VM_LOCKED) || sbinfo->noswap))
1440 goto redirty;
1441
1442 if (!total_swap_pages)
1443 goto redirty;
1444
1445 /*
1446 * If /sys/kernel/mm/transparent_hugepage/shmem_enabled is "always" or
1447 * "force", drivers/gpu/drm/i915/gem/i915_gem_shmem.c gets huge pages,
1448 * and its shmem_writeback() needs them to be split when swapping.
1449 */
1450 if (folio_test_large(folio)) {
1451 /* Ensure the subpages are still dirty */
1452 folio_test_set_dirty(folio);
1453 if (split_huge_page(page) < 0)
1454 goto redirty;
1455 folio = page_folio(page);
1456 folio_clear_dirty(folio);
1457 }
1458
1459 index = folio->index;
1460
1461 /*
1462 * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
1463 * value into swapfile.c, the only way we can correctly account for a
1464 * fallocated folio arriving here is now to initialize it and write it.
1465 *
1466 * That's okay for a folio already fallocated earlier, but if we have
1467 * not yet completed the fallocation, then (a) we want to keep track
1468 * of this folio in case we have to undo it, and (b) it may not be a
1469 * good idea to continue anyway, once we're pushing into swap. So
1470 * reactivate the folio, and let shmem_fallocate() quit when too many.
1471 */
1472 if (!folio_test_uptodate(folio)) {
1473 if (inode->i_private) {
1474 struct shmem_falloc *shmem_falloc;
1475 spin_lock(&inode->i_lock);
1476 shmem_falloc = inode->i_private;
1477 if (shmem_falloc &&
1478 !shmem_falloc->waitq &&
1479 index >= shmem_falloc->start &&
1480 index < shmem_falloc->next)
1481 shmem_falloc->nr_unswapped++;
1482 else
1483 shmem_falloc = NULL;
1484 spin_unlock(&inode->i_lock);
1485 if (shmem_falloc)
1486 goto redirty;
1487 }
1488 folio_zero_range(folio, 0, folio_size(folio));
1489 flush_dcache_folio(folio);
1490 folio_mark_uptodate(folio);
1491 }
1492
1493 swap = folio_alloc_swap(folio);
1494 if (!swap.val)
1495 goto redirty;
1496
1497 /*
1498 * Add inode to shmem_unuse()'s list of swapped-out inodes,
1499 * if it's not already there. Do it now before the folio is
1500 * moved to swap cache, when its pagelock no longer protects
1501 * the inode from eviction. But don't unlock the mutex until
1502 * we've incremented swapped, because shmem_unuse_inode() will
1503 * prune a !swapped inode from the swaplist under this mutex.
1504 */
1505 mutex_lock(&shmem_swaplist_mutex);
1506 if (list_empty(&info->swaplist))
1507 list_add(&info->swaplist, &shmem_swaplist);
1508
1509 if (add_to_swap_cache(folio, swap,
1510 __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN,
1511 NULL) == 0) {
1512 shmem_recalc_inode(inode, 0, 1);
1513 swap_shmem_alloc(swap);
1514 shmem_delete_from_page_cache(folio, swp_to_radix_entry(swap));
1515
1516 mutex_unlock(&shmem_swaplist_mutex);
1517 BUG_ON(folio_mapped(folio));
1518 return swap_writepage(&folio->page, wbc);
1519 }
1520
1521 mutex_unlock(&shmem_swaplist_mutex);
1522 put_swap_folio(folio, swap);
1523redirty:
1524 folio_mark_dirty(folio);
1525 if (wbc->for_reclaim)
1526 return AOP_WRITEPAGE_ACTIVATE; /* Return with folio locked */
1527 folio_unlock(folio);
1528 return 0;
1529}
1530
1531#if defined(CONFIG_NUMA) && defined(CONFIG_TMPFS)
1532static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1533{
1534 char buffer[64];
1535
1536 if (!mpol || mpol->mode == MPOL_DEFAULT)
1537 return; /* show nothing */
1538
1539 mpol_to_str(buffer, sizeof(buffer), mpol);
1540
1541 seq_printf(seq, ",mpol=%s", buffer);
1542}
1543
1544static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1545{
1546 struct mempolicy *mpol = NULL;
1547 if (sbinfo->mpol) {
1548 raw_spin_lock(&sbinfo->stat_lock); /* prevent replace/use races */
1549 mpol = sbinfo->mpol;
1550 mpol_get(mpol);
1551 raw_spin_unlock(&sbinfo->stat_lock);
1552 }
1553 return mpol;
1554}
1555#else /* !CONFIG_NUMA || !CONFIG_TMPFS */
1556static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1557{
1558}
1559static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1560{
1561 return NULL;
1562}
1563#endif /* CONFIG_NUMA && CONFIG_TMPFS */
1564
1565static struct mempolicy *shmem_get_pgoff_policy(struct shmem_inode_info *info,
1566 pgoff_t index, unsigned int order, pgoff_t *ilx);
1567
1568static struct folio *shmem_swapin_cluster(swp_entry_t swap, gfp_t gfp,
1569 struct shmem_inode_info *info, pgoff_t index)
1570{
1571 struct mempolicy *mpol;
1572 pgoff_t ilx;
1573 struct folio *folio;
1574
1575 mpol = shmem_get_pgoff_policy(info, index, 0, &ilx);
1576 folio = swap_cluster_readahead(swap, gfp, mpol, ilx);
1577 mpol_cond_put(mpol);
1578
1579 return folio;
1580}
1581
1582/*
1583 * Make sure huge_gfp is always more limited than limit_gfp.
1584 * Some of the flags set permissions, while others set limitations.
1585 */
1586static gfp_t limit_gfp_mask(gfp_t huge_gfp, gfp_t limit_gfp)
1587{
1588 gfp_t allowflags = __GFP_IO | __GFP_FS | __GFP_RECLAIM;
1589 gfp_t denyflags = __GFP_NOWARN | __GFP_NORETRY;
1590 gfp_t zoneflags = limit_gfp & GFP_ZONEMASK;
1591 gfp_t result = huge_gfp & ~(allowflags | GFP_ZONEMASK);
1592
1593 /* Allow allocations only from the originally specified zones. */
1594 result |= zoneflags;
1595
1596 /*
1597 * Minimize the result gfp by taking the union with the deny flags,
1598 * and the intersection of the allow flags.
1599 */
1600 result |= (limit_gfp & denyflags);
1601 result |= (huge_gfp & limit_gfp) & allowflags;
1602
1603 return result;
1604}
1605
1606static struct folio *shmem_alloc_hugefolio(gfp_t gfp,
1607 struct shmem_inode_info *info, pgoff_t index)
1608{
1609 struct mempolicy *mpol;
1610 pgoff_t ilx;
1611 struct page *page;
1612
1613 mpol = shmem_get_pgoff_policy(info, index, HPAGE_PMD_ORDER, &ilx);
1614 page = alloc_pages_mpol(gfp, HPAGE_PMD_ORDER, mpol, ilx, numa_node_id());
1615 mpol_cond_put(mpol);
1616
1617 return page_rmappable_folio(page);
1618}
1619
1620static struct folio *shmem_alloc_folio(gfp_t gfp,
1621 struct shmem_inode_info *info, pgoff_t index)
1622{
1623 struct mempolicy *mpol;
1624 pgoff_t ilx;
1625 struct page *page;
1626
1627 mpol = shmem_get_pgoff_policy(info, index, 0, &ilx);
1628 page = alloc_pages_mpol(gfp, 0, mpol, ilx, numa_node_id());
1629 mpol_cond_put(mpol);
1630
1631 return (struct folio *)page;
1632}
1633
1634static struct folio *shmem_alloc_and_add_folio(gfp_t gfp,
1635 struct inode *inode, pgoff_t index,
1636 struct mm_struct *fault_mm, bool huge)
1637{
1638 struct address_space *mapping = inode->i_mapping;
1639 struct shmem_inode_info *info = SHMEM_I(inode);
1640 struct folio *folio;
1641 long pages;
1642 int error;
1643
1644 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1645 huge = false;
1646
1647 if (huge) {
1648 pages = HPAGE_PMD_NR;
1649 index = round_down(index, HPAGE_PMD_NR);
1650
1651 /*
1652 * Check for conflict before waiting on a huge allocation.
1653 * Conflict might be that a huge page has just been allocated
1654 * and added to page cache by a racing thread, or that there
1655 * is already at least one small page in the huge extent.
1656 * Be careful to retry when appropriate, but not forever!
1657 * Elsewhere -EEXIST would be the right code, but not here.
1658 */
1659 if (xa_find(&mapping->i_pages, &index,
1660 index + HPAGE_PMD_NR - 1, XA_PRESENT))
1661 return ERR_PTR(-E2BIG);
1662
1663 folio = shmem_alloc_hugefolio(gfp, info, index);
1664 if (!folio)
1665 count_vm_event(THP_FILE_FALLBACK);
1666 } else {
1667 pages = 1;
1668 folio = shmem_alloc_folio(gfp, info, index);
1669 }
1670 if (!folio)
1671 return ERR_PTR(-ENOMEM);
1672
1673 __folio_set_locked(folio);
1674 __folio_set_swapbacked(folio);
1675
1676 gfp &= GFP_RECLAIM_MASK;
1677 error = mem_cgroup_charge(folio, fault_mm, gfp);
1678 if (error) {
1679 if (xa_find(&mapping->i_pages, &index,
1680 index + pages - 1, XA_PRESENT)) {
1681 error = -EEXIST;
1682 } else if (huge) {
1683 count_vm_event(THP_FILE_FALLBACK);
1684 count_vm_event(THP_FILE_FALLBACK_CHARGE);
1685 }
1686 goto unlock;
1687 }
1688
1689 error = shmem_add_to_page_cache(folio, mapping, index, NULL, gfp);
1690 if (error)
1691 goto unlock;
1692
1693 error = shmem_inode_acct_blocks(inode, pages);
1694 if (error) {
1695 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1696 long freed;
1697 /*
1698 * Try to reclaim some space by splitting a few
1699 * large folios beyond i_size on the filesystem.
1700 */
1701 shmem_unused_huge_shrink(sbinfo, NULL, 2);
1702 /*
1703 * And do a shmem_recalc_inode() to account for freed pages:
1704 * except our folio is there in cache, so not quite balanced.
1705 */
1706 spin_lock(&info->lock);
1707 freed = pages + info->alloced - info->swapped -
1708 READ_ONCE(mapping->nrpages);
1709 if (freed > 0)
1710 info->alloced -= freed;
1711 spin_unlock(&info->lock);
1712 if (freed > 0)
1713 shmem_inode_unacct_blocks(inode, freed);
1714 error = shmem_inode_acct_blocks(inode, pages);
1715 if (error) {
1716 filemap_remove_folio(folio);
1717 goto unlock;
1718 }
1719 }
1720
1721 shmem_recalc_inode(inode, pages, 0);
1722 folio_add_lru(folio);
1723 return folio;
1724
1725unlock:
1726 folio_unlock(folio);
1727 folio_put(folio);
1728 return ERR_PTR(error);
1729}
1730
1731/*
1732 * When a page is moved from swapcache to shmem filecache (either by the
1733 * usual swapin of shmem_get_folio_gfp(), or by the less common swapoff of
1734 * shmem_unuse_inode()), it may have been read in earlier from swap, in
1735 * ignorance of the mapping it belongs to. If that mapping has special
1736 * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
1737 * we may need to copy to a suitable page before moving to filecache.
1738 *
1739 * In a future release, this may well be extended to respect cpuset and
1740 * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
1741 * but for now it is a simple matter of zone.
1742 */
1743static bool shmem_should_replace_folio(struct folio *folio, gfp_t gfp)
1744{
1745 return folio_zonenum(folio) > gfp_zone(gfp);
1746}
1747
1748static int shmem_replace_folio(struct folio **foliop, gfp_t gfp,
1749 struct shmem_inode_info *info, pgoff_t index)
1750{
1751 struct folio *old, *new;
1752 struct address_space *swap_mapping;
1753 swp_entry_t entry;
1754 pgoff_t swap_index;
1755 int error;
1756
1757 old = *foliop;
1758 entry = old->swap;
1759 swap_index = swp_offset(entry);
1760 swap_mapping = swap_address_space(entry);
1761
1762 /*
1763 * We have arrived here because our zones are constrained, so don't
1764 * limit chance of success by further cpuset and node constraints.
1765 */
1766 gfp &= ~GFP_CONSTRAINT_MASK;
1767 VM_BUG_ON_FOLIO(folio_test_large(old), old);
1768 new = shmem_alloc_folio(gfp, info, index);
1769 if (!new)
1770 return -ENOMEM;
1771
1772 folio_get(new);
1773 folio_copy(new, old);
1774 flush_dcache_folio(new);
1775
1776 __folio_set_locked(new);
1777 __folio_set_swapbacked(new);
1778 folio_mark_uptodate(new);
1779 new->swap = entry;
1780 folio_set_swapcache(new);
1781
1782 /*
1783 * Our caller will very soon move newpage out of swapcache, but it's
1784 * a nice clean interface for us to replace oldpage by newpage there.
1785 */
1786 xa_lock_irq(&swap_mapping->i_pages);
1787 error = shmem_replace_entry(swap_mapping, swap_index, old, new);
1788 if (!error) {
1789 mem_cgroup_migrate(old, new);
1790 __lruvec_stat_mod_folio(new, NR_FILE_PAGES, 1);
1791 __lruvec_stat_mod_folio(new, NR_SHMEM, 1);
1792 __lruvec_stat_mod_folio(old, NR_FILE_PAGES, -1);
1793 __lruvec_stat_mod_folio(old, NR_SHMEM, -1);
1794 }
1795 xa_unlock_irq(&swap_mapping->i_pages);
1796
1797 if (unlikely(error)) {
1798 /*
1799 * Is this possible? I think not, now that our callers check
1800 * both PageSwapCache and page_private after getting page lock;
1801 * but be defensive. Reverse old to newpage for clear and free.
1802 */
1803 old = new;
1804 } else {
1805 folio_add_lru(new);
1806 *foliop = new;
1807 }
1808
1809 folio_clear_swapcache(old);
1810 old->private = NULL;
1811
1812 folio_unlock(old);
1813 folio_put_refs(old, 2);
1814 return error;
1815}
1816
1817static void shmem_set_folio_swapin_error(struct inode *inode, pgoff_t index,
1818 struct folio *folio, swp_entry_t swap)
1819{
1820 struct address_space *mapping = inode->i_mapping;
1821 swp_entry_t swapin_error;
1822 void *old;
1823
1824 swapin_error = make_poisoned_swp_entry();
1825 old = xa_cmpxchg_irq(&mapping->i_pages, index,
1826 swp_to_radix_entry(swap),
1827 swp_to_radix_entry(swapin_error), 0);
1828 if (old != swp_to_radix_entry(swap))
1829 return;
1830
1831 folio_wait_writeback(folio);
1832 delete_from_swap_cache(folio);
1833 /*
1834 * Don't treat swapin error folio as alloced. Otherwise inode->i_blocks
1835 * won't be 0 when inode is released and thus trigger WARN_ON(i_blocks)
1836 * in shmem_evict_inode().
1837 */
1838 shmem_recalc_inode(inode, -1, -1);
1839 swap_free(swap);
1840}
1841
1842/*
1843 * Swap in the folio pointed to by *foliop.
1844 * Caller has to make sure that *foliop contains a valid swapped folio.
1845 * Returns 0 and the folio in foliop if success. On failure, returns the
1846 * error code and NULL in *foliop.
