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;
249const 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 __mod_lruvec_page_state(page, NR_SHMEM_THPS, nr);
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 index = 0;
846
847 pagevec_init(&pvec);
848 /*
849 * Minor point, but we might as well stop if someone else SHM_LOCKs it.
850 */
851 while (!mapping_unevictable(mapping)) {
852 if (!pagevec_lookup(&pvec, mapping, &index))
853 break;
854 check_move_unevictable_pages(&pvec);
855 pagevec_release(&pvec);
856 cond_resched();
857 }
858}
859
860/*
861 * Check whether a hole-punch or truncation needs to split a huge page,
862 * returning true if no split was required, or the split has been successful.
863 *
864 * Eviction (or truncation to 0 size) should never need to split a huge page;
865 * but in rare cases might do so, if shmem_undo_range() failed to trylock on
866 * head, and then succeeded to trylock on tail.
867 *
868 * A split can only succeed when there are no additional references on the
869 * huge page: so the split below relies upon find_get_entries() having stopped
870 * when it found a subpage of the huge page, without getting further references.
871 */
872static bool shmem_punch_compound(struct page *page, pgoff_t start, pgoff_t end)
873{
874 if (!PageTransCompound(page))
875 return true;
876
877 /* Just proceed to delete a huge page wholly within the range punched */
878 if (PageHead(page) &&
879 page->index >= start && page->index + HPAGE_PMD_NR <= end)
880 return true;
881
882 /* Try to split huge page, so we can truly punch the hole or truncate */
883 return split_huge_page(page) >= 0;
884}
885
886/*
887 * Remove range of pages and swap entries from page cache, and free them.
888 * If !unfalloc, truncate or punch hole; if unfalloc, undo failed fallocate.
889 */
890static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
891 bool unfalloc)
892{
893 struct address_space *mapping = inode->i_mapping;
894 struct shmem_inode_info *info = SHMEM_I(inode);
895 pgoff_t start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
896 pgoff_t end = (lend + 1) >> PAGE_SHIFT;
897 unsigned int partial_start = lstart & (PAGE_SIZE - 1);
898 unsigned int partial_end = (lend + 1) & (PAGE_SIZE - 1);
899 struct pagevec pvec;
900 pgoff_t indices[PAGEVEC_SIZE];
901 long nr_swaps_freed = 0;
902 pgoff_t index;
903 int i;
904
905 if (lend == -1)
906 end = -1; /* unsigned, so actually very big */
907
908 pagevec_init(&pvec);
909 index = start;
910 while (index < end && find_lock_entries(mapping, index, end - 1,
911 &pvec, indices)) {
912 for (i = 0; i < pagevec_count(&pvec); i++) {
913 struct page *page = pvec.pages[i];
914
915 index = indices[i];
916
917 if (xa_is_value(page)) {
918 if (unfalloc)
919 continue;
920 nr_swaps_freed += !shmem_free_swap(mapping,
921 index, page);
922 continue;
923 }
924 index += thp_nr_pages(page) - 1;
925
926 if (!unfalloc || !PageUptodate(page))
927 truncate_inode_page(mapping, page);
928 unlock_page(page);
929 }
930 pagevec_remove_exceptionals(&pvec);
931 pagevec_release(&pvec);
932 cond_resched();
933 index++;
934 }
935
936 if (partial_start) {
937 struct page *page = NULL;
938 shmem_getpage(inode, start - 1, &page, SGP_READ);
939 if (page) {
940 unsigned int top = PAGE_SIZE;
941 if (start > end) {
942 top = partial_end;
943 partial_end = 0;
944 }
945 zero_user_segment(page, partial_start, top);
946 set_page_dirty(page);
947 unlock_page(page);
948 put_page(page);
949 }
950 }
951 if (partial_end) {
952 struct page *page = NULL;
953 shmem_getpage(inode, end, &page, SGP_READ);
954 if (page) {
955 zero_user_segment(page, 0, partial_end);
956 set_page_dirty(page);
957 unlock_page(page);
958 put_page(page);
959 }
960 }
961 if (start >= end)
962 return;
963
964 index = start;
965 while (index < end) {
966 cond_resched();
967
968 if (!find_get_entries(mapping, index, end - 1, &pvec,
969 indices)) {
970 /* If all gone or hole-punch or unfalloc, we're done */
971 if (index == start || end != -1)
972 break;
973 /* But if truncating, restart to make sure all gone */
974 index = start;
975 continue;
976 }
977 for (i = 0; i < pagevec_count(&pvec); i++) {
978 struct page *page = pvec.pages[i];
979
980 index = indices[i];
981 if (xa_is_value(page)) {
982 if (unfalloc)
983 continue;
984 if (shmem_free_swap(mapping, index, page)) {
985 /* Swap was replaced by page: retry */
986 index--;
987 break;
988 }
989 nr_swaps_freed++;
990 continue;
991 }
992
993 lock_page(page);
994
995 if (!unfalloc || !PageUptodate(page)) {
996 if (page_mapping(page) != mapping) {
997 /* Page was replaced by swap: retry */
998 unlock_page(page);
999 index--;
1000 break;
1001 }
1002 VM_BUG_ON_PAGE(PageWriteback(page), page);
1003 if (shmem_punch_compound(page, start, end))
1004 truncate_inode_page(mapping, page);
1005 else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
1006 /* Wipe the page and don't get stuck */
1007 clear_highpage(page);
1008 flush_dcache_page(page);
1009 set_page_dirty(page);
1010 if (index <
1011 round_up(start, HPAGE_PMD_NR))
1012 start = index + 1;
1013 }
1014 }
1015 unlock_page(page);
1016 }
1017 pagevec_remove_exceptionals(&pvec);
1018 pagevec_release(&pvec);
1019 index++;
1020 }
1021
1022 spin_lock_irq(&info->lock);
1023 info->swapped -= nr_swaps_freed;
1024 shmem_recalc_inode(inode);
1025 spin_unlock_irq(&info->lock);
1026}
1027
1028void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
1029{
1030 shmem_undo_range(inode, lstart, lend, false);
1031 inode->i_ctime = inode->i_mtime = current_time(inode);
1032}
1033EXPORT_SYMBOL_GPL(shmem_truncate_range);
1034
1035static int shmem_getattr(struct user_namespace *mnt_userns,
1036 const struct path *path, struct kstat *stat,
1037 u32 request_mask, unsigned int query_flags)
1038{
1039 struct inode *inode = path->dentry->d_inode;
1040 struct shmem_inode_info *info = SHMEM_I(inode);
1041 struct shmem_sb_info *sb_info = SHMEM_SB(inode->i_sb);
1042
1043 if (info->alloced - info->swapped != inode->i_mapping->nrpages) {
1044 spin_lock_irq(&info->lock);
1045 shmem_recalc_inode(inode);
1046 spin_unlock_irq(&info->lock);
1047 }
1048 generic_fillattr(&init_user_ns, inode, stat);
1049
1050 if (is_huge_enabled(sb_info))
1051 stat->blksize = HPAGE_PMD_SIZE;
1052
1053 return 0;
1054}
1055
1056static int shmem_setattr(struct user_namespace *mnt_userns,
1057 struct dentry *dentry, struct iattr *attr)
1058{
1059 struct inode *inode = d_inode(dentry);
1060 struct shmem_inode_info *info = SHMEM_I(inode);
1061 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1062 int error;
1063
1064 error = setattr_prepare(&init_user_ns, dentry, attr);
1065 if (error)
1066 return error;
1067
1068 if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
1069 loff_t oldsize = inode->i_size;
1070 loff_t newsize = attr->ia_size;
1071
1072 /* protected by i_mutex */
1073 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
1074 (newsize > oldsize && (info->seals & F_SEAL_GROW)))
1075 return -EPERM;
1076
1077 if (newsize != oldsize) {
1078 error = shmem_reacct_size(SHMEM_I(inode)->flags,
1079 oldsize, newsize);
1080 if (error)
1081 return error;
1082 i_size_write(inode, newsize);
1083 inode->i_ctime = inode->i_mtime = current_time(inode);
1084 }
1085 if (newsize <= oldsize) {
1086 loff_t holebegin = round_up(newsize, PAGE_SIZE);
1087 if (oldsize > holebegin)
1088 unmap_mapping_range(inode->i_mapping,
1089 holebegin, 0, 1);
1090 if (info->alloced)
1091 shmem_truncate_range(inode,
1092 newsize, (loff_t)-1);
1093 /* unmap again to remove racily COWed private pages */
1094 if (oldsize > holebegin)
1095 unmap_mapping_range(inode->i_mapping,
1096 holebegin, 0, 1);
1097
1098 /*
1099 * Part of the huge page can be beyond i_size: subject
1100 * to shrink under memory pressure.
1101 */
1102 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
1103 spin_lock(&sbinfo->shrinklist_lock);
1104 /*
1105 * _careful to defend against unlocked access to
1106 * ->shrink_list in shmem_unused_huge_shrink()
1107 */
1108 if (list_empty_careful(&info->shrinklist)) {
1109 list_add_tail(&info->shrinklist,
1110 &sbinfo->shrinklist);
1111 sbinfo->shrinklist_len++;
1112 }
1113 spin_unlock(&sbinfo->shrinklist_lock);
1114 }
1115 }
1116 }
1117
1118 setattr_copy(&init_user_ns, inode, attr);
1119 if (attr->ia_valid & ATTR_MODE)
1120 error = posix_acl_chmod(&init_user_ns, inode, inode->i_mode);
1121 return error;
1122}
1123
1124static void shmem_evict_inode(struct inode *inode)
1125{
1126 struct shmem_inode_info *info = SHMEM_I(inode);
1127 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1128
1129 if (shmem_mapping(inode->i_mapping)) {
1130 shmem_unacct_size(info->flags, inode->i_size);
1131 inode->i_size = 0;
1132 shmem_truncate_range(inode, 0, (loff_t)-1);
1133 if (!list_empty(&info->shrinklist)) {
1134 spin_lock(&sbinfo->shrinklist_lock);
1135 if (!list_empty(&info->shrinklist)) {
1136 list_del_init(&info->shrinklist);
1137 sbinfo->shrinklist_len--;
1138 }
1139 spin_unlock(&sbinfo->shrinklist_lock);
1140 }
1141 while (!list_empty(&info->swaplist)) {
1142 /* Wait while shmem_unuse() is scanning this inode... */
1143 wait_var_event(&info->stop_eviction,
1144 !atomic_read(&info->stop_eviction));
1145 mutex_lock(&shmem_swaplist_mutex);
1146 /* ...but beware of the race if we peeked too early */
1147 if (!atomic_read(&info->stop_eviction))
1148 list_del_init(&info->swaplist);
1149 mutex_unlock(&shmem_swaplist_mutex);
1150 }
1151 }
1152
1153 simple_xattrs_free(&info->xattrs);
1154 WARN_ON(inode->i_blocks);
1155 shmem_free_inode(inode->i_sb);
1156 clear_inode(inode);
1157}
1158
1159extern struct swap_info_struct *swap_info[];
1160
1161static int shmem_find_swap_entries(struct address_space *mapping,
1162 pgoff_t start, unsigned int nr_entries,
1163 struct page **entries, pgoff_t *indices,
1164 unsigned int type, bool frontswap)
1165{
1166 XA_STATE(xas, &mapping->i_pages, start);
1167 struct page *page;
1168 swp_entry_t entry;
1169 unsigned int ret = 0;
1170
1171 if (!nr_entries)
1172 return 0;
1173
1174 rcu_read_lock();
1175 xas_for_each(&xas, page, ULONG_MAX) {
1176 if (xas_retry(&xas, page))
1177 continue;
1178
1179 if (!xa_is_value(page))
1180 continue;
1181
1182 entry = radix_to_swp_entry(page);
1183 if (swp_type(entry) != type)
1184 continue;
1185 if (frontswap &&
1186 !frontswap_test(swap_info[type], swp_offset(entry)))
1187 continue;
1188
1189 indices[ret] = xas.xa_index;
1190 entries[ret] = page;
1191
1192 if (need_resched()) {
1193 xas_pause(&xas);
1194 cond_resched_rcu();
1195 }
1196 if (++ret == nr_entries)
1197 break;
1198 }
1199 rcu_read_unlock();
1200
1201 return ret;
1202}
1203
1204/*
1205 * Move the swapped pages for an inode to page cache. Returns the count
1206 * of pages swapped in, or the error in case of failure.
1207 */
1208static int shmem_unuse_swap_entries(struct inode *inode, struct pagevec pvec,
1209 pgoff_t *indices)
1210{
1211 int i = 0;
1212 int ret = 0;
1213 int error = 0;
1214 struct address_space *mapping = inode->i_mapping;
1215
1216 for (i = 0; i < pvec.nr; i++) {
1217 struct page *page = pvec.pages[i];
1218
1219 if (!xa_is_value(page))
1220 continue;
1221 error = shmem_swapin_page(inode, indices[i],
1222 &page, SGP_CACHE,
1223 mapping_gfp_mask(mapping),
1224 NULL, NULL);
1225 if (error == 0) {
1226 unlock_page(page);
1227 put_page(page);
1228 ret++;
1229 }
1230 if (error == -ENOMEM)
1231 break;
1232 error = 0;
1233 }
1234 return error ? error : ret;
1235}
1236
1237/*
1238 * If swap found in inode, free it and move page from swapcache to filecache.
1239 */
1240static int shmem_unuse_inode(struct inode *inode, unsigned int type,
1241 bool frontswap, unsigned long *fs_pages_to_unuse)
1242{
1243 struct address_space *mapping = inode->i_mapping;
1244 pgoff_t start = 0;
1245 struct pagevec pvec;
1246 pgoff_t indices[PAGEVEC_SIZE];
1247 bool frontswap_partial = (frontswap && *fs_pages_to_unuse > 0);
1248 int ret = 0;
1249
1250 pagevec_init(&pvec);
1251 do {
1252 unsigned int nr_entries = PAGEVEC_SIZE;
1253
1254 if (frontswap_partial && *fs_pages_to_unuse < PAGEVEC_SIZE)
1255 nr_entries = *fs_pages_to_unuse;
1256
1257 pvec.nr = shmem_find_swap_entries(mapping, start, nr_entries,
1258 pvec.pages, indices,
1259 type, frontswap);
1260 if (pvec.nr == 0) {
1261 ret = 0;
1262 break;
1263 }
1264
1265 ret = shmem_unuse_swap_entries(inode, pvec, indices);
1266 if (ret < 0)
1267 break;
1268
1269 if (frontswap_partial) {
1270 *fs_pages_to_unuse -= ret;
1271 if (*fs_pages_to_unuse == 0) {
1272 ret = FRONTSWAP_PAGES_UNUSED;
1273 break;
1274 }
1275 }
1276
1277 start = indices[pvec.nr - 1];
1278 } while (true);
1279
1280 return ret;
1281}
1282
1283/*
1284 * Read all the shared memory data that resides in the swap
1285 * device 'type' back into memory, so the swap device can be
1286 * unused.
1287 */
1288int shmem_unuse(unsigned int type, bool frontswap,
1289 unsigned long *fs_pages_to_unuse)
1290{
1291 struct shmem_inode_info *info, *next;
1292 int error = 0;
1293
1294 if (list_empty(&shmem_swaplist))
1295 return 0;
1296
1297 mutex_lock(&shmem_swaplist_mutex);
1298 list_for_each_entry_safe(info, next, &shmem_swaplist, swaplist) {
1299 if (!info->swapped) {
1300 list_del_init(&info->swaplist);
1301 continue;
1302 }
1303 /*
1304 * Drop the swaplist mutex while searching the inode for swap;
1305 * but before doing so, make sure shmem_evict_inode() will not
1306 * remove placeholder inode from swaplist, nor let it be freed
1307 * (igrab() would protect from unlink, but not from unmount).
1308 */
1309 atomic_inc(&info->stop_eviction);
1310 mutex_unlock(&shmem_swaplist_mutex);
1311
1312 error = shmem_unuse_inode(&info->vfs_inode, type, frontswap,
1313 fs_pages_to_unuse);
1314 cond_resched();
1315
1316 mutex_lock(&shmem_swaplist_mutex);
1317 next = list_next_entry(info, swaplist);
1318 if (!info->swapped)
1319 list_del_init(&info->swaplist);
1320 if (atomic_dec_and_test(&info->stop_eviction))
1321 wake_up_var(&info->stop_eviction);
1322 if (error)
1323 break;
1324 }
1325 mutex_unlock(&shmem_swaplist_mutex);
1326
1327 return error;
1328}
1329
1330/*
1331 * Move the page from the page cache to the swap cache.
1332 */
1333static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1334{
1335 struct shmem_inode_info *info;
1336 struct address_space *mapping;
1337 struct inode *inode;
1338 swp_entry_t swap;
1339 pgoff_t index;
1340
1341 VM_BUG_ON_PAGE(PageCompound(page), page);
1342 BUG_ON(!PageLocked(page));
1343 mapping = page->mapping;
1344 index = page->index;
1345 inode = mapping->host;
1346 info = SHMEM_I(inode);
1347 if (info->flags & VM_LOCKED)
1348 goto redirty;
1349 if (!total_swap_pages)
1350 goto redirty;
1351
1352 /*
1353 * Our capabilities prevent regular writeback or sync from ever calling
1354 * shmem_writepage; but a stacking filesystem might use ->writepage of
1355 * its underlying filesystem, in which case tmpfs should write out to
1356 * swap only in response to memory pressure, and not for the writeback
1357 * threads or sync.
1358 */
1359 if (!wbc->for_reclaim) {
1360 WARN_ON_ONCE(1); /* Still happens? Tell us about it! */
1361 goto redirty;
1362 }
1363
1364 /*
1365 * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
1366 * value into swapfile.c, the only way we can correctly account for a
1367 * fallocated page arriving here is now to initialize it and write it.
1368 *
1369 * That's okay for a page already fallocated earlier, but if we have
1370 * not yet completed the fallocation, then (a) we want to keep track
1371 * of this page in case we have to undo it, and (b) it may not be a
1372 * good idea to continue anyway, once we're pushing into swap. So
1373 * reactivate the page, and let shmem_fallocate() quit when too many.
1374 */
1375 if (!PageUptodate(page)) {
1376 if (inode->i_private) {
1377 struct shmem_falloc *shmem_falloc;
1378 spin_lock(&inode->i_lock);
1379 shmem_falloc = inode->i_private;
1380 if (shmem_falloc &&
1381 !shmem_falloc->waitq &&
1382 index >= shmem_falloc->start &&
1383 index < shmem_falloc->next)
1384 shmem_falloc->nr_unswapped++;
1385 else
1386 shmem_falloc = NULL;
1387 spin_unlock(&inode->i_lock);
1388 if (shmem_falloc)
1389 goto redirty;
1390 }
1391 clear_highpage(page);
1392 flush_dcache_page(page);
1393 SetPageUptodate(page);
1394 }
1395
1396 swap = get_swap_page(page);
1397 if (!swap.val)
1398 goto redirty;
1399
1400 /*
1401 * Add inode to shmem_unuse()'s list of swapped-out inodes,
1402 * if it's not already there. Do it now before the page is
1403 * moved to swap cache, when its pagelock no longer protects
1404 * the inode from eviction. But don't unlock the mutex until
1405 * we've incremented swapped, because shmem_unuse_inode() will
1406 * prune a !swapped inode from the swaplist under this mutex.
1407 */
1408 mutex_lock(&shmem_swaplist_mutex);
1409 if (list_empty(&info->swaplist))
1410 list_add(&info->swaplist, &shmem_swaplist);
1411
1412 if (add_to_swap_cache(page, swap,
1413 __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN,
1414 NULL) == 0) {
1415 spin_lock_irq(&info->lock);
1416 shmem_recalc_inode(inode);
1417 info->swapped++;
1418 spin_unlock_irq(&info->lock);
1419
1420 swap_shmem_alloc(swap);
1421 shmem_delete_from_page_cache(page, swp_to_radix_entry(swap));
1422
1423 mutex_unlock(&shmem_swaplist_mutex);
1424 BUG_ON(page_mapped(page));
1425 swap_writepage(page, wbc);
1426 return 0;
1427 }
1428
1429 mutex_unlock(&shmem_swaplist_mutex);
1430 put_swap_page(page, swap);
1431redirty:
1432 set_page_dirty(page);
1433 if (wbc->for_reclaim)
1434 return AOP_WRITEPAGE_ACTIVATE; /* Return with page locked */
1435 unlock_page(page);
1436 return 0;
1437}
1438
1439#if defined(CONFIG_NUMA) && defined(CONFIG_TMPFS)
1440static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1441{
1442 char buffer[64];
1443
1444 if (!mpol || mpol->mode == MPOL_DEFAULT)
1445 return; /* show nothing */
1446
1447 mpol_to_str(buffer, sizeof(buffer), mpol);
1448
1449 seq_printf(seq, ",mpol=%s", buffer);
1450}
1451
1452static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1453{
1454 struct mempolicy *mpol = NULL;
1455 if (sbinfo->mpol) {
1456 spin_lock(&sbinfo->stat_lock); /* prevent replace/use races */
1457 mpol = sbinfo->mpol;
1458 mpol_get(mpol);
1459 spin_unlock(&sbinfo->stat_lock);
1460 }
1461 return mpol;
1462}
1463#else /* !CONFIG_NUMA || !CONFIG_TMPFS */
1464static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1465{
1466}
1467static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1468{
1469 return NULL;
1470}
1471#endif /* CONFIG_NUMA && CONFIG_TMPFS */
1472#ifndef CONFIG_NUMA
1473#define vm_policy vm_private_data
1474#endif
1475
1476static void shmem_pseudo_vma_init(struct vm_area_struct *vma,
1477 struct shmem_inode_info *info, pgoff_t index)
1478{
1479 /* Create a pseudo vma that just contains the policy */
1480 vma_init(vma, NULL);
1481 /* Bias interleave by inode number to distribute better across nodes */
1482 vma->vm_pgoff = index + info->vfs_inode.i_ino;
1483 vma->vm_policy = mpol_shared_policy_lookup(&info->policy, index);
1484}
1485
1486static void shmem_pseudo_vma_destroy(struct vm_area_struct *vma)
1487{
1488 /* Drop reference taken by mpol_shared_policy_lookup() */
1489 mpol_cond_put(vma->vm_policy);
1490}
1491
1492static struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
1493 struct shmem_inode_info *info, pgoff_t index)
1494{
1495 struct vm_area_struct pvma;
1496 struct page *page;
1497 struct vm_fault vmf = {
1498 .vma = &pvma,
1499 };
1500
1501 shmem_pseudo_vma_init(&pvma, info, index);
1502 page = swap_cluster_readahead(swap, gfp, &vmf);
1503 shmem_pseudo_vma_destroy(&pvma);
1504
1505 return page;
1506}
1507
1508/*
1509 * Make sure huge_gfp is always more limited than limit_gfp.
1510 * Some of the flags set permissions, while others set limitations.
1511 */
1512static gfp_t limit_gfp_mask(gfp_t huge_gfp, gfp_t limit_gfp)
1513{
1514 gfp_t allowflags = __GFP_IO | __GFP_FS | __GFP_RECLAIM;
1515 gfp_t denyflags = __GFP_NOWARN | __GFP_NORETRY;
1516 gfp_t zoneflags = limit_gfp & GFP_ZONEMASK;
1517 gfp_t result = huge_gfp & ~(allowflags | GFP_ZONEMASK);
1518
1519 /* Allow allocations only from the originally specified zones. */
1520 result |= zoneflags;
1521
1522 /*
1523 * Minimize the result gfp by taking the union with the deny flags,
1524 * and the intersection of the allow flags.
1525 */
1526 result |= (limit_gfp & denyflags);
1527 result |= (huge_gfp & limit_gfp) & allowflags;
1528
1529 return result;
1530}
1531
1532static struct page *shmem_alloc_hugepage(gfp_t gfp,
1533 struct shmem_inode_info *info, pgoff_t index)
1534{
1535 struct vm_area_struct pvma;
1536 struct address_space *mapping = info->vfs_inode.i_mapping;
1537 pgoff_t hindex;
1538 struct page *page;
1539
1540 hindex = round_down(index, HPAGE_PMD_NR);
1541 if (xa_find(&mapping->i_pages, &hindex, hindex + HPAGE_PMD_NR - 1,
1542 XA_PRESENT))
1543 return NULL;
1544
1545 shmem_pseudo_vma_init(&pvma, info, hindex);
1546 page = alloc_pages_vma(gfp, HPAGE_PMD_ORDER, &pvma, 0, numa_node_id(),
1547 true);
1548 shmem_pseudo_vma_destroy(&pvma);
1549 if (page)
1550 prep_transhuge_page(page);
1551 else
1552 count_vm_event(THP_FILE_FALLBACK);
1553 return page;
1554}
1555
1556static struct page *shmem_alloc_page(gfp_t gfp,
1557 struct shmem_inode_info *info, pgoff_t index)
1558{
1559 struct vm_area_struct pvma;
1560 struct page *page;
1561
1562 shmem_pseudo_vma_init(&pvma, info, index);
1563 page = alloc_page_vma(gfp, &pvma, 0);
1564 shmem_pseudo_vma_destroy(&pvma);
1565
1566 return page;
1567}
1568
1569static struct page *shmem_alloc_and_acct_page(gfp_t gfp,
1570 struct inode *inode,
1571 pgoff_t index, bool huge)
1572{
1573 struct shmem_inode_info *info = SHMEM_I(inode);
1574 struct page *page;
1575 int nr;
1576 int err = -ENOSPC;
1577
1578 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1579 huge = false;
1580 nr = huge ? HPAGE_PMD_NR : 1;
1581
1582 if (!shmem_inode_acct_block(inode, nr))
1583 goto failed;
1584
1585 if (huge)
1586 page = shmem_alloc_hugepage(gfp, info, index);
1587 else
1588 page = shmem_alloc_page(gfp, info, index);
1589 if (page) {
1590 __SetPageLocked(page);
1591 __SetPageSwapBacked(page);
1592 return page;
1593 }
1594
1595 err = -ENOMEM;
1596 shmem_inode_unacct_blocks(inode, nr);
1597failed:
1598 return ERR_PTR(err);
1599}
1600
1601/*
1602 * When a page is moved from swapcache to shmem filecache (either by the
1603 * usual swapin of shmem_getpage_gfp(), or by the less common swapoff of
1604 * shmem_unuse_inode()), it may have been read in earlier from swap, in
1605 * ignorance of the mapping it belongs to. If that mapping has special
1606 * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
1607 * we may need to copy to a suitable page before moving to filecache.
1608 *
1609 * In a future release, this may well be extended to respect cpuset and
1610 * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
1611 * but for now it is a simple matter of zone.
1612 */
1613static bool shmem_should_replace_page(struct page *page, gfp_t gfp)
1614{
1615 return page_zonenum(page) > gfp_zone(gfp);
1616}
1617
1618static int shmem_replace_page(struct page **pagep, gfp_t gfp,
1619 struct shmem_inode_info *info, pgoff_t index)
1620{
1621 struct page *oldpage, *newpage;
1622 struct address_space *swap_mapping;
1623 swp_entry_t entry;
1624 pgoff_t swap_index;
1625 int error;
1626
1627 oldpage = *pagep;
1628 entry.val = page_private(oldpage);
1629 swap_index = swp_offset(entry);
1630 swap_mapping = page_mapping(oldpage);
1631
1632 /*
1633 * We have arrived here because our zones are constrained, so don't
1634 * limit chance of success by further cpuset and node constraints.
1635 */
1636 gfp &= ~GFP_CONSTRAINT_MASK;
1637 newpage = shmem_alloc_page(gfp, info, index);
1638 if (!newpage)
1639 return -ENOMEM;
1640
1641 get_page(newpage);
1642 copy_highpage(newpage, oldpage);
1643 flush_dcache_page(newpage);
1644
1645 __SetPageLocked(newpage);
1646 __SetPageSwapBacked(newpage);
1647 SetPageUptodate(newpage);
1648 set_page_private(newpage, entry.val);
1649 SetPageSwapCache(newpage);
1650
1651 /*
1652 * Our caller will very soon move newpage out of swapcache, but it's
1653 * a nice clean interface for us to replace oldpage by newpage there.
1654 */
1655 xa_lock_irq(&swap_mapping->i_pages);
1656 error = shmem_replace_entry(swap_mapping, swap_index, oldpage, newpage);
1657 if (!error) {
1658 mem_cgroup_migrate(oldpage, newpage);
1659 __inc_lruvec_page_state(newpage, NR_FILE_PAGES);
1660 __dec_lruvec_page_state(oldpage, NR_FILE_PAGES);
1661 }
1662 xa_unlock_irq(&swap_mapping->i_pages);
1663
1664 if (unlikely(error)) {
1665 /*
1666 * Is this possible? I think not, now that our callers check
1667 * both PageSwapCache and page_private after getting page lock;
1668 * but be defensive. Reverse old to newpage for clear and free.
1669 */
1670 oldpage = newpage;
1671 } else {
1672 lru_cache_add(newpage);
1673 *pagep = newpage;
1674 }
1675
1676 ClearPageSwapCache(oldpage);
1677 set_page_private(oldpage, 0);
1678
1679 unlock_page(oldpage);
1680 put_page(oldpage);
1681 put_page(oldpage);
1682 return error;
1683}
1684
1685/*
1686 * Swap in the page pointed to by *pagep.
1687 * Caller has to make sure that *pagep contains a valid swapped page.
1688 * Returns 0 and the page in pagep if success. On failure, returns the
1689 * error code and NULL in *pagep.
1690 */
1691static int shmem_swapin_page(struct inode *inode, pgoff_t index,
1692 struct page **pagep, enum sgp_type sgp,
1693 gfp_t gfp, struct vm_area_struct *vma,
1694 vm_fault_t *fault_type)
1695{
1696 struct address_space *mapping = inode->i_mapping;
1697 struct shmem_inode_info *info = SHMEM_I(inode);
1698 struct mm_struct *charge_mm = vma ? vma->vm_mm : NULL;
1699 struct page *page;
1700 swp_entry_t swap;
1701 int error;
1702
1703 VM_BUG_ON(!*pagep || !xa_is_value(*pagep));
1704 swap = radix_to_swp_entry(*pagep);
1705 *pagep = NULL;
1706
1707 /* Look it up and read it in.. */
1708 page = lookup_swap_cache(swap, NULL, 0);
1709 if (!page) {
1710 /* Or update major stats only when swapin succeeds?? */
1711 if (fault_type) {
1712 *fault_type |= VM_FAULT_MAJOR;
1713 count_vm_event(PGMAJFAULT);
1714 count_memcg_event_mm(charge_mm, PGMAJFAULT);
1715 }
1716 /* Here we actually start the io */
1717 page = shmem_swapin(swap, gfp, info, index);
1718 if (!page) {
1719 error = -ENOMEM;
1720 goto failed;
1721 }
1722 }
1723
1724 /* We have to do this with page locked to prevent races */
1725 lock_page(page);
1726 if (!PageSwapCache(page) || page_private(page) != swap.val ||
1727 !shmem_confirm_swap(mapping, index, swap)) {
1728 error = -EEXIST;
1729 goto unlock;
1730 }
1731 if (!PageUptodate(page)) {
1732 error = -EIO;
1733 goto failed;
1734 }
1735 wait_on_page_writeback(page);
1736
1737 /*
1738 * Some architectures may have to restore extra metadata to the
1739 * physical page after reading from swap.
1740 */
1741 arch_swap_restore(swap, page);
1742
1743 if (shmem_should_replace_page(page, gfp)) {
1744 error = shmem_replace_page(&page, gfp, info, index);
1745 if (error)
1746 goto failed;
1747 }
1748
1749 error = shmem_add_to_page_cache(page, mapping, index,
1750 swp_to_radix_entry(swap), gfp,
1751 charge_mm);
1752 if (error)
1753 goto failed;
1754
1755 spin_lock_irq(&info->lock);
1756 info->swapped--;
1757 shmem_recalc_inode(inode);
1758 spin_unlock_irq(&info->lock);
1759
1760 if (sgp == SGP_WRITE)
1761 mark_page_accessed(page);
1762
1763 delete_from_swap_cache(page);
1764 set_page_dirty(page);
1765 swap_free(swap);
1766
1767 *pagep = page;
1768 return 0;
1769failed:
1770 if (!shmem_confirm_swap(mapping, index, swap))
1771 error = -EEXIST;
1772unlock:
1773 if (page) {
1774 unlock_page(page);
1775 put_page(page);
1776 }
1777
1778 return error;
1779}
1780
1781/*
1782 * shmem_getpage_gfp - find page in cache, or get from swap, or allocate
1783 *
1784 * If we allocate a new one we do not mark it dirty. That's up to the
1785 * vm. If we swap it in we mark it dirty since we also free the swap
1786 * entry since a page cannot live in both the swap and page cache.