1847 */
1848static int shmem_swapin_folio(struct inode *inode, pgoff_t index,
1849 struct folio **foliop, enum sgp_type sgp,
1850 gfp_t gfp, struct mm_struct *fault_mm,
1851 vm_fault_t *fault_type)
1852{
1853 struct address_space *mapping = inode->i_mapping;
1854 struct shmem_inode_info *info = SHMEM_I(inode);
1855 struct swap_info_struct *si;
1856 struct folio *folio = NULL;
1857 swp_entry_t swap;
1858 int error;
1859
1860 VM_BUG_ON(!*foliop || !xa_is_value(*foliop));
1861 swap = radix_to_swp_entry(*foliop);
1862 *foliop = NULL;
1863
1864 if (is_poisoned_swp_entry(swap))
1865 return -EIO;
1866
1867 si = get_swap_device(swap);
1868 if (!si) {
1869 if (!shmem_confirm_swap(mapping, index, swap))
1870 return -EEXIST;
1871 else
1872 return -EINVAL;
1873 }
1874
1875 /* Look it up and read it in.. */
1876 folio = swap_cache_get_folio(swap, NULL, 0);
1877 if (!folio) {
1878 /* Or update major stats only when swapin succeeds?? */
1879 if (fault_type) {
1880 *fault_type |= VM_FAULT_MAJOR;
1881 count_vm_event(PGMAJFAULT);
1882 count_memcg_event_mm(fault_mm, PGMAJFAULT);
1883 }
1884 /* Here we actually start the io */
1885 folio = shmem_swapin_cluster(swap, gfp, info, index);
1886 if (!folio) {
1887 error = -ENOMEM;
1888 goto failed;
1889 }
1890 }
1891
1892 /* We have to do this with folio locked to prevent races */
1893 folio_lock(folio);
1894 if (!folio_test_swapcache(folio) ||
1895 folio->swap.val != swap.val ||
1896 !shmem_confirm_swap(mapping, index, swap)) {
1897 error = -EEXIST;
1898 goto unlock;
1899 }
1900 if (!folio_test_uptodate(folio)) {
1901 error = -EIO;
1902 goto failed;
1903 }
1904 folio_wait_writeback(folio);
1905
1906 /*
1907 * Some architectures may have to restore extra metadata to the
1908 * folio after reading from swap.
1909 */
1910 arch_swap_restore(swap, folio);
1911
1912 if (shmem_should_replace_folio(folio, gfp)) {
1913 error = shmem_replace_folio(&folio, gfp, info, index);
1914 if (error)
1915 goto failed;
1916 }
1917
1918 error = shmem_add_to_page_cache(folio, mapping, index,
1919 swp_to_radix_entry(swap), gfp);
1920 if (error)
1921 goto failed;
1922
1923 shmem_recalc_inode(inode, 0, -1);
1924
1925 if (sgp == SGP_WRITE)
1926 folio_mark_accessed(folio);
1927
1928 delete_from_swap_cache(folio);
1929 folio_mark_dirty(folio);
1930 swap_free(swap);
1931 put_swap_device(si);
1932
1933 *foliop = folio;
1934 return 0;
1935failed:
1936 if (!shmem_confirm_swap(mapping, index, swap))
1937 error = -EEXIST;
1938 if (error == -EIO)
1939 shmem_set_folio_swapin_error(inode, index, folio, swap);
1940unlock:
1941 if (folio) {
1942 folio_unlock(folio);
1943 folio_put(folio);
1944 }
1945 put_swap_device(si);
1946
1947 return error;
1948}
1949
1950/*
1951 * shmem_get_folio_gfp - find page in cache, or get from swap, or allocate
1952 *
1953 * If we allocate a new one we do not mark it dirty. That's up to the
1954 * vm. If we swap it in we mark it dirty since we also free the swap
1955 * entry since a page cannot live in both the swap and page cache.
1956 *
1957 * vmf and fault_type are only supplied by shmem_fault: otherwise they are NULL.
1958 */
1959static int shmem_get_folio_gfp(struct inode *inode, pgoff_t index,
1960 struct folio **foliop, enum sgp_type sgp, gfp_t gfp,
1961 struct vm_fault *vmf, vm_fault_t *fault_type)
1962{
1963 struct vm_area_struct *vma = vmf ? vmf->vma : NULL;
1964 struct mm_struct *fault_mm;
1965 struct folio *folio;
1966 int error;
1967 bool alloced;
1968
1969 if (index > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
1970 return -EFBIG;
1971repeat:
1972 if (sgp <= SGP_CACHE &&
1973 ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode))
1974 return -EINVAL;
1975
1976 alloced = false;
1977 fault_mm = vma ? vma->vm_mm : NULL;
1978
1979 folio = filemap_get_entry(inode->i_mapping, index);
1980 if (folio && vma && userfaultfd_minor(vma)) {
1981 if (!xa_is_value(folio))
1982 folio_put(folio);
1983 *fault_type = handle_userfault(vmf, VM_UFFD_MINOR);
1984 return 0;
1985 }
1986
1987 if (xa_is_value(folio)) {
1988 error = shmem_swapin_folio(inode, index, &folio,
1989 sgp, gfp, fault_mm, fault_type);
1990 if (error == -EEXIST)
1991 goto repeat;
1992
1993 *foliop = folio;
1994 return error;
1995 }
1996
1997 if (folio) {
1998 folio_lock(folio);
1999
2000 /* Has the folio been truncated or swapped out? */
2001 if (unlikely(folio->mapping != inode->i_mapping)) {
2002 folio_unlock(folio);
2003 folio_put(folio);
2004 goto repeat;
2005 }
2006 if (sgp == SGP_WRITE)
2007 folio_mark_accessed(folio);
2008 if (folio_test_uptodate(folio))
2009 goto out;
2010 /* fallocated folio */
2011 if (sgp != SGP_READ)
2012 goto clear;
2013 folio_unlock(folio);
2014 folio_put(folio);
2015 }
2016
2017 /*
2018 * SGP_READ: succeed on hole, with NULL folio, letting caller zero.
2019 * SGP_NOALLOC: fail on hole, with NULL folio, letting caller fail.
2020 */
2021 *foliop = NULL;
2022 if (sgp == SGP_READ)
2023 return 0;
2024 if (sgp == SGP_NOALLOC)
2025 return -ENOENT;
2026
2027 /*
2028 * Fast cache lookup and swap lookup did not find it: allocate.
2029 */
2030
2031 if (vma && userfaultfd_missing(vma)) {
2032 *fault_type = handle_userfault(vmf, VM_UFFD_MISSING);
2033 return 0;
2034 }
2035
2036 if (shmem_is_huge(inode, index, false, fault_mm,
2037 vma ? vma->vm_flags : 0)) {
2038 gfp_t huge_gfp;
2039
2040 huge_gfp = vma_thp_gfp_mask(vma);
2041 huge_gfp = limit_gfp_mask(huge_gfp, gfp);
2042 folio = shmem_alloc_and_add_folio(huge_gfp,
2043 inode, index, fault_mm, true);
2044 if (!IS_ERR(folio)) {
2045 count_vm_event(THP_FILE_ALLOC);
2046 goto alloced;
2047 }
2048 if (PTR_ERR(folio) == -EEXIST)
2049 goto repeat;
2050 }
2051
2052 folio = shmem_alloc_and_add_folio(gfp, inode, index, fault_mm, false);
2053 if (IS_ERR(folio)) {
2054 error = PTR_ERR(folio);
2055 if (error == -EEXIST)
2056 goto repeat;
2057 folio = NULL;
2058 goto unlock;
2059 }
2060
2061alloced:
2062 alloced = true;
2063 if (folio_test_pmd_mappable(folio) &&
2064 DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE) <
2065 folio_next_index(folio) - 1) {
2066 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
2067 struct shmem_inode_info *info = SHMEM_I(inode);
2068 /*
2069 * Part of the large folio is beyond i_size: subject
2070 * to shrink under memory pressure.
2071 */
2072 spin_lock(&sbinfo->shrinklist_lock);
2073 /*
2074 * _careful to defend against unlocked access to
2075 * ->shrink_list in shmem_unused_huge_shrink()
2076 */
2077 if (list_empty_careful(&info->shrinklist)) {
2078 list_add_tail(&info->shrinklist,
2079 &sbinfo->shrinklist);
2080 sbinfo->shrinklist_len++;
2081 }
2082 spin_unlock(&sbinfo->shrinklist_lock);
2083 }
2084
2085 if (sgp == SGP_WRITE)
2086 folio_set_referenced(folio);
2087 /*
2088 * Let SGP_FALLOC use the SGP_WRITE optimization on a new folio.
2089 */
2090 if (sgp == SGP_FALLOC)
2091 sgp = SGP_WRITE;
2092clear:
2093 /*
2094 * Let SGP_WRITE caller clear ends if write does not fill folio;
2095 * but SGP_FALLOC on a folio fallocated earlier must initialize
2096 * it now, lest undo on failure cancel our earlier guarantee.
2097 */
2098 if (sgp != SGP_WRITE && !folio_test_uptodate(folio)) {
2099 long i, n = folio_nr_pages(folio);
2100
2101 for (i = 0; i < n; i++)
2102 clear_highpage(folio_page(folio, i));
2103 flush_dcache_folio(folio);
2104 folio_mark_uptodate(folio);
2105 }
2106
2107 /* Perhaps the file has been truncated since we checked */
2108 if (sgp <= SGP_CACHE &&
2109 ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
2110 error = -EINVAL;
2111 goto unlock;
2112 }
2113out:
2114 *foliop = folio;
2115 return 0;
2116
2117 /*
2118 * Error recovery.
2119 */
2120unlock:
2121 if (alloced)
2122 filemap_remove_folio(folio);
2123 shmem_recalc_inode(inode, 0, 0);
2124 if (folio) {
2125 folio_unlock(folio);
2126 folio_put(folio);
2127 }
2128 return error;
2129}
2130
2131int shmem_get_folio(struct inode *inode, pgoff_t index, struct folio **foliop,
2132 enum sgp_type sgp)
2133{
2134 return shmem_get_folio_gfp(inode, index, foliop, sgp,
2135 mapping_gfp_mask(inode->i_mapping), NULL, NULL);
2136}
2137
2138/*
2139 * This is like autoremove_wake_function, but it removes the wait queue
2140 * entry unconditionally - even if something else had already woken the
2141 * target.
2142 */
2143static int synchronous_wake_function(wait_queue_entry_t *wait,
2144 unsigned int mode, int sync, void *key)
2145{
2146 int ret = default_wake_function(wait, mode, sync, key);
2147 list_del_init(&wait->entry);
2148 return ret;
2149}
2150
2151/*
2152 * Trinity finds that probing a hole which tmpfs is punching can
2153 * prevent the hole-punch from ever completing: which in turn
2154 * locks writers out with its hold on i_rwsem. So refrain from
2155 * faulting pages into the hole while it's being punched. Although
2156 * shmem_undo_range() does remove the additions, it may be unable to
2157 * keep up, as each new page needs its own unmap_mapping_range() call,
2158 * and the i_mmap tree grows ever slower to scan if new vmas are added.
2159 *
2160 * It does not matter if we sometimes reach this check just before the
2161 * hole-punch begins, so that one fault then races with the punch:
2162 * we just need to make racing faults a rare case.
2163 *
2164 * The implementation below would be much simpler if we just used a
2165 * standard mutex or completion: but we cannot take i_rwsem in fault,
2166 * and bloating every shmem inode for this unlikely case would be sad.
2167 */
2168static vm_fault_t shmem_falloc_wait(struct vm_fault *vmf, struct inode *inode)
2169{
2170 struct shmem_falloc *shmem_falloc;
2171 struct file *fpin = NULL;
2172 vm_fault_t ret = 0;
2173
2174 spin_lock(&inode->i_lock);
2175 shmem_falloc = inode->i_private;
2176 if (shmem_falloc &&
2177 shmem_falloc->waitq &&
2178 vmf->pgoff >= shmem_falloc->start &&
2179 vmf->pgoff < shmem_falloc->next) {
2180 wait_queue_head_t *shmem_falloc_waitq;
2181 DEFINE_WAIT_FUNC(shmem_fault_wait, synchronous_wake_function);
2182
2183 ret = VM_FAULT_NOPAGE;
2184 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2185 shmem_falloc_waitq = shmem_falloc->waitq;
2186 prepare_to_wait(shmem_falloc_waitq, &shmem_fault_wait,
2187 TASK_UNINTERRUPTIBLE);
2188 spin_unlock(&inode->i_lock);
2189 schedule();
2190
2191 /*
2192 * shmem_falloc_waitq points into the shmem_fallocate()
2193 * stack of the hole-punching task: shmem_falloc_waitq
2194 * is usually invalid by the time we reach here, but
2195 * finish_wait() does not dereference it in that case;
2196 * though i_lock needed lest racing with wake_up_all().
2197 */
2198 spin_lock(&inode->i_lock);
2199 finish_wait(shmem_falloc_waitq, &shmem_fault_wait);
2200 }
2201 spin_unlock(&inode->i_lock);
2202 if (fpin) {
2203 fput(fpin);
2204 ret = VM_FAULT_RETRY;
2205 }
2206 return ret;
2207}
2208
2209static vm_fault_t shmem_fault(struct vm_fault *vmf)
2210{
2211 struct inode *inode = file_inode(vmf->vma->vm_file);
2212 gfp_t gfp = mapping_gfp_mask(inode->i_mapping);
2213 struct folio *folio = NULL;
2214 vm_fault_t ret = 0;
2215 int err;
2216
2217 /*
2218 * Trinity finds that probing a hole which tmpfs is punching can
2219 * prevent the hole-punch from ever completing: noted in i_private.
2220 */
2221 if (unlikely(inode->i_private)) {
2222 ret = shmem_falloc_wait(vmf, inode);
2223 if (ret)
2224 return ret;
2225 }
2226
2227 WARN_ON_ONCE(vmf->page != NULL);
2228 err = shmem_get_folio_gfp(inode, vmf->pgoff, &folio, SGP_CACHE,
2229 gfp, vmf, &ret);
2230 if (err)
2231 return vmf_error(err);
2232 if (folio) {
2233 vmf->page = folio_file_page(folio, vmf->pgoff);
2234 ret |= VM_FAULT_LOCKED;
2235 }
2236 return ret;
2237}
2238
2239unsigned long shmem_get_unmapped_area(struct file *file,
2240 unsigned long uaddr, unsigned long len,
2241 unsigned long pgoff, unsigned long flags)
2242{
2243 unsigned long (*get_area)(struct file *,
2244 unsigned long, unsigned long, unsigned long, unsigned long);
2245 unsigned long addr;
2246 unsigned long offset;
2247 unsigned long inflated_len;
2248 unsigned long inflated_addr;
2249 unsigned long inflated_offset;
2250
2251 if (len > TASK_SIZE)
2252 return -ENOMEM;
2253
2254 get_area = current->mm->get_unmapped_area;
2255 addr = get_area(file, uaddr, len, pgoff, flags);
2256
2257 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
2258 return addr;
2259 if (IS_ERR_VALUE(addr))
2260 return addr;
2261 if (addr & ~PAGE_MASK)
2262 return addr;
2263 if (addr > TASK_SIZE - len)
2264 return addr;
2265
2266 if (shmem_huge == SHMEM_HUGE_DENY)
2267 return addr;
2268 if (len < HPAGE_PMD_SIZE)
2269 return addr;
2270 if (flags & MAP_FIXED)
2271 return addr;
2272 /*
2273 * Our priority is to support MAP_SHARED mapped hugely;
2274 * and support MAP_PRIVATE mapped hugely too, until it is COWed.
2275 * But if caller specified an address hint and we allocated area there
2276 * successfully, respect that as before.
2277 */
2278 if (uaddr == addr)
2279 return addr;
2280
2281 if (shmem_huge != SHMEM_HUGE_FORCE) {
2282 struct super_block *sb;
2283
2284 if (file) {
2285 VM_BUG_ON(file->f_op != &shmem_file_operations);
2286 sb = file_inode(file)->i_sb;
2287 } else {
2288 /*
2289 * Called directly from mm/mmap.c, or drivers/char/mem.c
2290 * for "/dev/zero", to create a shared anonymous object.