1787 *
1788 * vma, vmf, and fault_type are only supplied by shmem_fault:
1789 * otherwise they are NULL.
1790 */
1791static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
1792 struct page **pagep, enum sgp_type sgp, gfp_t gfp,
1793 struct vm_area_struct *vma, struct vm_fault *vmf,
1794 vm_fault_t *fault_type)
1795{
1796 struct address_space *mapping = inode->i_mapping;
1797 struct shmem_inode_info *info = SHMEM_I(inode);
1798 struct shmem_sb_info *sbinfo;
1799 struct mm_struct *charge_mm;
1800 struct page *page;
1801 enum sgp_type sgp_huge = sgp;
1802 pgoff_t hindex = index;
1803 gfp_t huge_gfp;
1804 int error;
1805 int once = 0;
1806 int alloced = 0;
1807
1808 if (index > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
1809 return -EFBIG;
1810 if (sgp == SGP_NOHUGE || sgp == SGP_HUGE)
1811 sgp = SGP_CACHE;
1812repeat:
1813 if (sgp <= SGP_CACHE &&
1814 ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1815 return -EINVAL;
1816 }
1817
1818 sbinfo = SHMEM_SB(inode->i_sb);
1819 charge_mm = vma ? vma->vm_mm : NULL;
1820
1821 page = pagecache_get_page(mapping, index,
1822 FGP_ENTRY | FGP_HEAD | FGP_LOCK, 0);
1823
1824 if (page && vma && userfaultfd_minor(vma)) {
1825 if (!xa_is_value(page)) {
1826 unlock_page(page);
1827 put_page(page);
1828 }
1829 *fault_type = handle_userfault(vmf, VM_UFFD_MINOR);
1830 return 0;
1831 }
1832
1833 if (xa_is_value(page)) {
1834 error = shmem_swapin_page(inode, index, &page,
1835 sgp, gfp, vma, fault_type);
1836 if (error == -EEXIST)
1837 goto repeat;
1838
1839 *pagep = page;
1840 return error;
1841 }
1842
1843 if (page)
1844 hindex = page->index;
1845 if (page && sgp == SGP_WRITE)
1846 mark_page_accessed(page);
1847
1848 /* fallocated page? */
1849 if (page && !PageUptodate(page)) {
1850 if (sgp != SGP_READ)
1851 goto clear;
1852 unlock_page(page);
1853 put_page(page);
1854 page = NULL;
1855 hindex = index;
1856 }
1857 if (page || sgp == SGP_READ)
1858 goto out;
1859
1860 /*
1861 * Fast cache lookup did not find it:
1862 * bring it back from swap or allocate.
1863 */
1864
1865 if (vma && userfaultfd_missing(vma)) {
1866 *fault_type = handle_userfault(vmf, VM_UFFD_MISSING);
1867 return 0;
1868 }
1869
1870 /* shmem_symlink() */
1871 if (!shmem_mapping(mapping))
1872 goto alloc_nohuge;
1873 if (shmem_huge == SHMEM_HUGE_DENY || sgp_huge == SGP_NOHUGE)
1874 goto alloc_nohuge;
1875 if (shmem_huge == SHMEM_HUGE_FORCE)
1876 goto alloc_huge;
1877 switch (sbinfo->huge) {
1878 case SHMEM_HUGE_NEVER:
1879 goto alloc_nohuge;
1880 case SHMEM_HUGE_WITHIN_SIZE: {
1881 loff_t i_size;
1882 pgoff_t off;
1883
1884 off = round_up(index, HPAGE_PMD_NR);
1885 i_size = round_up(i_size_read(inode), PAGE_SIZE);
1886 if (i_size >= HPAGE_PMD_SIZE &&
1887 i_size >> PAGE_SHIFT >= off)
1888 goto alloc_huge;
1889
1890 fallthrough;
1891 }
1892 case SHMEM_HUGE_ADVISE:
1893 if (sgp_huge == SGP_HUGE)
1894 goto alloc_huge;
1895 /* TODO: implement fadvise() hints */
1896 goto alloc_nohuge;
1897 }
1898
1899alloc_huge:
1900 huge_gfp = vma_thp_gfp_mask(vma);
1901 huge_gfp = limit_gfp_mask(huge_gfp, gfp);
1902 page = shmem_alloc_and_acct_page(huge_gfp, inode, index, true);
1903 if (IS_ERR(page)) {
1904alloc_nohuge:
1905 page = shmem_alloc_and_acct_page(gfp, inode,
1906 index, false);
1907 }
1908 if (IS_ERR(page)) {
1909 int retry = 5;
1910
1911 error = PTR_ERR(page);
1912 page = NULL;
1913 if (error != -ENOSPC)
1914 goto unlock;
1915 /*
1916 * Try to reclaim some space by splitting a huge page
1917 * beyond i_size on the filesystem.
1918 */
1919 while (retry--) {
1920 int ret;
1921
1922 ret = shmem_unused_huge_shrink(sbinfo, NULL, 1);
1923 if (ret == SHRINK_STOP)
1924 break;
1925 if (ret)
1926 goto alloc_nohuge;
1927 }
1928 goto unlock;
1929 }
1930
1931 if (PageTransHuge(page))
1932 hindex = round_down(index, HPAGE_PMD_NR);
1933 else
1934 hindex = index;
1935
1936 if (sgp == SGP_WRITE)
1937 __SetPageReferenced(page);
1938
1939 error = shmem_add_to_page_cache(page, mapping, hindex,
1940 NULL, gfp & GFP_RECLAIM_MASK,
1941 charge_mm);
1942 if (error)
1943 goto unacct;
1944 lru_cache_add(page);
1945
1946 spin_lock_irq(&info->lock);
1947 info->alloced += compound_nr(page);
1948 inode->i_blocks += BLOCKS_PER_PAGE << compound_order(page);
1949 shmem_recalc_inode(inode);
1950 spin_unlock_irq(&info->lock);
1951 alloced = true;
1952
1953 if (PageTransHuge(page) &&
1954 DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE) <
1955 hindex + HPAGE_PMD_NR - 1) {
1956 /*
1957 * Part of the huge page is beyond i_size: subject
1958 * to shrink under memory pressure.
1959 */
1960 spin_lock(&sbinfo->shrinklist_lock);
1961 /*
1962 * _careful to defend against unlocked access to
1963 * ->shrink_list in shmem_unused_huge_shrink()
1964 */
1965 if (list_empty_careful(&info->shrinklist)) {
1966 list_add_tail(&info->shrinklist,
1967 &sbinfo->shrinklist);
1968 sbinfo->shrinklist_len++;
1969 }
1970 spin_unlock(&sbinfo->shrinklist_lock);
1971 }
1972
1973 /*
1974 * Let SGP_FALLOC use the SGP_WRITE optimization on a new page.
1975 */
1976 if (sgp == SGP_FALLOC)
1977 sgp = SGP_WRITE;
1978clear:
1979 /*
1980 * Let SGP_WRITE caller clear ends if write does not fill page;
1981 * but SGP_FALLOC on a page fallocated earlier must initialize
1982 * it now, lest undo on failure cancel our earlier guarantee.
1983 */
1984 if (sgp != SGP_WRITE && !PageUptodate(page)) {
1985 int i;
1986
1987 for (i = 0; i < compound_nr(page); i++) {
1988 clear_highpage(page + i);
1989 flush_dcache_page(page + i);
1990 }
1991 SetPageUptodate(page);
1992 }
1993
1994 /* Perhaps the file has been truncated since we checked */
1995 if (sgp <= SGP_CACHE &&
1996 ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1997 if (alloced) {
1998 ClearPageDirty(page);
1999 delete_from_page_cache(page);
2000 spin_lock_irq(&info->lock);
2001 shmem_recalc_inode(inode);
2002 spin_unlock_irq(&info->lock);
2003 }
2004 error = -EINVAL;
2005 goto unlock;
2006 }
2007out:
2008 *pagep = page + index - hindex;
2009 return 0;
2010
2011 /*
2012 * Error recovery.
2013 */
2014unacct:
2015 shmem_inode_unacct_blocks(inode, compound_nr(page));
2016
2017 if (PageTransHuge(page)) {
2018 unlock_page(page);
2019 put_page(page);
2020 goto alloc_nohuge;
2021 }
2022unlock:
2023 if (page) {
2024 unlock_page(page);
2025 put_page(page);
2026 }
2027 if (error == -ENOSPC && !once++) {
2028 spin_lock_irq(&info->lock);
2029 shmem_recalc_inode(inode);
2030 spin_unlock_irq(&info->lock);
2031 goto repeat;
2032 }
2033 if (error == -EEXIST)
2034 goto repeat;
2035 return error;
2036}
2037
2038/*
2039 * This is like autoremove_wake_function, but it removes the wait queue
2040 * entry unconditionally - even if something else had already woken the
2041 * target.
2042 */
2043static int synchronous_wake_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
2044{
2045 int ret = default_wake_function(wait, mode, sync, key);
2046 list_del_init(&wait->entry);
2047 return ret;
2048}
2049
2050static vm_fault_t shmem_fault(struct vm_fault *vmf)
2051{
2052 struct vm_area_struct *vma = vmf->vma;
2053 struct inode *inode = file_inode(vma->vm_file);
2054 gfp_t gfp = mapping_gfp_mask(inode->i_mapping);
2055 enum sgp_type sgp;
2056 int err;
2057 vm_fault_t ret = VM_FAULT_LOCKED;
2058
2059 /*
2060 * Trinity finds that probing a hole which tmpfs is punching can
2061 * prevent the hole-punch from ever completing: which in turn
2062 * locks writers out with its hold on i_mutex. So refrain from
2063 * faulting pages into the hole while it's being punched. Although
2064 * shmem_undo_range() does remove the additions, it may be unable to
2065 * keep up, as each new page needs its own unmap_mapping_range() call,
2066 * and the i_mmap tree grows ever slower to scan if new vmas are added.
2067 *
2068 * It does not matter if we sometimes reach this check just before the
2069 * hole-punch begins, so that one fault then races with the punch:
2070 * we just need to make racing faults a rare case.
2071 *
2072 * The implementation below would be much simpler if we just used a
2073 * standard mutex or completion: but we cannot take i_mutex in fault,
2074 * and bloating every shmem inode for this unlikely case would be sad.
2075 */
2076 if (unlikely(inode->i_private)) {
2077 struct shmem_falloc *shmem_falloc;
2078
2079 spin_lock(&inode->i_lock);
2080 shmem_falloc = inode->i_private;
2081 if (shmem_falloc &&
2082 shmem_falloc->waitq &&
2083 vmf->pgoff >= shmem_falloc->start &&
2084 vmf->pgoff < shmem_falloc->next) {
2085 struct file *fpin;
2086 wait_queue_head_t *shmem_falloc_waitq;
2087 DEFINE_WAIT_FUNC(shmem_fault_wait, synchronous_wake_function);
2088
2089 ret = VM_FAULT_NOPAGE;
2090 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2091 if (fpin)
2092 ret = VM_FAULT_RETRY;
2093
2094 shmem_falloc_waitq = shmem_falloc->waitq;
2095 prepare_to_wait(shmem_falloc_waitq, &shmem_fault_wait,
2096 TASK_UNINTERRUPTIBLE);
2097 spin_unlock(&inode->i_lock);
2098 schedule();
2099
2100 /*
2101 * shmem_falloc_waitq points into the shmem_fallocate()
2102 * stack of the hole-punching task: shmem_falloc_waitq
2103 * is usually invalid by the time we reach here, but
2104 * finish_wait() does not dereference it in that case;
2105 * though i_lock needed lest racing with wake_up_all().
2106 */
2107 spin_lock(&inode->i_lock);
2108 finish_wait(shmem_falloc_waitq, &shmem_fault_wait);
2109 spin_unlock(&inode->i_lock);
2110
2111 if (fpin)
2112 fput(fpin);
2113 return ret;
2114 }
2115 spin_unlock(&inode->i_lock);
2116 }
2117
2118 sgp = SGP_CACHE;
2119
2120 if ((vma->vm_flags & VM_NOHUGEPAGE) ||
2121 test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
2122 sgp = SGP_NOHUGE;
2123 else if (vma->vm_flags & VM_HUGEPAGE)
2124 sgp = SGP_HUGE;
2125
2126 err = shmem_getpage_gfp(inode, vmf->pgoff, &vmf->page, sgp,
2127 gfp, vma, vmf, &ret);
2128 if (err)
2129 return vmf_error(err);
2130 return ret;
2131}
2132
2133unsigned long shmem_get_unmapped_area(struct file *file,
2134 unsigned long uaddr, unsigned long len,
2135 unsigned long pgoff, unsigned long flags)
2136{
2137 unsigned long (*get_area)(struct file *,
2138 unsigned long, unsigned long, unsigned long, unsigned long);
2139 unsigned long addr;
2140 unsigned long offset;
2141 unsigned long inflated_len;
2142 unsigned long inflated_addr;
2143 unsigned long inflated_offset;
2144
2145 if (len > TASK_SIZE)
2146 return -ENOMEM;
2147
2148 get_area = current->mm->get_unmapped_area;
2149 addr = get_area(file, uaddr, len, pgoff, flags);
2150
2151 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
2152 return addr;
2153 if (IS_ERR_VALUE(addr))
2154 return addr;
2155 if (addr & ~PAGE_MASK)
2156 return addr;
2157 if (addr > TASK_SIZE - len)
2158 return addr;
2159
2160 if (shmem_huge == SHMEM_HUGE_DENY)
2161 return addr;
2162 if (len < HPAGE_PMD_SIZE)
2163 return addr;
2164 if (flags & MAP_FIXED)
2165 return addr;
2166 /*
2167 * Our priority is to support MAP_SHARED mapped hugely;
2168 * and support MAP_PRIVATE mapped hugely too, until it is COWed.
2169 * But if caller specified an address hint and we allocated area there
2170 * successfully, respect that as before.
2171 */
2172 if (uaddr == addr)
2173 return addr;
2174
2175 if (shmem_huge != SHMEM_HUGE_FORCE) {
2176 struct super_block *sb;
2177
2178 if (file) {
2179 VM_BUG_ON(file->f_op != &shmem_file_operations);
2180 sb = file_inode(file)->i_sb;
2181 } else {
2182 /*
2183 * Called directly from mm/mmap.c, or drivers/char/mem.c
2184 * for "/dev/zero", to create a shared anonymous object.
2185 */
2186 if (IS_ERR(shm_mnt))
2187 return addr;
2188 sb = shm_mnt->mnt_sb;
2189 }
2190 if (SHMEM_SB(sb)->huge == SHMEM_HUGE_NEVER)
2191 return addr;
2192 }
2193
2194 offset = (pgoff << PAGE_SHIFT) & (HPAGE_PMD_SIZE-1);
2195 if (offset && offset + len < 2 * HPAGE_PMD_SIZE)
2196 return addr;
2197 if ((addr & (HPAGE_PMD_SIZE-1)) == offset)
2198 return addr;
2199
2200 inflated_len = len + HPAGE_PMD_SIZE - PAGE_SIZE;
2201 if (inflated_len > TASK_SIZE)
2202 return addr;
2203 if (inflated_len < len)
2204 return addr;
2205
2206 inflated_addr = get_area(NULL, uaddr, inflated_len, 0, flags);
2207 if (IS_ERR_VALUE(inflated_addr))
2208 return addr;
2209 if (inflated_addr & ~PAGE_MASK)
2210 return addr;
2211
2212 inflated_offset = inflated_addr & (HPAGE_PMD_SIZE-1);
2213 inflated_addr += offset - inflated_offset;
2214 if (inflated_offset > offset)
2215 inflated_addr += HPAGE_PMD_SIZE;
2216
2217 if (inflated_addr > TASK_SIZE - len)
2218 return addr;
2219 return inflated_addr;
2220}
2221
2222#ifdef CONFIG_NUMA
2223static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
2224{
2225 struct inode *inode = file_inode(vma->vm_file);
2226 return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
2227}
2228
2229static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
2230 unsigned long addr)
2231{
2232 struct inode *inode = file_inode(vma->vm_file);
2233 pgoff_t index;
2234
2235 index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2236 return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
2237}
2238#endif
2239
2240int shmem_lock(struct file *file, int lock, struct ucounts *ucounts)
2241{
2242 struct inode *inode = file_inode(file);
2243 struct shmem_inode_info *info = SHMEM_I(inode);
2244 int retval = -ENOMEM;
2245
2246 /*
2247 * What serializes the accesses to info->flags?
2248 * ipc_lock_object() when called from shmctl_do_lock(),
2249 * no serialization needed when called from shm_destroy().
2250 */
2251 if (lock && !(info->flags & VM_LOCKED)) {
2252 if (!user_shm_lock(inode->i_size, ucounts))
2253 goto out_nomem;
2254 info->flags |= VM_LOCKED;
2255 mapping_set_unevictable(file->f_mapping);
2256 }
2257 if (!lock && (info->flags & VM_LOCKED) && ucounts) {
2258 user_shm_unlock(inode->i_size, ucounts);
2259 info->flags &= ~VM_LOCKED;
2260 mapping_clear_unevictable(file->f_mapping);
2261 }
2262 retval = 0;
2263
2264out_nomem:
2265 return retval;
2266}
2267
2268static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
2269{
2270 struct shmem_inode_info *info = SHMEM_I(file_inode(file));
2271 int ret;
2272
2273 ret = seal_check_future_write(info->seals, vma);
2274 if (ret)
2275 return ret;
2276
2277 /* arm64 - allow memory tagging on RAM-based files */
2278 vma->vm_flags |= VM_MTE_ALLOWED;
2279
2280 file_accessed(file);
2281 vma->vm_ops = &shmem_vm_ops;
2282 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
2283 ((vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK) <
2284 (vma->vm_end & HPAGE_PMD_MASK)) {
2285 khugepaged_enter(vma, vma->vm_flags);
2286 }
2287 return 0;
2288}
2289
2290static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
2291 umode_t mode, dev_t dev, unsigned long flags)
2292{
2293 struct inode *inode;
2294 struct shmem_inode_info *info;
2295 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2296 ino_t ino;
2297
2298 if (shmem_reserve_inode(sb, &ino))
2299 return NULL;
2300
2301 inode = new_inode(sb);
2302 if (inode) {
2303 inode->i_ino = ino;
2304 inode_init_owner(&init_user_ns, inode, dir, mode);
2305 inode->i_blocks = 0;
2306 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
2307 inode->i_generation = prandom_u32();
2308 info = SHMEM_I(inode);
2309 memset(info, 0, (char *)inode - (char *)info);
2310 spin_lock_init(&info->lock);
2311 atomic_set(&info->stop_eviction, 0);
2312 info->seals = F_SEAL_SEAL;
2313 info->flags = flags & VM_NORESERVE;
2314 INIT_LIST_HEAD(&info->shrinklist);
2315 INIT_LIST_HEAD(&info->swaplist);
2316 simple_xattrs_init(&info->xattrs);
2317 cache_no_acl(inode);
2318
2319 switch (mode & S_IFMT) {
2320 default:
2321 inode->i_op = &shmem_special_inode_operations;
2322 init_special_inode(inode, mode, dev);
2323 break;
2324 case S_IFREG:
2325 inode->i_mapping->a_ops = &shmem_aops;
2326 inode->i_op = &shmem_inode_operations;
2327 inode->i_fop = &shmem_file_operations;
2328 mpol_shared_policy_init(&info->policy,
2329 shmem_get_sbmpol(sbinfo));
2330 break;
2331 case S_IFDIR:
2332 inc_nlink(inode);
2333 /* Some things misbehave if size == 0 on a directory */
2334 inode->i_size = 2 * BOGO_DIRENT_SIZE;
2335 inode->i_op = &shmem_dir_inode_operations;
2336 inode->i_fop = &simple_dir_operations;
2337 break;
2338 case S_IFLNK:
2339 /*
2340 * Must not load anything in the rbtree,
2341 * mpol_free_shared_policy will not be called.
2342 */
2343 mpol_shared_policy_init(&info->policy, NULL);
2344 break;
2345 }
2346
2347 lockdep_annotate_inode_mutex_key(inode);
2348 } else
2349 shmem_free_inode(sb);
2350 return inode;
2351}
2352
2353#ifdef CONFIG_USERFAULTFD
2354int shmem_mfill_atomic_pte(struct mm_struct *dst_mm,
2355 pmd_t *dst_pmd,
2356 struct vm_area_struct *dst_vma,
2357 unsigned long dst_addr,
2358 unsigned long src_addr,
2359 bool zeropage,
2360 struct page **pagep)
2361{
2362 struct inode *inode = file_inode(dst_vma->vm_file);
2363 struct shmem_inode_info *info = SHMEM_I(inode);
2364 struct address_space *mapping = inode->i_mapping;
2365 gfp_t gfp = mapping_gfp_mask(mapping);
2366 pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
2367 void *page_kaddr;
2368 struct page *page;
2369 int ret;
2370 pgoff_t max_off;
2371
2372 if (!shmem_inode_acct_block(inode, 1)) {
2373 /*
2374 * We may have got a page, returned -ENOENT triggering a retry,
2375 * and now we find ourselves with -ENOMEM. Release the page, to
2376 * avoid a BUG_ON in our caller.
2377 */
2378 if (unlikely(*pagep)) {
2379 put_page(*pagep);
2380 *pagep = NULL;
2381 }
2382 return -ENOMEM;
2383 }
2384
2385 if (!*pagep) {
2386 ret = -ENOMEM;
2387 page = shmem_alloc_page(gfp, info, pgoff);
2388 if (!page)
2389 goto out_unacct_blocks;
2390
2391 if (!zeropage) { /* COPY */
2392 page_kaddr = kmap_atomic(page);
2393 ret = copy_from_user(page_kaddr,
2394 (const void __user *)src_addr,
2395 PAGE_SIZE);
2396 kunmap_atomic(page_kaddr);
2397
2398 /* fallback to copy_from_user outside mmap_lock */
2399 if (unlikely(ret)) {
2400 *pagep = page;
2401 ret = -ENOENT;
2402 /* don't free the page */
2403 goto out_unacct_blocks;
2404 }
2405 } else { /* ZEROPAGE */
2406 clear_highpage(page);
2407 }
2408 } else {
2409 page = *pagep;
2410 *pagep = NULL;
2411 }
2412
2413 VM_BUG_ON(PageLocked(page));
2414 VM_BUG_ON(PageSwapBacked(page));
2415 __SetPageLocked(page);
2416 __SetPageSwapBacked(page);
2417 __SetPageUptodate(page);
2418
2419 ret = -EFAULT;
2420 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2421 if (unlikely(pgoff >= max_off))
2422 goto out_release;
2423
2424 ret = shmem_add_to_page_cache(page, mapping, pgoff, NULL,
2425 gfp & GFP_RECLAIM_MASK, dst_mm);
2426 if (ret)
2427 goto out_release;
2428
2429 ret = mfill_atomic_install_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
2430 page, true, false);
2431 if (ret)
2432 goto out_delete_from_cache;
2433
2434 spin_lock_irq(&info->lock);
2435 info->alloced++;
2436 inode->i_blocks += BLOCKS_PER_PAGE;
2437 shmem_recalc_inode(inode);
2438 spin_unlock_irq(&info->lock);
2439
2440 SetPageDirty(page);
2441 unlock_page(page);
2442 return 0;
2443out_delete_from_cache:
2444 delete_from_page_cache(page);
2445out_release:
2446 unlock_page(page);
2447 put_page(page);
2448out_unacct_blocks:
2449 shmem_inode_unacct_blocks(inode, 1);
2450 return ret;
2451}
2452#endif /* CONFIG_USERFAULTFD */
2453
2454#ifdef CONFIG_TMPFS
2455static const struct inode_operations shmem_symlink_inode_operations;
2456static const struct inode_operations shmem_short_symlink_operations;
2457
2458#ifdef CONFIG_TMPFS_XATTR
2459static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
2460#else
2461#define shmem_initxattrs NULL
2462#endif
2463
2464static int
2465shmem_write_begin(struct file *file, struct address_space *mapping,
2466 loff_t pos, unsigned len, unsigned flags,
2467 struct page **pagep, void **fsdata)
2468{
2469 struct inode *inode = mapping->host;
2470 struct shmem_inode_info *info = SHMEM_I(inode);
2471 pgoff_t index = pos >> PAGE_SHIFT;
2472
2473 /* i_mutex is held by caller */
2474 if (unlikely(info->seals & (F_SEAL_GROW |
2475 F_SEAL_WRITE | F_SEAL_FUTURE_WRITE))) {
2476 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE))
2477 return -EPERM;
2478 if ((info->seals & F_SEAL_GROW) && pos + len > inode->i_size)
2479 return -EPERM;
2480 }
2481
2482 return shmem_getpage(inode, index, pagep, SGP_WRITE);
2483}
2484
2485static int
2486shmem_write_end(struct file *file, struct address_space *mapping,
2487 loff_t pos, unsigned len, unsigned copied,
2488 struct page *page, void *fsdata)
2489{
2490 struct inode *inode = mapping->host;
2491
2492 if (pos + copied > inode->i_size)
2493 i_size_write(inode, pos + copied);
2494
2495 if (!PageUptodate(page)) {
2496 struct page *head = compound_head(page);
2497 if (PageTransCompound(page)) {
2498 int i;
2499
2500 for (i = 0; i < HPAGE_PMD_NR; i++) {
2501 if (head + i == page)
2502 continue;
2503 clear_highpage(head + i);
2504 flush_dcache_page(head + i);
2505 }
2506 }
2507 if (copied < PAGE_SIZE) {
2508 unsigned from = pos & (PAGE_SIZE - 1);
2509 zero_user_segments(page, 0, from,
2510 from + copied, PAGE_SIZE);
2511 }
2512 SetPageUptodate(head);
2513 }
2514 set_page_dirty(page);
2515 unlock_page(page);
2516 put_page(page);
2517
2518 return copied;
2519}
2520
2521static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
2522{
2523 struct file *file = iocb->ki_filp;
2524 struct inode *inode = file_inode(file);
2525 struct address_space *mapping = inode->i_mapping;
2526 pgoff_t index;
2527 unsigned long offset;
2528 enum sgp_type sgp = SGP_READ;
2529 int error = 0;
2530 ssize_t retval = 0;
2531 loff_t *ppos = &iocb->ki_pos;
2532
2533 /*
2534 * Might this read be for a stacking filesystem? Then when reading
2535 * holes of a sparse file, we actually need to allocate those pages,
2536 * and even mark them dirty, so it cannot exceed the max_blocks limit.
2537 */
2538 if (!iter_is_iovec(to))
2539 sgp = SGP_CACHE;
2540
2541 index = *ppos >> PAGE_SHIFT;
2542 offset = *ppos & ~PAGE_MASK;
2543
2544 for (;;) {
2545 struct page *page = NULL;
2546 pgoff_t end_index;
2547 unsigned long nr, ret;
2548 loff_t i_size = i_size_read(inode);
2549
2550 end_index = i_size >> PAGE_SHIFT;
2551 if (index > end_index)
2552 break;
2553 if (index == end_index) {
2554 nr = i_size & ~PAGE_MASK;
2555 if (nr <= offset)
2556 break;
2557 }
2558
2559 error = shmem_getpage(inode, index, &page, sgp);
2560 if (error) {
2561 if (error == -EINVAL)
2562 error = 0;
2563 break;
2564 }
2565 if (page) {
2566 if (sgp == SGP_CACHE)
2567 set_page_dirty(page);
2568 unlock_page(page);
2569 }
2570
2571 /*
2572 * We must evaluate after, since reads (unlike writes)
2573 * are called without i_mutex protection against truncate
2574 */
2575 nr = PAGE_SIZE;
2576 i_size = i_size_read(inode);
2577 end_index = i_size >> PAGE_SHIFT;
2578 if (index == end_index) {
2579 nr = i_size & ~PAGE_MASK;
2580 if (nr <= offset) {
2581 if (page)
2582 put_page(page);
2583 break;
2584 }
2585 }
2586 nr -= offset;
2587
2588 if (page) {
2589 /*
2590 * If users can be writing to this page using arbitrary
2591 * virtual addresses, take care about potential aliasing
2592 * before reading the page on the kernel side.
2593 */
2594 if (mapping_writably_mapped(mapping))
2595 flush_dcache_page(page);
2596 /*
2597 * Mark the page accessed if we read the beginning.
2598 */
2599 if (!offset)
2600 mark_page_accessed(page);
2601 } else {
2602 page = ZERO_PAGE(0);
2603 get_page(page);
2604 }
2605
2606 /*
2607 * Ok, we have the page, and it's up-to-date, so
2608 * now we can copy it to user space...
2609 */
2610 ret = copy_page_to_iter(page, offset, nr, to);
2611 retval += ret;
2612 offset += ret;
2613 index += offset >> PAGE_SHIFT;
2614 offset &= ~PAGE_MASK;
2615
2616 put_page(page);
2617 if (!iov_iter_count(to))
2618 break;
2619 if (ret < nr) {
2620 error = -EFAULT;
2621 break;
2622 }
2623 cond_resched();
2624 }
2625
2626 *ppos = ((loff_t) index << PAGE_SHIFT) + offset;
2627 file_accessed(file);
2628 return retval ? retval : error;
2629}
2630
2631static loff_t shmem_file_llseek(struct file *file, loff_t offset, int whence)
2632{
2633 struct address_space *mapping = file->f_mapping;
2634 struct inode *inode = mapping->host;
2635
2636 if (whence != SEEK_DATA && whence != SEEK_HOLE)
2637 return generic_file_llseek_size(file, offset, whence,
2638 MAX_LFS_FILESIZE, i_size_read(inode));
2639 if (offset < 0)
2640 return -ENXIO;
2641
2642 inode_lock(inode);
2643 /* We're holding i_mutex so we can access i_size directly */
2644 offset = mapping_seek_hole_data(mapping, offset, inode->i_size, whence);
2645 if (offset >= 0)
2646 offset = vfs_setpos(file, offset, MAX_LFS_FILESIZE);
2647 inode_unlock(inode);
2648 return offset;
2649}
2650
2651static long shmem_fallocate(struct file *file, int mode, loff_t offset,
2652 loff_t len)
2653{
2654 struct inode *inode = file_inode(file);
2655 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
2656 struct shmem_inode_info *info = SHMEM_I(inode);
2657 struct shmem_falloc shmem_falloc;
2658 pgoff_t start, index, end;
2659 int error;
2660
2661 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2662 return -EOPNOTSUPP;
2663
2664 inode_lock(inode);
2665
2666 if (mode & FALLOC_FL_PUNCH_HOLE) {
2667 struct address_space *mapping = file->f_mapping;
2668 loff_t unmap_start = round_up(offset, PAGE_SIZE);
2669 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
2670 DECLARE_WAIT_QUEUE_HEAD_ONSTACK(shmem_falloc_waitq);
2671
2672 /* protected by i_mutex */
2673 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
2674 error = -EPERM;
2675 goto out;
2676 }
2677
2678 shmem_falloc.waitq = &shmem_falloc_waitq;
2679 shmem_falloc.start = (u64)unmap_start >> PAGE_SHIFT;
2680 shmem_falloc.next = (unmap_end + 1) >> PAGE_SHIFT;
2681 spin_lock(&inode->i_lock);
2682 inode->i_private = &shmem_falloc;
2683 spin_unlock(&inode->i_lock);
2684
2685 if ((u64)unmap_end > (u64)unmap_start)
2686 unmap_mapping_range(mapping, unmap_start,
2687 1 + unmap_end - unmap_start, 0);
2688 shmem_truncate_range(inode, offset, offset + len - 1);
2689 /* No need to unmap again: hole-punching leaves COWed pages */
2690
2691 spin_lock(&inode->i_lock);
2692 inode->i_private = NULL;
2693 wake_up_all(&shmem_falloc_waitq);
2694 WARN_ON_ONCE(!list_empty(&shmem_falloc_waitq.head));
2695 spin_unlock(&inode->i_lock);
2696 error = 0;
2697 goto out;
2698 }
2699
2700 /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
2701 error = inode_newsize_ok(inode, offset + len);
2702 if (error)
2703 goto out;
2704
2705 if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
2706 error = -EPERM;
2707 goto out;
2708 }
2709
2710 start = offset >> PAGE_SHIFT;
2711 end = (offset + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
2712 /* Try to avoid a swapstorm if len is impossible to satisfy */
2713 if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
2714 error = -ENOSPC;
2715 goto out;
2716 }
2717
2718 shmem_falloc.waitq = NULL;
2719 shmem_falloc.start = start;
2720 shmem_falloc.next = start;
2721 shmem_falloc.nr_falloced = 0;
2722 shmem_falloc.nr_unswapped = 0;
2723 spin_lock(&inode->i_lock);
2724 inode->i_private = &shmem_falloc;
2725 spin_unlock(&inode->i_lock);
2726
2727 for (index = start; index < end; index++) {
2728 struct page *page;
2729
2730 /*
2731 * Good, the fallocate(2) manpage permits EINTR: we may have
2732 * been interrupted because we are using up too much memory.