2291 */
2292 if (IS_ERR(shm_mnt))
2293 return addr;
2294 sb = shm_mnt->mnt_sb;
2295 }
2296 if (SHMEM_SB(sb)->huge == SHMEM_HUGE_NEVER)
2297 return addr;
2298 }
2299
2300 offset = (pgoff << PAGE_SHIFT) & (HPAGE_PMD_SIZE-1);
2301 if (offset && offset + len < 2 * HPAGE_PMD_SIZE)
2302 return addr;
2303 if ((addr & (HPAGE_PMD_SIZE-1)) == offset)
2304 return addr;
2305
2306 inflated_len = len + HPAGE_PMD_SIZE - PAGE_SIZE;
2307 if (inflated_len > TASK_SIZE)
2308 return addr;
2309 if (inflated_len < len)
2310 return addr;
2311
2312 inflated_addr = get_area(NULL, uaddr, inflated_len, 0, flags);
2313 if (IS_ERR_VALUE(inflated_addr))
2314 return addr;
2315 if (inflated_addr & ~PAGE_MASK)
2316 return addr;
2317
2318 inflated_offset = inflated_addr & (HPAGE_PMD_SIZE-1);
2319 inflated_addr += offset - inflated_offset;
2320 if (inflated_offset > offset)
2321 inflated_addr += HPAGE_PMD_SIZE;
2322
2323 if (inflated_addr > TASK_SIZE - len)
2324 return addr;
2325 return inflated_addr;
2326}
2327
2328#ifdef CONFIG_NUMA
2329static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
2330{
2331 struct inode *inode = file_inode(vma->vm_file);
2332 return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
2333}
2334
2335static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
2336 unsigned long addr, pgoff_t *ilx)
2337{
2338 struct inode *inode = file_inode(vma->vm_file);
2339 pgoff_t index;
2340
2341 /*
2342 * Bias interleave by inode number to distribute better across nodes;
2343 * but this interface is independent of which page order is used, so
2344 * supplies only that bias, letting caller apply the offset (adjusted
2345 * by page order, as in shmem_get_pgoff_policy() and get_vma_policy()).
2346 */
2347 *ilx = inode->i_ino;
2348 index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2349 return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
2350}
2351
2352static struct mempolicy *shmem_get_pgoff_policy(struct shmem_inode_info *info,
2353 pgoff_t index, unsigned int order, pgoff_t *ilx)
2354{
2355 struct mempolicy *mpol;
2356
2357 /* Bias interleave by inode number to distribute better across nodes */
2358 *ilx = info->vfs_inode.i_ino + (index >> order);
2359
2360 mpol = mpol_shared_policy_lookup(&info->policy, index);
2361 return mpol ? mpol : get_task_policy(current);
2362}
2363#else
2364static struct mempolicy *shmem_get_pgoff_policy(struct shmem_inode_info *info,
2365 pgoff_t index, unsigned int order, pgoff_t *ilx)
2366{
2367 *ilx = 0;
2368 return NULL;
2369}
2370#endif /* CONFIG_NUMA */
2371
2372int shmem_lock(struct file *file, int lock, struct ucounts *ucounts)
2373{
2374 struct inode *inode = file_inode(file);
2375 struct shmem_inode_info *info = SHMEM_I(inode);
2376 int retval = -ENOMEM;
2377
2378 /*
2379 * What serializes the accesses to info->flags?
2380 * ipc_lock_object() when called from shmctl_do_lock(),
2381 * no serialization needed when called from shm_destroy().
2382 */
2383 if (lock && !(info->flags & VM_LOCKED)) {
2384 if (!user_shm_lock(inode->i_size, ucounts))
2385 goto out_nomem;
2386 info->flags |= VM_LOCKED;
2387 mapping_set_unevictable(file->f_mapping);
2388 }
2389 if (!lock && (info->flags & VM_LOCKED) && ucounts) {
2390 user_shm_unlock(inode->i_size, ucounts);
2391 info->flags &= ~VM_LOCKED;
2392 mapping_clear_unevictable(file->f_mapping);
2393 }
2394 retval = 0;
2395
2396out_nomem:
2397 return retval;
2398}
2399
2400static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
2401{
2402 struct inode *inode = file_inode(file);
2403 struct shmem_inode_info *info = SHMEM_I(inode);
2404 int ret;
2405
2406 ret = seal_check_write(info->seals, vma);
2407 if (ret)
2408 return ret;
2409
2410 /* arm64 - allow memory tagging on RAM-based files */
2411 vm_flags_set(vma, VM_MTE_ALLOWED);
2412
2413 file_accessed(file);
2414 /* This is anonymous shared memory if it is unlinked at the time of mmap */
2415 if (inode->i_nlink)
2416 vma->vm_ops = &shmem_vm_ops;
2417 else
2418 vma->vm_ops = &shmem_anon_vm_ops;
2419 return 0;
2420}
2421
2422static int shmem_file_open(struct inode *inode, struct file *file)
2423{
2424 file->f_mode |= FMODE_CAN_ODIRECT;
2425 return generic_file_open(inode, file);
2426}
2427
2428#ifdef CONFIG_TMPFS_XATTR
2429static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
2430
2431/*
2432 * chattr's fsflags are unrelated to extended attributes,
2433 * but tmpfs has chosen to enable them under the same config option.
2434 */
2435static void shmem_set_inode_flags(struct inode *inode, unsigned int fsflags)
2436{
2437 unsigned int i_flags = 0;
2438
2439 if (fsflags & FS_NOATIME_FL)
2440 i_flags |= S_NOATIME;
2441 if (fsflags & FS_APPEND_FL)
2442 i_flags |= S_APPEND;
2443 if (fsflags & FS_IMMUTABLE_FL)
2444 i_flags |= S_IMMUTABLE;
2445 /*
2446 * But FS_NODUMP_FL does not require any action in i_flags.
2447 */
2448 inode_set_flags(inode, i_flags, S_NOATIME | S_APPEND | S_IMMUTABLE);
2449}
2450#else
2451static void shmem_set_inode_flags(struct inode *inode, unsigned int fsflags)
2452{
2453}
2454#define shmem_initxattrs NULL
2455#endif
2456
2457static struct offset_ctx *shmem_get_offset_ctx(struct inode *inode)
2458{
2459 return &SHMEM_I(inode)->dir_offsets;
2460}
2461
2462static struct inode *__shmem_get_inode(struct mnt_idmap *idmap,
2463 struct super_block *sb,
2464 struct inode *dir, umode_t mode,
2465 dev_t dev, unsigned long flags)
2466{
2467 struct inode *inode;
2468 struct shmem_inode_info *info;
2469 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2470 ino_t ino;
2471 int err;
2472
2473 err = shmem_reserve_inode(sb, &ino);
2474 if (err)
2475 return ERR_PTR(err);
2476
2477 inode = new_inode(sb);
2478 if (!inode) {
2479 shmem_free_inode(sb, 0);
2480 return ERR_PTR(-ENOSPC);
2481 }
2482
2483 inode->i_ino = ino;
2484 inode_init_owner(idmap, inode, dir, mode);
2485 inode->i_blocks = 0;
2486 simple_inode_init_ts(inode);
2487 inode->i_generation = get_random_u32();
2488 info = SHMEM_I(inode);
2489 memset(info, 0, (char *)inode - (char *)info);
2490 spin_lock_init(&info->lock);
2491 atomic_set(&info->stop_eviction, 0);
2492 info->seals = F_SEAL_SEAL;
2493 info->flags = flags & VM_NORESERVE;
2494 info->i_crtime = inode_get_mtime(inode);
2495 info->fsflags = (dir == NULL) ? 0 :
2496 SHMEM_I(dir)->fsflags & SHMEM_FL_INHERITED;
2497 if (info->fsflags)
2498 shmem_set_inode_flags(inode, info->fsflags);
2499 INIT_LIST_HEAD(&info->shrinklist);
2500 INIT_LIST_HEAD(&info->swaplist);
2501 simple_xattrs_init(&info->xattrs);
2502 cache_no_acl(inode);
2503 if (sbinfo->noswap)
2504 mapping_set_unevictable(inode->i_mapping);
2505 mapping_set_large_folios(inode->i_mapping);
2506
2507 switch (mode & S_IFMT) {
2508 default:
2509 inode->i_op = &shmem_special_inode_operations;
2510 init_special_inode(inode, mode, dev);
2511 break;
2512 case S_IFREG:
2513 inode->i_mapping->a_ops = &shmem_aops;
2514 inode->i_op = &shmem_inode_operations;
2515 inode->i_fop = &shmem_file_operations;
2516 mpol_shared_policy_init(&info->policy,
2517 shmem_get_sbmpol(sbinfo));
2518 break;
2519 case S_IFDIR:
2520 inc_nlink(inode);
2521 /* Some things misbehave if size == 0 on a directory */
2522 inode->i_size = 2 * BOGO_DIRENT_SIZE;
2523 inode->i_op = &shmem_dir_inode_operations;
2524 inode->i_fop = &simple_offset_dir_operations;
2525 simple_offset_init(shmem_get_offset_ctx(inode));
2526 break;
2527 case S_IFLNK:
2528 /*
2529 * Must not load anything in the rbtree,
2530 * mpol_free_shared_policy will not be called.
2531 */
2532 mpol_shared_policy_init(&info->policy, NULL);
2533 break;
2534 }
2535
2536 lockdep_annotate_inode_mutex_key(inode);
2537 return inode;
2538}
2539
2540#ifdef CONFIG_TMPFS_QUOTA
2541static struct inode *shmem_get_inode(struct mnt_idmap *idmap,
2542 struct super_block *sb, struct inode *dir,
2543 umode_t mode, dev_t dev, unsigned long flags)
2544{
2545 int err;
2546 struct inode *inode;
2547
2548 inode = __shmem_get_inode(idmap, sb, dir, mode, dev, flags);
2549 if (IS_ERR(inode))
2550 return inode;
2551
2552 err = dquot_initialize(inode);
2553 if (err)
2554 goto errout;
2555
2556 err = dquot_alloc_inode(inode);
2557 if (err) {
2558 dquot_drop(inode);
2559 goto errout;
2560 }
2561 return inode;
2562
2563errout:
2564 inode->i_flags |= S_NOQUOTA;
2565 iput(inode);
2566 return ERR_PTR(err);
2567}
2568#else
2569static inline struct inode *shmem_get_inode(struct mnt_idmap *idmap,
2570 struct super_block *sb, struct inode *dir,
2571 umode_t mode, dev_t dev, unsigned long flags)
2572{
2573 return __shmem_get_inode(idmap, sb, dir, mode, dev, flags);
2574}
2575#endif /* CONFIG_TMPFS_QUOTA */
2576
2577#ifdef CONFIG_USERFAULTFD
2578int shmem_mfill_atomic_pte(pmd_t *dst_pmd,
2579 struct vm_area_struct *dst_vma,
2580 unsigned long dst_addr,
2581 unsigned long src_addr,
2582 uffd_flags_t flags,
2583 struct folio **foliop)
2584{
2585 struct inode *inode = file_inode(dst_vma->vm_file);
2586 struct shmem_inode_info *info = SHMEM_I(inode);
2587 struct address_space *mapping = inode->i_mapping;
2588 gfp_t gfp = mapping_gfp_mask(mapping);
2589 pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
2590 void *page_kaddr;
2591 struct folio *folio;
2592 int ret;
2593 pgoff_t max_off;
2594
2595 if (shmem_inode_acct_blocks(inode, 1)) {
2596 /*
2597 * We may have got a page, returned -ENOENT triggering a retry,
2598 * and now we find ourselves with -ENOMEM. Release the page, to
2599 * avoid a BUG_ON in our caller.
2600 */
2601 if (unlikely(*foliop)) {
2602 folio_put(*foliop);
2603 *foliop = NULL;
2604 }
2605 return -ENOMEM;
2606 }
2607
2608 if (!*foliop) {
2609 ret = -ENOMEM;
2610 folio = shmem_alloc_folio(gfp, info, pgoff);
2611 if (!folio)
2612 goto out_unacct_blocks;
2613
2614 if (uffd_flags_mode_is(flags, MFILL_ATOMIC_COPY)) {
2615 page_kaddr = kmap_local_folio(folio, 0);
2616 /*
2617 * The read mmap_lock is held here. Despite the
2618 * mmap_lock being read recursive a deadlock is still
2619 * possible if a writer has taken a lock. For example:
2620 *
2621 * process A thread 1 takes read lock on own mmap_lock
2622 * process A thread 2 calls mmap, blocks taking write lock
2623 * process B thread 1 takes page fault, read lock on own mmap lock
2624 * process B thread 2 calls mmap, blocks taking write lock
2625 * process A thread 1 blocks taking read lock on process B
2626 * process B thread 1 blocks taking read lock on process A
2627 *
2628 * Disable page faults to prevent potential deadlock
2629 * and retry the copy outside the mmap_lock.
2630 */
2631 pagefault_disable();
2632 ret = copy_from_user(page_kaddr,
2633 (const void __user *)src_addr,
2634 PAGE_SIZE);
2635 pagefault_enable();
2636 kunmap_local(page_kaddr);
2637
2638 /* fallback to copy_from_user outside mmap_lock */
2639 if (unlikely(ret)) {
2640 *foliop = folio;
2641 ret = -ENOENT;
2642 /* don't free the page */
2643 goto out_unacct_blocks;
2644 }
2645
2646 flush_dcache_folio(folio);
2647 } else { /* ZEROPAGE */
2648 clear_user_highpage(&folio->page, dst_addr);
2649 }
2650 } else {
2651 folio = *foliop;
2652 VM_BUG_ON_FOLIO(folio_test_large(folio), folio);
2653 *foliop = NULL;
2654 }
2655
2656 VM_BUG_ON(folio_test_locked(folio));
2657 VM_BUG_ON(folio_test_swapbacked(folio));
2658 __folio_set_locked(folio);
2659 __folio_set_swapbacked(folio);
2660 __folio_mark_uptodate(folio);
2661
2662 ret = -EFAULT;
2663 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2664 if (unlikely(pgoff >= max_off))
2665 goto out_release;
2666
2667 ret = mem_cgroup_charge(folio, dst_vma->vm_mm, gfp);
2668 if (ret)
2669 goto out_release;
2670 ret = shmem_add_to_page_cache(folio, mapping, pgoff, NULL, gfp);
2671 if (ret)
2672 goto out_release;
2673
2674 ret = mfill_atomic_install_pte(dst_pmd, dst_vma, dst_addr,
2675 &folio->page, true, flags);
2676 if (ret)
2677 goto out_delete_from_cache;
2678
2679 shmem_recalc_inode(inode, 1, 0);
2680 folio_unlock(folio);
2681 return 0;
2682out_delete_from_cache:
2683 filemap_remove_folio(folio);
2684out_release:
2685 folio_unlock(folio);
2686 folio_put(folio);
2687out_unacct_blocks:
2688 shmem_inode_unacct_blocks(inode, 1);
2689 return ret;
2690}
2691#endif /* CONFIG_USERFAULTFD */
2692
2693#ifdef CONFIG_TMPFS
2694static const struct inode_operations shmem_symlink_inode_operations;
2695static const struct inode_operations shmem_short_symlink_operations;
2696
2697static int
2698shmem_write_begin(struct file *file, struct address_space *mapping,
2699 loff_t pos, unsigned len,
2700 struct page **pagep, void **fsdata)
2701{
2702 struct inode *inode = mapping->host;
2703 struct shmem_inode_info *info = SHMEM_I(inode);
2704 pgoff_t index = pos >> PAGE_SHIFT;
2705 struct folio *folio;
2706 int ret = 0;
2707
2708 /* i_rwsem is held by caller */
2709 if (unlikely(info->seals & (F_SEAL_GROW |
2710 F_SEAL_WRITE | F_SEAL_FUTURE_WRITE))) {
2711 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE))
2712 return -EPERM;
2713 if ((info->seals & F_SEAL_GROW) && pos + len > inode->i_size)
2714 return -EPERM;
2715 }
2716
2717 ret = shmem_get_folio(inode, index, &folio, SGP_WRITE);
2718 if (ret)
2719 return ret;
2720
2721 *pagep = folio_file_page(folio, index);
2722 if (PageHWPoison(*pagep)) {
2723 folio_unlock(folio);
2724 folio_put(folio);
2725 *pagep = NULL;
2726 return -EIO;
2727 }
2728
2729 return 0;
2730}
2731
2732static int
2733shmem_write_end(struct file *file, struct address_space *mapping,
2734 loff_t pos, unsigned len, unsigned copied,
2735 struct page *page, void *fsdata)
2736{
2737 struct folio *folio = page_folio(page);
2738 struct inode *inode = mapping->host;
2739
2740 if (pos + copied > inode->i_size)
2741 i_size_write(inode, pos + copied);
2742
2743 if (!folio_test_uptodate(folio)) {
2744 if (copied < folio_size(folio)) {
2745 size_t from = offset_in_folio(folio, pos);
2746 folio_zero_segments(folio, 0, from,
2747 from + copied, folio_size(folio));
2748 }
2749 folio_mark_uptodate(folio);
2750 }
2751 folio_mark_dirty(folio);
2752 folio_unlock(folio);
2753 folio_put(folio);
2754
2755 return copied;
2756}
2757
2758static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
2759{
2760 struct file *file = iocb->ki_filp;
2761 struct inode *inode = file_inode(file);
2762 struct address_space *mapping = inode->i_mapping;
2763 pgoff_t index;
2764 unsigned long offset;
2765 int error = 0;
2766 ssize_t retval = 0;
2767 loff_t *ppos = &iocb->ki_pos;
2768
2769 index = *ppos >> PAGE_SHIFT;
2770 offset = *ppos & ~PAGE_MASK;
2771
2772 for (;;) {
2773 struct folio *folio = NULL;
2774 struct page *page = NULL;
2775 pgoff_t end_index;
2776 unsigned long nr, ret;
2777 loff_t i_size = i_size_read(inode);
2778
2779 end_index = i_size >> PAGE_SHIFT;
2780 if (index > end_index)
2781 break;
2782 if (index == end_index) {
2783 nr = i_size & ~PAGE_MASK;
2784 if (nr <= offset)
2785 break;
2786 }
2787
2788 error = shmem_get_folio(inode, index, &folio, SGP_READ);
2789 if (error) {
2790 if (error == -EINVAL)
2791 error = 0;
2792 break;
2793 }
2794 if (folio) {
2795 folio_unlock(folio);
2796
2797 page = folio_file_page(folio, index);
2798 if (PageHWPoison(page)) {
2799 folio_put(folio);
2800 error = -EIO;
2801 break;
2802 }
2803 }
2804
2805 /*
2806 * We must evaluate after, since reads (unlike writes)
2807 * are called without i_rwsem protection against truncate
2808 */
2809 nr = PAGE_SIZE;
2810 i_size = i_size_read(inode);
2811 end_index = i_size >> PAGE_SHIFT;
2812 if (index == end_index) {
2813 nr = i_size & ~PAGE_MASK;
2814 if (nr <= offset) {
2815 if (folio)
2816 folio_put(folio);
2817 break;
2818 }
2819 }
2820 nr -= offset;
2821
2822 if (folio) {
2823 /*
2824 * If users can be writing to this page using arbitrary
2825 * virtual addresses, take care about potential aliasing
2826 * before reading the page on the kernel side.
2827 */
2828 if (mapping_writably_mapped(mapping))
2829 flush_dcache_page(page);
2830 /*
2831 * Mark the page accessed if we read the beginning.
2832 */
2833 if (!offset)
2834 folio_mark_accessed(folio);
2835 /*
2836 * Ok, we have the page, and it's up-to-date, so
2837 * now we can copy it to user space...
2838 */
2839 ret = copy_page_to_iter(page, offset, nr, to);
2840 folio_put(folio);
2841
2842 } else if (user_backed_iter(to)) {
2843 /*
2844 * Copy to user tends to be so well optimized, but
2845 * clear_user() not so much, that it is noticeably
2846 * faster to copy the zero page instead of clearing.
2847 */
2848 ret = copy_page_to_iter(ZERO_PAGE(0), offset, nr, to);
2849 } else {
2850 /*
2851 * But submitting the same page twice in a row to
2852 * splice() - or others? - can result in confusion:
2853 * so don't attempt that optimization on pipes etc.
2854 */
2855 ret = iov_iter_zero(nr, to);
2856 }
2857
2858 retval += ret;
2859 offset += ret;
2860 index += offset >> PAGE_SHIFT;
2861 offset &= ~PAGE_MASK;
2862
2863 if (!iov_iter_count(to))
2864 break;
2865 if (ret < nr) {
2866 error = -EFAULT;
2867 break;
2868 }
2869 cond_resched();
2870 }
2871
2872 *ppos = ((loff_t) index << PAGE_SHIFT) + offset;
2873 file_accessed(file);
2874 return retval ? retval : error;
2875}
2876
2877static ssize_t shmem_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
2878{
2879 struct file *file = iocb->ki_filp;
2880 struct inode *inode = file->f_mapping->host;
2881 ssize_t ret;
2882
2883 inode_lock(inode);
2884 ret = generic_write_checks(iocb, from);
2885 if (ret <= 0)
2886 goto unlock;
2887 ret = file_remove_privs(file);
2888 if (ret)
2889 goto unlock;
2890 ret = file_update_time(file);
2891 if (ret)
2892 goto unlock;
2893 ret = generic_perform_write(iocb, from);
2894unlock:
2895 inode_unlock(inode);
2896 return ret;
2897}
2898
2899static bool zero_pipe_buf_get(struct pipe_inode_info *pipe,
2900 struct pipe_buffer *buf)
2901{
2902 return true;
2903}
2904
2905static void zero_pipe_buf_release(struct pipe_inode_info *pipe,
2906 struct pipe_buffer *buf)
2907{
2908}
2909
2910static bool zero_pipe_buf_try_steal(struct pipe_inode_info *pipe,
2911 struct pipe_buffer *buf)
2912{
2913 return false;
2914}
2915
2916static const struct pipe_buf_operations zero_pipe_buf_ops = {
2917 .release = zero_pipe_buf_release,
2918 .try_steal = zero_pipe_buf_try_steal,
2919 .get = zero_pipe_buf_get,
2920};
2921
2922static size_t splice_zeropage_into_pipe(struct pipe_inode_info *pipe,
2923 loff_t fpos, size_t size)
2924{
2925 size_t offset = fpos & ~PAGE_MASK;
2926
2927 size = min_t(size_t, size, PAGE_SIZE - offset);
2928
2929 if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage)) {
2930 struct pipe_buffer *buf = pipe_head_buf(pipe);
2931
2932 *buf = (struct pipe_buffer) {
2933 .ops = &zero_pipe_buf_ops,
2934 .page = ZERO_PAGE(0),
2935 .offset = offset,
2936 .len = size,
2937 };
2938 pipe->head++;
2939 }
2940
2941 return size;
2942}
2943
2944static ssize_t shmem_file_splice_read(struct file *in, loff_t *ppos,
2945 struct pipe_inode_info *pipe,
2946 size_t len, unsigned int flags)
2947{
2948 struct inode *inode = file_inode(in);
2949 struct address_space *mapping = inode->i_mapping;
2950 struct folio *folio = NULL;
2951 size_t total_spliced = 0, used, npages, n, part;
2952 loff_t isize;
2953 int error = 0;
2954
2955 /* Work out how much data we can actually add into the pipe */
2956 used = pipe_occupancy(pipe->head, pipe->tail);
2957 npages = max_t(ssize_t, pipe->max_usage - used, 0);
2958 len = min_t(size_t, len, npages * PAGE_SIZE);
2959
2960 do {
2961 if (*ppos >= i_size_read(inode))
2962 break;
2963
2964 error = shmem_get_folio(inode, *ppos / PAGE_SIZE, &folio,
2965 SGP_READ);
2966 if (error) {
2967 if (error == -EINVAL)
2968 error = 0;
2969 break;
2970 }
2971 if (folio) {
2972 folio_unlock(folio);
2973
2974 if (folio_test_hwpoison(folio) ||
2975 (folio_test_large(folio) &&
2976 folio_test_has_hwpoisoned(folio))) {
2977 error = -EIO;
2978 break;
2979 }
2980 }
2981
2982 /*
2983 * i_size must be checked after we know the pages are Uptodate.
2984 *
2985 * Checking i_size after the check allows us to calculate
2986 * the correct value for "nr", which means the zero-filled
2987 * part of the page is not copied back to userspace (unless
2988 * another truncate extends the file - this is desired though).
2989 */
2990 isize = i_size_read(inode);
2991 if (unlikely(*ppos >= isize))
2992 break;
2993 part = min_t(loff_t, isize - *ppos, len);
2994
2995 if (folio) {
2996 /*
2997 * If users can be writing to this page using arbitrary
2998 * virtual addresses, take care about potential aliasing
2999 * before reading the page on the kernel side.
3000 */
3001 if (mapping_writably_mapped(mapping))
3002 flush_dcache_folio(folio);
3003 folio_mark_accessed(folio);
3004 /*
3005 * Ok, we have the page, and it's up-to-date, so we can
3006 * now splice it into the pipe.
3007 */
3008 n = splice_folio_into_pipe(pipe, folio, *ppos, part);
3009 folio_put(folio);
3010 folio = NULL;
3011 } else {
3012 n = splice_zeropage_into_pipe(pipe, *ppos, part);
3013 }
3014
3015 if (!n)
3016 break;
3017 len -= n;
3018 total_spliced += n;
3019 *ppos += n;
3020 in->f_ra.prev_pos = *ppos;
3021 if (pipe_full(pipe->head, pipe->tail, pipe->max_usage))
3022 break;
3023
3024 cond_resched();
3025 } while (len);
3026
3027 if (folio)
3028 folio_put(folio);
3029
3030 file_accessed(in);
3031 return total_spliced ? total_spliced : error;
3032}
3033
3034static loff_t shmem_file_llseek(struct file *file, loff_t offset, int whence)
3035{
3036 struct address_space *mapping = file->f_mapping;
3037 struct inode *inode = mapping->host;
3038
3039 if (whence != SEEK_DATA && whence != SEEK_HOLE)
3040 return generic_file_llseek_size(file, offset, whence,
3041 MAX_LFS_FILESIZE, i_size_read(inode));
3042 if (offset < 0)
3043 return -ENXIO;
3044
3045 inode_lock(inode);
3046 /* We're holding i_rwsem so we can access i_size directly */
3047 offset = mapping_seek_hole_data(mapping, offset, inode->i_size, whence);
3048 if (offset >= 0)
3049 offset = vfs_setpos(file, offset, MAX_LFS_FILESIZE);
3050 inode_unlock(inode);
3051 return offset;
3052}
3053
3054static long shmem_fallocate(struct file *file, int mode, loff_t offset,
3055 loff_t len)
3056{
3057 struct inode *inode = file_inode(file);
3058 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
3059 struct shmem_inode_info *info = SHMEM_I(inode);
3060 struct shmem_falloc shmem_falloc;
3061 pgoff_t start, index, end, undo_fallocend;
3062 int error;
3063
3064 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
3065 return -EOPNOTSUPP;
3066
3067 inode_lock(inode);
3068
3069 if (mode & FALLOC_FL_PUNCH_HOLE) {
3070 struct address_space *mapping = file->f_mapping;
3071 loff_t unmap_start = round_up(offset, PAGE_SIZE);
3072 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
3073 DECLARE_WAIT_QUEUE_HEAD_ONSTACK(shmem_falloc_waitq);
3074
3075 /* protected by i_rwsem */
3076 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
3077 error = -EPERM;
3078 goto out;
3079 }
3080
3081 shmem_falloc.waitq = &shmem_falloc_waitq;
3082 shmem_falloc.start = (u64)unmap_start >> PAGE_SHIFT;
3083 shmem_falloc.next = (unmap_end + 1) >> PAGE_SHIFT;
3084 spin_lock(&inode->i_lock);
3085 inode->i_private = &shmem_falloc;
3086 spin_unlock(&inode->i_lock);
3087
3088 if ((u64)unmap_end > (u64)unmap_start)
3089 unmap_mapping_range(mapping, unmap_start,
3090 1 + unmap_end - unmap_start, 0);
3091 shmem_truncate_range(inode, offset, offset + len - 1);
3092 /* No need to unmap again: hole-punching leaves COWed pages */
3093
3094 spin_lock(&inode->i_lock);
3095 inode->i_private = NULL;
3096 wake_up_all(&shmem_falloc_waitq);
3097 WARN_ON_ONCE(!list_empty(&shmem_falloc_waitq.head));
3098 spin_unlock(&inode->i_lock);
3099 error = 0;
3100 goto out;
3101 }
3102
3103 /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
3104 error = inode_newsize_ok(inode, offset + len);
3105 if (error)
3106 goto out;
3107
3108 if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
3109 error = -EPERM;
3110 goto out;
3111 }
3112
3113 start = offset >> PAGE_SHIFT;
3114 end = (offset + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
3115 /* Try to avoid a swapstorm if len is impossible to satisfy */
3116 if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
3117 error = -ENOSPC;
3118 goto out;
3119 }
3120
3121 shmem_falloc.waitq = NULL;
3122 shmem_falloc.start = start;
3123 shmem_falloc.next = start;
3124 shmem_falloc.nr_falloced = 0;
3125 shmem_falloc.nr_unswapped = 0;
3126 spin_lock(&inode->i_lock);
3127 inode->i_private = &shmem_falloc;
3128 spin_unlock(&inode->i_lock);
3129
3130 /*
3131 * info->fallocend is only relevant when huge pages might be
3132 * involved: to prevent split_huge_page() freeing fallocated
3133 * pages when FALLOC_FL_KEEP_SIZE committed beyond i_size.
3134 */
3135 undo_fallocend = info->fallocend;
3136 if (info->fallocend < end)
3137 info->fallocend = end;
3138
3139 for (index = start; index < end; ) {
3140 struct folio *folio;
3141
3142 /*
3143 * Good, the fallocate(2) manpage permits EINTR: we may have
3144 * been interrupted because we are using up too much memory.
3145 */
3146 if (signal_pending(current))
3147 error = -EINTR;
3148 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
3149 error = -ENOMEM;
3150 else
3151 error = shmem_get_folio(inode, index, &folio,
3152 SGP_FALLOC);
3153 if (error) {
3154 info->fallocend = undo_fallocend;
3155 /* Remove the !uptodate folios we added */
3156 if (index > start) {
3157 shmem_undo_range(inode,
3158 (loff_t)start << PAGE_SHIFT,
3159 ((loff_t)index << PAGE_SHIFT) - 1, true);
3160 }
3161 goto undone;
3162 }
3163
3164 /*
3165 * Here is a more important optimization than it appears:
3166 * a second SGP_FALLOC on the same large folio will clear it,
3167 * making it uptodate and un-undoable if we fail later.
3168 */
3169 index = folio_next_index(folio);
3170 /* Beware 32-bit wraparound */
3171 if (!index)
3172 index--;
3173
3174 /*
3175 * Inform shmem_writepage() how far we have reached.
3176 * No need for lock or barrier: we have the page lock.
3177 */
3178 if (!folio_test_uptodate(folio))
3179 shmem_falloc.nr_falloced += index - shmem_falloc.next;
3180 shmem_falloc.next = index;
3181
3182 /*
3183 * If !uptodate, leave it that way so that freeable folios
3184 * can be recognized if we need to rollback on error later.
3185 * But mark it dirty so that memory pressure will swap rather
3186 * than free the folios we are allocating (and SGP_CACHE folios
3187 * might still be clean: we now need to mark those dirty too).
3188 */
3189 folio_mark_dirty(folio);
3190 folio_unlock(folio);
3191 folio_put(folio);
3192 cond_resched();
3193 }
3194
3195 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
3196 i_size_write(inode, offset + len);
3197undone:
3198 spin_lock(&inode->i_lock);
3199 inode->i_private = NULL;
3200 spin_unlock(&inode->i_lock);
3201out:
3202 if (!error)
3203 file_modified(file);
3204 inode_unlock(inode);
3205 return error;
3206}
3207
3208static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
3209{
3210 struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
3211
3212 buf->f_type = TMPFS_MAGIC;
3213 buf->f_bsize = PAGE_SIZE;
3214 buf->f_namelen = NAME_MAX;
3215 if (sbinfo->max_blocks) {
3216 buf->f_blocks = sbinfo->max_blocks;
3217 buf->f_bavail =
3218 buf->f_bfree = sbinfo->max_blocks -
3219 percpu_counter_sum(&sbinfo->used_blocks);
3220 }
3221 if (sbinfo->max_inodes) {
3222 buf->f_files = sbinfo->max_inodes;
3223 buf->f_ffree = sbinfo->free_ispace / BOGO_INODE_SIZE;
3224 }
3225 /* else leave those fields 0 like simple_statfs */
3226
3227 buf->f_fsid = uuid_to_fsid(dentry->d_sb->s_uuid.b);
3228
3229 return 0;
3230}
3231
3232/*
3233 * File creation. Allocate an inode, and we're done..