2733 */
2734 if (signal_pending(current))
2735 error = -EINTR;
2736 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
2737 error = -ENOMEM;
2738 else
2739 error = shmem_getpage(inode, index, &page, SGP_FALLOC);
2740 if (error) {
2741 /* Remove the !PageUptodate pages we added */
2742 if (index > start) {
2743 shmem_undo_range(inode,
2744 (loff_t)start << PAGE_SHIFT,
2745 ((loff_t)index << PAGE_SHIFT) - 1, true);
2746 }
2747 goto undone;
2748 }
2749
2750 /*
2751 * Inform shmem_writepage() how far we have reached.
2752 * No need for lock or barrier: we have the page lock.
2753 */
2754 shmem_falloc.next++;
2755 if (!PageUptodate(page))
2756 shmem_falloc.nr_falloced++;
2757
2758 /*
2759 * If !PageUptodate, leave it that way so that freeable pages
2760 * can be recognized if we need to rollback on error later.
2761 * But set_page_dirty so that memory pressure will swap rather
2762 * than free the pages we are allocating (and SGP_CACHE pages
2763 * might still be clean: we now need to mark those dirty too).
2764 */
2765 set_page_dirty(page);
2766 unlock_page(page);
2767 put_page(page);
2768 cond_resched();
2769 }
2770
2771 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
2772 i_size_write(inode, offset + len);
2773 inode->i_ctime = current_time(inode);
2774undone:
2775 spin_lock(&inode->i_lock);
2776 inode->i_private = NULL;
2777 spin_unlock(&inode->i_lock);
2778out:
2779 inode_unlock(inode);
2780 return error;
2781}
2782
2783static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
2784{
2785 struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
2786
2787 buf->f_type = TMPFS_MAGIC;
2788 buf->f_bsize = PAGE_SIZE;
2789 buf->f_namelen = NAME_MAX;
2790 if (sbinfo->max_blocks) {
2791 buf->f_blocks = sbinfo->max_blocks;
2792 buf->f_bavail =
2793 buf->f_bfree = sbinfo->max_blocks -
2794 percpu_counter_sum(&sbinfo->used_blocks);
2795 }
2796 if (sbinfo->max_inodes) {
2797 buf->f_files = sbinfo->max_inodes;
2798 buf->f_ffree = sbinfo->free_inodes;
2799 }
2800 /* else leave those fields 0 like simple_statfs */
2801
2802 buf->f_fsid = uuid_to_fsid(dentry->d_sb->s_uuid.b);
2803
2804 return 0;
2805}
2806
2807/*
2808 * File creation. Allocate an inode, and we're done..
2809 */
2810static int
2811shmem_mknod(struct user_namespace *mnt_userns, struct inode *dir,
2812 struct dentry *dentry, umode_t mode, dev_t dev)
2813{
2814 struct inode *inode;
2815 int error = -ENOSPC;
2816
2817 inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
2818 if (inode) {
2819 error = simple_acl_create(dir, inode);
2820 if (error)
2821 goto out_iput;
2822 error = security_inode_init_security(inode, dir,
2823 &dentry->d_name,
2824 shmem_initxattrs, NULL);
2825 if (error && error != -EOPNOTSUPP)
2826 goto out_iput;
2827
2828 error = 0;
2829 dir->i_size += BOGO_DIRENT_SIZE;
2830 dir->i_ctime = dir->i_mtime = current_time(dir);
2831 d_instantiate(dentry, inode);
2832 dget(dentry); /* Extra count - pin the dentry in core */
2833 }
2834 return error;
2835out_iput:
2836 iput(inode);
2837 return error;
2838}
2839
2840static int
2841shmem_tmpfile(struct user_namespace *mnt_userns, struct inode *dir,
2842 struct dentry *dentry, umode_t mode)
2843{
2844 struct inode *inode;
2845 int error = -ENOSPC;
2846
2847 inode = shmem_get_inode(dir->i_sb, dir, mode, 0, VM_NORESERVE);
2848 if (inode) {
2849 error = security_inode_init_security(inode, dir,
2850 NULL,
2851 shmem_initxattrs, NULL);
2852 if (error && error != -EOPNOTSUPP)
2853 goto out_iput;
2854 error = simple_acl_create(dir, inode);
2855 if (error)
2856 goto out_iput;
2857 d_tmpfile(dentry, inode);
2858 }
2859 return error;
2860out_iput:
2861 iput(inode);
2862 return error;
2863}
2864
2865static int shmem_mkdir(struct user_namespace *mnt_userns, struct inode *dir,
2866 struct dentry *dentry, umode_t mode)
2867{
2868 int error;
2869
2870 if ((error = shmem_mknod(&init_user_ns, dir, dentry,
2871 mode | S_IFDIR, 0)))
2872 return error;
2873 inc_nlink(dir);
2874 return 0;
2875}
2876
2877static int shmem_create(struct user_namespace *mnt_userns, struct inode *dir,
2878 struct dentry *dentry, umode_t mode, bool excl)
2879{
2880 return shmem_mknod(&init_user_ns, dir, dentry, mode | S_IFREG, 0);
2881}
2882
2883/*
2884 * Link a file..
2885 */
2886static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
2887{
2888 struct inode *inode = d_inode(old_dentry);
2889 int ret = 0;
2890
2891 /*
2892 * No ordinary (disk based) filesystem counts links as inodes;
2893 * but each new link needs a new dentry, pinning lowmem, and
2894 * tmpfs dentries cannot be pruned until they are unlinked.
2895 * But if an O_TMPFILE file is linked into the tmpfs, the
2896 * first link must skip that, to get the accounting right.
2897 */
2898 if (inode->i_nlink) {
2899 ret = shmem_reserve_inode(inode->i_sb, NULL);
2900 if (ret)
2901 goto out;
2902 }
2903
2904 dir->i_size += BOGO_DIRENT_SIZE;
2905 inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
2906 inc_nlink(inode);
2907 ihold(inode); /* New dentry reference */
2908 dget(dentry); /* Extra pinning count for the created dentry */
2909 d_instantiate(dentry, inode);
2910out:
2911 return ret;
2912}
2913
2914static int shmem_unlink(struct inode *dir, struct dentry *dentry)
2915{
2916 struct inode *inode = d_inode(dentry);
2917
2918 if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
2919 shmem_free_inode(inode->i_sb);
2920
2921 dir->i_size -= BOGO_DIRENT_SIZE;
2922 inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
2923 drop_nlink(inode);
2924 dput(dentry); /* Undo the count from "create" - this does all the work */
2925 return 0;
2926}
2927
2928static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
2929{
2930 if (!simple_empty(dentry))
2931 return -ENOTEMPTY;
2932
2933 drop_nlink(d_inode(dentry));
2934 drop_nlink(dir);
2935 return shmem_unlink(dir, dentry);
2936}
2937
2938static int shmem_exchange(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
2939{
2940 bool old_is_dir = d_is_dir(old_dentry);
2941 bool new_is_dir = d_is_dir(new_dentry);
2942
2943 if (old_dir != new_dir && old_is_dir != new_is_dir) {
2944 if (old_is_dir) {
2945 drop_nlink(old_dir);
2946 inc_nlink(new_dir);
2947 } else {
2948 drop_nlink(new_dir);
2949 inc_nlink(old_dir);
2950 }
2951 }
2952 old_dir->i_ctime = old_dir->i_mtime =
2953 new_dir->i_ctime = new_dir->i_mtime =
2954 d_inode(old_dentry)->i_ctime =
2955 d_inode(new_dentry)->i_ctime = current_time(old_dir);
2956
2957 return 0;
2958}
2959
2960static int shmem_whiteout(struct user_namespace *mnt_userns,
2961 struct inode *old_dir, struct dentry *old_dentry)
2962{
2963 struct dentry *whiteout;
2964 int error;
2965
2966 whiteout = d_alloc(old_dentry->d_parent, &old_dentry->d_name);
2967 if (!whiteout)
2968 return -ENOMEM;
2969
2970 error = shmem_mknod(&init_user_ns, old_dir, whiteout,
2971 S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
2972 dput(whiteout);
2973 if (error)
2974 return error;
2975
2976 /*
2977 * Cheat and hash the whiteout while the old dentry is still in
2978 * place, instead of playing games with FS_RENAME_DOES_D_MOVE.
2979 *
2980 * d_lookup() will consistently find one of them at this point,
2981 * not sure which one, but that isn't even important.
2982 */
2983 d_rehash(whiteout);
2984 return 0;
2985}
2986
2987/*
2988 * The VFS layer already does all the dentry stuff for rename,
2989 * we just have to decrement the usage count for the target if
2990 * it exists so that the VFS layer correctly free's it when it
2991 * gets overwritten.
2992 */
2993static int shmem_rename2(struct user_namespace *mnt_userns,
2994 struct inode *old_dir, struct dentry *old_dentry,
2995 struct inode *new_dir, struct dentry *new_dentry,
2996 unsigned int flags)
2997{
2998 struct inode *inode = d_inode(old_dentry);
2999 int they_are_dirs = S_ISDIR(inode->i_mode);
3000
3001 if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
3002 return -EINVAL;
3003
3004 if (flags & RENAME_EXCHANGE)
3005 return shmem_exchange(old_dir, old_dentry, new_dir, new_dentry);
3006
3007 if (!simple_empty(new_dentry))
3008 return -ENOTEMPTY;
3009
3010 if (flags & RENAME_WHITEOUT) {
3011 int error;
3012
3013 error = shmem_whiteout(&init_user_ns, old_dir, old_dentry);
3014 if (error)
3015 return error;
3016 }
3017
3018 if (d_really_is_positive(new_dentry)) {
3019 (void) shmem_unlink(new_dir, new_dentry);
3020 if (they_are_dirs) {
3021 drop_nlink(d_inode(new_dentry));
3022 drop_nlink(old_dir);
3023 }
3024 } else if (they_are_dirs) {
3025 drop_nlink(old_dir);
3026 inc_nlink(new_dir);
3027 }
3028
3029 old_dir->i_size -= BOGO_DIRENT_SIZE;
3030 new_dir->i_size += BOGO_DIRENT_SIZE;
3031 old_dir->i_ctime = old_dir->i_mtime =
3032 new_dir->i_ctime = new_dir->i_mtime =
3033 inode->i_ctime = current_time(old_dir);
3034 return 0;
3035}
3036
3037static int shmem_symlink(struct user_namespace *mnt_userns, struct inode *dir,
3038 struct dentry *dentry, const char *symname)
3039{
3040 int error;
3041 int len;
3042 struct inode *inode;
3043 struct page *page;
3044
3045 len = strlen(symname) + 1;
3046 if (len > PAGE_SIZE)
3047 return -ENAMETOOLONG;
3048
3049 inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK | 0777, 0,
3050 VM_NORESERVE);
3051 if (!inode)
3052 return -ENOSPC;
3053
3054 error = security_inode_init_security(inode, dir, &dentry->d_name,
3055 shmem_initxattrs, NULL);
3056 if (error && error != -EOPNOTSUPP) {
3057 iput(inode);
3058 return error;
3059 }
3060
3061 inode->i_size = len-1;
3062 if (len <= SHORT_SYMLINK_LEN) {
3063 inode->i_link = kmemdup(symname, len, GFP_KERNEL);
3064 if (!inode->i_link) {
3065 iput(inode);
3066 return -ENOMEM;
3067 }
3068 inode->i_op = &shmem_short_symlink_operations;
3069 } else {
3070 inode_nohighmem(inode);
3071 error = shmem_getpage(inode, 0, &page, SGP_WRITE);
3072 if (error) {
3073 iput(inode);
3074 return error;
3075 }
3076 inode->i_mapping->a_ops = &shmem_aops;
3077 inode->i_op = &shmem_symlink_inode_operations;
3078 memcpy(page_address(page), symname, len);
3079 SetPageUptodate(page);
3080 set_page_dirty(page);
3081 unlock_page(page);
3082 put_page(page);
3083 }
3084 dir->i_size += BOGO_DIRENT_SIZE;
3085 dir->i_ctime = dir->i_mtime = current_time(dir);
3086 d_instantiate(dentry, inode);
3087 dget(dentry);
3088 return 0;
3089}
3090
3091static void shmem_put_link(void *arg)
3092{
3093 mark_page_accessed(arg);
3094 put_page(arg);
3095}
3096
3097static const char *shmem_get_link(struct dentry *dentry,
3098 struct inode *inode,
3099 struct delayed_call *done)
3100{
3101 struct page *page = NULL;
3102 int error;
3103 if (!dentry) {
3104 page = find_get_page(inode->i_mapping, 0);
3105 if (!page)
3106 return ERR_PTR(-ECHILD);
3107 if (!PageUptodate(page)) {
3108 put_page(page);
3109 return ERR_PTR(-ECHILD);
3110 }
3111 } else {
3112 error = shmem_getpage(inode, 0, &page, SGP_READ);
3113 if (error)
3114 return ERR_PTR(error);
3115 unlock_page(page);
3116 }
3117 set_delayed_call(done, shmem_put_link, page);
3118 return page_address(page);
3119}
3120
3121#ifdef CONFIG_TMPFS_XATTR
3122/*
3123 * Superblocks without xattr inode operations may get some security.* xattr
3124 * support from the LSM "for free". As soon as we have any other xattrs
3125 * like ACLs, we also need to implement the security.* handlers at
3126 * filesystem level, though.
3127 */
3128
3129/*
3130 * Callback for security_inode_init_security() for acquiring xattrs.
3131 */
3132static int shmem_initxattrs(struct inode *inode,
3133 const struct xattr *xattr_array,
3134 void *fs_info)
3135{
3136 struct shmem_inode_info *info = SHMEM_I(inode);
3137 const struct xattr *xattr;
3138 struct simple_xattr *new_xattr;
3139 size_t len;
3140
3141 for (xattr = xattr_array; xattr->name != NULL; xattr++) {
3142 new_xattr = simple_xattr_alloc(xattr->value, xattr->value_len);
3143 if (!new_xattr)
3144 return -ENOMEM;
3145
3146 len = strlen(xattr->name) + 1;
3147 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
3148 GFP_KERNEL);
3149 if (!new_xattr->name) {
3150 kvfree(new_xattr);
3151 return -ENOMEM;
3152 }
3153
3154 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
3155 XATTR_SECURITY_PREFIX_LEN);
3156 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
3157 xattr->name, len);
3158
3159 simple_xattr_list_add(&info->xattrs, new_xattr);
3160 }
3161
3162 return 0;
3163}
3164
3165static int shmem_xattr_handler_get(const struct xattr_handler *handler,
3166 struct dentry *unused, struct inode *inode,
3167 const char *name, void *buffer, size_t size)
3168{
3169 struct shmem_inode_info *info = SHMEM_I(inode);
3170
3171 name = xattr_full_name(handler, name);
3172 return simple_xattr_get(&info->xattrs, name, buffer, size);
3173}
3174
3175static int shmem_xattr_handler_set(const struct xattr_handler *handler,
3176 struct user_namespace *mnt_userns,
3177 struct dentry *unused, struct inode *inode,
3178 const char *name, const void *value,
3179 size_t size, int flags)
3180{
3181 struct shmem_inode_info *info = SHMEM_I(inode);
3182
3183 name = xattr_full_name(handler, name);
3184 return simple_xattr_set(&info->xattrs, name, value, size, flags, NULL);
3185}
3186
3187static const struct xattr_handler shmem_security_xattr_handler = {
3188 .prefix = XATTR_SECURITY_PREFIX,
3189 .get = shmem_xattr_handler_get,
3190 .set = shmem_xattr_handler_set,
3191};
3192
3193static const struct xattr_handler shmem_trusted_xattr_handler = {
3194 .prefix = XATTR_TRUSTED_PREFIX,
3195 .get = shmem_xattr_handler_get,
3196 .set = shmem_xattr_handler_set,
3197};
3198
3199static const struct xattr_handler *shmem_xattr_handlers[] = {
3200#ifdef CONFIG_TMPFS_POSIX_ACL
3201 &posix_acl_access_xattr_handler,
3202 &posix_acl_default_xattr_handler,
3203#endif
3204 &shmem_security_xattr_handler,
3205 &shmem_trusted_xattr_handler,
3206 NULL
3207};
3208
3209static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
3210{
3211 struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
3212 return simple_xattr_list(d_inode(dentry), &info->xattrs, buffer, size);
3213}
3214#endif /* CONFIG_TMPFS_XATTR */
3215
3216static const struct inode_operations shmem_short_symlink_operations = {
3217 .get_link = simple_get_link,
3218#ifdef CONFIG_TMPFS_XATTR
3219 .listxattr = shmem_listxattr,
3220#endif
3221};
3222
3223static const struct inode_operations shmem_symlink_inode_operations = {
3224 .get_link = shmem_get_link,
3225#ifdef CONFIG_TMPFS_XATTR
3226 .listxattr = shmem_listxattr,
3227#endif
3228};
3229
3230static struct dentry *shmem_get_parent(struct dentry *child)
3231{
3232 return ERR_PTR(-ESTALE);
3233}
3234
3235static int shmem_match(struct inode *ino, void *vfh)
3236{
3237 __u32 *fh = vfh;
3238 __u64 inum = fh[2];
3239 inum = (inum << 32) | fh[1];
3240 return ino->i_ino == inum && fh[0] == ino->i_generation;
3241}
3242
3243/* Find any alias of inode, but prefer a hashed alias */
3244static struct dentry *shmem_find_alias(struct inode *inode)
3245{
3246 struct dentry *alias = d_find_alias(inode);
3247
3248 return alias ?: d_find_any_alias(inode);
3249}
3250
3251
3252static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
3253 struct fid *fid, int fh_len, int fh_type)
3254{
3255 struct inode *inode;
3256 struct dentry *dentry = NULL;
3257 u64 inum;
3258
3259 if (fh_len < 3)
3260 return NULL;
3261
3262 inum = fid->raw[2];
3263 inum = (inum << 32) | fid->raw[1];
3264
3265 inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
3266 shmem_match, fid->raw);
3267 if (inode) {
3268 dentry = shmem_find_alias(inode);
3269 iput(inode);
3270 }
3271
3272 return dentry;
3273}
3274
3275static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
3276 struct inode *parent)
3277{
3278 if (*len < 3) {
3279 *len = 3;
3280 return FILEID_INVALID;
3281 }
3282
3283 if (inode_unhashed(inode)) {
3284 /* Unfortunately insert_inode_hash is not idempotent,
3285 * so as we hash inodes here rather than at creation
3286 * time, we need a lock to ensure we only try
3287 * to do it once
3288 */
3289 static DEFINE_SPINLOCK(lock);
3290 spin_lock(&lock);
3291 if (inode_unhashed(inode))
3292 __insert_inode_hash(inode,
3293 inode->i_ino + inode->i_generation);
3294 spin_unlock(&lock);
3295 }
3296
3297 fh[0] = inode->i_generation;
3298 fh[1] = inode->i_ino;
3299 fh[2] = ((__u64)inode->i_ino) >> 32;
3300
3301 *len = 3;
3302 return 1;
3303}
3304
3305static const struct export_operations shmem_export_ops = {
3306 .get_parent = shmem_get_parent,
3307 .encode_fh = shmem_encode_fh,
3308 .fh_to_dentry = shmem_fh_to_dentry,
3309};
3310
3311enum shmem_param {
3312 Opt_gid,
3313 Opt_huge,
3314 Opt_mode,
3315 Opt_mpol,
3316 Opt_nr_blocks,
3317 Opt_nr_inodes,
3318 Opt_size,
3319 Opt_uid,
3320 Opt_inode32,
3321 Opt_inode64,
3322};
3323
3324static const struct constant_table shmem_param_enums_huge[] = {
3325 {"never", SHMEM_HUGE_NEVER },
3326 {"always", SHMEM_HUGE_ALWAYS },
3327 {"within_size", SHMEM_HUGE_WITHIN_SIZE },
3328 {"advise", SHMEM_HUGE_ADVISE },
3329 {}
3330};
3331
3332const struct fs_parameter_spec shmem_fs_parameters[] = {
3333 fsparam_u32 ("gid", Opt_gid),
3334 fsparam_enum ("huge", Opt_huge, shmem_param_enums_huge),
3335 fsparam_u32oct("mode", Opt_mode),
3336 fsparam_string("mpol", Opt_mpol),
3337 fsparam_string("nr_blocks", Opt_nr_blocks),
3338 fsparam_string("nr_inodes", Opt_nr_inodes),
3339 fsparam_string("size", Opt_size),
3340 fsparam_u32 ("uid", Opt_uid),
3341 fsparam_flag ("inode32", Opt_inode32),
3342 fsparam_flag ("inode64", Opt_inode64),
3343 {}
3344};
3345
3346static int shmem_parse_one(struct fs_context *fc, struct fs_parameter *param)
3347{
3348 struct shmem_options *ctx = fc->fs_private;
3349 struct fs_parse_result result;
3350 unsigned long long size;
3351 char *rest;
3352 int opt;
3353
3354 opt = fs_parse(fc, shmem_fs_parameters, param, &result);
3355 if (opt < 0)
3356 return opt;
3357
3358 switch (opt) {
3359 case Opt_size:
3360 size = memparse(param->string, &rest);
3361 if (*rest == '%') {
3362 size <<= PAGE_SHIFT;
3363 size *= totalram_pages();
3364 do_div(size, 100);
3365 rest++;
3366 }
3367 if (*rest)
3368 goto bad_value;
3369 ctx->blocks = DIV_ROUND_UP(size, PAGE_SIZE);
3370 ctx->seen |= SHMEM_SEEN_BLOCKS;
3371 break;
3372 case Opt_nr_blocks:
3373 ctx->blocks = memparse(param->string, &rest);
3374 if (*rest)
3375 goto bad_value;
3376 ctx->seen |= SHMEM_SEEN_BLOCKS;
3377 break;
3378 case Opt_nr_inodes:
3379 ctx->inodes = memparse(param->string, &rest);
3380 if (*rest)
3381 goto bad_value;
3382 ctx->seen |= SHMEM_SEEN_INODES;
3383 break;
3384 case Opt_mode:
3385 ctx->mode = result.uint_32 & 07777;
3386 break;
3387 case Opt_uid:
3388 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
3389 if (!uid_valid(ctx->uid))
3390 goto bad_value;
3391 break;
3392 case Opt_gid:
3393 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
3394 if (!gid_valid(ctx->gid))
3395 goto bad_value;
3396 break;
3397 case Opt_huge:
3398 ctx->huge = result.uint_32;
3399 if (ctx->huge != SHMEM_HUGE_NEVER &&
3400 !(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
3401 has_transparent_hugepage()))
3402 goto unsupported_parameter;
3403 ctx->seen |= SHMEM_SEEN_HUGE;
3404 break;
3405 case Opt_mpol:
3406 if (IS_ENABLED(CONFIG_NUMA)) {
3407 mpol_put(ctx->mpol);
3408 ctx->mpol = NULL;
3409 if (mpol_parse_str(param->string, &ctx->mpol))
3410 goto bad_value;
3411 break;
3412 }
3413 goto unsupported_parameter;
3414 case Opt_inode32:
3415 ctx->full_inums = false;
3416 ctx->seen |= SHMEM_SEEN_INUMS;
3417 break;
3418 case Opt_inode64:
3419 if (sizeof(ino_t) < 8) {
3420 return invalfc(fc,
3421 "Cannot use inode64 with <64bit inums in kernel\n");
3422 }
3423 ctx->full_inums = true;
3424 ctx->seen |= SHMEM_SEEN_INUMS;
3425 break;
3426 }
3427 return 0;
3428
3429unsupported_parameter:
3430 return invalfc(fc, "Unsupported parameter '%s'", param->key);
3431bad_value:
3432 return invalfc(fc, "Bad value for '%s'", param->key);
3433}
3434
3435static int shmem_parse_options(struct fs_context *fc, void *data)
3436{
3437 char *options = data;
3438
3439 if (options) {
3440 int err = security_sb_eat_lsm_opts(options, &fc->security);
3441 if (err)
3442 return err;
3443 }
3444
3445 while (options != NULL) {
3446 char *this_char = options;
3447 for (;;) {
3448 /*
3449 * NUL-terminate this option: unfortunately,
3450 * mount options form a comma-separated list,
3451 * but mpol's nodelist may also contain commas.
3452 */
3453 options = strchr(options, ',');
3454 if (options == NULL)
3455 break;
3456 options++;
3457 if (!isdigit(*options)) {
3458 options[-1] = '\0';
3459 break;
3460 }
3461 }
3462 if (*this_char) {
3463 char *value = strchr(this_char, '=');
3464 size_t len = 0;
3465 int err;
3466
3467 if (value) {
3468 *value++ = '\0';
3469 len = strlen(value);
3470 }
3471 err = vfs_parse_fs_string(fc, this_char, value, len);
3472 if (err < 0)
3473 return err;
3474 }
3475 }
3476 return 0;
3477}
3478
3479/*
3480 * Reconfigure a shmem filesystem.
3481 *
3482 * Note that we disallow change from limited->unlimited blocks/inodes while any
3483 * are in use; but we must separately disallow unlimited->limited, because in
3484 * that case we have no record of how much is already in use.
3485 */
3486static int shmem_reconfigure(struct fs_context *fc)
3487{
3488 struct shmem_options *ctx = fc->fs_private;
3489 struct shmem_sb_info *sbinfo = SHMEM_SB(fc->root->d_sb);
3490 unsigned long inodes;
3491 const char *err;
3492
3493 spin_lock(&sbinfo->stat_lock);
3494 inodes = sbinfo->max_inodes - sbinfo->free_inodes;
3495 if ((ctx->seen & SHMEM_SEEN_BLOCKS) && ctx->blocks) {
3496 if (!sbinfo->max_blocks) {
3497 err = "Cannot retroactively limit size";
3498 goto out;
3499 }
3500 if (percpu_counter_compare(&sbinfo->used_blocks,
3501 ctx->blocks) > 0) {
3502 err = "Too small a size for current use";
3503 goto out;
3504 }
3505 }
3506 if ((ctx->seen & SHMEM_SEEN_INODES) && ctx->inodes) {
3507 if (!sbinfo->max_inodes) {
3508 err = "Cannot retroactively limit inodes";
3509 goto out;
3510 }
3511 if (ctx->inodes < inodes) {
3512 err = "Too few inodes for current use";
3513 goto out;
3514 }
3515 }
3516
3517 if ((ctx->seen & SHMEM_SEEN_INUMS) && !ctx->full_inums &&
3518 sbinfo->next_ino > UINT_MAX) {
3519 err = "Current inum too high to switch to 32-bit inums";
3520 goto out;
3521 }
3522
3523 if (ctx->seen & SHMEM_SEEN_HUGE)
3524 sbinfo->huge = ctx->huge;
3525 if (ctx->seen & SHMEM_SEEN_INUMS)
3526 sbinfo->full_inums = ctx->full_inums;
3527 if (ctx->seen & SHMEM_SEEN_BLOCKS)
3528 sbinfo->max_blocks = ctx->blocks;
3529 if (ctx->seen & SHMEM_SEEN_INODES) {
3530 sbinfo->max_inodes = ctx->inodes;
3531 sbinfo->free_inodes = ctx->inodes - inodes;
3532 }
3533
3534 /*
3535 * Preserve previous mempolicy unless mpol remount option was specified.
3536 */
3537 if (ctx->mpol) {
3538 mpol_put(sbinfo->mpol);
3539 sbinfo->mpol = ctx->mpol; /* transfers initial ref */
3540 ctx->mpol = NULL;
3541 }
3542 spin_unlock(&sbinfo->stat_lock);
3543 return 0;
3544out:
3545 spin_unlock(&sbinfo->stat_lock);
3546 return invalfc(fc, "%s", err);
3547}
3548
3549static int shmem_show_options(struct seq_file *seq, struct dentry *root)
3550{
3551 struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
3552
3553 if (sbinfo->max_blocks != shmem_default_max_blocks())
3554 seq_printf(seq, ",size=%luk",
3555 sbinfo->max_blocks << (PAGE_SHIFT - 10));
3556 if (sbinfo->max_inodes != shmem_default_max_inodes())
3557 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
3558 if (sbinfo->mode != (0777 | S_ISVTX))
3559 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
3560 if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
3561 seq_printf(seq, ",uid=%u",
3562 from_kuid_munged(&init_user_ns, sbinfo->uid));
3563 if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
3564 seq_printf(seq, ",gid=%u",
3565 from_kgid_munged(&init_user_ns, sbinfo->gid));
3566
3567 /*
3568 * Showing inode{64,32} might be useful even if it's the system default,
3569 * since then people don't have to resort to checking both here and
3570 * /proc/config.gz to confirm 64-bit inums were successfully applied
3571 * (which may not even exist if IKCONFIG_PROC isn't enabled).
3572 *
3573 * We hide it when inode64 isn't the default and we are using 32-bit
3574 * inodes, since that probably just means the feature isn't even under
3575 * consideration.
3576 *
3577 * As such:
3578 *
3579 * +-----------------+-----------------+
3580 * | TMPFS_INODE64=y | TMPFS_INODE64=n |
3581 * +------------------+-----------------+-----------------+
3582 * | full_inums=true | show | show |
3583 * | full_inums=false | show | hide |
3584 * +------------------+-----------------+-----------------+
3585 *
3586 */
3587 if (IS_ENABLED(CONFIG_TMPFS_INODE64) || sbinfo->full_inums)
3588 seq_printf(seq, ",inode%d", (sbinfo->full_inums ? 64 : 32));
3589#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3590 /* Rightly or wrongly, show huge mount option unmasked by shmem_huge */
3591 if (sbinfo->huge)
3592 seq_printf(seq, ",huge=%s", shmem_format_huge(sbinfo->huge));
3593#endif
3594 shmem_show_mpol(seq, sbinfo->mpol);
3595 return 0;
3596}
3597
3598#endif /* CONFIG_TMPFS */
3599
3600static void shmem_put_super(struct super_block *sb)
3601{
3602 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
3603
3604 free_percpu(sbinfo->ino_batch);
3605 percpu_counter_destroy(&sbinfo->used_blocks);
3606 mpol_put(sbinfo->mpol);
3607 kfree(sbinfo);
3608 sb->s_fs_info = NULL;
3609}
3610
3611static int shmem_fill_super(struct super_block *sb, struct fs_context *fc)
3612{
3613 struct shmem_options *ctx = fc->fs_private;
3614 struct inode *inode;
3615 struct shmem_sb_info *sbinfo;
3616 int err = -ENOMEM;
3617
3618 /* Round up to L1_CACHE_BYTES to resist false sharing */
3619 sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
3620 L1_CACHE_BYTES), GFP_KERNEL);
3621 if (!sbinfo)
3622 return -ENOMEM;
3623
3624 sb->s_fs_info = sbinfo;
3625
3626#ifdef CONFIG_TMPFS
3627 /*
3628 * Per default we only allow half of the physical ram per
3629 * tmpfs instance, limiting inodes to one per page of lowmem;
3630 * but the internal instance is left unlimited.