3234 */
3235static int
3236shmem_mknod(struct mnt_idmap *idmap, struct inode *dir,
3237 struct dentry *dentry, umode_t mode, dev_t dev)
3238{
3239 struct inode *inode;
3240 int error;
3241
3242 inode = shmem_get_inode(idmap, dir->i_sb, dir, mode, dev, VM_NORESERVE);
3243 if (IS_ERR(inode))
3244 return PTR_ERR(inode);
3245
3246 error = simple_acl_create(dir, inode);
3247 if (error)
3248 goto out_iput;
3249 error = security_inode_init_security(inode, dir, &dentry->d_name,
3250 shmem_initxattrs, NULL);
3251 if (error && error != -EOPNOTSUPP)
3252 goto out_iput;
3253
3254 error = simple_offset_add(shmem_get_offset_ctx(dir), dentry);
3255 if (error)
3256 goto out_iput;
3257
3258 dir->i_size += BOGO_DIRENT_SIZE;
3259 inode_set_mtime_to_ts(dir, inode_set_ctime_current(dir));
3260 inode_inc_iversion(dir);
3261 d_instantiate(dentry, inode);
3262 dget(dentry); /* Extra count - pin the dentry in core */
3263 return error;
3264
3265out_iput:
3266 iput(inode);
3267 return error;
3268}
3269
3270static int
3271shmem_tmpfile(struct mnt_idmap *idmap, struct inode *dir,
3272 struct file *file, umode_t mode)
3273{
3274 struct inode *inode;
3275 int error;
3276
3277 inode = shmem_get_inode(idmap, dir->i_sb, dir, mode, 0, VM_NORESERVE);
3278 if (IS_ERR(inode)) {
3279 error = PTR_ERR(inode);
3280 goto err_out;
3281 }
3282 error = security_inode_init_security(inode, dir, NULL,
3283 shmem_initxattrs, NULL);
3284 if (error && error != -EOPNOTSUPP)
3285 goto out_iput;
3286 error = simple_acl_create(dir, inode);
3287 if (error)
3288 goto out_iput;
3289 d_tmpfile(file, inode);
3290
3291err_out:
3292 return finish_open_simple(file, error);
3293out_iput:
3294 iput(inode);
3295 return error;
3296}
3297
3298static int shmem_mkdir(struct mnt_idmap *idmap, struct inode *dir,
3299 struct dentry *dentry, umode_t mode)
3300{
3301 int error;
3302
3303 error = shmem_mknod(idmap, dir, dentry, mode | S_IFDIR, 0);
3304 if (error)
3305 return error;
3306 inc_nlink(dir);
3307 return 0;
3308}
3309
3310static int shmem_create(struct mnt_idmap *idmap, struct inode *dir,
3311 struct dentry *dentry, umode_t mode, bool excl)
3312{
3313 return shmem_mknod(idmap, dir, dentry, mode | S_IFREG, 0);
3314}
3315
3316/*
3317 * Link a file..
3318 */
3319static int shmem_link(struct dentry *old_dentry, struct inode *dir,
3320 struct dentry *dentry)
3321{
3322 struct inode *inode = d_inode(old_dentry);
3323 int ret = 0;
3324
3325 /*
3326 * No ordinary (disk based) filesystem counts links as inodes;
3327 * but each new link needs a new dentry, pinning lowmem, and
3328 * tmpfs dentries cannot be pruned until they are unlinked.
3329 * But if an O_TMPFILE file is linked into the tmpfs, the
3330 * first link must skip that, to get the accounting right.
3331 */
3332 if (inode->i_nlink) {
3333 ret = shmem_reserve_inode(inode->i_sb, NULL);
3334 if (ret)
3335 goto out;
3336 }
3337
3338 ret = simple_offset_add(shmem_get_offset_ctx(dir), dentry);
3339 if (ret) {
3340 if (inode->i_nlink)
3341 shmem_free_inode(inode->i_sb, 0);
3342 goto out;
3343 }
3344
3345 dir->i_size += BOGO_DIRENT_SIZE;
3346 inode_set_mtime_to_ts(dir,
3347 inode_set_ctime_to_ts(dir, inode_set_ctime_current(inode)));
3348 inode_inc_iversion(dir);
3349 inc_nlink(inode);
3350 ihold(inode); /* New dentry reference */
3351 dget(dentry); /* Extra pinning count for the created dentry */
3352 d_instantiate(dentry, inode);
3353out:
3354 return ret;
3355}
3356
3357static int shmem_unlink(struct inode *dir, struct dentry *dentry)
3358{
3359 struct inode *inode = d_inode(dentry);
3360
3361 if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
3362 shmem_free_inode(inode->i_sb, 0);
3363
3364 simple_offset_remove(shmem_get_offset_ctx(dir), dentry);
3365
3366 dir->i_size -= BOGO_DIRENT_SIZE;
3367 inode_set_mtime_to_ts(dir,
3368 inode_set_ctime_to_ts(dir, inode_set_ctime_current(inode)));
3369 inode_inc_iversion(dir);
3370 drop_nlink(inode);
3371 dput(dentry); /* Undo the count from "create" - does all the work */
3372 return 0;
3373}
3374
3375static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
3376{
3377 if (!simple_empty(dentry))
3378 return -ENOTEMPTY;
3379
3380 drop_nlink(d_inode(dentry));
3381 drop_nlink(dir);
3382 return shmem_unlink(dir, dentry);
3383}
3384
3385static int shmem_whiteout(struct mnt_idmap *idmap,
3386 struct inode *old_dir, struct dentry *old_dentry)
3387{
3388 struct dentry *whiteout;
3389 int error;
3390
3391 whiteout = d_alloc(old_dentry->d_parent, &old_dentry->d_name);
3392 if (!whiteout)
3393 return -ENOMEM;
3394
3395 error = shmem_mknod(idmap, old_dir, whiteout,
3396 S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
3397 dput(whiteout);
3398 if (error)
3399 return error;
3400
3401 /*
3402 * Cheat and hash the whiteout while the old dentry is still in
3403 * place, instead of playing games with FS_RENAME_DOES_D_MOVE.
3404 *
3405 * d_lookup() will consistently find one of them at this point,
3406 * not sure which one, but that isn't even important.
3407 */
3408 d_rehash(whiteout);
3409 return 0;
3410}
3411
3412/*
3413 * The VFS layer already does all the dentry stuff for rename,
3414 * we just have to decrement the usage count for the target if
3415 * it exists so that the VFS layer correctly free's it when it
3416 * gets overwritten.
3417 */
3418static int shmem_rename2(struct mnt_idmap *idmap,
3419 struct inode *old_dir, struct dentry *old_dentry,
3420 struct inode *new_dir, struct dentry *new_dentry,
3421 unsigned int flags)
3422{
3423 struct inode *inode = d_inode(old_dentry);
3424 int they_are_dirs = S_ISDIR(inode->i_mode);
3425 int error;
3426
3427 if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
3428 return -EINVAL;
3429
3430 if (flags & RENAME_EXCHANGE)
3431 return simple_offset_rename_exchange(old_dir, old_dentry,
3432 new_dir, new_dentry);
3433
3434 if (!simple_empty(new_dentry))
3435 return -ENOTEMPTY;
3436
3437 if (flags & RENAME_WHITEOUT) {
3438 error = shmem_whiteout(idmap, old_dir, old_dentry);
3439 if (error)
3440 return error;
3441 }
3442
3443 simple_offset_remove(shmem_get_offset_ctx(old_dir), old_dentry);
3444 error = simple_offset_add(shmem_get_offset_ctx(new_dir), old_dentry);
3445 if (error)
3446 return error;
3447
3448 if (d_really_is_positive(new_dentry)) {
3449 (void) shmem_unlink(new_dir, new_dentry);
3450 if (they_are_dirs) {
3451 drop_nlink(d_inode(new_dentry));
3452 drop_nlink(old_dir);
3453 }
3454 } else if (they_are_dirs) {
3455 drop_nlink(old_dir);
3456 inc_nlink(new_dir);
3457 }
3458
3459 old_dir->i_size -= BOGO_DIRENT_SIZE;
3460 new_dir->i_size += BOGO_DIRENT_SIZE;
3461 simple_rename_timestamp(old_dir, old_dentry, new_dir, new_dentry);
3462 inode_inc_iversion(old_dir);
3463 inode_inc_iversion(new_dir);
3464 return 0;
3465}
3466
3467static int shmem_symlink(struct mnt_idmap *idmap, struct inode *dir,
3468 struct dentry *dentry, const char *symname)
3469{
3470 int error;
3471 int len;
3472 struct inode *inode;
3473 struct folio *folio;
3474
3475 len = strlen(symname) + 1;
3476 if (len > PAGE_SIZE)
3477 return -ENAMETOOLONG;
3478
3479 inode = shmem_get_inode(idmap, dir->i_sb, dir, S_IFLNK | 0777, 0,
3480 VM_NORESERVE);
3481 if (IS_ERR(inode))
3482 return PTR_ERR(inode);
3483
3484 error = security_inode_init_security(inode, dir, &dentry->d_name,
3485 shmem_initxattrs, NULL);
3486 if (error && error != -EOPNOTSUPP)
3487 goto out_iput;
3488
3489 error = simple_offset_add(shmem_get_offset_ctx(dir), dentry);
3490 if (error)
3491 goto out_iput;
3492
3493 inode->i_size = len-1;
3494 if (len <= SHORT_SYMLINK_LEN) {
3495 inode->i_link = kmemdup(symname, len, GFP_KERNEL);
3496 if (!inode->i_link) {
3497 error = -ENOMEM;
3498 goto out_remove_offset;
3499 }
3500 inode->i_op = &shmem_short_symlink_operations;
3501 } else {
3502 inode_nohighmem(inode);
3503 error = shmem_get_folio(inode, 0, &folio, SGP_WRITE);
3504 if (error)
3505 goto out_remove_offset;
3506 inode->i_mapping->a_ops = &shmem_aops;
3507 inode->i_op = &shmem_symlink_inode_operations;
3508 memcpy(folio_address(folio), symname, len);
3509 folio_mark_uptodate(folio);
3510 folio_mark_dirty(folio);
3511 folio_unlock(folio);
3512 folio_put(folio);
3513 }
3514 dir->i_size += BOGO_DIRENT_SIZE;
3515 inode_set_mtime_to_ts(dir, inode_set_ctime_current(dir));
3516 inode_inc_iversion(dir);
3517 d_instantiate(dentry, inode);
3518 dget(dentry);
3519 return 0;
3520
3521out_remove_offset:
3522 simple_offset_remove(shmem_get_offset_ctx(dir), dentry);
3523out_iput:
3524 iput(inode);
3525 return error;
3526}
3527
3528static void shmem_put_link(void *arg)
3529{
3530 folio_mark_accessed(arg);
3531 folio_put(arg);
3532}
3533
3534static const char *shmem_get_link(struct dentry *dentry, struct inode *inode,
3535 struct delayed_call *done)
3536{
3537 struct folio *folio = NULL;
3538 int error;
3539
3540 if (!dentry) {
3541 folio = filemap_get_folio(inode->i_mapping, 0);
3542 if (IS_ERR(folio))
3543 return ERR_PTR(-ECHILD);
3544 if (PageHWPoison(folio_page(folio, 0)) ||
3545 !folio_test_uptodate(folio)) {
3546 folio_put(folio);
3547 return ERR_PTR(-ECHILD);
3548 }
3549 } else {
3550 error = shmem_get_folio(inode, 0, &folio, SGP_READ);
3551 if (error)
3552 return ERR_PTR(error);
3553 if (!folio)
3554 return ERR_PTR(-ECHILD);
3555 if (PageHWPoison(folio_page(folio, 0))) {
3556 folio_unlock(folio);
3557 folio_put(folio);
3558 return ERR_PTR(-ECHILD);
3559 }
3560 folio_unlock(folio);
3561 }
3562 set_delayed_call(done, shmem_put_link, folio);
3563 return folio_address(folio);
3564}
3565
3566#ifdef CONFIG_TMPFS_XATTR
3567
3568static int shmem_fileattr_get(struct dentry *dentry, struct fileattr *fa)
3569{
3570 struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
3571
3572 fileattr_fill_flags(fa, info->fsflags & SHMEM_FL_USER_VISIBLE);
3573
3574 return 0;
3575}
3576
3577static int shmem_fileattr_set(struct mnt_idmap *idmap,
3578 struct dentry *dentry, struct fileattr *fa)
3579{
3580 struct inode *inode = d_inode(dentry);
3581 struct shmem_inode_info *info = SHMEM_I(inode);
3582
3583 if (fileattr_has_fsx(fa))
3584 return -EOPNOTSUPP;
3585 if (fa->flags & ~SHMEM_FL_USER_MODIFIABLE)
3586 return -EOPNOTSUPP;
3587
3588 info->fsflags = (info->fsflags & ~SHMEM_FL_USER_MODIFIABLE) |
3589 (fa->flags & SHMEM_FL_USER_MODIFIABLE);
3590
3591 shmem_set_inode_flags(inode, info->fsflags);
3592 inode_set_ctime_current(inode);
3593 inode_inc_iversion(inode);
3594 return 0;
3595}
3596
3597/*
3598 * Superblocks without xattr inode operations may get some security.* xattr
3599 * support from the LSM "for free". As soon as we have any other xattrs
3600 * like ACLs, we also need to implement the security.* handlers at
3601 * filesystem level, though.
3602 */
3603
3604/*
3605 * Callback for security_inode_init_security() for acquiring xattrs.