3631 */
3632 if (!(sb->s_flags & SB_KERNMOUNT)) {
3633 if (!(ctx->seen & SHMEM_SEEN_BLOCKS))
3634 ctx->blocks = shmem_default_max_blocks();
3635 if (!(ctx->seen & SHMEM_SEEN_INODES))
3636 ctx->inodes = shmem_default_max_inodes();
3637 if (!(ctx->seen & SHMEM_SEEN_INUMS))
3638 ctx->full_inums = IS_ENABLED(CONFIG_TMPFS_INODE64);
3639 } else {
3640 sb->s_flags |= SB_NOUSER;
3641 }
3642 sb->s_export_op = &shmem_export_ops;
3643 sb->s_flags |= SB_NOSEC;
3644#else
3645 sb->s_flags |= SB_NOUSER;
3646#endif
3647 sbinfo->max_blocks = ctx->blocks;
3648 sbinfo->free_inodes = sbinfo->max_inodes = ctx->inodes;
3649 if (sb->s_flags & SB_KERNMOUNT) {
3650 sbinfo->ino_batch = alloc_percpu(ino_t);
3651 if (!sbinfo->ino_batch)
3652 goto failed;
3653 }
3654 sbinfo->uid = ctx->uid;
3655 sbinfo->gid = ctx->gid;
3656 sbinfo->full_inums = ctx->full_inums;
3657 sbinfo->mode = ctx->mode;
3658 sbinfo->huge = ctx->huge;
3659 sbinfo->mpol = ctx->mpol;
3660 ctx->mpol = NULL;
3661
3662 spin_lock_init(&sbinfo->stat_lock);
3663 if (percpu_counter_init(&sbinfo->used_blocks, 0, GFP_KERNEL))
3664 goto failed;
3665 spin_lock_init(&sbinfo->shrinklist_lock);
3666 INIT_LIST_HEAD(&sbinfo->shrinklist);
3667
3668 sb->s_maxbytes = MAX_LFS_FILESIZE;
3669 sb->s_blocksize = PAGE_SIZE;
3670 sb->s_blocksize_bits = PAGE_SHIFT;
3671 sb->s_magic = TMPFS_MAGIC;
3672 sb->s_op = &shmem_ops;
3673 sb->s_time_gran = 1;
3674#ifdef CONFIG_TMPFS_XATTR
3675 sb->s_xattr = shmem_xattr_handlers;
3676#endif
3677#ifdef CONFIG_TMPFS_POSIX_ACL
3678 sb->s_flags |= SB_POSIXACL;
3679#endif
3680 uuid_gen(&sb->s_uuid);
3681
3682 inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
3683 if (!inode)
3684 goto failed;
3685 inode->i_uid = sbinfo->uid;
3686 inode->i_gid = sbinfo->gid;
3687 sb->s_root = d_make_root(inode);
3688 if (!sb->s_root)
3689 goto failed;
3690 return 0;
3691
3692failed:
3693 shmem_put_super(sb);
3694 return err;
3695}
3696
3697static int shmem_get_tree(struct fs_context *fc)
3698{
3699 return get_tree_nodev(fc, shmem_fill_super);
3700}
3701
3702static void shmem_free_fc(struct fs_context *fc)
3703{
3704 struct shmem_options *ctx = fc->fs_private;
3705
3706 if (ctx) {
3707 mpol_put(ctx->mpol);
3708 kfree(ctx);
3709 }
3710}
3711
3712static const struct fs_context_operations shmem_fs_context_ops = {
3713 .free = shmem_free_fc,
3714 .get_tree = shmem_get_tree,
3715#ifdef CONFIG_TMPFS
3716 .parse_monolithic = shmem_parse_options,
3717 .parse_param = shmem_parse_one,
3718 .reconfigure = shmem_reconfigure,
3719#endif
3720};
3721
3722static struct kmem_cache *shmem_inode_cachep;
3723
3724static struct inode *shmem_alloc_inode(struct super_block *sb)
3725{
3726 struct shmem_inode_info *info;
3727 info = kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
3728 if (!info)
3729 return NULL;
3730 return &info->vfs_inode;
3731}
3732
3733static void shmem_free_in_core_inode(struct inode *inode)
3734{
3735 if (S_ISLNK(inode->i_mode))
3736 kfree(inode->i_link);
3737 kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
3738}
3739
3740static void shmem_destroy_inode(struct inode *inode)
3741{
3742 if (S_ISREG(inode->i_mode))
3743 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
3744}
3745
3746static void shmem_init_inode(void *foo)
3747{
3748 struct shmem_inode_info *info = foo;
3749 inode_init_once(&info->vfs_inode);
3750}
3751
3752static void shmem_init_inodecache(void)
3753{
3754 shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
3755 sizeof(struct shmem_inode_info),
3756 0, SLAB_PANIC|SLAB_ACCOUNT, shmem_init_inode);
3757}
3758
3759static void shmem_destroy_inodecache(void)
3760{
3761 kmem_cache_destroy(shmem_inode_cachep);
3762}
3763
3764const struct address_space_operations shmem_aops = {
3765 .writepage = shmem_writepage,
3766 .set_page_dirty = __set_page_dirty_no_writeback,
3767#ifdef CONFIG_TMPFS
3768 .write_begin = shmem_write_begin,
3769 .write_end = shmem_write_end,
3770#endif
3771#ifdef CONFIG_MIGRATION
3772 .migratepage = migrate_page,
3773#endif
3774 .error_remove_page = generic_error_remove_page,
3775};
3776EXPORT_SYMBOL(shmem_aops);
3777
3778static const struct file_operations shmem_file_operations = {
3779 .mmap = shmem_mmap,
3780 .get_unmapped_area = shmem_get_unmapped_area,
3781#ifdef CONFIG_TMPFS
3782 .llseek = shmem_file_llseek,
3783 .read_iter = shmem_file_read_iter,
3784 .write_iter = generic_file_write_iter,
3785 .fsync = noop_fsync,
3786 .splice_read = generic_file_splice_read,
3787 .splice_write = iter_file_splice_write,
3788 .fallocate = shmem_fallocate,
3789#endif
3790};
3791
3792static const struct inode_operations shmem_inode_operations = {
3793 .getattr = shmem_getattr,
3794 .setattr = shmem_setattr,
3795#ifdef CONFIG_TMPFS_XATTR
3796 .listxattr = shmem_listxattr,
3797 .set_acl = simple_set_acl,
3798#endif
3799};
3800
3801static const struct inode_operations shmem_dir_inode_operations = {
3802#ifdef CONFIG_TMPFS
3803 .create = shmem_create,
3804 .lookup = simple_lookup,
3805 .link = shmem_link,
3806 .unlink = shmem_unlink,
3807 .symlink = shmem_symlink,
3808 .mkdir = shmem_mkdir,
3809 .rmdir = shmem_rmdir,
3810 .mknod = shmem_mknod,
3811 .rename = shmem_rename2,
3812 .tmpfile = shmem_tmpfile,
3813#endif
3814#ifdef CONFIG_TMPFS_XATTR
3815 .listxattr = shmem_listxattr,
3816#endif
3817#ifdef CONFIG_TMPFS_POSIX_ACL
3818 .setattr = shmem_setattr,
3819 .set_acl = simple_set_acl,
3820#endif
3821};
3822
3823static const struct inode_operations shmem_special_inode_operations = {
3824#ifdef CONFIG_TMPFS_XATTR
3825 .listxattr = shmem_listxattr,
3826#endif
3827#ifdef CONFIG_TMPFS_POSIX_ACL
3828 .setattr = shmem_setattr,
3829 .set_acl = simple_set_acl,
3830#endif
3831};
3832
3833static const struct super_operations shmem_ops = {
3834 .alloc_inode = shmem_alloc_inode,
3835 .free_inode = shmem_free_in_core_inode,
3836 .destroy_inode = shmem_destroy_inode,
3837#ifdef CONFIG_TMPFS
3838 .statfs = shmem_statfs,
3839 .show_options = shmem_show_options,
3840#endif
3841 .evict_inode = shmem_evict_inode,
3842 .drop_inode = generic_delete_inode,
3843 .put_super = shmem_put_super,
3844#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3845 .nr_cached_objects = shmem_unused_huge_count,
3846 .free_cached_objects = shmem_unused_huge_scan,
3847#endif
3848};
3849
3850static const struct vm_operations_struct shmem_vm_ops = {
3851 .fault = shmem_fault,
3852 .map_pages = filemap_map_pages,
3853#ifdef CONFIG_NUMA
3854 .set_policy = shmem_set_policy,
3855 .get_policy = shmem_get_policy,
3856#endif
3857};
3858
3859int shmem_init_fs_context(struct fs_context *fc)
3860{
3861 struct shmem_options *ctx;
3862
3863 ctx = kzalloc(sizeof(struct shmem_options), GFP_KERNEL);
3864 if (!ctx)
3865 return -ENOMEM;
3866
3867 ctx->mode = 0777 | S_ISVTX;
3868 ctx->uid = current_fsuid();
3869 ctx->gid = current_fsgid();
3870
3871 fc->fs_private = ctx;
3872 fc->ops = &shmem_fs_context_ops;
3873 return 0;
3874}
3875
3876static struct file_system_type shmem_fs_type = {
3877 .owner = THIS_MODULE,
3878 .name = "tmpfs",
3879 .init_fs_context = shmem_init_fs_context,
3880#ifdef CONFIG_TMPFS
3881 .parameters = shmem_fs_parameters,
3882#endif
3883 .kill_sb = kill_litter_super,
3884 .fs_flags = FS_USERNS_MOUNT | FS_THP_SUPPORT,
3885};
3886
3887int __init shmem_init(void)
3888{
3889 int error;
3890
3891 shmem_init_inodecache();
3892
3893 error = register_filesystem(&shmem_fs_type);
3894 if (error) {
3895 pr_err("Could not register tmpfs\n");
3896 goto out2;
3897 }
3898
3899 shm_mnt = kern_mount(&shmem_fs_type);
3900 if (IS_ERR(shm_mnt)) {
3901 error = PTR_ERR(shm_mnt);
3902 pr_err("Could not kern_mount tmpfs\n");
3903 goto out1;
3904 }
3905
3906#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3907 if (has_transparent_hugepage() && shmem_huge > SHMEM_HUGE_DENY)
3908 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
3909 else
3910 shmem_huge = 0; /* just in case it was patched */
3911#endif
3912 return 0;
3913
3914out1:
3915 unregister_filesystem(&shmem_fs_type);
3916out2:
3917 shmem_destroy_inodecache();
3918 shm_mnt = ERR_PTR(error);
3919 return error;
3920}
3921
3922#if defined(CONFIG_TRANSPARENT_HUGEPAGE) && defined(CONFIG_SYSFS)
3923static ssize_t shmem_enabled_show(struct kobject *kobj,
3924 struct kobj_attribute *attr, char *buf)
3925{
3926 static const int values[] = {
3927 SHMEM_HUGE_ALWAYS,
3928 SHMEM_HUGE_WITHIN_SIZE,
3929 SHMEM_HUGE_ADVISE,
3930 SHMEM_HUGE_NEVER,
3931 SHMEM_HUGE_DENY,
3932 SHMEM_HUGE_FORCE,
3933 };
3934 int len = 0;
3935 int i;
3936
3937 for (i = 0; i < ARRAY_SIZE(values); i++) {
3938 len += sysfs_emit_at(buf, len,
3939 shmem_huge == values[i] ? "%s[%s]" : "%s%s",
3940 i ? " " : "",
3941 shmem_format_huge(values[i]));
3942 }
3943
3944 len += sysfs_emit_at(buf, len, "\n");
3945
3946 return len;
3947}
3948
3949static ssize_t shmem_enabled_store(struct kobject *kobj,
3950 struct kobj_attribute *attr, const char *buf, size_t count)
3951{
3952 char tmp[16];
3953 int huge;
3954
3955 if (count + 1 > sizeof(tmp))
3956 return -EINVAL;
3957 memcpy(tmp, buf, count);
3958 tmp[count] = '\0';
3959 if (count && tmp[count - 1] == '\n')
3960 tmp[count - 1] = '\0';
3961
3962 huge = shmem_parse_huge(tmp);
3963 if (huge == -EINVAL)
3964 return -EINVAL;
3965 if (!has_transparent_hugepage() &&
3966 huge != SHMEM_HUGE_NEVER && huge != SHMEM_HUGE_DENY)
3967 return -EINVAL;
3968
3969 shmem_huge = huge;
3970 if (shmem_huge > SHMEM_HUGE_DENY)
3971 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
3972 return count;
3973}
3974
3975struct kobj_attribute shmem_enabled_attr =
3976 __ATTR(shmem_enabled, 0644, shmem_enabled_show, shmem_enabled_store);
3977#endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_SYSFS */
3978
3979#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3980bool shmem_huge_enabled(struct vm_area_struct *vma)
3981{
3982 struct inode *inode = file_inode(vma->vm_file);
3983 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
3984 loff_t i_size;
3985 pgoff_t off;
3986
3987 if (!transhuge_vma_enabled(vma, vma->vm_flags))
3988 return false;
3989 if (shmem_huge == SHMEM_HUGE_FORCE)
3990 return true;
3991 if (shmem_huge == SHMEM_HUGE_DENY)
3992 return false;
3993 switch (sbinfo->huge) {
3994 case SHMEM_HUGE_NEVER:
3995 return false;
3996 case SHMEM_HUGE_ALWAYS:
3997 return true;
3998 case SHMEM_HUGE_WITHIN_SIZE:
3999 off = round_up(vma->vm_pgoff, HPAGE_PMD_NR);
4000 i_size = round_up(i_size_read(inode), PAGE_SIZE);
4001 if (i_size >= HPAGE_PMD_SIZE &&
4002 i_size >> PAGE_SHIFT >= off)
4003 return true;
4004 fallthrough;
4005 case SHMEM_HUGE_ADVISE:
4006 /* TODO: implement fadvise() hints */
4007 return (vma->vm_flags & VM_HUGEPAGE);
4008 default:
4009 VM_BUG_ON(1);
4010 return false;
4011 }
4012}
4013#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4014
4015#else /* !CONFIG_SHMEM */
4016
4017/*
4018 * tiny-shmem: simple shmemfs and tmpfs using ramfs code
4019 *
4020 * This is intended for small system where the benefits of the full
4021 * shmem code (swap-backed and resource-limited) are outweighed by
4022 * their complexity. On systems without swap this code should be
4023 * effectively equivalent, but much lighter weight.
4024 */
4025
4026static struct file_system_type shmem_fs_type = {
4027 .name = "tmpfs",
4028 .init_fs_context = ramfs_init_fs_context,
4029 .parameters = ramfs_fs_parameters,
4030 .kill_sb = kill_litter_super,
4031 .fs_flags = FS_USERNS_MOUNT,
4032};
4033
4034int __init shmem_init(void)
4035{
4036 BUG_ON(register_filesystem(&shmem_fs_type) != 0);
4037
4038 shm_mnt = kern_mount(&shmem_fs_type);
4039 BUG_ON(IS_ERR(shm_mnt));
4040
4041 return 0;
4042}
4043
4044int shmem_unuse(unsigned int type, bool frontswap,
4045 unsigned long *fs_pages_to_unuse)
4046{
4047 return 0;
4048}
4049
4050int shmem_lock(struct file *file, int lock, struct ucounts *ucounts)
4051{
4052 return 0;
4053}
4054
4055void shmem_unlock_mapping(struct address_space *mapping)
4056{
4057}
4058
4059#ifdef CONFIG_MMU
4060unsigned long shmem_get_unmapped_area(struct file *file,
4061 unsigned long addr, unsigned long len,
4062 unsigned long pgoff, unsigned long flags)
4063{
4064 return current->mm->get_unmapped_area(file, addr, len, pgoff, flags);
4065}
4066#endif
4067
4068void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
4069{
4070 truncate_inode_pages_range(inode->i_mapping, lstart, lend);
4071}
4072EXPORT_SYMBOL_GPL(shmem_truncate_range);
4073
4074#define shmem_vm_ops generic_file_vm_ops
4075#define shmem_file_operations ramfs_file_operations
4076#define shmem_get_inode(sb, dir, mode, dev, flags) ramfs_get_inode(sb, dir, mode, dev)
4077#define shmem_acct_size(flags, size) 0
4078#define shmem_unacct_size(flags, size) do {} while (0)
4079
4080#endif /* CONFIG_SHMEM */
4081
4082/* common code */
4083
4084static struct file *__shmem_file_setup(struct vfsmount *mnt, const char *name, loff_t size,
4085 unsigned long flags, unsigned int i_flags)
4086{
4087 struct inode *inode;
4088 struct file *res;
4089
4090 if (IS_ERR(mnt))
4091 return ERR_CAST(mnt);
4092
4093 if (size < 0 || size > MAX_LFS_FILESIZE)
4094 return ERR_PTR(-EINVAL);
4095
4096 if (shmem_acct_size(flags, size))
4097 return ERR_PTR(-ENOMEM);
4098
4099 inode = shmem_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0,
4100 flags);
4101 if (unlikely(!inode)) {
4102 shmem_unacct_size(flags, size);
4103 return ERR_PTR(-ENOSPC);
4104 }
4105 inode->i_flags |= i_flags;
4106 inode->i_size = size;
4107 clear_nlink(inode); /* It is unlinked */
4108 res = ERR_PTR(ramfs_nommu_expand_for_mapping(inode, size));
4109 if (!IS_ERR(res))
4110 res = alloc_file_pseudo(inode, mnt, name, O_RDWR,
4111 &shmem_file_operations);
4112 if (IS_ERR(res))
4113 iput(inode);
4114 return res;
4115}
4116
4117/**
4118 * shmem_kernel_file_setup - get an unlinked file living in tmpfs which must be
4119 * kernel internal. There will be NO LSM permission checks against the
4120 * underlying inode. So users of this interface must do LSM checks at a
4121 * higher layer. The users are the big_key and shm implementations. LSM
4122 * checks are provided at the key or shm level rather than the inode.
4123 * @name: name for dentry (to be seen in /proc/<pid>/maps
4124 * @size: size to be set for the file
4125 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4126 */
4127struct file *shmem_kernel_file_setup(const char *name, loff_t size, unsigned long flags)
4128{
4129 return __shmem_file_setup(shm_mnt, name, size, flags, S_PRIVATE);
4130}
4131
4132/**
4133 * shmem_file_setup - get an unlinked file living in tmpfs
4134 * @name: name for dentry (to be seen in /proc/<pid>/maps
4135 * @size: size to be set for the file
4136 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4137 */
4138struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
4139{
4140 return __shmem_file_setup(shm_mnt, name, size, flags, 0);
4141}
4142EXPORT_SYMBOL_GPL(shmem_file_setup);
4143
4144/**
4145 * shmem_file_setup_with_mnt - get an unlinked file living in tmpfs
4146 * @mnt: the tmpfs mount where the file will be created
4147 * @name: name for dentry (to be seen in /proc/<pid>/maps
4148 * @size: size to be set for the file
4149 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4150 */
4151struct file *shmem_file_setup_with_mnt(struct vfsmount *mnt, const char *name,
4152 loff_t size, unsigned long flags)
4153{
4154 return __shmem_file_setup(mnt, name, size, flags, 0);
4155}
4156EXPORT_SYMBOL_GPL(shmem_file_setup_with_mnt);
4157
4158/**
4159 * shmem_zero_setup - setup a shared anonymous mapping
4160 * @vma: the vma to be mmapped is prepared by do_mmap
4161 */
4162int shmem_zero_setup(struct vm_area_struct *vma)
4163{
4164 struct file *file;
4165 loff_t size = vma->vm_end - vma->vm_start;
4166
4167 /*
4168 * Cloning a new file under mmap_lock leads to a lock ordering conflict
4169 * between XFS directory reading and selinux: since this file is only
4170 * accessible to the user through its mapping, use S_PRIVATE flag to
4171 * bypass file security, in the same way as shmem_kernel_file_setup().
4172 */
4173 file = shmem_kernel_file_setup("dev/zero", size, vma->vm_flags);
4174 if (IS_ERR(file))
4175 return PTR_ERR(file);
4176
4177 if (vma->vm_file)
4178 fput(vma->vm_file);
4179 vma->vm_file = file;
4180 vma->vm_ops = &shmem_vm_ops;
4181
4182 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
4183 ((vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK) <
4184 (vma->vm_end & HPAGE_PMD_MASK)) {
4185 khugepaged_enter(vma, vma->vm_flags);
4186 }
4187
4188 return 0;
4189}
4190
4191/**
4192 * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
4193 * @mapping: the page's address_space
4194 * @index: the page index
4195 * @gfp: the page allocator flags to use if allocating
4196 *
4197 * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
4198 * with any new page allocations done using the specified allocation flags.
4199 * But read_cache_page_gfp() uses the ->readpage() method: which does not
4200 * suit tmpfs, since it may have pages in swapcache, and needs to find those
4201 * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
4202 *
4203 * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
4204 * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
4205 */
4206struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
4207 pgoff_t index, gfp_t gfp)
4208{
4209#ifdef CONFIG_SHMEM
4210 struct inode *inode = mapping->host;
4211 struct page *page;
4212 int error;
4213
4214 BUG_ON(!shmem_mapping(mapping));
4215 error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE,
4216 gfp, NULL, NULL, NULL);
4217 if (error)
4218 page = ERR_PTR(error);
4219 else
4220 unlock_page(page);
4221 return page;
4222#else
4223 /*
4224 * The tiny !SHMEM case uses ramfs without swap
4225 */
4226 return read_cache_page_gfp(mapping, index, gfp);
4227#endif
4228}
4229EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);
1/*
2 * Resizable virtual memory filesystem for Linux.
3 *
4 * Copyright (C) 2000 Linus Torvalds.
5 * 2000 Transmeta Corp.
6 * 2000-2001 Christoph Rohland
7 * 2000-2001 SAP AG
8 * 2002 Red Hat Inc.
9 * Copyright (C) 2002-2011 Hugh Dickins.
10 * Copyright (C) 2011 Google Inc.
11 * Copyright (C) 2002-2005 VERITAS Software Corporation.
12 * Copyright (C) 2004 Andi Kleen, SuSE Labs
13 *
14 * Extended attribute support for tmpfs:
15 * Copyright (c) 2004, Luke Kenneth Casson Leighton <lkcl@lkcl.net>
16 * Copyright (c) 2004 Red Hat, Inc., James Morris <jmorris@redhat.com>
17 *
18 * tiny-shmem:
19 * Copyright (c) 2004, 2008 Matt Mackall <mpm@selenic.com>
20 *
21 * This file is released under the GPL.
22 */
23
24#include <linux/fs.h>
25#include <linux/init.h>
26#include <linux/vfs.h>
27#include <linux/mount.h>
28#include <linux/ramfs.h>
29#include <linux/pagemap.h>
30#include <linux/file.h>
31#include <linux/fileattr.h>
32#include <linux/mm.h>
33#include <linux/random.h>
34#include <linux/sched/signal.h>
35#include <linux/export.h>
36#include <linux/swap.h>
37#include <linux/uio.h>
38#include <linux/hugetlb.h>
39#include <linux/fs_parser.h>
40#include <linux/swapfile.h>
41#include <linux/iversion.h>
42#include "swap.h"
43
44static struct vfsmount *shm_mnt;
45
46#ifdef CONFIG_SHMEM
47/*
48 * This virtual memory filesystem is heavily based on the ramfs. It
49 * extends ramfs by the ability to use swap and honor resource limits
50 * which makes it a completely usable filesystem.
51 */
52
53#include <linux/xattr.h>
54#include <linux/exportfs.h>
55#include <linux/posix_acl.h>
56#include <linux/posix_acl_xattr.h>
57#include <linux/mman.h>
58#include <linux/string.h>
59#include <linux/slab.h>
60#include <linux/backing-dev.h>
61#include <linux/shmem_fs.h>
62#include <linux/writeback.h>
63#include <linux/pagevec.h>
64#include <linux/percpu_counter.h>
65#include <linux/falloc.h>
66#include <linux/splice.h>
67#include <linux/security.h>
68#include <linux/swapops.h>
69#include <linux/mempolicy.h>
70#include <linux/namei.h>
71#include <linux/ctype.h>
72#include <linux/migrate.h>
73#include <linux/highmem.h>
74#include <linux/seq_file.h>
75#include <linux/magic.h>
76#include <linux/syscalls.h>
77#include <linux/fcntl.h>
78#include <uapi/linux/memfd.h>
79#include <linux/userfaultfd_k.h>
80#include <linux/rmap.h>
81#include <linux/uuid.h>
82
83#include <linux/uaccess.h>
84
85#include "internal.h"
86
87#define BLOCKS_PER_PAGE (PAGE_SIZE/512)
88#define VM_ACCT(size) (PAGE_ALIGN(size) >> PAGE_SHIFT)
89
90/* Pretend that each entry is of this size in directory's i_size */
91#define BOGO_DIRENT_SIZE 20
92
93/* Symlink up to this size is kmalloc'ed instead of using a swappable page */
94#define SHORT_SYMLINK_LEN 128
95
96/*
97 * shmem_fallocate communicates with shmem_fault or shmem_writepage via
98 * inode->i_private (with i_rwsem making sure that it has only one user at
99 * a time): we would prefer not to enlarge the shmem inode just for that.
100 */
101struct shmem_falloc {
102 wait_queue_head_t *waitq; /* faults into hole wait for punch to end */
103 pgoff_t start; /* start of range currently being fallocated */
104 pgoff_t next; /* the next page offset to be fallocated */
105 pgoff_t nr_falloced; /* how many new pages have been fallocated */
106 pgoff_t nr_unswapped; /* how often writepage refused to swap out */
107};
108
109struct shmem_options {
110 unsigned long long blocks;
111 unsigned long long inodes;
112 struct mempolicy *mpol;
113 kuid_t uid;
114 kgid_t gid;
115 umode_t mode;
116 bool full_inums;
117 int huge;
118 int seen;
119#define SHMEM_SEEN_BLOCKS 1
120#define SHMEM_SEEN_INODES 2
121#define SHMEM_SEEN_HUGE 4
122#define SHMEM_SEEN_INUMS 8
123};
124
125#ifdef CONFIG_TMPFS
126static unsigned long shmem_default_max_blocks(void)
127{
128 return totalram_pages() / 2;
129}
130
131static unsigned long shmem_default_max_inodes(void)
132{
133 unsigned long nr_pages = totalram_pages();
134
135 return min(nr_pages - totalhigh_pages(), nr_pages / 2);
136}
137#endif
138
139static int shmem_swapin_folio(struct inode *inode, pgoff_t index,
140 struct folio **foliop, enum sgp_type sgp,
141 gfp_t gfp, struct vm_area_struct *vma,
142 vm_fault_t *fault_type);
143
144static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
145{
146 return sb->s_fs_info;
147}
148
149/*
150 * shmem_file_setup pre-accounts the whole fixed size of a VM object,
151 * for shared memory and for shared anonymous (/dev/zero) mappings
152 * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
153 * consistent with the pre-accounting of private mappings ...
154 */
155static inline int shmem_acct_size(unsigned long flags, loff_t size)
156{
157 return (flags & VM_NORESERVE) ?
158 0 : security_vm_enough_memory_mm(current->mm, VM_ACCT(size));
159}
160
161static inline void shmem_unacct_size(unsigned long flags, loff_t size)
162{
163 if (!(flags & VM_NORESERVE))
164 vm_unacct_memory(VM_ACCT(size));
165}
166
167static inline int shmem_reacct_size(unsigned long flags,
168 loff_t oldsize, loff_t newsize)
169{
170 if (!(flags & VM_NORESERVE)) {
171 if (VM_ACCT(newsize) > VM_ACCT(oldsize))
172 return security_vm_enough_memory_mm(current->mm,
173 VM_ACCT(newsize) - VM_ACCT(oldsize));
174 else if (VM_ACCT(newsize) < VM_ACCT(oldsize))
175 vm_unacct_memory(VM_ACCT(oldsize) - VM_ACCT(newsize));
176 }
177 return 0;
178}
179
180/*
181 * ... whereas tmpfs objects are accounted incrementally as
182 * pages are allocated, in order to allow large sparse files.
183 * shmem_get_folio reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
184 * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
185 */
186static inline int shmem_acct_block(unsigned long flags, long pages)
187{
188 if (!(flags & VM_NORESERVE))
189 return 0;
190
191 return security_vm_enough_memory_mm(current->mm,
192 pages * VM_ACCT(PAGE_SIZE));
193}
194
195static inline void shmem_unacct_blocks(unsigned long flags, long pages)
196{
197 if (flags & VM_NORESERVE)
198 vm_unacct_memory(pages * VM_ACCT(PAGE_SIZE));
199}
200
201static inline bool shmem_inode_acct_block(struct inode *inode, long pages)
202{
203 struct shmem_inode_info *info = SHMEM_I(inode);
204 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
205
206 if (shmem_acct_block(info->flags, pages))
207 return false;
208
209 if (sbinfo->max_blocks) {
210 if (percpu_counter_compare(&sbinfo->used_blocks,
211 sbinfo->max_blocks - pages) > 0)
212 goto unacct;
213 percpu_counter_add(&sbinfo->used_blocks, pages);
214 }
215
216 return true;
217
218unacct:
219 shmem_unacct_blocks(info->flags, pages);
220 return false;
221}
222
223static inline void shmem_inode_unacct_blocks(struct inode *inode, long pages)
224{
225 struct shmem_inode_info *info = SHMEM_I(inode);
226 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
227
228 if (sbinfo->max_blocks)
229 percpu_counter_sub(&sbinfo->used_blocks, pages);
230 shmem_unacct_blocks(info->flags, pages);
231}
232
233static const struct super_operations shmem_ops;
234const struct address_space_operations shmem_aops;
235static const struct file_operations shmem_file_operations;
236static const struct inode_operations shmem_inode_operations;
237static const struct inode_operations shmem_dir_inode_operations;
238static const struct inode_operations shmem_special_inode_operations;
239static const struct vm_operations_struct shmem_vm_ops;
240static const struct vm_operations_struct shmem_anon_vm_ops;
241static struct file_system_type shmem_fs_type;
242
243bool vma_is_anon_shmem(struct vm_area_struct *vma)
244{
245 return vma->vm_ops == &shmem_anon_vm_ops;
246}
247
248bool vma_is_shmem(struct vm_area_struct *vma)
249{
250 return vma_is_anon_shmem(vma) || vma->vm_ops == &shmem_vm_ops;
251}
252
253static LIST_HEAD(shmem_swaplist);
254static DEFINE_MUTEX(shmem_swaplist_mutex);
255
256/*
257 * shmem_reserve_inode() performs bookkeeping to reserve a shmem inode, and
258 * produces a novel ino for the newly allocated inode.
259 *
260 * It may also be called when making a hard link to permit the space needed by
261 * each dentry. However, in that case, no new inode number is needed since that
262 * internally draws from another pool of inode numbers (currently global
263 * get_next_ino()). This case is indicated by passing NULL as inop.
264 */
265#define SHMEM_INO_BATCH 1024
266static int shmem_reserve_inode(struct super_block *sb, ino_t *inop)
267{
268 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
269 ino_t ino;
270
271 if (!(sb->s_flags & SB_KERNMOUNT)) {
272 raw_spin_lock(&sbinfo->stat_lock);
273 if (sbinfo->max_inodes) {
274 if (!sbinfo->free_inodes) {
275 raw_spin_unlock(&sbinfo->stat_lock);
276 return -ENOSPC;
277 }
278 sbinfo->free_inodes--;
279 }
280 if (inop) {
281 ino = sbinfo->next_ino++;
282 if (unlikely(is_zero_ino(ino)))
283 ino = sbinfo->next_ino++;
284 if (unlikely(!sbinfo->full_inums &&
285 ino > UINT_MAX)) {
286 /*
287 * Emulate get_next_ino uint wraparound for
288 * compatibility
289 */
290 if (IS_ENABLED(CONFIG_64BIT))
291 pr_warn("%s: inode number overflow on device %d, consider using inode64 mount option\n",
292 __func__, MINOR(sb->s_dev));
293 sbinfo->next_ino = 1;
294 ino = sbinfo->next_ino++;
295 }
296 *inop = ino;
297 }
298 raw_spin_unlock(&sbinfo->stat_lock);
299 } else if (inop) {
300 /*
301 * __shmem_file_setup, one of our callers, is lock-free: it
302 * doesn't hold stat_lock in shmem_reserve_inode since
303 * max_inodes is always 0, and is called from potentially
304 * unknown contexts. As such, use a per-cpu batched allocator
305 * which doesn't require the per-sb stat_lock unless we are at
306 * the batch boundary.
307 *
308 * We don't need to worry about inode{32,64} since SB_KERNMOUNT
309 * shmem mounts are not exposed to userspace, so we don't need
310 * to worry about things like glibc compatibility.
311 */
312 ino_t *next_ino;
313
314 next_ino = per_cpu_ptr(sbinfo->ino_batch, get_cpu());
315 ino = *next_ino;
316 if (unlikely(ino % SHMEM_INO_BATCH == 0)) {
317 raw_spin_lock(&sbinfo->stat_lock);
318 ino = sbinfo->next_ino;
319 sbinfo->next_ino += SHMEM_INO_BATCH;
320 raw_spin_unlock(&sbinfo->stat_lock);
321 if (unlikely(is_zero_ino(ino)))
322 ino++;
323 }
324 *inop = ino;
325 *next_ino = ++ino;
326 put_cpu();
327 }
328
329 return 0;
330}
331
332static void shmem_free_inode(struct super_block *sb)
333{
334 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
335 if (sbinfo->max_inodes) {
336 raw_spin_lock(&sbinfo->stat_lock);
337 sbinfo->free_inodes++;
338 raw_spin_unlock(&sbinfo->stat_lock);
339 }
340}
341
342/**
343 * shmem_recalc_inode - recalculate the block usage of an inode
344 * @inode: inode to recalc
345 *
346 * We have to calculate the free blocks since the mm can drop
347 * undirtied hole pages behind our back.
348 *
349 * But normally info->alloced == inode->i_mapping->nrpages + info->swapped
350 * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
351 *
352 * It has to be called with the spinlock held.
353 */
354static void shmem_recalc_inode(struct inode *inode)
355{
356 struct shmem_inode_info *info = SHMEM_I(inode);
357 long freed;
358
359 freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
360 if (freed > 0) {
361 info->alloced -= freed;
362 inode->i_blocks -= freed * BLOCKS_PER_PAGE;
363 shmem_inode_unacct_blocks(inode, freed);
364 }
365}
366
367bool shmem_charge(struct inode *inode, long pages)
368{
369 struct shmem_inode_info *info = SHMEM_I(inode);
370 unsigned long flags;
371
372 if (!shmem_inode_acct_block(inode, pages))
373 return false;
374
375 /* nrpages adjustment first, then shmem_recalc_inode() when balanced */
376 inode->i_mapping->nrpages += pages;
377
378 spin_lock_irqsave(&info->lock, flags);
379 info->alloced += pages;
380 inode->i_blocks += pages * BLOCKS_PER_PAGE;
381 shmem_recalc_inode(inode);
382 spin_unlock_irqrestore(&info->lock, flags);
383
384 return true;
385}
386
387void shmem_uncharge(struct inode *inode, long pages)
388{
389 struct shmem_inode_info *info = SHMEM_I(inode);
390 unsigned long flags;
391
392 /* nrpages adjustment done by __filemap_remove_folio() or caller */
393
394 spin_lock_irqsave(&info->lock, flags);
395 info->alloced -= pages;
396 inode->i_blocks -= pages * BLOCKS_PER_PAGE;
397 shmem_recalc_inode(inode);
398 spin_unlock_irqrestore(&info->lock, flags);
399
400 shmem_inode_unacct_blocks(inode, pages);
401}
402
403/*
404 * Replace item expected in xarray by a new item, while holding xa_lock.
405 */
406static int shmem_replace_entry(struct address_space *mapping,
407 pgoff_t index, void *expected, void *replacement)
408{
409 XA_STATE(xas, &mapping->i_pages, index);
410 void *item;
411
412 VM_BUG_ON(!expected);
413 VM_BUG_ON(!replacement);
414 item = xas_load(&xas);
415 if (item != expected)
416 return -ENOENT;
417 xas_store(&xas, replacement);
418 return 0;
419}
420
421/*
422 * Sometimes, before we decide whether to proceed or to fail, we must check
423 * that an entry was not already brought back from swap by a racing thread.
424 *
425 * Checking page is not enough: by the time a SwapCache page is locked, it
426 * might be reused, and again be SwapCache, using the same swap as before.
427 */
428static bool shmem_confirm_swap(struct address_space *mapping,
429 pgoff_t index, swp_entry_t swap)
430{
431 return xa_load(&mapping->i_pages, index) == swp_to_radix_entry(swap);
432}
433
434/*
435 * Definitions for "huge tmpfs": tmpfs mounted with the huge= option
436 *
437 * SHMEM_HUGE_NEVER:
438 * disables huge pages for the mount;
439 * SHMEM_HUGE_ALWAYS:
440 * enables huge pages for the mount;
441 * SHMEM_HUGE_WITHIN_SIZE:
442 * only allocate huge pages if the page will be fully within i_size,
443 * also respect fadvise()/madvise() hints;
444 * SHMEM_HUGE_ADVISE:
445 * only allocate huge pages if requested with fadvise()/madvise();
446 */
447
448#define SHMEM_HUGE_NEVER 0
449#define SHMEM_HUGE_ALWAYS 1
450#define SHMEM_HUGE_WITHIN_SIZE 2
451#define SHMEM_HUGE_ADVISE 3
452
453/*
454 * Special values.
455 * Only can be set via /sys/kernel/mm/transparent_hugepage/shmem_enabled:
456 *
457 * SHMEM_HUGE_DENY:
458 * disables huge on shm_mnt and all mounts, for emergency use;
459 * SHMEM_HUGE_FORCE:
460 * enables huge on shm_mnt and all mounts, w/o needing option, for testing;
461 *
462 */
463#define SHMEM_HUGE_DENY (-1)
464#define SHMEM_HUGE_FORCE (-2)
465
466#ifdef CONFIG_TRANSPARENT_HUGEPAGE
467/* ifdef here to avoid bloating shmem.o when not necessary */
468
469static int shmem_huge __read_mostly = SHMEM_HUGE_NEVER;
470
471bool shmem_is_huge(struct vm_area_struct *vma, struct inode *inode,
472 pgoff_t index, bool shmem_huge_force)
473{
474 loff_t i_size;
475
476 if (!S_ISREG(inode->i_mode))
477 return false;
478 if (vma && ((vma->vm_flags & VM_NOHUGEPAGE) ||
479 test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags)))
480 return false;
481 if (shmem_huge == SHMEM_HUGE_DENY)
482 return false;
483 if (shmem_huge_force || shmem_huge == SHMEM_HUGE_FORCE)
484 return true;
485
486 switch (SHMEM_SB(inode->i_sb)->huge) {
487 case SHMEM_HUGE_ALWAYS:
488 return true;
489 case SHMEM_HUGE_WITHIN_SIZE:
490 index = round_up(index + 1, HPAGE_PMD_NR);
491 i_size = round_up(i_size_read(inode), PAGE_SIZE);
492 if (i_size >> PAGE_SHIFT >= index)
493 return true;
494 fallthrough;
495 case SHMEM_HUGE_ADVISE:
496 if (vma && (vma->vm_flags & VM_HUGEPAGE))
497 return true;
498 fallthrough;
499 default:
500 return false;
501 }
502}
503
504#if defined(CONFIG_SYSFS)
505static int shmem_parse_huge(const char *str)
506{
507 if (!strcmp(str, "never"))
508 return SHMEM_HUGE_NEVER;
509 if (!strcmp(str, "always"))
510 return SHMEM_HUGE_ALWAYS;
511 if (!strcmp(str, "within_size"))
512 return SHMEM_HUGE_WITHIN_SIZE;
513 if (!strcmp(str, "advise"))
514 return SHMEM_HUGE_ADVISE;
515 if (!strcmp(str, "deny"))
516 return SHMEM_HUGE_DENY;
517 if (!strcmp(str, "force"))
518 return SHMEM_HUGE_FORCE;
519 return -EINVAL;
520}
521#endif
522
523#if defined(CONFIG_SYSFS) || defined(CONFIG_TMPFS)
524static const char *shmem_format_huge(int huge)
525{
526 switch (huge) {
527 case SHMEM_HUGE_NEVER:
528 return "never";
529 case SHMEM_HUGE_ALWAYS:
530 return "always";
531 case SHMEM_HUGE_WITHIN_SIZE:
532 return "within_size";
533 case SHMEM_HUGE_ADVISE:
534 return "advise";
535 case SHMEM_HUGE_DENY:
536 return "deny";
537 case SHMEM_HUGE_FORCE:
538 return "force";
539 default:
540 VM_BUG_ON(1);
541 return "bad_val";
542 }
543}
544#endif
545
546static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
547 struct shrink_control *sc, unsigned long nr_to_split)
548{
549 LIST_HEAD(list), *pos, *next;
550 LIST_HEAD(to_remove);
551 struct inode *inode;
552 struct shmem_inode_info *info;
553 struct folio *folio;
554 unsigned long batch = sc ? sc->nr_to_scan : 128;
555 int split = 0;
556
557 if (list_empty(&sbinfo->shrinklist))
558 return SHRINK_STOP;
559
560 spin_lock(&sbinfo->shrinklist_lock);
561 list_for_each_safe(pos, next, &sbinfo->shrinklist) {
562 info = list_entry(pos, struct shmem_inode_info, shrinklist);
563
564 /* pin the inode */
565 inode = igrab(&info->vfs_inode);
566
567 /* inode is about to be evicted */
568 if (!inode) {
569 list_del_init(&info->shrinklist);
570 goto next;
571 }
572
573 /* Check if there's anything to gain */
574 if (round_up(inode->i_size, PAGE_SIZE) ==
575 round_up(inode->i_size, HPAGE_PMD_SIZE)) {
576 list_move(&info->shrinklist, &to_remove);
577 goto next;
578 }
579
580 list_move(&info->shrinklist, &list);
581next:
582 sbinfo->shrinklist_len--;
583 if (!--batch)
584 break;
585 }
586 spin_unlock(&sbinfo->shrinklist_lock);
587
588 list_for_each_safe(pos, next, &to_remove) {
589 info = list_entry(pos, struct shmem_inode_info, shrinklist);
590 inode = &info->vfs_inode;
591 list_del_init(&info->shrinklist);
592 iput(inode);
593 }
594
595 list_for_each_safe(pos, next, &list) {
596 int ret;
597 pgoff_t index;
598
599 info = list_entry(pos, struct shmem_inode_info, shrinklist);
600 inode = &info->vfs_inode;
601
602 if (nr_to_split && split >= nr_to_split)
603 goto move_back;
604
605 index = (inode->i_size & HPAGE_PMD_MASK) >> PAGE_SHIFT;
606 folio = filemap_get_folio(inode->i_mapping, index);
607 if (!folio)
608 goto drop;
609
610 /* No huge page at the end of the file: nothing to split */
611 if (!folio_test_large(folio)) {
612 folio_put(folio);
613 goto drop;
614 }
615
616 /*
617 * Move the inode on the list back to shrinklist if we failed
618 * to lock the page at this time.
619 *
620 * Waiting for the lock may lead to deadlock in the
621 * reclaim path.
622 */
623 if (!folio_trylock(folio)) {
624 folio_put(folio);
625 goto move_back;
626 }
627
628 ret = split_folio(folio);
629 folio_unlock(folio);
630 folio_put(folio);
631
632 /* If split failed move the inode on the list back to shrinklist */
633 if (ret)
634 goto move_back;
635
636 split++;
637drop:
638 list_del_init(&info->shrinklist);
639 goto put;
640move_back:
641 /*
642 * Make sure the inode is either on the global list or deleted
643 * from any local list before iput() since it could be deleted
644 * in another thread once we put the inode (then the local list
645 * is corrupted).
646 */
647 spin_lock(&sbinfo->shrinklist_lock);
648 list_move(&info->shrinklist, &sbinfo->shrinklist);
649 sbinfo->shrinklist_len++;
650 spin_unlock(&sbinfo->shrinklist_lock);
651put:
652 iput(inode);
653 }
654
655 return split;
656}
657
658static long shmem_unused_huge_scan(struct super_block *sb,
659 struct shrink_control *sc)
660{
661 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
662
663 if (!READ_ONCE(sbinfo->shrinklist_len))
664 return SHRINK_STOP;
665
666 return shmem_unused_huge_shrink(sbinfo, sc, 0);
667}
668
669static long shmem_unused_huge_count(struct super_block *sb,
670 struct shrink_control *sc)
671{
672 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
673 return READ_ONCE(sbinfo->shrinklist_len);
674}
675#else /* !CONFIG_TRANSPARENT_HUGEPAGE */
676
677#define shmem_huge SHMEM_HUGE_DENY
678
679bool shmem_is_huge(struct vm_area_struct *vma, struct inode *inode,
680 pgoff_t index, bool shmem_huge_force)
681{
682 return false;
683}
684
685static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
686 struct shrink_control *sc, unsigned long nr_to_split)
687{
688 return 0;
689}
690#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
691
692/*
693 * Like filemap_add_folio, but error if expected item has gone.
694 */
695static int shmem_add_to_page_cache(struct folio *folio,
696 struct address_space *mapping,
697 pgoff_t index, void *expected, gfp_t gfp,
698 struct mm_struct *charge_mm)
699{
700 XA_STATE_ORDER(xas, &mapping->i_pages, index, folio_order(folio));
701 long nr = folio_nr_pages(folio);
702 int error;
703
704 VM_BUG_ON_FOLIO(index != round_down(index, nr), folio);
705 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
706 VM_BUG_ON_FOLIO(!folio_test_swapbacked(folio), folio);
707 VM_BUG_ON(expected && folio_test_large(folio));
708
709 folio_ref_add(folio, nr);
710 folio->mapping = mapping;
711 folio->index = index;
712
713 if (!folio_test_swapcache(folio)) {
714 error = mem_cgroup_charge(folio, charge_mm, gfp);
715 if (error) {
716 if (folio_test_pmd_mappable(folio)) {
717 count_vm_event(THP_FILE_FALLBACK);
718 count_vm_event(THP_FILE_FALLBACK_CHARGE);
719 }
720 goto error;
721 }
722 }
723 folio_throttle_swaprate(folio, gfp);
724
725 do {
726 xas_lock_irq(&xas);
727 if (expected != xas_find_conflict(&xas)) {
728 xas_set_err(&xas, -EEXIST);
729 goto unlock;
730 }
731 if (expected && xas_find_conflict(&xas)) {
732 xas_set_err(&xas, -EEXIST);
733 goto unlock;
734 }
735 xas_store(&xas, folio);
736 if (xas_error(&xas))
737 goto unlock;
738 if (folio_test_pmd_mappable(folio)) {
739 count_vm_event(THP_FILE_ALLOC);
740 __lruvec_stat_mod_folio(folio, NR_SHMEM_THPS, nr);
741 }
742 mapping->nrpages += nr;
743 __lruvec_stat_mod_folio(folio, NR_FILE_PAGES, nr);
744 __lruvec_stat_mod_folio(folio, NR_SHMEM, nr);
745unlock:
746 xas_unlock_irq(&xas);
747 } while (xas_nomem(&xas, gfp));
748
749 if (xas_error(&xas)) {
750 error = xas_error(&xas);
751 goto error;
752 }
753
754 return 0;
755error:
756 folio->mapping = NULL;
757 folio_ref_sub(folio, nr);
758 return error;
759}
760
761/*
762 * Like delete_from_page_cache, but substitutes swap for @folio.
763 */
764static void shmem_delete_from_page_cache(struct folio *folio, void *radswap)
765{
766 struct address_space *mapping = folio->mapping;
767 long nr = folio_nr_pages(folio);
768 int error;
769
770 xa_lock_irq(&mapping->i_pages);
771 error = shmem_replace_entry(mapping, folio->index, folio, radswap);
772 folio->mapping = NULL;
773 mapping->nrpages -= nr;
774 __lruvec_stat_mod_folio(folio, NR_FILE_PAGES, -nr);
775 __lruvec_stat_mod_folio(folio, NR_SHMEM, -nr);
776 xa_unlock_irq(&mapping->i_pages);
777 folio_put(folio);
778 BUG_ON(error);
779}
780
781/*
782 * Remove swap entry from page cache, free the swap and its page cache.
783 */
784static int shmem_free_swap(struct address_space *mapping,
785 pgoff_t index, void *radswap)
786{
787 void *old;
788
789 old = xa_cmpxchg_irq(&mapping->i_pages, index, radswap, NULL, 0);
790 if (old != radswap)
791 return -ENOENT;
792 free_swap_and_cache(radix_to_swp_entry(radswap));
793 return 0;
794}
795
796/*
797 * Determine (in bytes) how many of the shmem object's pages mapped by the
798 * given offsets are swapped out.
799 *
800 * This is safe to call without i_rwsem or the i_pages lock thanks to RCU,
801 * as long as the inode doesn't go away and racy results are not a problem.
802 */
803unsigned long shmem_partial_swap_usage(struct address_space *mapping,
804 pgoff_t start, pgoff_t end)
805{
806 XA_STATE(xas, &mapping->i_pages, start);
807 struct page *page;
808 unsigned long swapped = 0;
809
810 rcu_read_lock();
811 xas_for_each(&xas, page, end - 1) {
812 if (xas_retry(&xas, page))
813 continue;
814 if (xa_is_value(page))
815 swapped++;
816
817 if (need_resched()) {
818 xas_pause(&xas);
819 cond_resched_rcu();
820 }
821 }
822
823 rcu_read_unlock();
824
825 return swapped << PAGE_SHIFT;
826}
827
828/*
829 * Determine (in bytes) how many of the shmem object's pages mapped by the
830 * given vma is swapped out.
831 *
832 * This is safe to call without i_rwsem or the i_pages lock thanks to RCU,
833 * as long as the inode doesn't go away and racy results are not a problem.
834 */
835unsigned long shmem_swap_usage(struct vm_area_struct *vma)
836{
837 struct inode *inode = file_inode(vma->vm_file);
838 struct shmem_inode_info *info = SHMEM_I(inode);
839 struct address_space *mapping = inode->i_mapping;
840 unsigned long swapped;
841
842 /* Be careful as we don't hold info->lock */
843 swapped = READ_ONCE(info->swapped);
844
845 /*
846 * The easier cases are when the shmem object has nothing in swap, or
847 * the vma maps it whole. Then we can simply use the stats that we
848 * already track.
849 */
850 if (!swapped)
851 return 0;
852
853 if (!vma->vm_pgoff && vma->vm_end - vma->vm_start >= inode->i_size)
854 return swapped << PAGE_SHIFT;
855
856 /* Here comes the more involved part */
857 return shmem_partial_swap_usage(mapping, vma->vm_pgoff,
858 vma->vm_pgoff + vma_pages(vma));
859}
860
861/*
862 * SysV IPC SHM_UNLOCK restore Unevictable pages to their evictable lists.
863 */
864void shmem_unlock_mapping(struct address_space *mapping)
865{
866 struct folio_batch fbatch;
867 pgoff_t index = 0;
868
869 folio_batch_init(&fbatch);
870 /*
871 * Minor point, but we might as well stop if someone else SHM_LOCKs it.
872 */
873 while (!mapping_unevictable(mapping) &&
874 filemap_get_folios(mapping, &index, ~0UL, &fbatch)) {
875 check_move_unevictable_folios(&fbatch);
876 folio_batch_release(&fbatch);
877 cond_resched();
878 }
879}
880
881static struct folio *shmem_get_partial_folio(struct inode *inode, pgoff_t index)
882{
883 struct folio *folio;
884
885 /*
886 * At first avoid shmem_get_folio(,,,SGP_READ): that fails
887 * beyond i_size, and reports fallocated pages as holes.
888 */
889 folio = __filemap_get_folio(inode->i_mapping, index,
890 FGP_ENTRY | FGP_LOCK, 0);
891 if (!xa_is_value(folio))
892 return folio;
893 /*
894 * But read a page back from swap if any of it is within i_size
895 * (although in some cases this is just a waste of time).
896 */
897 folio = NULL;
898 shmem_get_folio(inode, index, &folio, SGP_READ);
899 return folio;
900}
901
902/*
903 * Remove range of pages and swap entries from page cache, and free them.
904 * If !unfalloc, truncate or punch hole; if unfalloc, undo failed fallocate.
905 */
906static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
907 bool unfalloc)
908{
909 struct address_space *mapping = inode->i_mapping;
910 struct shmem_inode_info *info = SHMEM_I(inode);
911 pgoff_t start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
912 pgoff_t end = (lend + 1) >> PAGE_SHIFT;
913 struct folio_batch fbatch;
914 pgoff_t indices[PAGEVEC_SIZE];
915 struct folio *folio;
916 bool same_folio;
917 long nr_swaps_freed = 0;
918 pgoff_t index;
919 int i;
920
921 if (lend == -1)
922 end = -1; /* unsigned, so actually very big */
923
924 if (info->fallocend > start && info->fallocend <= end && !unfalloc)
925 info->fallocend = start;
926
927 folio_batch_init(&fbatch);
928 index = start;
929 while (index < end && find_lock_entries(mapping, &index, end - 1,
930 &fbatch, indices)) {
931 for (i = 0; i < folio_batch_count(&fbatch); i++) {
932 folio = fbatch.folios[i];
933
934 if (xa_is_value(folio)) {
935 if (unfalloc)
936 continue;
937 nr_swaps_freed += !shmem_free_swap(mapping,
938 indices[i], folio);
939 continue;
940 }
941
942 if (!unfalloc || !folio_test_uptodate(folio))
943 truncate_inode_folio(mapping, folio);
944 folio_unlock(folio);
945 }
946 folio_batch_remove_exceptionals(&fbatch);
947 folio_batch_release(&fbatch);
948 cond_resched();
949 }
950
951 /*
952 * When undoing a failed fallocate, we want none of the partial folio
953 * zeroing and splitting below, but shall want to truncate the whole
954 * folio when !uptodate indicates that it was added by this fallocate,
955 * even when [lstart, lend] covers only a part of the folio.
956 */
957 if (unfalloc)
958 goto whole_folios;
959
960 same_folio = (lstart >> PAGE_SHIFT) == (lend >> PAGE_SHIFT);
961 folio = shmem_get_partial_folio(inode, lstart >> PAGE_SHIFT);
962 if (folio) {
963 same_folio = lend < folio_pos(folio) + folio_size(folio);
964 folio_mark_dirty(folio);
965 if (!truncate_inode_partial_folio(folio, lstart, lend)) {
966 start = folio->index + folio_nr_pages(folio);
967 if (same_folio)
968 end = folio->index;
969 }
970 folio_unlock(folio);
971 folio_put(folio);
972 folio = NULL;
973 }
974
975 if (!same_folio)
976 folio = shmem_get_partial_folio(inode, lend >> PAGE_SHIFT);
977 if (folio) {
978 folio_mark_dirty(folio);
979 if (!truncate_inode_partial_folio(folio, lstart, lend))
980 end = folio->index;
981 folio_unlock(folio);
982 folio_put(folio);
983 }
984
985whole_folios:
986
987 index = start;
988 while (index < end) {
989 cond_resched();
990
991 if (!find_get_entries(mapping, &index, end - 1, &fbatch,
992 indices)) {
993 /* If all gone or hole-punch or unfalloc, we're done */
994 if (index == start || end != -1)
995 break;
996 /* But if truncating, restart to make sure all gone */
997 index = start;
998 continue;
999 }
1000 for (i = 0; i < folio_batch_count(&fbatch); i++) {
1001 folio = fbatch.folios[i];
1002
1003 if (xa_is_value(folio)) {
1004 if (unfalloc)
1005 continue;
1006 if (shmem_free_swap(mapping, indices[i], folio)) {
1007 /* Swap was replaced by page: retry */
1008 index = indices[i];
1009 break;
1010 }
1011 nr_swaps_freed++;
1012 continue;
1013 }
1014
1015 folio_lock(folio);
1016
1017 if (!unfalloc || !folio_test_uptodate(folio)) {
1018 if (folio_mapping(folio) != mapping) {
1019 /* Page was replaced by swap: retry */
1020 folio_unlock(folio);
1021 index = indices[i];
1022 break;
1023 }
1024 VM_BUG_ON_FOLIO(folio_test_writeback(folio),
1025 folio);
1026 truncate_inode_folio(mapping, folio);
1027 }
1028 folio_unlock(folio);
1029 }
1030 folio_batch_remove_exceptionals(&fbatch);
1031 folio_batch_release(&fbatch);
1032 }
1033
1034 spin_lock_irq(&info->lock);
1035 info->swapped -= nr_swaps_freed;
1036 shmem_recalc_inode(inode);
1037 spin_unlock_irq(&info->lock);
1038}
1039
1040void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
1041{
1042 shmem_undo_range(inode, lstart, lend, false);
1043 inode->i_ctime = inode->i_mtime = current_time(inode);
1044 inode_inc_iversion(inode);
1045}
1046EXPORT_SYMBOL_GPL(shmem_truncate_range);
1047
1048static int shmem_getattr(struct user_namespace *mnt_userns,
1049 const struct path *path, struct kstat *stat,
1050 u32 request_mask, unsigned int query_flags)
1051{
1052 struct inode *inode = path->dentry->d_inode;
1053 struct shmem_inode_info *info = SHMEM_I(inode);
1054
1055 if (info->alloced - info->swapped != inode->i_mapping->nrpages) {
1056 spin_lock_irq(&info->lock);
1057 shmem_recalc_inode(inode);
1058 spin_unlock_irq(&info->lock);
1059 }
1060 if (info->fsflags & FS_APPEND_FL)
1061 stat->attributes |= STATX_ATTR_APPEND;
1062 if (info->fsflags & FS_IMMUTABLE_FL)
1063 stat->attributes |= STATX_ATTR_IMMUTABLE;
1064 if (info->fsflags & FS_NODUMP_FL)
1065 stat->attributes |= STATX_ATTR_NODUMP;
1066 stat->attributes_mask |= (STATX_ATTR_APPEND |
1067 STATX_ATTR_IMMUTABLE |
1068 STATX_ATTR_NODUMP);
1069 generic_fillattr(&init_user_ns, inode, stat);
1070
1071 if (shmem_is_huge(NULL, inode, 0, false))
1072 stat->blksize = HPAGE_PMD_SIZE;
1073
1074 if (request_mask & STATX_BTIME) {
1075 stat->result_mask |= STATX_BTIME;
1076 stat->btime.tv_sec = info->i_crtime.tv_sec;
1077 stat->btime.tv_nsec = info->i_crtime.tv_nsec;
1078 }
1079
1080 return 0;
1081}
1082
1083static int shmem_setattr(struct user_namespace *mnt_userns,
1084 struct dentry *dentry, struct iattr *attr)
1085{
1086 struct inode *inode = d_inode(dentry);
1087 struct shmem_inode_info *info = SHMEM_I(inode);
1088 int error;
1089 bool update_mtime = false;
1090 bool update_ctime = true;
1091
1092 error = setattr_prepare(&init_user_ns, dentry, attr);
1093 if (error)
1094 return error;
1095
1096 if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
1097 loff_t oldsize = inode->i_size;
1098 loff_t newsize = attr->ia_size;
1099
1100 /* protected by i_rwsem */
1101 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
1102 (newsize > oldsize && (info->seals & F_SEAL_GROW)))
1103 return -EPERM;
1104
1105 if (newsize != oldsize) {
1106 error = shmem_reacct_size(SHMEM_I(inode)->flags,
1107 oldsize, newsize);
1108 if (error)
1109 return error;
1110 i_size_write(inode, newsize);
1111 update_mtime = true;
1112 } else {
1113 update_ctime = false;
1114 }
1115 if (newsize <= oldsize) {
1116 loff_t holebegin = round_up(newsize, PAGE_SIZE);
1117 if (oldsize > holebegin)
1118 unmap_mapping_range(inode->i_mapping,
1119 holebegin, 0, 1);
1120 if (info->alloced)
1121 shmem_truncate_range(inode,
1122 newsize, (loff_t)-1);
1123 /* unmap again to remove racily COWed private pages */
1124 if (oldsize > holebegin)
1125 unmap_mapping_range(inode->i_mapping,
1126 holebegin, 0, 1);
1127 }
1128 }
1129
1130 setattr_copy(&init_user_ns, inode, attr);
1131 if (attr->ia_valid & ATTR_MODE)
1132 error = posix_acl_chmod(&init_user_ns, dentry, inode->i_mode);
1133 if (!error && update_ctime) {
1134 inode->i_ctime = current_time(inode);
1135 if (update_mtime)
1136 inode->i_mtime = inode->i_ctime;
1137 inode_inc_iversion(inode);
1138 }
1139 return error;
1140}
1141
1142static void shmem_evict_inode(struct inode *inode)
1143{
1144 struct shmem_inode_info *info = SHMEM_I(inode);
1145 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1146
1147 if (shmem_mapping(inode->i_mapping)) {
1148 shmem_unacct_size(info->flags, inode->i_size);
1149 inode->i_size = 0;
1150 mapping_set_exiting(inode->i_mapping);
1151 shmem_truncate_range(inode, 0, (loff_t)-1);
1152 if (!list_empty(&info->shrinklist)) {
1153 spin_lock(&sbinfo->shrinklist_lock);
1154 if (!list_empty(&info->shrinklist)) {
1155 list_del_init(&info->shrinklist);
1156 sbinfo->shrinklist_len--;
1157 }
1158 spin_unlock(&sbinfo->shrinklist_lock);
1159 }
1160 while (!list_empty(&info->swaplist)) {
1161 /* Wait while shmem_unuse() is scanning this inode... */
1162 wait_var_event(&info->stop_eviction,
1163 !atomic_read(&info->stop_eviction));
1164 mutex_lock(&shmem_swaplist_mutex);
1165 /* ...but beware of the race if we peeked too early */
1166 if (!atomic_read(&info->stop_eviction))
1167 list_del_init(&info->swaplist);
1168 mutex_unlock(&shmem_swaplist_mutex);
1169 }
1170 }
1171
1172 simple_xattrs_free(&info->xattrs);
1173 WARN_ON(inode->i_blocks);
1174 shmem_free_inode(inode->i_sb);
1175 clear_inode(inode);
1176}
1177
1178static int shmem_find_swap_entries(struct address_space *mapping,
1179 pgoff_t start, struct folio_batch *fbatch,
1180 pgoff_t *indices, unsigned int type)
1181{
1182 XA_STATE(xas, &mapping->i_pages, start);
1183 struct folio *folio;
1184 swp_entry_t entry;
1185
1186 rcu_read_lock();
1187 xas_for_each(&xas, folio, ULONG_MAX) {
1188 if (xas_retry(&xas, folio))
1189 continue;
1190
1191 if (!xa_is_value(folio))
1192 continue;
1193
1194 entry = radix_to_swp_entry(folio);
1195 /*
1196 * swapin error entries can be found in the mapping. But they're
1197 * deliberately ignored here as we've done everything we can do.
1198 */
1199 if (swp_type(entry) != type)
1200 continue;
1201
1202 indices[folio_batch_count(fbatch)] = xas.xa_index;
1203 if (!folio_batch_add(fbatch, folio))
1204 break;
1205
1206 if (need_resched()) {
1207 xas_pause(&xas);
1208 cond_resched_rcu();
1209 }
1210 }
1211 rcu_read_unlock();
1212
1213 return xas.xa_index;
1214}
1215
1216/*
1217 * Move the swapped pages for an inode to page cache. Returns the count
1218 * of pages swapped in, or the error in case of failure.