3606 */
3607static int shmem_initxattrs(struct inode *inode,
3608 const struct xattr *xattr_array, void *fs_info)
3609{
3610 struct shmem_inode_info *info = SHMEM_I(inode);
3611 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
3612 const struct xattr *xattr;
3613 struct simple_xattr *new_xattr;
3614 size_t ispace = 0;
3615 size_t len;
3616
3617 if (sbinfo->max_inodes) {
3618 for (xattr = xattr_array; xattr->name != NULL; xattr++) {
3619 ispace += simple_xattr_space(xattr->name,
3620 xattr->value_len + XATTR_SECURITY_PREFIX_LEN);
3621 }
3622 if (ispace) {
3623 raw_spin_lock(&sbinfo->stat_lock);
3624 if (sbinfo->free_ispace < ispace)
3625 ispace = 0;
3626 else
3627 sbinfo->free_ispace -= ispace;
3628 raw_spin_unlock(&sbinfo->stat_lock);
3629 if (!ispace)
3630 return -ENOSPC;
3631 }
3632 }
3633
3634 for (xattr = xattr_array; xattr->name != NULL; xattr++) {
3635 new_xattr = simple_xattr_alloc(xattr->value, xattr->value_len);
3636 if (!new_xattr)
3637 break;
3638
3639 len = strlen(xattr->name) + 1;
3640 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
3641 GFP_KERNEL_ACCOUNT);
3642 if (!new_xattr->name) {
3643 kvfree(new_xattr);
3644 break;
3645 }
3646
3647 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
3648 XATTR_SECURITY_PREFIX_LEN);
3649 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
3650 xattr->name, len);
3651
3652 simple_xattr_add(&info->xattrs, new_xattr);
3653 }
3654
3655 if (xattr->name != NULL) {
3656 if (ispace) {
3657 raw_spin_lock(&sbinfo->stat_lock);
3658 sbinfo->free_ispace += ispace;
3659 raw_spin_unlock(&sbinfo->stat_lock);
3660 }
3661 simple_xattrs_free(&info->xattrs, NULL);
3662 return -ENOMEM;
3663 }
3664
3665 return 0;
3666}
3667
3668static int shmem_xattr_handler_get(const struct xattr_handler *handler,
3669 struct dentry *unused, struct inode *inode,
3670 const char *name, void *buffer, size_t size)
3671{
3672 struct shmem_inode_info *info = SHMEM_I(inode);
3673
3674 name = xattr_full_name(handler, name);
3675 return simple_xattr_get(&info->xattrs, name, buffer, size);
3676}
3677
3678static int shmem_xattr_handler_set(const struct xattr_handler *handler,
3679 struct mnt_idmap *idmap,
3680 struct dentry *unused, struct inode *inode,
3681 const char *name, const void *value,
3682 size_t size, int flags)
3683{
3684 struct shmem_inode_info *info = SHMEM_I(inode);
3685 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
3686 struct simple_xattr *old_xattr;
3687 size_t ispace = 0;
3688
3689 name = xattr_full_name(handler, name);
3690 if (value && sbinfo->max_inodes) {
3691 ispace = simple_xattr_space(name, size);
3692 raw_spin_lock(&sbinfo->stat_lock);
3693 if (sbinfo->free_ispace < ispace)
3694 ispace = 0;
3695 else
3696 sbinfo->free_ispace -= ispace;
3697 raw_spin_unlock(&sbinfo->stat_lock);
3698 if (!ispace)
3699 return -ENOSPC;
3700 }
3701
3702 old_xattr = simple_xattr_set(&info->xattrs, name, value, size, flags);
3703 if (!IS_ERR(old_xattr)) {
3704 ispace = 0;
3705 if (old_xattr && sbinfo->max_inodes)
3706 ispace = simple_xattr_space(old_xattr->name,
3707 old_xattr->size);
3708 simple_xattr_free(old_xattr);
3709 old_xattr = NULL;
3710 inode_set_ctime_current(inode);
3711 inode_inc_iversion(inode);
3712 }
3713 if (ispace) {
3714 raw_spin_lock(&sbinfo->stat_lock);
3715 sbinfo->free_ispace += ispace;
3716 raw_spin_unlock(&sbinfo->stat_lock);
3717 }
3718 return PTR_ERR(old_xattr);
3719}
3720
3721static const struct xattr_handler shmem_security_xattr_handler = {
3722 .prefix = XATTR_SECURITY_PREFIX,
3723 .get = shmem_xattr_handler_get,
3724 .set = shmem_xattr_handler_set,
3725};
3726
3727static const struct xattr_handler shmem_trusted_xattr_handler = {
3728 .prefix = XATTR_TRUSTED_PREFIX,
3729 .get = shmem_xattr_handler_get,
3730 .set = shmem_xattr_handler_set,
3731};
3732
3733static const struct xattr_handler shmem_user_xattr_handler = {
3734 .prefix = XATTR_USER_PREFIX,
3735 .get = shmem_xattr_handler_get,
3736 .set = shmem_xattr_handler_set,
3737};
3738
3739static const struct xattr_handler * const shmem_xattr_handlers[] = {
3740 &shmem_security_xattr_handler,
3741 &shmem_trusted_xattr_handler,
3742 &shmem_user_xattr_handler,
3743 NULL
3744};
3745
3746static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
3747{
3748 struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
3749 return simple_xattr_list(d_inode(dentry), &info->xattrs, buffer, size);
3750}
3751#endif /* CONFIG_TMPFS_XATTR */
3752
3753static const struct inode_operations shmem_short_symlink_operations = {
3754 .getattr = shmem_getattr,
3755 .setattr = shmem_setattr,
3756 .get_link = simple_get_link,
3757#ifdef CONFIG_TMPFS_XATTR
3758 .listxattr = shmem_listxattr,
3759#endif
3760};
3761
3762static const struct inode_operations shmem_symlink_inode_operations = {
3763 .getattr = shmem_getattr,
3764 .setattr = shmem_setattr,
3765 .get_link = shmem_get_link,
3766#ifdef CONFIG_TMPFS_XATTR
3767 .listxattr = shmem_listxattr,
3768#endif
3769};
3770
3771static struct dentry *shmem_get_parent(struct dentry *child)
3772{
3773 return ERR_PTR(-ESTALE);
3774}
3775
3776static int shmem_match(struct inode *ino, void *vfh)
3777{
3778 __u32 *fh = vfh;
3779 __u64 inum = fh[2];
3780 inum = (inum << 32) | fh[1];
3781 return ino->i_ino == inum && fh[0] == ino->i_generation;
3782}
3783
3784/* Find any alias of inode, but prefer a hashed alias */
3785static struct dentry *shmem_find_alias(struct inode *inode)
3786{
3787 struct dentry *alias = d_find_alias(inode);
3788
3789 return alias ?: d_find_any_alias(inode);
3790}
3791
3792static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
3793 struct fid *fid, int fh_len, int fh_type)
3794{
3795 struct inode *inode;
3796 struct dentry *dentry = NULL;
3797 u64 inum;
3798
3799 if (fh_len < 3)
3800 return NULL;
3801
3802 inum = fid->raw[2];
3803 inum = (inum << 32) | fid->raw[1];
3804
3805 inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
3806 shmem_match, fid->raw);
3807 if (inode) {
3808 dentry = shmem_find_alias(inode);
3809 iput(inode);
3810 }
3811
3812 return dentry;
3813}
3814
3815static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
3816 struct inode *parent)
3817{
3818 if (*len < 3) {
3819 *len = 3;
3820 return FILEID_INVALID;
3821 }
3822
3823 if (inode_unhashed(inode)) {
3824 /* Unfortunately insert_inode_hash is not idempotent,
3825 * so as we hash inodes here rather than at creation
3826 * time, we need a lock to ensure we only try
3827 * to do it once
3828 */
3829 static DEFINE_SPINLOCK(lock);
3830 spin_lock(&lock);
3831 if (inode_unhashed(inode))
3832 __insert_inode_hash(inode,
3833 inode->i_ino + inode->i_generation);
3834 spin_unlock(&lock);
3835 }
3836
3837 fh[0] = inode->i_generation;
3838 fh[1] = inode->i_ino;
3839 fh[2] = ((__u64)inode->i_ino) >> 32;
3840
3841 *len = 3;
3842 return 1;
3843}
3844
3845static const struct export_operations shmem_export_ops = {
3846 .get_parent = shmem_get_parent,
3847 .encode_fh = shmem_encode_fh,
3848 .fh_to_dentry = shmem_fh_to_dentry,
3849};
3850
3851enum shmem_param {
3852 Opt_gid,
3853 Opt_huge,
3854 Opt_mode,
3855 Opt_mpol,
3856 Opt_nr_blocks,
3857 Opt_nr_inodes,
3858 Opt_size,
3859 Opt_uid,
3860 Opt_inode32,
3861 Opt_inode64,
3862 Opt_noswap,
3863 Opt_quota,
3864 Opt_usrquota,
3865 Opt_grpquota,
3866 Opt_usrquota_block_hardlimit,
3867 Opt_usrquota_inode_hardlimit,
3868 Opt_grpquota_block_hardlimit,
3869 Opt_grpquota_inode_hardlimit,
3870};
3871
3872static const struct constant_table shmem_param_enums_huge[] = {
3873 {"never", SHMEM_HUGE_NEVER },
3874 {"always", SHMEM_HUGE_ALWAYS },
3875 {"within_size", SHMEM_HUGE_WITHIN_SIZE },
3876 {"advise", SHMEM_HUGE_ADVISE },
3877 {}
3878};
3879
3880const struct fs_parameter_spec shmem_fs_parameters[] = {
3881 fsparam_u32 ("gid", Opt_gid),
3882 fsparam_enum ("huge", Opt_huge, shmem_param_enums_huge),
3883 fsparam_u32oct("mode", Opt_mode),
3884 fsparam_string("mpol", Opt_mpol),
3885 fsparam_string("nr_blocks", Opt_nr_blocks),
3886 fsparam_string("nr_inodes", Opt_nr_inodes),
3887 fsparam_string("size", Opt_size),
3888 fsparam_u32 ("uid", Opt_uid),
3889 fsparam_flag ("inode32", Opt_inode32),
3890 fsparam_flag ("inode64", Opt_inode64),
3891 fsparam_flag ("noswap", Opt_noswap),
3892#ifdef CONFIG_TMPFS_QUOTA
3893 fsparam_flag ("quota", Opt_quota),
3894 fsparam_flag ("usrquota", Opt_usrquota),
3895 fsparam_flag ("grpquota", Opt_grpquota),
3896 fsparam_string("usrquota_block_hardlimit", Opt_usrquota_block_hardlimit),
3897 fsparam_string("usrquota_inode_hardlimit", Opt_usrquota_inode_hardlimit),
3898 fsparam_string("grpquota_block_hardlimit", Opt_grpquota_block_hardlimit),
3899 fsparam_string("grpquota_inode_hardlimit", Opt_grpquota_inode_hardlimit),
3900#endif
3901 {}
3902};
3903
3904static int shmem_parse_one(struct fs_context *fc, struct fs_parameter *param)
3905{
3906 struct shmem_options *ctx = fc->fs_private;
3907 struct fs_parse_result result;
3908 unsigned long long size;
3909 char *rest;
3910 int opt;
3911 kuid_t kuid;
3912 kgid_t kgid;
3913
3914 opt = fs_parse(fc, shmem_fs_parameters, param, &result);
3915 if (opt < 0)
3916 return opt;
3917
3918 switch (opt) {
3919 case Opt_size:
3920 size = memparse(param->string, &rest);
3921 if (*rest == '%') {
3922 size <<= PAGE_SHIFT;
3923 size *= totalram_pages();
3924 do_div(size, 100);
3925 rest++;
3926 }
3927 if (*rest)
3928 goto bad_value;
3929 ctx->blocks = DIV_ROUND_UP(size, PAGE_SIZE);
3930 ctx->seen |= SHMEM_SEEN_BLOCKS;
3931 break;
3932 case Opt_nr_blocks:
3933 ctx->blocks = memparse(param->string, &rest);
3934 if (*rest || ctx->blocks > LONG_MAX)
3935 goto bad_value;
3936 ctx->seen |= SHMEM_SEEN_BLOCKS;
3937 break;
3938 case Opt_nr_inodes:
3939 ctx->inodes = memparse(param->string, &rest);
3940 if (*rest || ctx->inodes > ULONG_MAX / BOGO_INODE_SIZE)
3941 goto bad_value;
3942 ctx->seen |= SHMEM_SEEN_INODES;
3943 break;
3944 case Opt_mode:
3945 ctx->mode = result.uint_32 & 07777;
3946 break;
3947 case Opt_uid:
3948 kuid = make_kuid(current_user_ns(), result.uint_32);
3949 if (!uid_valid(kuid))
3950 goto bad_value;
3951
3952 /*
3953 * The requested uid must be representable in the
3954 * filesystem's idmapping.
3955 */
3956 if (!kuid_has_mapping(fc->user_ns, kuid))
3957 goto bad_value;
3958
3959 ctx->uid = kuid;
3960 break;
3961 case Opt_gid:
3962 kgid = make_kgid(current_user_ns(), result.uint_32);
3963 if (!gid_valid(kgid))
3964 goto bad_value;
3965
3966 /*
3967 * The requested gid must be representable in the
3968 * filesystem's idmapping.
3969 */
3970 if (!kgid_has_mapping(fc->user_ns, kgid))
3971 goto bad_value;
3972
3973 ctx->gid = kgid;
3974 break;
3975 case Opt_huge:
3976 ctx->huge = result.uint_32;
3977 if (ctx->huge != SHMEM_HUGE_NEVER &&
3978 !(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
3979 has_transparent_hugepage()))
3980 goto unsupported_parameter;
3981 ctx->seen |= SHMEM_SEEN_HUGE;
3982 break;
3983 case Opt_mpol:
3984 if (IS_ENABLED(CONFIG_NUMA)) {
3985 mpol_put(ctx->mpol);
3986 ctx->mpol = NULL;
3987 if (mpol_parse_str(param->string, &ctx->mpol))
3988 goto bad_value;
3989 break;
3990 }
3991 goto unsupported_parameter;
3992 case Opt_inode32:
3993 ctx->full_inums = false;
3994 ctx->seen |= SHMEM_SEEN_INUMS;
3995 break;
3996 case Opt_inode64:
3997 if (sizeof(ino_t) < 8) {
3998 return invalfc(fc,
3999 "Cannot use inode64 with <64bit inums in kernel\n");
4000 }
4001 ctx->full_inums = true;
4002 ctx->seen |= SHMEM_SEEN_INUMS;
4003 break;
4004 case Opt_noswap:
4005 if ((fc->user_ns != &init_user_ns) || !capable(CAP_SYS_ADMIN)) {
4006 return invalfc(fc,
4007 "Turning off swap in unprivileged tmpfs mounts unsupported");
4008 }
4009 ctx->noswap = true;
4010 ctx->seen |= SHMEM_SEEN_NOSWAP;
4011 break;
4012 case Opt_quota:
4013 if (fc->user_ns != &init_user_ns)
4014 return invalfc(fc, "Quotas in unprivileged tmpfs mounts are unsupported");
4015 ctx->seen |= SHMEM_SEEN_QUOTA;
4016 ctx->quota_types |= (QTYPE_MASK_USR | QTYPE_MASK_GRP);
4017 break;
4018 case Opt_usrquota:
4019 if (fc->user_ns != &init_user_ns)
4020 return invalfc(fc, "Quotas in unprivileged tmpfs mounts are unsupported");
4021 ctx->seen |= SHMEM_SEEN_QUOTA;
4022 ctx->quota_types |= QTYPE_MASK_USR;
4023 break;
4024 case Opt_grpquota:
4025 if (fc->user_ns != &init_user_ns)
4026 return invalfc(fc, "Quotas in unprivileged tmpfs mounts are unsupported");
4027 ctx->seen |= SHMEM_SEEN_QUOTA;
4028 ctx->quota_types |= QTYPE_MASK_GRP;
4029 break;
4030 case Opt_usrquota_block_hardlimit:
4031 size = memparse(param->string, &rest);
4032 if (*rest || !size)
4033 goto bad_value;
4034 if (size > SHMEM_QUOTA_MAX_SPC_LIMIT)
4035 return invalfc(fc,
4036 "User quota block hardlimit too large.");
4037 ctx->qlimits.usrquota_bhardlimit = size;
4038 break;
4039 case Opt_grpquota_block_hardlimit:
4040 size = memparse(param->string, &rest);
4041 if (*rest || !size)
4042 goto bad_value;
4043 if (size > SHMEM_QUOTA_MAX_SPC_LIMIT)
4044 return invalfc(fc,
4045 "Group quota block hardlimit too large.");
4046 ctx->qlimits.grpquota_bhardlimit = size;
4047 break;
4048 case Opt_usrquota_inode_hardlimit:
4049 size = memparse(param->string, &rest);
4050 if (*rest || !size)
4051 goto bad_value;
4052 if (size > SHMEM_QUOTA_MAX_INO_LIMIT)
4053 return invalfc(fc,
4054 "User quota inode hardlimit too large.");
4055 ctx->qlimits.usrquota_ihardlimit = size;
4056 break;
4057 case Opt_grpquota_inode_hardlimit:
4058 size = memparse(param->string, &rest);
4059 if (*rest || !size)
4060 goto bad_value;
4061 if (size > SHMEM_QUOTA_MAX_INO_LIMIT)
4062 return invalfc(fc,
4063 "Group quota inode hardlimit too large.");
4064 ctx->qlimits.grpquota_ihardlimit = size;
4065 break;
4066 }
4067 return 0;
4068
4069unsupported_parameter:
4070 return invalfc(fc, "Unsupported parameter '%s'", param->key);
4071bad_value:
4072 return invalfc(fc, "Bad value for '%s'", param->key);
4073}
4074
4075static int shmem_parse_options(struct fs_context *fc, void *data)
4076{
4077 char *options = data;
4078
4079 if (options) {
4080 int err = security_sb_eat_lsm_opts(options, &fc->security);
4081 if (err)
4082 return err;
4083 }
4084
4085 while (options != NULL) {
4086 char *this_char = options;
4087 for (;;) {
4088 /*
4089 * NUL-terminate this option: unfortunately,
4090 * mount options form a comma-separated list,
4091 * but mpol's nodelist may also contain commas.