1219 */
1220static int shmem_unuse_swap_entries(struct inode *inode,
1221 struct folio_batch *fbatch, pgoff_t *indices)
1222{
1223 int i = 0;
1224 int ret = 0;
1225 int error = 0;
1226 struct address_space *mapping = inode->i_mapping;
1227
1228 for (i = 0; i < folio_batch_count(fbatch); i++) {
1229 struct folio *folio = fbatch->folios[i];
1230
1231 if (!xa_is_value(folio))
1232 continue;
1233 error = shmem_swapin_folio(inode, indices[i],
1234 &folio, SGP_CACHE,
1235 mapping_gfp_mask(mapping),
1236 NULL, NULL);
1237 if (error == 0) {
1238 folio_unlock(folio);
1239 folio_put(folio);
1240 ret++;
1241 }
1242 if (error == -ENOMEM)
1243 break;
1244 error = 0;
1245 }
1246 return error ? error : ret;
1247}
1248
1249/*
1250 * If swap found in inode, free it and move page from swapcache to filecache.
1251 */
1252static int shmem_unuse_inode(struct inode *inode, unsigned int type)
1253{
1254 struct address_space *mapping = inode->i_mapping;
1255 pgoff_t start = 0;
1256 struct folio_batch fbatch;
1257 pgoff_t indices[PAGEVEC_SIZE];
1258 int ret = 0;
1259
1260 do {
1261 folio_batch_init(&fbatch);
1262 shmem_find_swap_entries(mapping, start, &fbatch, indices, type);
1263 if (folio_batch_count(&fbatch) == 0) {
1264 ret = 0;
1265 break;
1266 }
1267
1268 ret = shmem_unuse_swap_entries(inode, &fbatch, indices);
1269 if (ret < 0)
1270 break;
1271
1272 start = indices[folio_batch_count(&fbatch) - 1];
1273 } while (true);
1274
1275 return ret;
1276}
1277
1278/*
1279 * Read all the shared memory data that resides in the swap
1280 * device 'type' back into memory, so the swap device can be
1281 * unused.
1282 */
1283int shmem_unuse(unsigned int type)
1284{
1285 struct shmem_inode_info *info, *next;
1286 int error = 0;
1287
1288 if (list_empty(&shmem_swaplist))
1289 return 0;
1290
1291 mutex_lock(&shmem_swaplist_mutex);
1292 list_for_each_entry_safe(info, next, &shmem_swaplist, swaplist) {
1293 if (!info->swapped) {
1294 list_del_init(&info->swaplist);
1295 continue;
1296 }
1297 /*
1298 * Drop the swaplist mutex while searching the inode for swap;
1299 * but before doing so, make sure shmem_evict_inode() will not
1300 * remove placeholder inode from swaplist, nor let it be freed
1301 * (igrab() would protect from unlink, but not from unmount).
1302 */
1303 atomic_inc(&info->stop_eviction);
1304 mutex_unlock(&shmem_swaplist_mutex);
1305
1306 error = shmem_unuse_inode(&info->vfs_inode, type);
1307 cond_resched();
1308
1309 mutex_lock(&shmem_swaplist_mutex);
1310 next = list_next_entry(info, swaplist);
1311 if (!info->swapped)
1312 list_del_init(&info->swaplist);
1313 if (atomic_dec_and_test(&info->stop_eviction))
1314 wake_up_var(&info->stop_eviction);
1315 if (error)
1316 break;
1317 }
1318 mutex_unlock(&shmem_swaplist_mutex);
1319
1320 return error;
1321}
1322
1323/*
1324 * Move the page from the page cache to the swap cache.
1325 */
1326static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1327{
1328 struct folio *folio = page_folio(page);
1329 struct shmem_inode_info *info;
1330 struct address_space *mapping;
1331 struct inode *inode;
1332 swp_entry_t swap;
1333 pgoff_t index;
1334
1335 /*
1336 * If /sys/kernel/mm/transparent_hugepage/shmem_enabled is "always" or
1337 * "force", drivers/gpu/drm/i915/gem/i915_gem_shmem.c gets huge pages,
1338 * and its shmem_writeback() needs them to be split when swapping.
1339 */
1340 if (folio_test_large(folio)) {
1341 /* Ensure the subpages are still dirty */
1342 folio_test_set_dirty(folio);
1343 if (split_huge_page(page) < 0)
1344 goto redirty;
1345 folio = page_folio(page);
1346 folio_clear_dirty(folio);
1347 }
1348
1349 BUG_ON(!folio_test_locked(folio));
1350 mapping = folio->mapping;
1351 index = folio->index;
1352 inode = mapping->host;
1353 info = SHMEM_I(inode);
1354 if (info->flags & VM_LOCKED)
1355 goto redirty;
1356 if (!total_swap_pages)
1357 goto redirty;
1358
1359 /*
1360 * Our capabilities prevent regular writeback or sync from ever calling
1361 * shmem_writepage; but a stacking filesystem might use ->writepage of
1362 * its underlying filesystem, in which case tmpfs should write out to
1363 * swap only in response to memory pressure, and not for the writeback
1364 * threads or sync.
1365 */
1366 if (!wbc->for_reclaim) {
1367 WARN_ON_ONCE(1); /* Still happens? Tell us about it! */
1368 goto redirty;
1369 }
1370
1371 /*
1372 * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
1373 * value into swapfile.c, the only way we can correctly account for a
1374 * fallocated folio arriving here is now to initialize it and write it.
1375 *
1376 * That's okay for a folio already fallocated earlier, but if we have
1377 * not yet completed the fallocation, then (a) we want to keep track
1378 * of this folio in case we have to undo it, and (b) it may not be a
1379 * good idea to continue anyway, once we're pushing into swap. So
1380 * reactivate the folio, and let shmem_fallocate() quit when too many.
1381 */
1382 if (!folio_test_uptodate(folio)) {
1383 if (inode->i_private) {
1384 struct shmem_falloc *shmem_falloc;
1385 spin_lock(&inode->i_lock);
1386 shmem_falloc = inode->i_private;
1387 if (shmem_falloc &&
1388 !shmem_falloc->waitq &&
1389 index >= shmem_falloc->start &&
1390 index < shmem_falloc->next)
1391 shmem_falloc->nr_unswapped++;
1392 else
1393 shmem_falloc = NULL;
1394 spin_unlock(&inode->i_lock);
1395 if (shmem_falloc)
1396 goto redirty;
1397 }
1398 folio_zero_range(folio, 0, folio_size(folio));
1399 flush_dcache_folio(folio);
1400 folio_mark_uptodate(folio);
1401 }
1402
1403 swap = folio_alloc_swap(folio);
1404 if (!swap.val)
1405 goto redirty;
1406
1407 /*
1408 * Add inode to shmem_unuse()'s list of swapped-out inodes,
1409 * if it's not already there. Do it now before the folio is
1410 * moved to swap cache, when its pagelock no longer protects
1411 * the inode from eviction. But don't unlock the mutex until
1412 * we've incremented swapped, because shmem_unuse_inode() will
1413 * prune a !swapped inode from the swaplist under this mutex.
1414 */
1415 mutex_lock(&shmem_swaplist_mutex);
1416 if (list_empty(&info->swaplist))
1417 list_add(&info->swaplist, &shmem_swaplist);
1418
1419 if (add_to_swap_cache(folio, swap,
1420 __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN,
1421 NULL) == 0) {
1422 spin_lock_irq(&info->lock);
1423 shmem_recalc_inode(inode);
1424 info->swapped++;
1425 spin_unlock_irq(&info->lock);
1426
1427 swap_shmem_alloc(swap);
1428 shmem_delete_from_page_cache(folio, swp_to_radix_entry(swap));
1429
1430 mutex_unlock(&shmem_swaplist_mutex);
1431 BUG_ON(folio_mapped(folio));
1432 swap_writepage(&folio->page, wbc);
1433 return 0;
1434 }
1435
1436 mutex_unlock(&shmem_swaplist_mutex);
1437 put_swap_folio(folio, swap);
1438redirty:
1439 folio_mark_dirty(folio);
1440 if (wbc->for_reclaim)
1441 return AOP_WRITEPAGE_ACTIVATE; /* Return with folio locked */
1442 folio_unlock(folio);
1443 return 0;
1444}
1445
1446#if defined(CONFIG_NUMA) && defined(CONFIG_TMPFS)
1447static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1448{
1449 char buffer[64];
1450
1451 if (!mpol || mpol->mode == MPOL_DEFAULT)
1452 return; /* show nothing */
1453
1454 mpol_to_str(buffer, sizeof(buffer), mpol);
1455
1456 seq_printf(seq, ",mpol=%s", buffer);
1457}
1458
1459static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1460{
1461 struct mempolicy *mpol = NULL;
1462 if (sbinfo->mpol) {
1463 raw_spin_lock(&sbinfo->stat_lock); /* prevent replace/use races */
1464 mpol = sbinfo->mpol;
1465 mpol_get(mpol);
1466 raw_spin_unlock(&sbinfo->stat_lock);
1467 }
1468 return mpol;
1469}
1470#else /* !CONFIG_NUMA || !CONFIG_TMPFS */
1471static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1472{
1473}
1474static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1475{
1476 return NULL;
1477}
1478#endif /* CONFIG_NUMA && CONFIG_TMPFS */
1479#ifndef CONFIG_NUMA
1480#define vm_policy vm_private_data
1481#endif
1482
1483static void shmem_pseudo_vma_init(struct vm_area_struct *vma,
1484 struct shmem_inode_info *info, pgoff_t index)
1485{
1486 /* Create a pseudo vma that just contains the policy */
1487 vma_init(vma, NULL);
1488 /* Bias interleave by inode number to distribute better across nodes */
1489 vma->vm_pgoff = index + info->vfs_inode.i_ino;
1490 vma->vm_policy = mpol_shared_policy_lookup(&info->policy, index);
1491}
1492
1493static void shmem_pseudo_vma_destroy(struct vm_area_struct *vma)
1494{
1495 /* Drop reference taken by mpol_shared_policy_lookup() */
1496 mpol_cond_put(vma->vm_policy);
1497}
1498
1499static struct folio *shmem_swapin(swp_entry_t swap, gfp_t gfp,
1500 struct shmem_inode_info *info, pgoff_t index)
1501{
1502 struct vm_area_struct pvma;
1503 struct page *page;
1504 struct vm_fault vmf = {
1505 .vma = &pvma,
1506 };
1507
1508 shmem_pseudo_vma_init(&pvma, info, index);
1509 page = swap_cluster_readahead(swap, gfp, &vmf);
1510 shmem_pseudo_vma_destroy(&pvma);
1511
1512 if (!page)
1513 return NULL;
1514 return page_folio(page);
1515}
1516
1517/*
1518 * Make sure huge_gfp is always more limited than limit_gfp.
1519 * Some of the flags set permissions, while others set limitations.
1520 */
1521static gfp_t limit_gfp_mask(gfp_t huge_gfp, gfp_t limit_gfp)
1522{
1523 gfp_t allowflags = __GFP_IO | __GFP_FS | __GFP_RECLAIM;
1524 gfp_t denyflags = __GFP_NOWARN | __GFP_NORETRY;
1525 gfp_t zoneflags = limit_gfp & GFP_ZONEMASK;
1526 gfp_t result = huge_gfp & ~(allowflags | GFP_ZONEMASK);
1527
1528 /* Allow allocations only from the originally specified zones. */
1529 result |= zoneflags;
1530
1531 /*
1532 * Minimize the result gfp by taking the union with the deny flags,
1533 * and the intersection of the allow flags.
1534 */
1535 result |= (limit_gfp & denyflags);
1536 result |= (huge_gfp & limit_gfp) & allowflags;
1537
1538 return result;
1539}
1540
1541static struct folio *shmem_alloc_hugefolio(gfp_t gfp,
1542 struct shmem_inode_info *info, pgoff_t index)
1543{
1544 struct vm_area_struct pvma;
1545 struct address_space *mapping = info->vfs_inode.i_mapping;
1546 pgoff_t hindex;
1547 struct folio *folio;
1548
1549 hindex = round_down(index, HPAGE_PMD_NR);
1550 if (xa_find(&mapping->i_pages, &hindex, hindex + HPAGE_PMD_NR - 1,
1551 XA_PRESENT))
1552 return NULL;
1553
1554 shmem_pseudo_vma_init(&pvma, info, hindex);
1555 folio = vma_alloc_folio(gfp, HPAGE_PMD_ORDER, &pvma, 0, true);
1556 shmem_pseudo_vma_destroy(&pvma);
1557 if (!folio)
1558 count_vm_event(THP_FILE_FALLBACK);
1559 return folio;
1560}
1561
1562static struct folio *shmem_alloc_folio(gfp_t gfp,
1563 struct shmem_inode_info *info, pgoff_t index)
1564{
1565 struct vm_area_struct pvma;
1566 struct folio *folio;
1567
1568 shmem_pseudo_vma_init(&pvma, info, index);
1569 folio = vma_alloc_folio(gfp, 0, &pvma, 0, false);
1570 shmem_pseudo_vma_destroy(&pvma);
1571
1572 return folio;
1573}
1574
1575static struct folio *shmem_alloc_and_acct_folio(gfp_t gfp, struct inode *inode,
1576 pgoff_t index, bool huge)
1577{
1578 struct shmem_inode_info *info = SHMEM_I(inode);
1579 struct folio *folio;
1580 int nr;
1581 int err = -ENOSPC;
1582
1583 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1584 huge = false;
1585 nr = huge ? HPAGE_PMD_NR : 1;
1586
1587 if (!shmem_inode_acct_block(inode, nr))
1588 goto failed;
1589
1590 if (huge)
1591 folio = shmem_alloc_hugefolio(gfp, info, index);
1592 else
1593 folio = shmem_alloc_folio(gfp, info, index);
1594 if (folio) {
1595 __folio_set_locked(folio);
1596 __folio_set_swapbacked(folio);
1597 return folio;
1598 }
1599
1600 err = -ENOMEM;
1601 shmem_inode_unacct_blocks(inode, nr);
1602failed:
1603 return ERR_PTR(err);
1604}
1605
1606/*
1607 * When a page is moved from swapcache to shmem filecache (either by the
1608 * usual swapin of shmem_get_folio_gfp(), or by the less common swapoff of
1609 * shmem_unuse_inode()), it may have been read in earlier from swap, in
1610 * ignorance of the mapping it belongs to. If that mapping has special
1611 * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
1612 * we may need to copy to a suitable page before moving to filecache.
1613 *
1614 * In a future release, this may well be extended to respect cpuset and
1615 * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
1616 * but for now it is a simple matter of zone.
1617 */
1618static bool shmem_should_replace_folio(struct folio *folio, gfp_t gfp)
1619{
1620 return folio_zonenum(folio) > gfp_zone(gfp);
1621}
1622
1623static int shmem_replace_folio(struct folio **foliop, gfp_t gfp,
1624 struct shmem_inode_info *info, pgoff_t index)
1625{
1626 struct folio *old, *new;
1627 struct address_space *swap_mapping;
1628 swp_entry_t entry;
1629 pgoff_t swap_index;
1630 int error;
1631
1632 old = *foliop;
1633 entry = folio_swap_entry(old);
1634 swap_index = swp_offset(entry);
1635 swap_mapping = swap_address_space(entry);
1636
1637 /*
1638 * We have arrived here because our zones are constrained, so don't
1639 * limit chance of success by further cpuset and node constraints.
1640 */
1641 gfp &= ~GFP_CONSTRAINT_MASK;
1642 VM_BUG_ON_FOLIO(folio_test_large(old), old);
1643 new = shmem_alloc_folio(gfp, info, index);
1644 if (!new)
1645 return -ENOMEM;
1646
1647 folio_get(new);
1648 folio_copy(new, old);
1649 flush_dcache_folio(new);
1650
1651 __folio_set_locked(new);
1652 __folio_set_swapbacked(new);
1653 folio_mark_uptodate(new);
1654 folio_set_swap_entry(new, entry);
1655 folio_set_swapcache(new);
1656
1657 /*
1658 * Our caller will very soon move newpage out of swapcache, but it's
1659 * a nice clean interface for us to replace oldpage by newpage there.
1660 */
1661 xa_lock_irq(&swap_mapping->i_pages);
1662 error = shmem_replace_entry(swap_mapping, swap_index, old, new);
1663 if (!error) {
1664 mem_cgroup_migrate(old, new);
1665 __lruvec_stat_mod_folio(new, NR_FILE_PAGES, 1);
1666 __lruvec_stat_mod_folio(new, NR_SHMEM, 1);
1667 __lruvec_stat_mod_folio(old, NR_FILE_PAGES, -1);
1668 __lruvec_stat_mod_folio(old, NR_SHMEM, -1);
1669 }
1670 xa_unlock_irq(&swap_mapping->i_pages);
1671
1672 if (unlikely(error)) {
1673 /*
1674 * Is this possible? I think not, now that our callers check
1675 * both PageSwapCache and page_private after getting page lock;
1676 * but be defensive. Reverse old to newpage for clear and free.
1677 */
1678 old = new;
1679 } else {
1680 folio_add_lru(new);
1681 *foliop = new;
1682 }
1683
1684 folio_clear_swapcache(old);
1685 old->private = NULL;
1686
1687 folio_unlock(old);
1688 folio_put_refs(old, 2);
1689 return error;
1690}
1691
1692static void shmem_set_folio_swapin_error(struct inode *inode, pgoff_t index,
1693 struct folio *folio, swp_entry_t swap)
1694{
1695 struct address_space *mapping = inode->i_mapping;
1696 struct shmem_inode_info *info = SHMEM_I(inode);
1697 swp_entry_t swapin_error;
1698 void *old;
1699
1700 swapin_error = make_swapin_error_entry();
1701 old = xa_cmpxchg_irq(&mapping->i_pages, index,
1702 swp_to_radix_entry(swap),
1703 swp_to_radix_entry(swapin_error), 0);
1704 if (old != swp_to_radix_entry(swap))
1705 return;
1706
1707 folio_wait_writeback(folio);
1708 delete_from_swap_cache(folio);
1709 spin_lock_irq(&info->lock);
1710 /*
1711 * Don't treat swapin error folio as alloced. Otherwise inode->i_blocks won't
1712 * be 0 when inode is released and thus trigger WARN_ON(inode->i_blocks) in
1713 * shmem_evict_inode.
1714 */
1715 info->alloced--;
1716 info->swapped--;
1717 shmem_recalc_inode(inode);
1718 spin_unlock_irq(&info->lock);
1719 swap_free(swap);
1720}
1721
1722/*
1723 * Swap in the folio pointed to by *foliop.
1724 * Caller has to make sure that *foliop contains a valid swapped folio.
1725 * Returns 0 and the folio in foliop if success. On failure, returns the
1726 * error code and NULL in *foliop.
1727 */
1728static int shmem_swapin_folio(struct inode *inode, pgoff_t index,
1729 struct folio **foliop, enum sgp_type sgp,
1730 gfp_t gfp, struct vm_area_struct *vma,
1731 vm_fault_t *fault_type)
1732{
1733 struct address_space *mapping = inode->i_mapping;
1734 struct shmem_inode_info *info = SHMEM_I(inode);
1735 struct mm_struct *charge_mm = vma ? vma->vm_mm : NULL;
1736 struct folio *folio = NULL;
1737 swp_entry_t swap;
1738 int error;
1739
1740 VM_BUG_ON(!*foliop || !xa_is_value(*foliop));
1741 swap = radix_to_swp_entry(*foliop);
1742 *foliop = NULL;
1743
1744 if (is_swapin_error_entry(swap))
1745 return -EIO;
1746
1747 /* Look it up and read it in.. */
1748 folio = swap_cache_get_folio(swap, NULL, 0);
1749 if (!folio) {
1750 /* Or update major stats only when swapin succeeds?? */
1751 if (fault_type) {
1752 *fault_type |= VM_FAULT_MAJOR;
1753 count_vm_event(PGMAJFAULT);
1754 count_memcg_event_mm(charge_mm, PGMAJFAULT);
1755 }
1756 /* Here we actually start the io */
1757 folio = shmem_swapin(swap, gfp, info, index);
1758 if (!folio) {
1759 error = -ENOMEM;
1760 goto failed;
1761 }
1762 }
1763
1764 /* We have to do this with folio locked to prevent races */
1765 folio_lock(folio);
1766 if (!folio_test_swapcache(folio) ||
1767 folio_swap_entry(folio).val != swap.val ||
1768 !shmem_confirm_swap(mapping, index, swap)) {
1769 error = -EEXIST;
1770 goto unlock;
1771 }
1772 if (!folio_test_uptodate(folio)) {
1773 error = -EIO;
1774 goto failed;
1775 }
1776 folio_wait_writeback(folio);
1777
1778 /*
1779 * Some architectures may have to restore extra metadata to the
1780 * folio after reading from swap.
1781 */
1782 arch_swap_restore(swap, folio);
1783
1784 if (shmem_should_replace_folio(folio, gfp)) {
1785 error = shmem_replace_folio(&folio, gfp, info, index);
1786 if (error)
1787 goto failed;
1788 }
1789
1790 error = shmem_add_to_page_cache(folio, mapping, index,
1791 swp_to_radix_entry(swap), gfp,
1792 charge_mm);
1793 if (error)
1794 goto failed;
1795
1796 spin_lock_irq(&info->lock);
1797 info->swapped--;
1798 shmem_recalc_inode(inode);
1799 spin_unlock_irq(&info->lock);
1800
1801 if (sgp == SGP_WRITE)
1802 folio_mark_accessed(folio);
1803
1804 delete_from_swap_cache(folio);
1805 folio_mark_dirty(folio);
1806 swap_free(swap);
1807
1808 *foliop = folio;
1809 return 0;
1810failed:
1811 if (!shmem_confirm_swap(mapping, index, swap))
1812 error = -EEXIST;
1813 if (error == -EIO)
1814 shmem_set_folio_swapin_error(inode, index, folio, swap);
1815unlock:
1816 if (folio) {
1817 folio_unlock(folio);
1818 folio_put(folio);
1819 }
1820
1821 return error;
1822}
1823
1824/*
1825 * shmem_get_folio_gfp - find page in cache, or get from swap, or allocate
1826 *
1827 * If we allocate a new one we do not mark it dirty. That's up to the
1828 * vm. If we swap it in we mark it dirty since we also free the swap
1829 * entry since a page cannot live in both the swap and page cache.
1830 *
1831 * vma, vmf, and fault_type are only supplied by shmem_fault:
1832 * otherwise they are NULL.
1833 */
1834static int shmem_get_folio_gfp(struct inode *inode, pgoff_t index,
1835 struct folio **foliop, enum sgp_type sgp, gfp_t gfp,
1836 struct vm_area_struct *vma, struct vm_fault *vmf,
1837 vm_fault_t *fault_type)
1838{
1839 struct address_space *mapping = inode->i_mapping;
1840 struct shmem_inode_info *info = SHMEM_I(inode);
1841 struct shmem_sb_info *sbinfo;
1842 struct mm_struct *charge_mm;
1843 struct folio *folio;
1844 pgoff_t hindex;
1845 gfp_t huge_gfp;
1846 int error;
1847 int once = 0;
1848 int alloced = 0;
1849
1850 if (index > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
1851 return -EFBIG;
1852repeat:
1853 if (sgp <= SGP_CACHE &&
1854 ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1855 return -EINVAL;
1856 }
1857
1858 sbinfo = SHMEM_SB(inode->i_sb);
1859 charge_mm = vma ? vma->vm_mm : NULL;
1860
1861 folio = __filemap_get_folio(mapping, index, FGP_ENTRY | FGP_LOCK, 0);
1862 if (folio && vma && userfaultfd_minor(vma)) {
1863 if (!xa_is_value(folio)) {
1864 folio_unlock(folio);
1865 folio_put(folio);
1866 }
1867 *fault_type = handle_userfault(vmf, VM_UFFD_MINOR);
1868 return 0;
1869 }
1870
1871 if (xa_is_value(folio)) {
1872 error = shmem_swapin_folio(inode, index, &folio,
1873 sgp, gfp, vma, fault_type);
1874 if (error == -EEXIST)
1875 goto repeat;
1876
1877 *foliop = folio;
1878 return error;
1879 }
1880
1881 if (folio) {
1882 if (sgp == SGP_WRITE)
1883 folio_mark_accessed(folio);
1884 if (folio_test_uptodate(folio))
1885 goto out;
1886 /* fallocated folio */
1887 if (sgp != SGP_READ)
1888 goto clear;
1889 folio_unlock(folio);
1890 folio_put(folio);
1891 }
1892
1893 /*
1894 * SGP_READ: succeed on hole, with NULL folio, letting caller zero.
1895 * SGP_NOALLOC: fail on hole, with NULL folio, letting caller fail.
1896 */
1897 *foliop = NULL;
1898 if (sgp == SGP_READ)
1899 return 0;
1900 if (sgp == SGP_NOALLOC)
1901 return -ENOENT;
1902
1903 /*
1904 * Fast cache lookup and swap lookup did not find it: allocate.
1905 */
1906
1907 if (vma && userfaultfd_missing(vma)) {
1908 *fault_type = handle_userfault(vmf, VM_UFFD_MISSING);
1909 return 0;
1910 }
1911
1912 if (!shmem_is_huge(vma, inode, index, false))
1913 goto alloc_nohuge;
1914
1915 huge_gfp = vma_thp_gfp_mask(vma);
1916 huge_gfp = limit_gfp_mask(huge_gfp, gfp);
1917 folio = shmem_alloc_and_acct_folio(huge_gfp, inode, index, true);
1918 if (IS_ERR(folio)) {
1919alloc_nohuge:
1920 folio = shmem_alloc_and_acct_folio(gfp, inode, index, false);
1921 }
1922 if (IS_ERR(folio)) {
1923 int retry = 5;
1924
1925 error = PTR_ERR(folio);
1926 folio = NULL;
1927 if (error != -ENOSPC)
1928 goto unlock;
1929 /*
1930 * Try to reclaim some space by splitting a large folio
1931 * beyond i_size on the filesystem.
1932 */
1933 while (retry--) {
1934 int ret;
1935
1936 ret = shmem_unused_huge_shrink(sbinfo, NULL, 1);
1937 if (ret == SHRINK_STOP)
1938 break;
1939 if (ret)
1940 goto alloc_nohuge;
1941 }
1942 goto unlock;
1943 }
1944
1945 hindex = round_down(index, folio_nr_pages(folio));
1946
1947 if (sgp == SGP_WRITE)
1948 __folio_set_referenced(folio);
1949
1950 error = shmem_add_to_page_cache(folio, mapping, hindex,
1951 NULL, gfp & GFP_RECLAIM_MASK,
1952 charge_mm);
1953 if (error)
1954 goto unacct;
1955 folio_add_lru(folio);
1956
1957 spin_lock_irq(&info->lock);
1958 info->alloced += folio_nr_pages(folio);
1959 inode->i_blocks += (blkcnt_t)BLOCKS_PER_PAGE << folio_order(folio);
1960 shmem_recalc_inode(inode);
1961 spin_unlock_irq(&info->lock);
1962 alloced = true;
1963
1964 if (folio_test_pmd_mappable(folio) &&
1965 DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE) <
1966 folio_next_index(folio) - 1) {
1967 /*
1968 * Part of the large folio is beyond i_size: subject
1969 * to shrink under memory pressure.
1970 */
1971 spin_lock(&sbinfo->shrinklist_lock);
1972 /*
1973 * _careful to defend against unlocked access to
1974 * ->shrink_list in shmem_unused_huge_shrink()
1975 */
1976 if (list_empty_careful(&info->shrinklist)) {
1977 list_add_tail(&info->shrinklist,
1978 &sbinfo->shrinklist);
1979 sbinfo->shrinklist_len++;
1980 }
1981 spin_unlock(&sbinfo->shrinklist_lock);
1982 }
1983
1984 /*
1985 * Let SGP_FALLOC use the SGP_WRITE optimization on a new folio.
1986 */
1987 if (sgp == SGP_FALLOC)
1988 sgp = SGP_WRITE;
1989clear:
1990 /*
1991 * Let SGP_WRITE caller clear ends if write does not fill folio;
1992 * but SGP_FALLOC on a folio fallocated earlier must initialize
1993 * it now, lest undo on failure cancel our earlier guarantee.
1994 */
1995 if (sgp != SGP_WRITE && !folio_test_uptodate(folio)) {
1996 long i, n = folio_nr_pages(folio);
1997
1998 for (i = 0; i < n; i++)
1999 clear_highpage(folio_page(folio, i));
2000 flush_dcache_folio(folio);
2001 folio_mark_uptodate(folio);
2002 }
2003
2004 /* Perhaps the file has been truncated since we checked */
2005 if (sgp <= SGP_CACHE &&
2006 ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
2007 if (alloced) {
2008 folio_clear_dirty(folio);
2009 filemap_remove_folio(folio);
2010 spin_lock_irq(&info->lock);
2011 shmem_recalc_inode(inode);
2012 spin_unlock_irq(&info->lock);
2013 }
2014 error = -EINVAL;
2015 goto unlock;
2016 }
2017out:
2018 *foliop = folio;
2019 return 0;
2020
2021 /*
2022 * Error recovery.
2023 */
2024unacct:
2025 shmem_inode_unacct_blocks(inode, folio_nr_pages(folio));
2026
2027 if (folio_test_large(folio)) {
2028 folio_unlock(folio);
2029 folio_put(folio);
2030 goto alloc_nohuge;
2031 }
2032unlock:
2033 if (folio) {
2034 folio_unlock(folio);
2035 folio_put(folio);
2036 }
2037 if (error == -ENOSPC && !once++) {
2038 spin_lock_irq(&info->lock);
2039 shmem_recalc_inode(inode);
2040 spin_unlock_irq(&info->lock);
2041 goto repeat;
2042 }
2043 if (error == -EEXIST)
2044 goto repeat;
2045 return error;
2046}
2047
2048int shmem_get_folio(struct inode *inode, pgoff_t index, struct folio **foliop,
2049 enum sgp_type sgp)
2050{
2051 return shmem_get_folio_gfp(inode, index, foliop, sgp,
2052 mapping_gfp_mask(inode->i_mapping), NULL, NULL, NULL);
2053}
2054
2055/*
2056 * This is like autoremove_wake_function, but it removes the wait queue
2057 * entry unconditionally - even if something else had already woken the
2058 * target.
2059 */
2060static int synchronous_wake_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
2061{
2062 int ret = default_wake_function(wait, mode, sync, key);
2063 list_del_init(&wait->entry);
2064 return ret;
2065}
2066
2067static vm_fault_t shmem_fault(struct vm_fault *vmf)
2068{
2069 struct vm_area_struct *vma = vmf->vma;
2070 struct inode *inode = file_inode(vma->vm_file);
2071 gfp_t gfp = mapping_gfp_mask(inode->i_mapping);
2072 struct folio *folio = NULL;
2073 int err;
2074 vm_fault_t ret = VM_FAULT_LOCKED;
2075
2076 /*
2077 * Trinity finds that probing a hole which tmpfs is punching can
2078 * prevent the hole-punch from ever completing: which in turn
2079 * locks writers out with its hold on i_rwsem. So refrain from
2080 * faulting pages into the hole while it's being punched. Although
2081 * shmem_undo_range() does remove the additions, it may be unable to
2082 * keep up, as each new page needs its own unmap_mapping_range() call,
2083 * and the i_mmap tree grows ever slower to scan if new vmas are added.
2084 *
2085 * It does not matter if we sometimes reach this check just before the
2086 * hole-punch begins, so that one fault then races with the punch:
2087 * we just need to make racing faults a rare case.
2088 *
2089 * The implementation below would be much simpler if we just used a
2090 * standard mutex or completion: but we cannot take i_rwsem in fault,
2091 * and bloating every shmem inode for this unlikely case would be sad.
2092 */
2093 if (unlikely(inode->i_private)) {
2094 struct shmem_falloc *shmem_falloc;
2095
2096 spin_lock(&inode->i_lock);
2097 shmem_falloc = inode->i_private;
2098 if (shmem_falloc &&
2099 shmem_falloc->waitq &&
2100 vmf->pgoff >= shmem_falloc->start &&
2101 vmf->pgoff < shmem_falloc->next) {
2102 struct file *fpin;
2103 wait_queue_head_t *shmem_falloc_waitq;
2104 DEFINE_WAIT_FUNC(shmem_fault_wait, synchronous_wake_function);
2105
2106 ret = VM_FAULT_NOPAGE;
2107 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2108 if (fpin)
2109 ret = VM_FAULT_RETRY;
2110
2111 shmem_falloc_waitq = shmem_falloc->waitq;
2112 prepare_to_wait(shmem_falloc_waitq, &shmem_fault_wait,
2113 TASK_UNINTERRUPTIBLE);
2114 spin_unlock(&inode->i_lock);
2115 schedule();
2116
2117 /*
2118 * shmem_falloc_waitq points into the shmem_fallocate()
2119 * stack of the hole-punching task: shmem_falloc_waitq
2120 * is usually invalid by the time we reach here, but
2121 * finish_wait() does not dereference it in that case;
2122 * though i_lock needed lest racing with wake_up_all().