4092 */
4093 options = strchr(options, ',');
4094 if (options == NULL)
4095 break;
4096 options++;
4097 if (!isdigit(*options)) {
4098 options[-1] = '\0';
4099 break;
4100 }
4101 }
4102 if (*this_char) {
4103 char *value = strchr(this_char, '=');
4104 size_t len = 0;
4105 int err;
4106
4107 if (value) {
4108 *value++ = '\0';
4109 len = strlen(value);
4110 }
4111 err = vfs_parse_fs_string(fc, this_char, value, len);
4112 if (err < 0)
4113 return err;
4114 }
4115 }
4116 return 0;
4117}
4118
4119/*
4120 * Reconfigure a shmem filesystem.
4121 */
4122static int shmem_reconfigure(struct fs_context *fc)
4123{
4124 struct shmem_options *ctx = fc->fs_private;
4125 struct shmem_sb_info *sbinfo = SHMEM_SB(fc->root->d_sb);
4126 unsigned long used_isp;
4127 struct mempolicy *mpol = NULL;
4128 const char *err;
4129
4130 raw_spin_lock(&sbinfo->stat_lock);
4131 used_isp = sbinfo->max_inodes * BOGO_INODE_SIZE - sbinfo->free_ispace;
4132
4133 if ((ctx->seen & SHMEM_SEEN_BLOCKS) && ctx->blocks) {
4134 if (!sbinfo->max_blocks) {
4135 err = "Cannot retroactively limit size";
4136 goto out;
4137 }
4138 if (percpu_counter_compare(&sbinfo->used_blocks,
4139 ctx->blocks) > 0) {
4140 err = "Too small a size for current use";
4141 goto out;
4142 }
4143 }
4144 if ((ctx->seen & SHMEM_SEEN_INODES) && ctx->inodes) {
4145 if (!sbinfo->max_inodes) {
4146 err = "Cannot retroactively limit inodes";
4147 goto out;
4148 }
4149 if (ctx->inodes * BOGO_INODE_SIZE < used_isp) {
4150 err = "Too few inodes for current use";
4151 goto out;
4152 }
4153 }
4154
4155 if ((ctx->seen & SHMEM_SEEN_INUMS) && !ctx->full_inums &&
4156 sbinfo->next_ino > UINT_MAX) {
4157 err = "Current inum too high to switch to 32-bit inums";
4158 goto out;
4159 }
4160 if ((ctx->seen & SHMEM_SEEN_NOSWAP) && ctx->noswap && !sbinfo->noswap) {
4161 err = "Cannot disable swap on remount";
4162 goto out;
4163 }
4164 if (!(ctx->seen & SHMEM_SEEN_NOSWAP) && !ctx->noswap && sbinfo->noswap) {
4165 err = "Cannot enable swap on remount if it was disabled on first mount";
4166 goto out;
4167 }
4168
4169 if (ctx->seen & SHMEM_SEEN_QUOTA &&
4170 !sb_any_quota_loaded(fc->root->d_sb)) {
4171 err = "Cannot enable quota on remount";
4172 goto out;
4173 }
4174
4175#ifdef CONFIG_TMPFS_QUOTA
4176#define CHANGED_LIMIT(name) \
4177 (ctx->qlimits.name## hardlimit && \
4178 (ctx->qlimits.name## hardlimit != sbinfo->qlimits.name## hardlimit))
4179
4180 if (CHANGED_LIMIT(usrquota_b) || CHANGED_LIMIT(usrquota_i) ||
4181 CHANGED_LIMIT(grpquota_b) || CHANGED_LIMIT(grpquota_i)) {
4182 err = "Cannot change global quota limit on remount";
4183 goto out;
4184 }
4185#endif /* CONFIG_TMPFS_QUOTA */
4186
4187 if (ctx->seen & SHMEM_SEEN_HUGE)
4188 sbinfo->huge = ctx->huge;
4189 if (ctx->seen & SHMEM_SEEN_INUMS)
4190 sbinfo->full_inums = ctx->full_inums;
4191 if (ctx->seen & SHMEM_SEEN_BLOCKS)
4192 sbinfo->max_blocks = ctx->blocks;
4193 if (ctx->seen & SHMEM_SEEN_INODES) {
4194 sbinfo->max_inodes = ctx->inodes;
4195 sbinfo->free_ispace = ctx->inodes * BOGO_INODE_SIZE - used_isp;
4196 }
4197
4198 /*
4199 * Preserve previous mempolicy unless mpol remount option was specified.
4200 */
4201 if (ctx->mpol) {
4202 mpol = sbinfo->mpol;
4203 sbinfo->mpol = ctx->mpol; /* transfers initial ref */
4204 ctx->mpol = NULL;
4205 }
4206
4207 if (ctx->noswap)
4208 sbinfo->noswap = true;
4209
4210 raw_spin_unlock(&sbinfo->stat_lock);
4211 mpol_put(mpol);
4212 return 0;
4213out:
4214 raw_spin_unlock(&sbinfo->stat_lock);
4215 return invalfc(fc, "%s", err);
4216}
4217
4218static int shmem_show_options(struct seq_file *seq, struct dentry *root)
4219{
4220 struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
4221 struct mempolicy *mpol;
4222
4223 if (sbinfo->max_blocks != shmem_default_max_blocks())
4224 seq_printf(seq, ",size=%luk", K(sbinfo->max_blocks));
4225 if (sbinfo->max_inodes != shmem_default_max_inodes())
4226 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
4227 if (sbinfo->mode != (0777 | S_ISVTX))
4228 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
4229 if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
4230 seq_printf(seq, ",uid=%u",
4231 from_kuid_munged(&init_user_ns, sbinfo->uid));
4232 if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
4233 seq_printf(seq, ",gid=%u",
4234 from_kgid_munged(&init_user_ns, sbinfo->gid));
4235
4236 /*
4237 * Showing inode{64,32} might be useful even if it's the system default,
4238 * since then people don't have to resort to checking both here and
4239 * /proc/config.gz to confirm 64-bit inums were successfully applied
4240 * (which may not even exist if IKCONFIG_PROC isn't enabled).
4241 *
4242 * We hide it when inode64 isn't the default and we are using 32-bit
4243 * inodes, since that probably just means the feature isn't even under
4244 * consideration.
4245 *
4246 * As such:
4247 *
4248 * +-----------------+-----------------+
4249 * | TMPFS_INODE64=y | TMPFS_INODE64=n |
4250 * +------------------+-----------------+-----------------+
4251 * | full_inums=true | show | show |
4252 * | full_inums=false | show | hide |
4253 * +------------------+-----------------+-----------------+
4254 *
4255 */
4256 if (IS_ENABLED(CONFIG_TMPFS_INODE64) || sbinfo->full_inums)
4257 seq_printf(seq, ",inode%d", (sbinfo->full_inums ? 64 : 32));
4258#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4259 /* Rightly or wrongly, show huge mount option unmasked by shmem_huge */
4260 if (sbinfo->huge)
4261 seq_printf(seq, ",huge=%s", shmem_format_huge(sbinfo->huge));
4262#endif
4263 mpol = shmem_get_sbmpol(sbinfo);
4264 shmem_show_mpol(seq, mpol);
4265 mpol_put(mpol);
4266 if (sbinfo->noswap)
4267 seq_printf(seq, ",noswap");
4268 return 0;
4269}
4270
4271#endif /* CONFIG_TMPFS */
4272
4273static void shmem_put_super(struct super_block *sb)
4274{
4275 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
4276
4277#ifdef CONFIG_TMPFS_QUOTA
4278 shmem_disable_quotas(sb);
4279#endif
4280 free_percpu(sbinfo->ino_batch);
4281 percpu_counter_destroy(&sbinfo->used_blocks);
4282 mpol_put(sbinfo->mpol);
4283 kfree(sbinfo);
4284 sb->s_fs_info = NULL;
4285}
4286
4287static int shmem_fill_super(struct super_block *sb, struct fs_context *fc)
4288{
4289 struct shmem_options *ctx = fc->fs_private;
4290 struct inode *inode;
4291 struct shmem_sb_info *sbinfo;
4292 int error = -ENOMEM;
4293
4294 /* Round up to L1_CACHE_BYTES to resist false sharing */
4295 sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
4296 L1_CACHE_BYTES), GFP_KERNEL);
4297 if (!sbinfo)
4298 return error;
4299
4300 sb->s_fs_info = sbinfo;
4301
4302#ifdef CONFIG_TMPFS
4303 /*
4304 * Per default we only allow half of the physical ram per
4305 * tmpfs instance, limiting inodes to one per page of lowmem;
4306 * but the internal instance is left unlimited.
4307 */
4308 if (!(sb->s_flags & SB_KERNMOUNT)) {
4309 if (!(ctx->seen & SHMEM_SEEN_BLOCKS))
4310 ctx->blocks = shmem_default_max_blocks();
4311 if (!(ctx->seen & SHMEM_SEEN_INODES))
4312 ctx->inodes = shmem_default_max_inodes();
4313 if (!(ctx->seen & SHMEM_SEEN_INUMS))
4314 ctx->full_inums = IS_ENABLED(CONFIG_TMPFS_INODE64);
4315 sbinfo->noswap = ctx->noswap;
4316 } else {
4317 sb->s_flags |= SB_NOUSER;
4318 }
4319 sb->s_export_op = &shmem_export_ops;
4320 sb->s_flags |= SB_NOSEC | SB_I_VERSION;
4321#else
4322 sb->s_flags |= SB_NOUSER;
4323#endif
4324 sbinfo->max_blocks = ctx->blocks;
4325 sbinfo->max_inodes = ctx->inodes;
4326 sbinfo->free_ispace = sbinfo->max_inodes * BOGO_INODE_SIZE;
4327 if (sb->s_flags & SB_KERNMOUNT) {
4328 sbinfo->ino_batch = alloc_percpu(ino_t);
4329 if (!sbinfo->ino_batch)
4330 goto failed;
4331 }
4332 sbinfo->uid = ctx->uid;
4333 sbinfo->gid = ctx->gid;
4334 sbinfo->full_inums = ctx->full_inums;
4335 sbinfo->mode = ctx->mode;
4336 sbinfo->huge = ctx->huge;
4337 sbinfo->mpol = ctx->mpol;
4338 ctx->mpol = NULL;
4339
4340 raw_spin_lock_init(&sbinfo->stat_lock);
4341 if (percpu_counter_init(&sbinfo->used_blocks, 0, GFP_KERNEL))
4342 goto failed;
4343 spin_lock_init(&sbinfo->shrinklist_lock);
4344 INIT_LIST_HEAD(&sbinfo->shrinklist);
4345
4346 sb->s_maxbytes = MAX_LFS_FILESIZE;
4347 sb->s_blocksize = PAGE_SIZE;
4348 sb->s_blocksize_bits = PAGE_SHIFT;
4349 sb->s_magic = TMPFS_MAGIC;
4350 sb->s_op = &shmem_ops;
4351 sb->s_time_gran = 1;
4352#ifdef CONFIG_TMPFS_XATTR
4353 sb->s_xattr = shmem_xattr_handlers;
4354#endif
4355#ifdef CONFIG_TMPFS_POSIX_ACL
4356 sb->s_flags |= SB_POSIXACL;
4357#endif
4358 uuid_gen(&sb->s_uuid);
4359
4360#ifdef CONFIG_TMPFS_QUOTA
4361 if (ctx->seen & SHMEM_SEEN_QUOTA) {
4362 sb->dq_op = &shmem_quota_operations;
4363 sb->s_qcop = &dquot_quotactl_sysfile_ops;
4364 sb->s_quota_types = QTYPE_MASK_USR | QTYPE_MASK_GRP;
4365
4366 /* Copy the default limits from ctx into sbinfo */
4367 memcpy(&sbinfo->qlimits, &ctx->qlimits,
4368 sizeof(struct shmem_quota_limits));
4369
4370 if (shmem_enable_quotas(sb, ctx->quota_types))
4371 goto failed;
4372 }
4373#endif /* CONFIG_TMPFS_QUOTA */
4374
4375 inode = shmem_get_inode(&nop_mnt_idmap, sb, NULL,
4376 S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
4377 if (IS_ERR(inode)) {
4378 error = PTR_ERR(inode);
4379 goto failed;
4380 }
4381 inode->i_uid = sbinfo->uid;
4382 inode->i_gid = sbinfo->gid;
4383 sb->s_root = d_make_root(inode);
4384 if (!sb->s_root)
4385 goto failed;
4386 return 0;
4387
4388failed:
4389 shmem_put_super(sb);
4390 return error;
4391}
4392
4393static int shmem_get_tree(struct fs_context *fc)
4394{
4395 return get_tree_nodev(fc, shmem_fill_super);
4396}
4397
4398static void shmem_free_fc(struct fs_context *fc)
4399{
4400 struct shmem_options *ctx = fc->fs_private;
4401
4402 if (ctx) {
4403 mpol_put(ctx->mpol);
4404 kfree(ctx);
4405 }
4406}
4407
4408static const struct fs_context_operations shmem_fs_context_ops = {
4409 .free = shmem_free_fc,
4410 .get_tree = shmem_get_tree,
4411#ifdef CONFIG_TMPFS
4412 .parse_monolithic = shmem_parse_options,
4413 .parse_param = shmem_parse_one,
4414 .reconfigure = shmem_reconfigure,
4415#endif
4416};
4417
4418static struct kmem_cache *shmem_inode_cachep __ro_after_init;
4419
4420static struct inode *shmem_alloc_inode(struct super_block *sb)
4421{
4422 struct shmem_inode_info *info;
4423 info = alloc_inode_sb(sb, shmem_inode_cachep, GFP_KERNEL);
4424 if (!info)
4425 return NULL;
4426 return &info->vfs_inode;
4427}
4428
4429static void shmem_free_in_core_inode(struct inode *inode)
4430{
4431 if (S_ISLNK(inode->i_mode))
4432 kfree(inode->i_link);
4433 kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
4434}
4435
4436static void shmem_destroy_inode(struct inode *inode)
4437{
4438 if (S_ISREG(inode->i_mode))
4439 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
4440 if (S_ISDIR(inode->i_mode))
4441 simple_offset_destroy(shmem_get_offset_ctx(inode));
4442}
4443
4444static void shmem_init_inode(void *foo)
4445{
4446 struct shmem_inode_info *info = foo;
4447 inode_init_once(&info->vfs_inode);
4448}
4449
4450static void __init shmem_init_inodecache(void)
4451{
4452 shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
4453 sizeof(struct shmem_inode_info),
4454 0, SLAB_PANIC|SLAB_ACCOUNT, shmem_init_inode);
4455}
4456
4457static void __init shmem_destroy_inodecache(void)
4458{
4459 kmem_cache_destroy(shmem_inode_cachep);
4460}
4461
4462/* Keep the page in page cache instead of truncating it */
4463static int shmem_error_remove_folio(struct address_space *mapping,
4464 struct folio *folio)
4465{
4466 return 0;
4467}
4468
4469const struct address_space_operations shmem_aops = {
4470 .writepage = shmem_writepage,
4471 .dirty_folio = noop_dirty_folio,
4472#ifdef CONFIG_TMPFS
4473 .write_begin = shmem_write_begin,
4474 .write_end = shmem_write_end,
4475#endif
4476#ifdef CONFIG_MIGRATION
4477 .migrate_folio = migrate_folio,
4478#endif
4479 .error_remove_folio = shmem_error_remove_folio,
4480};
4481EXPORT_SYMBOL(shmem_aops);
4482
4483static const struct file_operations shmem_file_operations = {
4484 .mmap = shmem_mmap,
4485 .open = shmem_file_open,
4486 .get_unmapped_area = shmem_get_unmapped_area,
4487#ifdef CONFIG_TMPFS
4488 .llseek = shmem_file_llseek,
4489 .read_iter = shmem_file_read_iter,
4490 .write_iter = shmem_file_write_iter,
4491 .fsync = noop_fsync,
4492 .splice_read = shmem_file_splice_read,
4493 .splice_write = iter_file_splice_write,
4494 .fallocate = shmem_fallocate,
4495#endif
4496};
4497
4498static const struct inode_operations shmem_inode_operations = {
4499 .getattr = shmem_getattr,
4500 .setattr = shmem_setattr,
4501#ifdef CONFIG_TMPFS_XATTR
4502 .listxattr = shmem_listxattr,
4503 .set_acl = simple_set_acl,
4504 .fileattr_get = shmem_fileattr_get,
4505 .fileattr_set = shmem_fileattr_set,
4506#endif
4507};
4508
4509static const struct inode_operations shmem_dir_inode_operations = {
4510#ifdef CONFIG_TMPFS
4511 .getattr = shmem_getattr,
4512 .create = shmem_create,
4513 .lookup = simple_lookup,
4514 .link = shmem_link,
4515 .unlink = shmem_unlink,
4516 .symlink = shmem_symlink,
4517 .mkdir = shmem_mkdir,