2123 */
2124 spin_lock(&inode->i_lock);
2125 finish_wait(shmem_falloc_waitq, &shmem_fault_wait);
2126 spin_unlock(&inode->i_lock);
2127
2128 if (fpin)
2129 fput(fpin);
2130 return ret;
2131 }
2132 spin_unlock(&inode->i_lock);
2133 }
2134
2135 err = shmem_get_folio_gfp(inode, vmf->pgoff, &folio, SGP_CACHE,
2136 gfp, vma, vmf, &ret);
2137 if (err)
2138 return vmf_error(err);
2139 if (folio)
2140 vmf->page = folio_file_page(folio, vmf->pgoff);
2141 return ret;
2142}
2143
2144unsigned long shmem_get_unmapped_area(struct file *file,
2145 unsigned long uaddr, unsigned long len,
2146 unsigned long pgoff, unsigned long flags)
2147{
2148 unsigned long (*get_area)(struct file *,
2149 unsigned long, unsigned long, unsigned long, unsigned long);
2150 unsigned long addr;
2151 unsigned long offset;
2152 unsigned long inflated_len;
2153 unsigned long inflated_addr;
2154 unsigned long inflated_offset;
2155
2156 if (len > TASK_SIZE)
2157 return -ENOMEM;
2158
2159 get_area = current->mm->get_unmapped_area;
2160 addr = get_area(file, uaddr, len, pgoff, flags);
2161
2162 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
2163 return addr;
2164 if (IS_ERR_VALUE(addr))
2165 return addr;
2166 if (addr & ~PAGE_MASK)
2167 return addr;
2168 if (addr > TASK_SIZE - len)
2169 return addr;
2170
2171 if (shmem_huge == SHMEM_HUGE_DENY)
2172 return addr;
2173 if (len < HPAGE_PMD_SIZE)
2174 return addr;
2175 if (flags & MAP_FIXED)
2176 return addr;
2177 /*
2178 * Our priority is to support MAP_SHARED mapped hugely;
2179 * and support MAP_PRIVATE mapped hugely too, until it is COWed.
2180 * But if caller specified an address hint and we allocated area there
2181 * successfully, respect that as before.
2182 */
2183 if (uaddr == addr)
2184 return addr;
2185
2186 if (shmem_huge != SHMEM_HUGE_FORCE) {
2187 struct super_block *sb;
2188
2189 if (file) {
2190 VM_BUG_ON(file->f_op != &shmem_file_operations);
2191 sb = file_inode(file)->i_sb;
2192 } else {
2193 /*
2194 * Called directly from mm/mmap.c, or drivers/char/mem.c
2195 * for "/dev/zero", to create a shared anonymous object.
2196 */
2197 if (IS_ERR(shm_mnt))
2198 return addr;
2199 sb = shm_mnt->mnt_sb;
2200 }
2201 if (SHMEM_SB(sb)->huge == SHMEM_HUGE_NEVER)
2202 return addr;
2203 }
2204
2205 offset = (pgoff << PAGE_SHIFT) & (HPAGE_PMD_SIZE-1);
2206 if (offset && offset + len < 2 * HPAGE_PMD_SIZE)
2207 return addr;
2208 if ((addr & (HPAGE_PMD_SIZE-1)) == offset)
2209 return addr;
2210
2211 inflated_len = len + HPAGE_PMD_SIZE - PAGE_SIZE;
2212 if (inflated_len > TASK_SIZE)
2213 return addr;
2214 if (inflated_len < len)
2215 return addr;
2216
2217 inflated_addr = get_area(NULL, uaddr, inflated_len, 0, flags);
2218 if (IS_ERR_VALUE(inflated_addr))
2219 return addr;
2220 if (inflated_addr & ~PAGE_MASK)
2221 return addr;
2222
2223 inflated_offset = inflated_addr & (HPAGE_PMD_SIZE-1);
2224 inflated_addr += offset - inflated_offset;
2225 if (inflated_offset > offset)
2226 inflated_addr += HPAGE_PMD_SIZE;
2227
2228 if (inflated_addr > TASK_SIZE - len)
2229 return addr;
2230 return inflated_addr;
2231}
2232
2233#ifdef CONFIG_NUMA
2234static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
2235{
2236 struct inode *inode = file_inode(vma->vm_file);
2237 return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
2238}
2239
2240static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
2241 unsigned long addr)
2242{
2243 struct inode *inode = file_inode(vma->vm_file);
2244 pgoff_t index;
2245
2246 index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2247 return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
2248}
2249#endif
2250
2251int shmem_lock(struct file *file, int lock, struct ucounts *ucounts)
2252{
2253 struct inode *inode = file_inode(file);
2254 struct shmem_inode_info *info = SHMEM_I(inode);
2255 int retval = -ENOMEM;
2256
2257 /*
2258 * What serializes the accesses to info->flags?
2259 * ipc_lock_object() when called from shmctl_do_lock(),
2260 * no serialization needed when called from shm_destroy().
2261 */
2262 if (lock && !(info->flags & VM_LOCKED)) {
2263 if (!user_shm_lock(inode->i_size, ucounts))
2264 goto out_nomem;
2265 info->flags |= VM_LOCKED;
2266 mapping_set_unevictable(file->f_mapping);
2267 }
2268 if (!lock && (info->flags & VM_LOCKED) && ucounts) {
2269 user_shm_unlock(inode->i_size, ucounts);
2270 info->flags &= ~VM_LOCKED;
2271 mapping_clear_unevictable(file->f_mapping);
2272 }
2273 retval = 0;
2274
2275out_nomem:
2276 return retval;
2277}
2278
2279static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
2280{
2281 struct inode *inode = file_inode(file);
2282 struct shmem_inode_info *info = SHMEM_I(inode);
2283 int ret;
2284
2285 ret = seal_check_future_write(info->seals, vma);
2286 if (ret)
2287 return ret;
2288
2289 /* arm64 - allow memory tagging on RAM-based files */
2290 vma->vm_flags |= VM_MTE_ALLOWED;
2291
2292 file_accessed(file);
2293 /* This is anonymous shared memory if it is unlinked at the time of mmap */
2294 if (inode->i_nlink)
2295 vma->vm_ops = &shmem_vm_ops;
2296 else
2297 vma->vm_ops = &shmem_anon_vm_ops;
2298 return 0;
2299}
2300
2301#ifdef CONFIG_TMPFS_XATTR
2302static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
2303
2304/*
2305 * chattr's fsflags are unrelated to extended attributes,
2306 * but tmpfs has chosen to enable them under the same config option.
2307 */
2308static void shmem_set_inode_flags(struct inode *inode, unsigned int fsflags)
2309{
2310 unsigned int i_flags = 0;
2311
2312 if (fsflags & FS_NOATIME_FL)
2313 i_flags |= S_NOATIME;
2314 if (fsflags & FS_APPEND_FL)
2315 i_flags |= S_APPEND;
2316 if (fsflags & FS_IMMUTABLE_FL)
2317 i_flags |= S_IMMUTABLE;
2318 /*
2319 * But FS_NODUMP_FL does not require any action in i_flags.
2320 */
2321 inode_set_flags(inode, i_flags, S_NOATIME | S_APPEND | S_IMMUTABLE);
2322}
2323#else
2324static void shmem_set_inode_flags(struct inode *inode, unsigned int fsflags)
2325{
2326}
2327#define shmem_initxattrs NULL
2328#endif
2329
2330static struct inode *shmem_get_inode(struct super_block *sb, struct inode *dir,
2331 umode_t mode, dev_t dev, unsigned long flags)
2332{
2333 struct inode *inode;
2334 struct shmem_inode_info *info;
2335 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2336 ino_t ino;
2337
2338 if (shmem_reserve_inode(sb, &ino))
2339 return NULL;
2340
2341 inode = new_inode(sb);
2342 if (inode) {
2343 inode->i_ino = ino;
2344 inode_init_owner(&init_user_ns, inode, dir, mode);
2345 inode->i_blocks = 0;
2346 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
2347 inode->i_generation = get_random_u32();
2348 info = SHMEM_I(inode);
2349 memset(info, 0, (char *)inode - (char *)info);
2350 spin_lock_init(&info->lock);
2351 atomic_set(&info->stop_eviction, 0);
2352 info->seals = F_SEAL_SEAL;
2353 info->flags = flags & VM_NORESERVE;
2354 info->i_crtime = inode->i_mtime;
2355 info->fsflags = (dir == NULL) ? 0 :
2356 SHMEM_I(dir)->fsflags & SHMEM_FL_INHERITED;
2357 if (info->fsflags)
2358 shmem_set_inode_flags(inode, info->fsflags);
2359 INIT_LIST_HEAD(&info->shrinklist);
2360 INIT_LIST_HEAD(&info->swaplist);
2361 simple_xattrs_init(&info->xattrs);
2362 cache_no_acl(inode);
2363 mapping_set_large_folios(inode->i_mapping);
2364
2365 switch (mode & S_IFMT) {
2366 default:
2367 inode->i_op = &shmem_special_inode_operations;
2368 init_special_inode(inode, mode, dev);
2369 break;
2370 case S_IFREG:
2371 inode->i_mapping->a_ops = &shmem_aops;
2372 inode->i_op = &shmem_inode_operations;
2373 inode->i_fop = &shmem_file_operations;
2374 mpol_shared_policy_init(&info->policy,
2375 shmem_get_sbmpol(sbinfo));
2376 break;
2377 case S_IFDIR:
2378 inc_nlink(inode);
2379 /* Some things misbehave if size == 0 on a directory */
2380 inode->i_size = 2 * BOGO_DIRENT_SIZE;
2381 inode->i_op = &shmem_dir_inode_operations;
2382 inode->i_fop = &simple_dir_operations;
2383 break;
2384 case S_IFLNK:
2385 /*
2386 * Must not load anything in the rbtree,
2387 * mpol_free_shared_policy will not be called.
2388 */
2389 mpol_shared_policy_init(&info->policy, NULL);
2390 break;
2391 }
2392
2393 lockdep_annotate_inode_mutex_key(inode);
2394 } else
2395 shmem_free_inode(sb);
2396 return inode;
2397}
2398
2399#ifdef CONFIG_USERFAULTFD
2400int shmem_mfill_atomic_pte(struct mm_struct *dst_mm,
2401 pmd_t *dst_pmd,
2402 struct vm_area_struct *dst_vma,
2403 unsigned long dst_addr,
2404 unsigned long src_addr,
2405 bool zeropage, bool wp_copy,
2406 struct page **pagep)
2407{
2408 struct inode *inode = file_inode(dst_vma->vm_file);
2409 struct shmem_inode_info *info = SHMEM_I(inode);
2410 struct address_space *mapping = inode->i_mapping;
2411 gfp_t gfp = mapping_gfp_mask(mapping);
2412 pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
2413 void *page_kaddr;
2414 struct folio *folio;
2415 int ret;
2416 pgoff_t max_off;
2417
2418 if (!shmem_inode_acct_block(inode, 1)) {
2419 /*
2420 * We may have got a page, returned -ENOENT triggering a retry,
2421 * and now we find ourselves with -ENOMEM. Release the page, to
2422 * avoid a BUG_ON in our caller.
2423 */
2424 if (unlikely(*pagep)) {
2425 put_page(*pagep);
2426 *pagep = NULL;
2427 }
2428 return -ENOMEM;
2429 }
2430
2431 if (!*pagep) {
2432 ret = -ENOMEM;
2433 folio = shmem_alloc_folio(gfp, info, pgoff);
2434 if (!folio)
2435 goto out_unacct_blocks;
2436
2437 if (!zeropage) { /* COPY */
2438 page_kaddr = kmap_local_folio(folio, 0);
2439 /*
2440 * The read mmap_lock is held here. Despite the
2441 * mmap_lock being read recursive a deadlock is still
2442 * possible if a writer has taken a lock. For example:
2443 *
2444 * process A thread 1 takes read lock on own mmap_lock
2445 * process A thread 2 calls mmap, blocks taking write lock
2446 * process B thread 1 takes page fault, read lock on own mmap lock
2447 * process B thread 2 calls mmap, blocks taking write lock
2448 * process A thread 1 blocks taking read lock on process B
2449 * process B thread 1 blocks taking read lock on process A
2450 *
2451 * Disable page faults to prevent potential deadlock
2452 * and retry the copy outside the mmap_lock.
2453 */
2454 pagefault_disable();
2455 ret = copy_from_user(page_kaddr,
2456 (const void __user *)src_addr,
2457 PAGE_SIZE);
2458 pagefault_enable();
2459 kunmap_local(page_kaddr);
2460
2461 /* fallback to copy_from_user outside mmap_lock */
2462 if (unlikely(ret)) {
2463 *pagep = &folio->page;
2464 ret = -ENOENT;
2465 /* don't free the page */
2466 goto out_unacct_blocks;
2467 }
2468
2469 flush_dcache_folio(folio);
2470 } else { /* ZEROPAGE */
2471 clear_user_highpage(&folio->page, dst_addr);
2472 }
2473 } else {
2474 folio = page_folio(*pagep);
2475 VM_BUG_ON_FOLIO(folio_test_large(folio), folio);
2476 *pagep = NULL;
2477 }
2478
2479 VM_BUG_ON(folio_test_locked(folio));
2480 VM_BUG_ON(folio_test_swapbacked(folio));
2481 __folio_set_locked(folio);
2482 __folio_set_swapbacked(folio);
2483 __folio_mark_uptodate(folio);
2484
2485 ret = -EFAULT;
2486 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2487 if (unlikely(pgoff >= max_off))
2488 goto out_release;
2489
2490 ret = shmem_add_to_page_cache(folio, mapping, pgoff, NULL,
2491 gfp & GFP_RECLAIM_MASK, dst_mm);
2492 if (ret)
2493 goto out_release;
2494
2495 ret = mfill_atomic_install_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
2496 &folio->page, true, wp_copy);
2497 if (ret)
2498 goto out_delete_from_cache;
2499
2500 spin_lock_irq(&info->lock);
2501 info->alloced++;
2502 inode->i_blocks += BLOCKS_PER_PAGE;
2503 shmem_recalc_inode(inode);
2504 spin_unlock_irq(&info->lock);
2505
2506 folio_unlock(folio);
2507 return 0;
2508out_delete_from_cache:
2509 filemap_remove_folio(folio);
2510out_release:
2511 folio_unlock(folio);
2512 folio_put(folio);
2513out_unacct_blocks:
2514 shmem_inode_unacct_blocks(inode, 1);
2515 return ret;
2516}
2517#endif /* CONFIG_USERFAULTFD */
2518
2519#ifdef CONFIG_TMPFS
2520static const struct inode_operations shmem_symlink_inode_operations;
2521static const struct inode_operations shmem_short_symlink_operations;
2522
2523static int
2524shmem_write_begin(struct file *file, struct address_space *mapping,
2525 loff_t pos, unsigned len,
2526 struct page **pagep, void **fsdata)
2527{
2528 struct inode *inode = mapping->host;
2529 struct shmem_inode_info *info = SHMEM_I(inode);
2530 pgoff_t index = pos >> PAGE_SHIFT;
2531 struct folio *folio;
2532 int ret = 0;
2533
2534 /* i_rwsem is held by caller */
2535 if (unlikely(info->seals & (F_SEAL_GROW |
2536 F_SEAL_WRITE | F_SEAL_FUTURE_WRITE))) {
2537 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE))
2538 return -EPERM;
2539 if ((info->seals & F_SEAL_GROW) && pos + len > inode->i_size)
2540 return -EPERM;
2541 }
2542
2543 ret = shmem_get_folio(inode, index, &folio, SGP_WRITE);
2544
2545 if (ret)
2546 return ret;
2547
2548 *pagep = folio_file_page(folio, index);
2549 if (PageHWPoison(*pagep)) {
2550 folio_unlock(folio);
2551 folio_put(folio);
2552 *pagep = NULL;
2553 return -EIO;
2554 }
2555
2556 return 0;
2557}
2558
2559static int
2560shmem_write_end(struct file *file, struct address_space *mapping,
2561 loff_t pos, unsigned len, unsigned copied,
2562 struct page *page, void *fsdata)
2563{
2564 struct inode *inode = mapping->host;
2565
2566 if (pos + copied > inode->i_size)
2567 i_size_write(inode, pos + copied);
2568
2569 if (!PageUptodate(page)) {
2570 struct page *head = compound_head(page);
2571 if (PageTransCompound(page)) {
2572 int i;
2573
2574 for (i = 0; i < HPAGE_PMD_NR; i++) {
2575 if (head + i == page)
2576 continue;
2577 clear_highpage(head + i);
2578 flush_dcache_page(head + i);
2579 }
2580 }
2581 if (copied < PAGE_SIZE) {
2582 unsigned from = pos & (PAGE_SIZE - 1);
2583 zero_user_segments(page, 0, from,
2584 from + copied, PAGE_SIZE);
2585 }
2586 SetPageUptodate(head);
2587 }
2588 set_page_dirty(page);
2589 unlock_page(page);
2590 put_page(page);
2591
2592 return copied;
2593}
2594
2595static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
2596{
2597 struct file *file = iocb->ki_filp;
2598 struct inode *inode = file_inode(file);
2599 struct address_space *mapping = inode->i_mapping;
2600 pgoff_t index;
2601 unsigned long offset;
2602 int error = 0;
2603 ssize_t retval = 0;
2604 loff_t *ppos = &iocb->ki_pos;
2605
2606 index = *ppos >> PAGE_SHIFT;
2607 offset = *ppos & ~PAGE_MASK;
2608
2609 for (;;) {
2610 struct folio *folio = NULL;
2611 struct page *page = NULL;
2612 pgoff_t end_index;
2613 unsigned long nr, ret;
2614 loff_t i_size = i_size_read(inode);
2615
2616 end_index = i_size >> PAGE_SHIFT;
2617 if (index > end_index)
2618 break;
2619 if (index == end_index) {
2620 nr = i_size & ~PAGE_MASK;
2621 if (nr <= offset)
2622 break;
2623 }
2624
2625 error = shmem_get_folio(inode, index, &folio, SGP_READ);
2626 if (error) {
2627 if (error == -EINVAL)
2628 error = 0;
2629 break;
2630 }
2631 if (folio) {
2632 folio_unlock(folio);
2633
2634 page = folio_file_page(folio, index);
2635 if (PageHWPoison(page)) {
2636 folio_put(folio);
2637 error = -EIO;
2638 break;
2639 }
2640 }
2641
2642 /*
2643 * We must evaluate after, since reads (unlike writes)
2644 * are called without i_rwsem protection against truncate
2645 */
2646 nr = PAGE_SIZE;
2647 i_size = i_size_read(inode);
2648 end_index = i_size >> PAGE_SHIFT;
2649 if (index == end_index) {
2650 nr = i_size & ~PAGE_MASK;
2651 if (nr <= offset) {
2652 if (folio)
2653 folio_put(folio);
2654 break;
2655 }
2656 }
2657 nr -= offset;
2658
2659 if (folio) {
2660 /*
2661 * If users can be writing to this page using arbitrary
2662 * virtual addresses, take care about potential aliasing
2663 * before reading the page on the kernel side.
2664 */
2665 if (mapping_writably_mapped(mapping))
2666 flush_dcache_page(page);
2667 /*
2668 * Mark the page accessed if we read the beginning.
2669 */
2670 if (!offset)
2671 folio_mark_accessed(folio);
2672 /*
2673 * Ok, we have the page, and it's up-to-date, so
2674 * now we can copy it to user space...
2675 */
2676 ret = copy_page_to_iter(page, offset, nr, to);
2677 folio_put(folio);
2678
2679 } else if (user_backed_iter(to)) {
2680 /*
2681 * Copy to user tends to be so well optimized, but
2682 * clear_user() not so much, that it is noticeably
2683 * faster to copy the zero page instead of clearing.
2684 */
2685 ret = copy_page_to_iter(ZERO_PAGE(0), offset, nr, to);
2686 } else {
2687 /*
2688 * But submitting the same page twice in a row to
2689 * splice() - or others? - can result in confusion:
2690 * so don't attempt that optimization on pipes etc.
2691 */
2692 ret = iov_iter_zero(nr, to);
2693 }
2694
2695 retval += ret;
2696 offset += ret;
2697 index += offset >> PAGE_SHIFT;
2698 offset &= ~PAGE_MASK;
2699
2700 if (!iov_iter_count(to))
2701 break;
2702 if (ret < nr) {
2703 error = -EFAULT;
2704 break;
2705 }
2706 cond_resched();
2707 }
2708
2709 *ppos = ((loff_t) index << PAGE_SHIFT) + offset;
2710 file_accessed(file);
2711 return retval ? retval : error;
2712}
2713
2714static loff_t shmem_file_llseek(struct file *file, loff_t offset, int whence)
2715{
2716 struct address_space *mapping = file->f_mapping;
2717 struct inode *inode = mapping->host;
2718
2719 if (whence != SEEK_DATA && whence != SEEK_HOLE)
2720 return generic_file_llseek_size(file, offset, whence,
2721 MAX_LFS_FILESIZE, i_size_read(inode));
2722 if (offset < 0)
2723 return -ENXIO;
2724
2725 inode_lock(inode);
2726 /* We're holding i_rwsem so we can access i_size directly */
2727 offset = mapping_seek_hole_data(mapping, offset, inode->i_size, whence);
2728 if (offset >= 0)
2729 offset = vfs_setpos(file, offset, MAX_LFS_FILESIZE);
2730 inode_unlock(inode);
2731 return offset;
2732}
2733
2734static long shmem_fallocate(struct file *file, int mode, loff_t offset,
2735 loff_t len)
2736{
2737 struct inode *inode = file_inode(file);
2738 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
2739 struct shmem_inode_info *info = SHMEM_I(inode);
2740 struct shmem_falloc shmem_falloc;
2741 pgoff_t start, index, end, undo_fallocend;
2742 int error;
2743
2744 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2745 return -EOPNOTSUPP;
2746
2747 inode_lock(inode);
2748
2749 if (mode & FALLOC_FL_PUNCH_HOLE) {
2750 struct address_space *mapping = file->f_mapping;
2751 loff_t unmap_start = round_up(offset, PAGE_SIZE);
2752 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
2753 DECLARE_WAIT_QUEUE_HEAD_ONSTACK(shmem_falloc_waitq);
2754
2755 /* protected by i_rwsem */
2756 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
2757 error = -EPERM;
2758 goto out;
2759 }
2760
2761 shmem_falloc.waitq = &shmem_falloc_waitq;
2762 shmem_falloc.start = (u64)unmap_start >> PAGE_SHIFT;
2763 shmem_falloc.next = (unmap_end + 1) >> PAGE_SHIFT;
2764 spin_lock(&inode->i_lock);
2765 inode->i_private = &shmem_falloc;
2766 spin_unlock(&inode->i_lock);
2767
2768 if ((u64)unmap_end > (u64)unmap_start)
2769 unmap_mapping_range(mapping, unmap_start,
2770 1 + unmap_end - unmap_start, 0);
2771 shmem_truncate_range(inode, offset, offset + len - 1);
2772 /* No need to unmap again: hole-punching leaves COWed pages */
2773
2774 spin_lock(&inode->i_lock);
2775 inode->i_private = NULL;
2776 wake_up_all(&shmem_falloc_waitq);
2777 WARN_ON_ONCE(!list_empty(&shmem_falloc_waitq.head));
2778 spin_unlock(&inode->i_lock);
2779 error = 0;
2780 goto out;
2781 }
2782
2783 /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
2784 error = inode_newsize_ok(inode, offset + len);
2785 if (error)
2786 goto out;
2787
2788 if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
2789 error = -EPERM;
2790 goto out;
2791 }
2792
2793 start = offset >> PAGE_SHIFT;
2794 end = (offset + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
2795 /* Try to avoid a swapstorm if len is impossible to satisfy */
2796 if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
2797 error = -ENOSPC;
2798 goto out;
2799 }
2800
2801 shmem_falloc.waitq = NULL;
2802 shmem_falloc.start = start;
2803 shmem_falloc.next = start;
2804 shmem_falloc.nr_falloced = 0;
2805 shmem_falloc.nr_unswapped = 0;
2806 spin_lock(&inode->i_lock);
2807 inode->i_private = &shmem_falloc;
2808 spin_unlock(&inode->i_lock);
2809
2810 /*
2811 * info->fallocend is only relevant when huge pages might be
2812 * involved: to prevent split_huge_page() freeing fallocated
2813 * pages when FALLOC_FL_KEEP_SIZE committed beyond i_size.
2814 */
2815 undo_fallocend = info->fallocend;
2816 if (info->fallocend < end)
2817 info->fallocend = end;
2818
2819 for (index = start; index < end; ) {
2820 struct folio *folio;
2821
2822 /*
2823 * Good, the fallocate(2) manpage permits EINTR: we may have
2824 * been interrupted because we are using up too much memory.
2825 */
2826 if (signal_pending(current))
2827 error = -EINTR;
2828 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
2829 error = -ENOMEM;
2830 else
2831 error = shmem_get_folio(inode, index, &folio,
2832 SGP_FALLOC);
2833 if (error) {
2834 info->fallocend = undo_fallocend;
2835 /* Remove the !uptodate folios we added */
2836 if (index > start) {
2837 shmem_undo_range(inode,
2838 (loff_t)start << PAGE_SHIFT,
2839 ((loff_t)index << PAGE_SHIFT) - 1, true);
2840 }
2841 goto undone;
2842 }
2843
2844 /*
2845 * Here is a more important optimization than it appears:
2846 * a second SGP_FALLOC on the same large folio will clear it,
2847 * making it uptodate and un-undoable if we fail later.
2848 */
2849 index = folio_next_index(folio);
2850 /* Beware 32-bit wraparound */
2851 if (!index)
2852 index--;
2853
2854 /*
2855 * Inform shmem_writepage() how far we have reached.
2856 * No need for lock or barrier: we have the page lock.
2857 */
2858 if (!folio_test_uptodate(folio))
2859 shmem_falloc.nr_falloced += index - shmem_falloc.next;
2860 shmem_falloc.next = index;
2861
2862 /*
2863 * If !uptodate, leave it that way so that freeable folios
2864 * can be recognized if we need to rollback on error later.
2865 * But mark it dirty so that memory pressure will swap rather
2866 * than free the folios we are allocating (and SGP_CACHE folios
2867 * might still be clean: we now need to mark those dirty too).
2868 */
2869 folio_mark_dirty(folio);
2870 folio_unlock(folio);
2871 folio_put(folio);
2872 cond_resched();
2873 }
2874
2875 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
2876 i_size_write(inode, offset + len);
2877undone:
2878 spin_lock(&inode->i_lock);
2879 inode->i_private = NULL;
2880 spin_unlock(&inode->i_lock);
2881out:
2882 if (!error)
2883 file_modified(file);
2884 inode_unlock(inode);
2885 return error;
2886}
2887
2888static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
2889{
2890 struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
2891
2892 buf->f_type = TMPFS_MAGIC;
2893 buf->f_bsize = PAGE_SIZE;
2894 buf->f_namelen = NAME_MAX;
2895 if (sbinfo->max_blocks) {
2896 buf->f_blocks = sbinfo->max_blocks;
2897 buf->f_bavail =
2898 buf->f_bfree = sbinfo->max_blocks -
2899 percpu_counter_sum(&sbinfo->used_blocks);
2900 }
2901 if (sbinfo->max_inodes) {
2902 buf->f_files = sbinfo->max_inodes;
2903 buf->f_ffree = sbinfo->free_inodes;
2904 }
2905 /* else leave those fields 0 like simple_statfs */
2906
2907 buf->f_fsid = uuid_to_fsid(dentry->d_sb->s_uuid.b);
2908
2909 return 0;
2910}
2911
2912/*
2913 * File creation. Allocate an inode, and we're done..
2914 */
2915static int
2916shmem_mknod(struct user_namespace *mnt_userns, struct inode *dir,
2917 struct dentry *dentry, umode_t mode, dev_t dev)
2918{
2919 struct inode *inode;
2920 int error = -ENOSPC;
2921
2922 inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
2923 if (inode) {
2924 error = simple_acl_create(dir, inode);
2925 if (error)
2926 goto out_iput;
2927 error = security_inode_init_security(inode, dir,
2928 &dentry->d_name,
2929 shmem_initxattrs, NULL);
2930 if (error && error != -EOPNOTSUPP)
2931 goto out_iput;
2932
2933 error = 0;
2934 dir->i_size += BOGO_DIRENT_SIZE;
2935 dir->i_ctime = dir->i_mtime = current_time(dir);
2936 inode_inc_iversion(dir);
2937 d_instantiate(dentry, inode);
2938 dget(dentry); /* Extra count - pin the dentry in core */
2939 }
2940 return error;
2941out_iput:
2942 iput(inode);
2943 return error;
2944}
2945
2946static int
2947shmem_tmpfile(struct user_namespace *mnt_userns, struct inode *dir,
2948 struct file *file, umode_t mode)
2949{
2950 struct inode *inode;
2951 int error = -ENOSPC;
2952
2953 inode = shmem_get_inode(dir->i_sb, dir, mode, 0, VM_NORESERVE);
2954 if (inode) {
2955 error = security_inode_init_security(inode, dir,
2956 NULL,
2957 shmem_initxattrs, NULL);
2958 if (error && error != -EOPNOTSUPP)
2959 goto out_iput;
2960 error = simple_acl_create(dir, inode);
2961 if (error)
2962 goto out_iput;
2963 d_tmpfile(file, inode);
2964 }
2965 return finish_open_simple(file, error);
2966out_iput:
2967 iput(inode);
2968 return error;
2969}
2970
2971static int shmem_mkdir(struct user_namespace *mnt_userns, struct inode *dir,
2972 struct dentry *dentry, umode_t mode)
2973{
2974 int error;
2975
2976 if ((error = shmem_mknod(&init_user_ns, dir, dentry,
2977 mode | S_IFDIR, 0)))
2978 return error;
2979 inc_nlink(dir);
2980 return 0;
2981}
2982
2983static int shmem_create(struct user_namespace *mnt_userns, struct inode *dir,
2984 struct dentry *dentry, umode_t mode, bool excl)
2985{
2986 return shmem_mknod(&init_user_ns, dir, dentry, mode | S_IFREG, 0);
2987}
2988
2989/*
2990 * Link a file..
2991 */
2992static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
2993{
2994 struct inode *inode = d_inode(old_dentry);
2995 int ret = 0;
2996
2997 /*
2998 * No ordinary (disk based) filesystem counts links as inodes;
2999 * but each new link needs a new dentry, pinning lowmem, and
3000 * tmpfs dentries cannot be pruned until they are unlinked.
3001 * But if an O_TMPFILE file is linked into the tmpfs, the
3002 * first link must skip that, to get the accounting right.
3003 */
3004 if (inode->i_nlink) {
3005 ret = shmem_reserve_inode(inode->i_sb, NULL);
3006 if (ret)
3007 goto out;
3008 }
3009
3010 dir->i_size += BOGO_DIRENT_SIZE;
3011 inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
3012 inode_inc_iversion(dir);
3013 inc_nlink(inode);
3014 ihold(inode); /* New dentry reference */
3015 dget(dentry); /* Extra pinning count for the created dentry */
3016 d_instantiate(dentry, inode);
3017out:
3018 return ret;
3019}
3020
3021static int shmem_unlink(struct inode *dir, struct dentry *dentry)
3022{
3023 struct inode *inode = d_inode(dentry);
3024
3025 if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
3026 shmem_free_inode(inode->i_sb);
3027
3028 dir->i_size -= BOGO_DIRENT_SIZE;
3029 inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
3030 inode_inc_iversion(dir);
3031 drop_nlink(inode);
3032 dput(dentry); /* Undo the count from "create" - this does all the work */
3033 return 0;
3034}
3035
3036static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
3037{
3038 if (!simple_empty(dentry))
3039 return -ENOTEMPTY;
3040
3041 drop_nlink(d_inode(dentry));
3042 drop_nlink(dir);
3043 return shmem_unlink(dir, dentry);
3044}
3045
3046static int shmem_whiteout(struct user_namespace *mnt_userns,
3047 struct inode *old_dir, struct dentry *old_dentry)
3048{
3049 struct dentry *whiteout;
3050 int error;
3051
3052 whiteout = d_alloc(old_dentry->d_parent, &old_dentry->d_name);
3053 if (!whiteout)
3054 return -ENOMEM;
3055
3056 error = shmem_mknod(&init_user_ns, old_dir, whiteout,
3057 S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
3058 dput(whiteout);
3059 if (error)
3060 return error;
3061
3062 /*
3063 * Cheat and hash the whiteout while the old dentry is still in
3064 * place, instead of playing games with FS_RENAME_DOES_D_MOVE.