4518 .rmdir = shmem_rmdir,
4519 .mknod = shmem_mknod,
4520 .rename = shmem_rename2,
4521 .tmpfile = shmem_tmpfile,
4522 .get_offset_ctx = shmem_get_offset_ctx,
4523#endif
4524#ifdef CONFIG_TMPFS_XATTR
4525 .listxattr = shmem_listxattr,
4526 .fileattr_get = shmem_fileattr_get,
4527 .fileattr_set = shmem_fileattr_set,
4528#endif
4529#ifdef CONFIG_TMPFS_POSIX_ACL
4530 .setattr = shmem_setattr,
4531 .set_acl = simple_set_acl,
4532#endif
4533};
4534
4535static const struct inode_operations shmem_special_inode_operations = {
4536 .getattr = shmem_getattr,
4537#ifdef CONFIG_TMPFS_XATTR
4538 .listxattr = shmem_listxattr,
4539#endif
4540#ifdef CONFIG_TMPFS_POSIX_ACL
4541 .setattr = shmem_setattr,
4542 .set_acl = simple_set_acl,
4543#endif
4544};
4545
4546static const struct super_operations shmem_ops = {
4547 .alloc_inode = shmem_alloc_inode,
4548 .free_inode = shmem_free_in_core_inode,
4549 .destroy_inode = shmem_destroy_inode,
4550#ifdef CONFIG_TMPFS
4551 .statfs = shmem_statfs,
4552 .show_options = shmem_show_options,
4553#endif
4554#ifdef CONFIG_TMPFS_QUOTA
4555 .get_dquots = shmem_get_dquots,
4556#endif
4557 .evict_inode = shmem_evict_inode,
4558 .drop_inode = generic_delete_inode,
4559 .put_super = shmem_put_super,
4560#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4561 .nr_cached_objects = shmem_unused_huge_count,
4562 .free_cached_objects = shmem_unused_huge_scan,
4563#endif
4564};
4565
4566static const struct vm_operations_struct shmem_vm_ops = {
4567 .fault = shmem_fault,
4568 .map_pages = filemap_map_pages,
4569#ifdef CONFIG_NUMA
4570 .set_policy = shmem_set_policy,
4571 .get_policy = shmem_get_policy,
4572#endif
4573};
4574
4575static const struct vm_operations_struct shmem_anon_vm_ops = {
4576 .fault = shmem_fault,
4577 .map_pages = filemap_map_pages,
4578#ifdef CONFIG_NUMA
4579 .set_policy = shmem_set_policy,
4580 .get_policy = shmem_get_policy,
4581#endif
4582};
4583
4584int shmem_init_fs_context(struct fs_context *fc)
4585{
4586 struct shmem_options *ctx;
4587
4588 ctx = kzalloc(sizeof(struct shmem_options), GFP_KERNEL);
4589 if (!ctx)
4590 return -ENOMEM;
4591
4592 ctx->mode = 0777 | S_ISVTX;
4593 ctx->uid = current_fsuid();
4594 ctx->gid = current_fsgid();
4595
4596 fc->fs_private = ctx;
4597 fc->ops = &shmem_fs_context_ops;
4598 return 0;
4599}
4600
4601static struct file_system_type shmem_fs_type = {
4602 .owner = THIS_MODULE,
4603 .name = "tmpfs",
4604 .init_fs_context = shmem_init_fs_context,
4605#ifdef CONFIG_TMPFS
4606 .parameters = shmem_fs_parameters,
4607#endif
4608 .kill_sb = kill_litter_super,
4609 .fs_flags = FS_USERNS_MOUNT | FS_ALLOW_IDMAP,
4610};
4611
4612void __init shmem_init(void)
4613{
4614 int error;
4615
4616 shmem_init_inodecache();
4617
4618#ifdef CONFIG_TMPFS_QUOTA
4619 error = register_quota_format(&shmem_quota_format);
4620 if (error < 0) {
4621 pr_err("Could not register quota format\n");
4622 goto out3;
4623 }
4624#endif
4625
4626 error = register_filesystem(&shmem_fs_type);
4627 if (error) {
4628 pr_err("Could not register tmpfs\n");
4629 goto out2;
4630 }
4631
4632 shm_mnt = kern_mount(&shmem_fs_type);
4633 if (IS_ERR(shm_mnt)) {
4634 error = PTR_ERR(shm_mnt);
4635 pr_err("Could not kern_mount tmpfs\n");
4636 goto out1;
4637 }
4638
4639#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4640 if (has_transparent_hugepage() && shmem_huge > SHMEM_HUGE_DENY)
4641 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
4642 else
4643 shmem_huge = SHMEM_HUGE_NEVER; /* just in case it was patched */
4644#endif
4645 return;
4646
4647out1:
4648 unregister_filesystem(&shmem_fs_type);
4649out2:
4650#ifdef CONFIG_TMPFS_QUOTA
4651 unregister_quota_format(&shmem_quota_format);
4652out3:
4653#endif
4654 shmem_destroy_inodecache();
4655 shm_mnt = ERR_PTR(error);
4656}
4657
4658#if defined(CONFIG_TRANSPARENT_HUGEPAGE) && defined(CONFIG_SYSFS)
4659static ssize_t shmem_enabled_show(struct kobject *kobj,
4660 struct kobj_attribute *attr, char *buf)
4661{
4662 static const int values[] = {
4663 SHMEM_HUGE_ALWAYS,
4664 SHMEM_HUGE_WITHIN_SIZE,
4665 SHMEM_HUGE_ADVISE,
4666 SHMEM_HUGE_NEVER,
4667 SHMEM_HUGE_DENY,
4668 SHMEM_HUGE_FORCE,
4669 };
4670 int len = 0;
4671 int i;
4672
4673 for (i = 0; i < ARRAY_SIZE(values); i++) {
4674 len += sysfs_emit_at(buf, len,
4675 shmem_huge == values[i] ? "%s[%s]" : "%s%s",
4676 i ? " " : "", shmem_format_huge(values[i]));
4677 }
4678 len += sysfs_emit_at(buf, len, "\n");
4679
4680 return len;
4681}
4682
4683static ssize_t shmem_enabled_store(struct kobject *kobj,
4684 struct kobj_attribute *attr, const char *buf, size_t count)
4685{
4686 char tmp[16];
4687 int huge;
4688
4689 if (count + 1 > sizeof(tmp))
4690 return -EINVAL;
4691 memcpy(tmp, buf, count);
4692 tmp[count] = '\0';
4693 if (count && tmp[count - 1] == '\n')
4694 tmp[count - 1] = '\0';
4695
4696 huge = shmem_parse_huge(tmp);
4697 if (huge == -EINVAL)
4698 return -EINVAL;
4699 if (!has_transparent_hugepage() &&
4700 huge != SHMEM_HUGE_NEVER && huge != SHMEM_HUGE_DENY)
4701 return -EINVAL;
4702
4703 shmem_huge = huge;
4704 if (shmem_huge > SHMEM_HUGE_DENY)
4705 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
4706 return count;
4707}
4708
4709struct kobj_attribute shmem_enabled_attr = __ATTR_RW(shmem_enabled);
4710#endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_SYSFS */
4711
4712#else /* !CONFIG_SHMEM */
4713
4714/*
4715 * tiny-shmem: simple shmemfs and tmpfs using ramfs code
4716 *
4717 * This is intended for small system where the benefits of the full
4718 * shmem code (swap-backed and resource-limited) are outweighed by
4719 * their complexity. On systems without swap this code should be
4720 * effectively equivalent, but much lighter weight.
4721 */
4722
4723static struct file_system_type shmem_fs_type = {
4724 .name = "tmpfs",
4725 .init_fs_context = ramfs_init_fs_context,
4726 .parameters = ramfs_fs_parameters,
4727 .kill_sb = ramfs_kill_sb,
4728 .fs_flags = FS_USERNS_MOUNT,
4729};
4730
4731void __init shmem_init(void)
4732{
4733 BUG_ON(register_filesystem(&shmem_fs_type) != 0);
4734
4735 shm_mnt = kern_mount(&shmem_fs_type);
4736 BUG_ON(IS_ERR(shm_mnt));
4737}
4738
4739int shmem_unuse(unsigned int type)
4740{
4741 return 0;
4742}
4743
4744int shmem_lock(struct file *file, int lock, struct ucounts *ucounts)
4745{
4746 return 0;
4747}
4748
4749void shmem_unlock_mapping(struct address_space *mapping)
4750{
4751}
4752
4753#ifdef CONFIG_MMU
4754unsigned long shmem_get_unmapped_area(struct file *file,
4755 unsigned long addr, unsigned long len,
4756 unsigned long pgoff, unsigned long flags)
4757{
4758 return current->mm->get_unmapped_area(file, addr, len, pgoff, flags);
4759}
4760#endif
4761
4762void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
4763{
4764 truncate_inode_pages_range(inode->i_mapping, lstart, lend);
4765}
4766EXPORT_SYMBOL_GPL(shmem_truncate_range);
4767
4768#define shmem_vm_ops generic_file_vm_ops
4769#define shmem_anon_vm_ops generic_file_vm_ops
4770#define shmem_file_operations ramfs_file_operations
4771#define shmem_acct_size(flags, size) 0
4772#define shmem_unacct_size(flags, size) do {} while (0)
4773
4774static inline struct inode *shmem_get_inode(struct mnt_idmap *idmap,
4775 struct super_block *sb, struct inode *dir,
4776 umode_t mode, dev_t dev, unsigned long flags)
4777{
4778 struct inode *inode = ramfs_get_inode(sb, dir, mode, dev);
4779 return inode ? inode : ERR_PTR(-ENOSPC);
4780}
4781
4782#endif /* CONFIG_SHMEM */
4783
4784/* common code */
4785
4786static struct file *__shmem_file_setup(struct vfsmount *mnt, const char *name,
4787 loff_t size, unsigned long flags, unsigned int i_flags)
4788{
4789 struct inode *inode;
4790 struct file *res;
4791
4792 if (IS_ERR(mnt))
4793 return ERR_CAST(mnt);
4794
4795 if (size < 0 || size > MAX_LFS_FILESIZE)
4796 return ERR_PTR(-EINVAL);
4797
4798 if (shmem_acct_size(flags, size))
4799 return ERR_PTR(-ENOMEM);
4800
4801 if (is_idmapped_mnt(mnt))
4802 return ERR_PTR(-EINVAL);
4803
4804 inode = shmem_get_inode(&nop_mnt_idmap, mnt->mnt_sb, NULL,
4805 S_IFREG | S_IRWXUGO, 0, flags);
4806 if (IS_ERR(inode)) {
4807 shmem_unacct_size(flags, size);
4808 return ERR_CAST(inode);
4809 }
4810 inode->i_flags |= i_flags;
4811 inode->i_size = size;
4812 clear_nlink(inode); /* It is unlinked */
4813 res = ERR_PTR(ramfs_nommu_expand_for_mapping(inode, size));
4814 if (!IS_ERR(res))
4815 res = alloc_file_pseudo(inode, mnt, name, O_RDWR,
4816 &shmem_file_operations);
4817 if (IS_ERR(res))
4818 iput(inode);
4819 return res;
4820}
4821
4822/**
4823 * shmem_kernel_file_setup - get an unlinked file living in tmpfs which must be
4824 * kernel internal. There will be NO LSM permission checks against the
4825 * underlying inode. So users of this interface must do LSM checks at a
4826 * higher layer. The users are the big_key and shm implementations. LSM
4827 * checks are provided at the key or shm level rather than the inode.
4828 * @name: name for dentry (to be seen in /proc/<pid>/maps
4829 * @size: size to be set for the file
4830 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4831 */
4832struct file *shmem_kernel_file_setup(const char *name, loff_t size, unsigned long flags)
4833{
4834 return __shmem_file_setup(shm_mnt, name, size, flags, S_PRIVATE);
4835}
4836
4837/**
4838 * shmem_file_setup - get an unlinked file living in tmpfs
4839 * @name: name for dentry (to be seen in /proc/<pid>/maps
4840 * @size: size to be set for the file
4841 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4842 */
4843struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
4844{
4845 return __shmem_file_setup(shm_mnt, name, size, flags, 0);
4846}
4847EXPORT_SYMBOL_GPL(shmem_file_setup);
4848
4849/**
4850 * shmem_file_setup_with_mnt - get an unlinked file living in tmpfs
4851 * @mnt: the tmpfs mount where the file will be created
4852 * @name: name for dentry (to be seen in /proc/<pid>/maps
4853 * @size: size to be set for the file
4854 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4855 */
4856struct file *shmem_file_setup_with_mnt(struct vfsmount *mnt, const char *name,
4857 loff_t size, unsigned long flags)
4858{
4859 return __shmem_file_setup(mnt, name, size, flags, 0);
4860}
4861EXPORT_SYMBOL_GPL(shmem_file_setup_with_mnt);
4862
4863/**
4864 * shmem_zero_setup - setup a shared anonymous mapping
4865 * @vma: the vma to be mmapped is prepared by do_mmap
4866 */
4867int shmem_zero_setup(struct vm_area_struct *vma)
4868{
4869 struct file *file;
4870 loff_t size = vma->vm_end - vma->vm_start;
4871
4872 /*
4873 * Cloning a new file under mmap_lock leads to a lock ordering conflict
4874 * between XFS directory reading and selinux: since this file is only
4875 * accessible to the user through its mapping, use S_PRIVATE flag to
4876 * bypass file security, in the same way as shmem_kernel_file_setup().
4877 */
4878 file = shmem_kernel_file_setup("dev/zero", size, vma->vm_flags);
4879 if (IS_ERR(file))
4880 return PTR_ERR(file);
4881
4882 if (vma->vm_file)
4883 fput(vma->vm_file);
4884 vma->vm_file = file;
4885 vma->vm_ops = &shmem_anon_vm_ops;
4886
4887 return 0;
4888}
4889
4890/**
4891 * shmem_read_folio_gfp - read into page cache, using specified page allocation flags.
4892 * @mapping: the folio's address_space
4893 * @index: the folio index
4894 * @gfp: the page allocator flags to use if allocating
4895 *
4896 * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
4897 * with any new page allocations done using the specified allocation flags.
4898 * But read_cache_page_gfp() uses the ->read_folio() method: which does not
4899 * suit tmpfs, since it may have pages in swapcache, and needs to find those
4900 * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
4901 *
4902 * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
4903 * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
4904 */
4905struct folio *shmem_read_folio_gfp(struct address_space *mapping,
4906 pgoff_t index, gfp_t gfp)
4907{
4908#ifdef CONFIG_SHMEM
4909 struct inode *inode = mapping->host;
4910 struct folio *folio;
4911 int error;
4912
4913 BUG_ON(!shmem_mapping(mapping));
4914 error = shmem_get_folio_gfp(inode, index, &folio, SGP_CACHE,
4915 gfp, NULL, NULL);
4916 if (error)
4917 return ERR_PTR(error);
4918
4919 folio_unlock(folio);
4920 return folio;
4921#else
4922 /*
4923 * The tiny !SHMEM case uses ramfs without swap
4924 */
4925 return mapping_read_folio_gfp(mapping, index, gfp);
4926#endif
4927}
4928EXPORT_SYMBOL_GPL(shmem_read_folio_gfp);
4929
4930struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
4931 pgoff_t index, gfp_t gfp)
4932{
4933 struct folio *folio = shmem_read_folio_gfp(mapping, index, gfp);
4934 struct page *page;
4935
4936 if (IS_ERR(folio))
4937 return &folio->page;
4938
4939 page = folio_file_page(folio, index);
4940 if (PageHWPoison(page)) {
4941 folio_put(folio);
4942 return ERR_PTR(-EIO);
4943 }
4944
4945 return page;
4946}
4947EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);