3065 *
3066 * d_lookup() will consistently find one of them at this point,
3067 * not sure which one, but that isn't even important.
3068 */
3069 d_rehash(whiteout);
3070 return 0;
3071}
3072
3073/*
3074 * The VFS layer already does all the dentry stuff for rename,
3075 * we just have to decrement the usage count for the target if
3076 * it exists so that the VFS layer correctly free's it when it
3077 * gets overwritten.
3078 */
3079static int shmem_rename2(struct user_namespace *mnt_userns,
3080 struct inode *old_dir, struct dentry *old_dentry,
3081 struct inode *new_dir, struct dentry *new_dentry,
3082 unsigned int flags)
3083{
3084 struct inode *inode = d_inode(old_dentry);
3085 int they_are_dirs = S_ISDIR(inode->i_mode);
3086
3087 if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
3088 return -EINVAL;
3089
3090 if (flags & RENAME_EXCHANGE)
3091 return simple_rename_exchange(old_dir, old_dentry, new_dir, new_dentry);
3092
3093 if (!simple_empty(new_dentry))
3094 return -ENOTEMPTY;
3095
3096 if (flags & RENAME_WHITEOUT) {
3097 int error;
3098
3099 error = shmem_whiteout(&init_user_ns, old_dir, old_dentry);
3100 if (error)
3101 return error;
3102 }
3103
3104 if (d_really_is_positive(new_dentry)) {
3105 (void) shmem_unlink(new_dir, new_dentry);
3106 if (they_are_dirs) {
3107 drop_nlink(d_inode(new_dentry));
3108 drop_nlink(old_dir);
3109 }
3110 } else if (they_are_dirs) {
3111 drop_nlink(old_dir);
3112 inc_nlink(new_dir);
3113 }
3114
3115 old_dir->i_size -= BOGO_DIRENT_SIZE;
3116 new_dir->i_size += BOGO_DIRENT_SIZE;
3117 old_dir->i_ctime = old_dir->i_mtime =
3118 new_dir->i_ctime = new_dir->i_mtime =
3119 inode->i_ctime = current_time(old_dir);
3120 inode_inc_iversion(old_dir);
3121 inode_inc_iversion(new_dir);
3122 return 0;
3123}
3124
3125static int shmem_symlink(struct user_namespace *mnt_userns, struct inode *dir,
3126 struct dentry *dentry, const char *symname)
3127{
3128 int error;
3129 int len;
3130 struct inode *inode;
3131 struct folio *folio;
3132
3133 len = strlen(symname) + 1;
3134 if (len > PAGE_SIZE)
3135 return -ENAMETOOLONG;
3136
3137 inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK | 0777, 0,
3138 VM_NORESERVE);
3139 if (!inode)
3140 return -ENOSPC;
3141
3142 error = security_inode_init_security(inode, dir, &dentry->d_name,
3143 shmem_initxattrs, NULL);
3144 if (error && error != -EOPNOTSUPP) {
3145 iput(inode);
3146 return error;
3147 }
3148
3149 inode->i_size = len-1;
3150 if (len <= SHORT_SYMLINK_LEN) {
3151 inode->i_link = kmemdup(symname, len, GFP_KERNEL);
3152 if (!inode->i_link) {
3153 iput(inode);
3154 return -ENOMEM;
3155 }
3156 inode->i_op = &shmem_short_symlink_operations;
3157 } else {
3158 inode_nohighmem(inode);
3159 error = shmem_get_folio(inode, 0, &folio, SGP_WRITE);
3160 if (error) {
3161 iput(inode);
3162 return error;
3163 }
3164 inode->i_mapping->a_ops = &shmem_aops;
3165 inode->i_op = &shmem_symlink_inode_operations;
3166 memcpy(folio_address(folio), symname, len);
3167 folio_mark_uptodate(folio);
3168 folio_mark_dirty(folio);
3169 folio_unlock(folio);
3170 folio_put(folio);
3171 }
3172 dir->i_size += BOGO_DIRENT_SIZE;
3173 dir->i_ctime = dir->i_mtime = current_time(dir);
3174 inode_inc_iversion(dir);
3175 d_instantiate(dentry, inode);
3176 dget(dentry);
3177 return 0;
3178}
3179
3180static void shmem_put_link(void *arg)
3181{
3182 folio_mark_accessed(arg);
3183 folio_put(arg);
3184}
3185
3186static const char *shmem_get_link(struct dentry *dentry,
3187 struct inode *inode,
3188 struct delayed_call *done)
3189{
3190 struct folio *folio = NULL;
3191 int error;
3192
3193 if (!dentry) {
3194 folio = filemap_get_folio(inode->i_mapping, 0);
3195 if (!folio)
3196 return ERR_PTR(-ECHILD);
3197 if (PageHWPoison(folio_page(folio, 0)) ||
3198 !folio_test_uptodate(folio)) {
3199 folio_put(folio);
3200 return ERR_PTR(-ECHILD);
3201 }
3202 } else {
3203 error = shmem_get_folio(inode, 0, &folio, SGP_READ);
3204 if (error)
3205 return ERR_PTR(error);
3206 if (!folio)
3207 return ERR_PTR(-ECHILD);
3208 if (PageHWPoison(folio_page(folio, 0))) {
3209 folio_unlock(folio);
3210 folio_put(folio);
3211 return ERR_PTR(-ECHILD);
3212 }
3213 folio_unlock(folio);
3214 }
3215 set_delayed_call(done, shmem_put_link, folio);
3216 return folio_address(folio);
3217}
3218
3219#ifdef CONFIG_TMPFS_XATTR
3220
3221static int shmem_fileattr_get(struct dentry *dentry, struct fileattr *fa)
3222{
3223 struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
3224
3225 fileattr_fill_flags(fa, info->fsflags & SHMEM_FL_USER_VISIBLE);
3226
3227 return 0;
3228}
3229
3230static int shmem_fileattr_set(struct user_namespace *mnt_userns,
3231 struct dentry *dentry, struct fileattr *fa)
3232{
3233 struct inode *inode = d_inode(dentry);
3234 struct shmem_inode_info *info = SHMEM_I(inode);
3235
3236 if (fileattr_has_fsx(fa))
3237 return -EOPNOTSUPP;
3238 if (fa->flags & ~SHMEM_FL_USER_MODIFIABLE)
3239 return -EOPNOTSUPP;
3240
3241 info->fsflags = (info->fsflags & ~SHMEM_FL_USER_MODIFIABLE) |
3242 (fa->flags & SHMEM_FL_USER_MODIFIABLE);
3243
3244 shmem_set_inode_flags(inode, info->fsflags);
3245 inode->i_ctime = current_time(inode);
3246 inode_inc_iversion(inode);
3247 return 0;
3248}
3249
3250/*
3251 * Superblocks without xattr inode operations may get some security.* xattr
3252 * support from the LSM "for free". As soon as we have any other xattrs
3253 * like ACLs, we also need to implement the security.* handlers at
3254 * filesystem level, though.
3255 */
3256
3257/*
3258 * Callback for security_inode_init_security() for acquiring xattrs.
3259 */
3260static int shmem_initxattrs(struct inode *inode,
3261 const struct xattr *xattr_array,
3262 void *fs_info)
3263{
3264 struct shmem_inode_info *info = SHMEM_I(inode);
3265 const struct xattr *xattr;
3266 struct simple_xattr *new_xattr;
3267 size_t len;
3268
3269 for (xattr = xattr_array; xattr->name != NULL; xattr++) {
3270 new_xattr = simple_xattr_alloc(xattr->value, xattr->value_len);
3271 if (!new_xattr)
3272 return -ENOMEM;
3273
3274 len = strlen(xattr->name) + 1;
3275 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
3276 GFP_KERNEL);
3277 if (!new_xattr->name) {
3278 kvfree(new_xattr);
3279 return -ENOMEM;
3280 }
3281
3282 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
3283 XATTR_SECURITY_PREFIX_LEN);
3284 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
3285 xattr->name, len);
3286
3287 simple_xattr_add(&info->xattrs, new_xattr);
3288 }
3289
3290 return 0;
3291}
3292
3293static int shmem_xattr_handler_get(const struct xattr_handler *handler,
3294 struct dentry *unused, struct inode *inode,
3295 const char *name, void *buffer, size_t size)
3296{
3297 struct shmem_inode_info *info = SHMEM_I(inode);
3298
3299 name = xattr_full_name(handler, name);
3300 return simple_xattr_get(&info->xattrs, name, buffer, size);
3301}
3302
3303static int shmem_xattr_handler_set(const struct xattr_handler *handler,
3304 struct user_namespace *mnt_userns,
3305 struct dentry *unused, struct inode *inode,
3306 const char *name, const void *value,
3307 size_t size, int flags)
3308{
3309 struct shmem_inode_info *info = SHMEM_I(inode);
3310 int err;
3311
3312 name = xattr_full_name(handler, name);
3313 err = simple_xattr_set(&info->xattrs, name, value, size, flags, NULL);
3314 if (!err) {
3315 inode->i_ctime = current_time(inode);
3316 inode_inc_iversion(inode);
3317 }
3318 return err;
3319}
3320
3321static const struct xattr_handler shmem_security_xattr_handler = {
3322 .prefix = XATTR_SECURITY_PREFIX,
3323 .get = shmem_xattr_handler_get,
3324 .set = shmem_xattr_handler_set,
3325};
3326
3327static const struct xattr_handler shmem_trusted_xattr_handler = {
3328 .prefix = XATTR_TRUSTED_PREFIX,
3329 .get = shmem_xattr_handler_get,
3330 .set = shmem_xattr_handler_set,
3331};
3332
3333static const struct xattr_handler *shmem_xattr_handlers[] = {
3334#ifdef CONFIG_TMPFS_POSIX_ACL
3335 &posix_acl_access_xattr_handler,
3336 &posix_acl_default_xattr_handler,
3337#endif
3338 &shmem_security_xattr_handler,
3339 &shmem_trusted_xattr_handler,
3340 NULL
3341};
3342
3343static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
3344{
3345 struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
3346 return simple_xattr_list(d_inode(dentry), &info->xattrs, buffer, size);
3347}
3348#endif /* CONFIG_TMPFS_XATTR */
3349
3350static const struct inode_operations shmem_short_symlink_operations = {
3351 .getattr = shmem_getattr,
3352 .get_link = simple_get_link,
3353#ifdef CONFIG_TMPFS_XATTR
3354 .listxattr = shmem_listxattr,
3355#endif
3356};
3357
3358static const struct inode_operations shmem_symlink_inode_operations = {
3359 .getattr = shmem_getattr,
3360 .get_link = shmem_get_link,
3361#ifdef CONFIG_TMPFS_XATTR
3362 .listxattr = shmem_listxattr,
3363#endif
3364};
3365
3366static struct dentry *shmem_get_parent(struct dentry *child)
3367{
3368 return ERR_PTR(-ESTALE);
3369}
3370
3371static int shmem_match(struct inode *ino, void *vfh)
3372{
3373 __u32 *fh = vfh;
3374 __u64 inum = fh[2];
3375 inum = (inum << 32) | fh[1];
3376 return ino->i_ino == inum && fh[0] == ino->i_generation;
3377}
3378
3379/* Find any alias of inode, but prefer a hashed alias */
3380static struct dentry *shmem_find_alias(struct inode *inode)
3381{
3382 struct dentry *alias = d_find_alias(inode);
3383
3384 return alias ?: d_find_any_alias(inode);
3385}
3386
3387
3388static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
3389 struct fid *fid, int fh_len, int fh_type)
3390{
3391 struct inode *inode;
3392 struct dentry *dentry = NULL;
3393 u64 inum;
3394
3395 if (fh_len < 3)
3396 return NULL;
3397
3398 inum = fid->raw[2];
3399 inum = (inum << 32) | fid->raw[1];
3400
3401 inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
3402 shmem_match, fid->raw);
3403 if (inode) {
3404 dentry = shmem_find_alias(inode);
3405 iput(inode);
3406 }
3407
3408 return dentry;
3409}
3410
3411static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
3412 struct inode *parent)
3413{
3414 if (*len < 3) {
3415 *len = 3;
3416 return FILEID_INVALID;
3417 }
3418
3419 if (inode_unhashed(inode)) {
3420 /* Unfortunately insert_inode_hash is not idempotent,
3421 * so as we hash inodes here rather than at creation
3422 * time, we need a lock to ensure we only try
3423 * to do it once
3424 */
3425 static DEFINE_SPINLOCK(lock);
3426 spin_lock(&lock);
3427 if (inode_unhashed(inode))
3428 __insert_inode_hash(inode,
3429 inode->i_ino + inode->i_generation);
3430 spin_unlock(&lock);
3431 }
3432
3433 fh[0] = inode->i_generation;
3434 fh[1] = inode->i_ino;
3435 fh[2] = ((__u64)inode->i_ino) >> 32;
3436
3437 *len = 3;
3438 return 1;
3439}
3440
3441static const struct export_operations shmem_export_ops = {
3442 .get_parent = shmem_get_parent,
3443 .encode_fh = shmem_encode_fh,
3444 .fh_to_dentry = shmem_fh_to_dentry,
3445};
3446
3447enum shmem_param {
3448 Opt_gid,
3449 Opt_huge,
3450 Opt_mode,
3451 Opt_mpol,
3452 Opt_nr_blocks,
3453 Opt_nr_inodes,
3454 Opt_size,
3455 Opt_uid,
3456 Opt_inode32,
3457 Opt_inode64,
3458};
3459
3460static const struct constant_table shmem_param_enums_huge[] = {
3461 {"never", SHMEM_HUGE_NEVER },
3462 {"always", SHMEM_HUGE_ALWAYS },
3463 {"within_size", SHMEM_HUGE_WITHIN_SIZE },
3464 {"advise", SHMEM_HUGE_ADVISE },
3465 {}
3466};
3467
3468const struct fs_parameter_spec shmem_fs_parameters[] = {
3469 fsparam_u32 ("gid", Opt_gid),
3470 fsparam_enum ("huge", Opt_huge, shmem_param_enums_huge),
3471 fsparam_u32oct("mode", Opt_mode),
3472 fsparam_string("mpol", Opt_mpol),
3473 fsparam_string("nr_blocks", Opt_nr_blocks),
3474 fsparam_string("nr_inodes", Opt_nr_inodes),
3475 fsparam_string("size", Opt_size),
3476 fsparam_u32 ("uid", Opt_uid),
3477 fsparam_flag ("inode32", Opt_inode32),
3478 fsparam_flag ("inode64", Opt_inode64),
3479 {}
3480};
3481
3482static int shmem_parse_one(struct fs_context *fc, struct fs_parameter *param)
3483{
3484 struct shmem_options *ctx = fc->fs_private;
3485 struct fs_parse_result result;
3486 unsigned long long size;
3487 char *rest;
3488 int opt;
3489
3490 opt = fs_parse(fc, shmem_fs_parameters, param, &result);
3491 if (opt < 0)
3492 return opt;
3493
3494 switch (opt) {
3495 case Opt_size:
3496 size = memparse(param->string, &rest);
3497 if (*rest == '%') {
3498 size <<= PAGE_SHIFT;
3499 size *= totalram_pages();
3500 do_div(size, 100);
3501 rest++;
3502 }
3503 if (*rest)
3504 goto bad_value;
3505 ctx->blocks = DIV_ROUND_UP(size, PAGE_SIZE);
3506 ctx->seen |= SHMEM_SEEN_BLOCKS;
3507 break;
3508 case Opt_nr_blocks:
3509 ctx->blocks = memparse(param->string, &rest);
3510 if (*rest || ctx->blocks > S64_MAX)
3511 goto bad_value;
3512 ctx->seen |= SHMEM_SEEN_BLOCKS;
3513 break;
3514 case Opt_nr_inodes:
3515 ctx->inodes = memparse(param->string, &rest);
3516 if (*rest)
3517 goto bad_value;
3518 ctx->seen |= SHMEM_SEEN_INODES;
3519 break;
3520 case Opt_mode:
3521 ctx->mode = result.uint_32 & 07777;
3522 break;
3523 case Opt_uid:
3524 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
3525 if (!uid_valid(ctx->uid))
3526 goto bad_value;
3527 break;
3528 case Opt_gid:
3529 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
3530 if (!gid_valid(ctx->gid))
3531 goto bad_value;
3532 break;
3533 case Opt_huge:
3534 ctx->huge = result.uint_32;
3535 if (ctx->huge != SHMEM_HUGE_NEVER &&
3536 !(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
3537 has_transparent_hugepage()))
3538 goto unsupported_parameter;
3539 ctx->seen |= SHMEM_SEEN_HUGE;
3540 break;
3541 case Opt_mpol:
3542 if (IS_ENABLED(CONFIG_NUMA)) {
3543 mpol_put(ctx->mpol);
3544 ctx->mpol = NULL;
3545 if (mpol_parse_str(param->string, &ctx->mpol))
3546 goto bad_value;
3547 break;
3548 }
3549 goto unsupported_parameter;
3550 case Opt_inode32:
3551 ctx->full_inums = false;
3552 ctx->seen |= SHMEM_SEEN_INUMS;
3553 break;
3554 case Opt_inode64:
3555 if (sizeof(ino_t) < 8) {
3556 return invalfc(fc,
3557 "Cannot use inode64 with <64bit inums in kernel\n");
3558 }
3559 ctx->full_inums = true;
3560 ctx->seen |= SHMEM_SEEN_INUMS;
3561 break;
3562 }
3563 return 0;
3564
3565unsupported_parameter:
3566 return invalfc(fc, "Unsupported parameter '%s'", param->key);
3567bad_value:
3568 return invalfc(fc, "Bad value for '%s'", param->key);
3569}
3570
3571static int shmem_parse_options(struct fs_context *fc, void *data)
3572{
3573 char *options = data;
3574
3575 if (options) {
3576 int err = security_sb_eat_lsm_opts(options, &fc->security);
3577 if (err)
3578 return err;
3579 }
3580
3581 while (options != NULL) {
3582 char *this_char = options;
3583 for (;;) {
3584 /*
3585 * NUL-terminate this option: unfortunately,
3586 * mount options form a comma-separated list,
3587 * but mpol's nodelist may also contain commas.
3588 */
3589 options = strchr(options, ',');
3590 if (options == NULL)
3591 break;
3592 options++;
3593 if (!isdigit(*options)) {
3594 options[-1] = '\0';
3595 break;
3596 }
3597 }
3598 if (*this_char) {
3599 char *value = strchr(this_char, '=');
3600 size_t len = 0;
3601 int err;
3602
3603 if (value) {
3604 *value++ = '\0';
3605 len = strlen(value);
3606 }
3607 err = vfs_parse_fs_string(fc, this_char, value, len);
3608 if (err < 0)
3609 return err;
3610 }
3611 }
3612 return 0;
3613}
3614
3615/*
3616 * Reconfigure a shmem filesystem.
3617 *
3618 * Note that we disallow change from limited->unlimited blocks/inodes while any
3619 * are in use; but we must separately disallow unlimited->limited, because in
3620 * that case we have no record of how much is already in use.
3621 */
3622static int shmem_reconfigure(struct fs_context *fc)
3623{
3624 struct shmem_options *ctx = fc->fs_private;
3625 struct shmem_sb_info *sbinfo = SHMEM_SB(fc->root->d_sb);
3626 unsigned long inodes;
3627 struct mempolicy *mpol = NULL;
3628 const char *err;
3629
3630 raw_spin_lock(&sbinfo->stat_lock);
3631 inodes = sbinfo->max_inodes - sbinfo->free_inodes;
3632
3633 if ((ctx->seen & SHMEM_SEEN_BLOCKS) && ctx->blocks) {
3634 if (!sbinfo->max_blocks) {
3635 err = "Cannot retroactively limit size";
3636 goto out;
3637 }
3638 if (percpu_counter_compare(&sbinfo->used_blocks,
3639 ctx->blocks) > 0) {
3640 err = "Too small a size for current use";
3641 goto out;
3642 }
3643 }
3644 if ((ctx->seen & SHMEM_SEEN_INODES) && ctx->inodes) {
3645 if (!sbinfo->max_inodes) {
3646 err = "Cannot retroactively limit inodes";
3647 goto out;
3648 }
3649 if (ctx->inodes < inodes) {
3650 err = "Too few inodes for current use";
3651 goto out;
3652 }
3653 }
3654
3655 if ((ctx->seen & SHMEM_SEEN_INUMS) && !ctx->full_inums &&
3656 sbinfo->next_ino > UINT_MAX) {
3657 err = "Current inum too high to switch to 32-bit inums";
3658 goto out;
3659 }
3660
3661 if (ctx->seen & SHMEM_SEEN_HUGE)
3662 sbinfo->huge = ctx->huge;
3663 if (ctx->seen & SHMEM_SEEN_INUMS)
3664 sbinfo->full_inums = ctx->full_inums;
3665 if (ctx->seen & SHMEM_SEEN_BLOCKS)
3666 sbinfo->max_blocks = ctx->blocks;
3667 if (ctx->seen & SHMEM_SEEN_INODES) {
3668 sbinfo->max_inodes = ctx->inodes;
3669 sbinfo->free_inodes = ctx->inodes - inodes;
3670 }
3671
3672 /*
3673 * Preserve previous mempolicy unless mpol remount option was specified.
3674 */
3675 if (ctx->mpol) {
3676 mpol = sbinfo->mpol;
3677 sbinfo->mpol = ctx->mpol; /* transfers initial ref */
3678 ctx->mpol = NULL;
3679 }
3680 raw_spin_unlock(&sbinfo->stat_lock);
3681 mpol_put(mpol);
3682 return 0;
3683out:
3684 raw_spin_unlock(&sbinfo->stat_lock);
3685 return invalfc(fc, "%s", err);
3686}
3687
3688static int shmem_show_options(struct seq_file *seq, struct dentry *root)
3689{
3690 struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
3691
3692 if (sbinfo->max_blocks != shmem_default_max_blocks())
3693 seq_printf(seq, ",size=%luk",
3694 sbinfo->max_blocks << (PAGE_SHIFT - 10));
3695 if (sbinfo->max_inodes != shmem_default_max_inodes())
3696 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
3697 if (sbinfo->mode != (0777 | S_ISVTX))
3698 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
3699 if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
3700 seq_printf(seq, ",uid=%u",
3701 from_kuid_munged(&init_user_ns, sbinfo->uid));
3702 if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
3703 seq_printf(seq, ",gid=%u",
3704 from_kgid_munged(&init_user_ns, sbinfo->gid));
3705
3706 /*
3707 * Showing inode{64,32} might be useful even if it's the system default,
3708 * since then people don't have to resort to checking both here and
3709 * /proc/config.gz to confirm 64-bit inums were successfully applied
3710 * (which may not even exist if IKCONFIG_PROC isn't enabled).
3711 *
3712 * We hide it when inode64 isn't the default and we are using 32-bit
3713 * inodes, since that probably just means the feature isn't even under
3714 * consideration.
3715 *
3716 * As such:
3717 *
3718 * +-----------------+-----------------+
3719 * | TMPFS_INODE64=y | TMPFS_INODE64=n |
3720 * +------------------+-----------------+-----------------+
3721 * | full_inums=true | show | show |
3722 * | full_inums=false | show | hide |
3723 * +------------------+-----------------+-----------------+
3724 *
3725 */
3726 if (IS_ENABLED(CONFIG_TMPFS_INODE64) || sbinfo->full_inums)
3727 seq_printf(seq, ",inode%d", (sbinfo->full_inums ? 64 : 32));
3728#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3729 /* Rightly or wrongly, show huge mount option unmasked by shmem_huge */
3730 if (sbinfo->huge)
3731 seq_printf(seq, ",huge=%s", shmem_format_huge(sbinfo->huge));
3732#endif
3733 shmem_show_mpol(seq, sbinfo->mpol);
3734 return 0;
3735}
3736
3737#endif /* CONFIG_TMPFS */
3738
3739static void shmem_put_super(struct super_block *sb)
3740{
3741 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
3742
3743 free_percpu(sbinfo->ino_batch);
3744 percpu_counter_destroy(&sbinfo->used_blocks);
3745 mpol_put(sbinfo->mpol);
3746 kfree(sbinfo);
3747 sb->s_fs_info = NULL;
3748}
3749
3750static int shmem_fill_super(struct super_block *sb, struct fs_context *fc)
3751{
3752 struct shmem_options *ctx = fc->fs_private;
3753 struct inode *inode;
3754 struct shmem_sb_info *sbinfo;
3755
3756 /* Round up to L1_CACHE_BYTES to resist false sharing */
3757 sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
3758 L1_CACHE_BYTES), GFP_KERNEL);
3759 if (!sbinfo)
3760 return -ENOMEM;
3761
3762 sb->s_fs_info = sbinfo;
3763
3764#ifdef CONFIG_TMPFS
3765 /*
3766 * Per default we only allow half of the physical ram per
3767 * tmpfs instance, limiting inodes to one per page of lowmem;
3768 * but the internal instance is left unlimited.
3769 */
3770 if (!(sb->s_flags & SB_KERNMOUNT)) {
3771 if (!(ctx->seen & SHMEM_SEEN_BLOCKS))
3772 ctx->blocks = shmem_default_max_blocks();
3773 if (!(ctx->seen & SHMEM_SEEN_INODES))
3774 ctx->inodes = shmem_default_max_inodes();
3775 if (!(ctx->seen & SHMEM_SEEN_INUMS))
3776 ctx->full_inums = IS_ENABLED(CONFIG_TMPFS_INODE64);
3777 } else {
3778 sb->s_flags |= SB_NOUSER;
3779 }
3780 sb->s_export_op = &shmem_export_ops;
3781 sb->s_flags |= SB_NOSEC | SB_I_VERSION;
3782#else
3783 sb->s_flags |= SB_NOUSER;
3784#endif
3785 sbinfo->max_blocks = ctx->blocks;
3786 sbinfo->free_inodes = sbinfo->max_inodes = ctx->inodes;
3787 if (sb->s_flags & SB_KERNMOUNT) {
3788 sbinfo->ino_batch = alloc_percpu(ino_t);
3789 if (!sbinfo->ino_batch)
3790 goto failed;
3791 }
3792 sbinfo->uid = ctx->uid;
3793 sbinfo->gid = ctx->gid;
3794 sbinfo->full_inums = ctx->full_inums;
3795 sbinfo->mode = ctx->mode;
3796 sbinfo->huge = ctx->huge;
3797 sbinfo->mpol = ctx->mpol;
3798 ctx->mpol = NULL;
3799
3800 raw_spin_lock_init(&sbinfo->stat_lock);
3801 if (percpu_counter_init(&sbinfo->used_blocks, 0, GFP_KERNEL))
3802 goto failed;
3803 spin_lock_init(&sbinfo->shrinklist_lock);
3804 INIT_LIST_HEAD(&sbinfo->shrinklist);
3805
3806 sb->s_maxbytes = MAX_LFS_FILESIZE;
3807 sb->s_blocksize = PAGE_SIZE;
3808 sb->s_blocksize_bits = PAGE_SHIFT;
3809 sb->s_magic = TMPFS_MAGIC;
3810 sb->s_op = &shmem_ops;
3811 sb->s_time_gran = 1;
3812#ifdef CONFIG_TMPFS_XATTR
3813 sb->s_xattr = shmem_xattr_handlers;
3814#endif
3815#ifdef CONFIG_TMPFS_POSIX_ACL
3816 sb->s_flags |= SB_POSIXACL;
3817#endif
3818 uuid_gen(&sb->s_uuid);
3819
3820 inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
3821 if (!inode)
3822 goto failed;
3823 inode->i_uid = sbinfo->uid;
3824 inode->i_gid = sbinfo->gid;
3825 sb->s_root = d_make_root(inode);
3826 if (!sb->s_root)
3827 goto failed;
3828 return 0;
3829
3830failed:
3831 shmem_put_super(sb);
3832 return -ENOMEM;
3833}
3834
3835static int shmem_get_tree(struct fs_context *fc)
3836{
3837 return get_tree_nodev(fc, shmem_fill_super);
3838}
3839
3840static void shmem_free_fc(struct fs_context *fc)
3841{
3842 struct shmem_options *ctx = fc->fs_private;
3843
3844 if (ctx) {
3845 mpol_put(ctx->mpol);
3846 kfree(ctx);
3847 }
3848}
3849
3850static const struct fs_context_operations shmem_fs_context_ops = {
3851 .free = shmem_free_fc,
3852 .get_tree = shmem_get_tree,
3853#ifdef CONFIG_TMPFS
3854 .parse_monolithic = shmem_parse_options,
3855 .parse_param = shmem_parse_one,
3856 .reconfigure = shmem_reconfigure,
3857#endif
3858};
3859
3860static struct kmem_cache *shmem_inode_cachep;
3861
3862static struct inode *shmem_alloc_inode(struct super_block *sb)
3863{
3864 struct shmem_inode_info *info;
3865 info = alloc_inode_sb(sb, shmem_inode_cachep, GFP_KERNEL);
3866 if (!info)
3867 return NULL;
3868 return &info->vfs_inode;
3869}
3870
3871static void shmem_free_in_core_inode(struct inode *inode)
3872{
3873 if (S_ISLNK(inode->i_mode))
3874 kfree(inode->i_link);
3875 kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
3876}
3877
3878static void shmem_destroy_inode(struct inode *inode)
3879{
3880 if (S_ISREG(inode->i_mode))
3881 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
3882}
3883
3884static void shmem_init_inode(void *foo)
3885{
3886 struct shmem_inode_info *info = foo;
3887 inode_init_once(&info->vfs_inode);
3888}
3889
3890static void shmem_init_inodecache(void)
3891{
3892 shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
3893 sizeof(struct shmem_inode_info),
3894 0, SLAB_PANIC|SLAB_ACCOUNT, shmem_init_inode);
3895}
3896
3897static void shmem_destroy_inodecache(void)
3898{
3899 kmem_cache_destroy(shmem_inode_cachep);
3900}
3901
3902/* Keep the page in page cache instead of truncating it */
3903static int shmem_error_remove_page(struct address_space *mapping,
3904 struct page *page)
3905{
3906 return 0;
3907}
3908
3909const struct address_space_operations shmem_aops = {
3910 .writepage = shmem_writepage,
3911 .dirty_folio = noop_dirty_folio,
3912#ifdef CONFIG_TMPFS
3913 .write_begin = shmem_write_begin,
3914 .write_end = shmem_write_end,
3915#endif
3916#ifdef CONFIG_MIGRATION
3917 .migrate_folio = migrate_folio,
3918#endif
3919 .error_remove_page = shmem_error_remove_page,
3920};
3921EXPORT_SYMBOL(shmem_aops);
3922
3923static const struct file_operations shmem_file_operations = {
3924 .mmap = shmem_mmap,
3925 .open = generic_file_open,
3926 .get_unmapped_area = shmem_get_unmapped_area,
3927#ifdef CONFIG_TMPFS
3928 .llseek = shmem_file_llseek,
3929 .read_iter = shmem_file_read_iter,
3930 .write_iter = generic_file_write_iter,
3931 .fsync = noop_fsync,
3932 .splice_read = generic_file_splice_read,
3933 .splice_write = iter_file_splice_write,
3934 .fallocate = shmem_fallocate,
3935#endif
3936};
3937
3938static const struct inode_operations shmem_inode_operations = {
3939 .getattr = shmem_getattr,
3940 .setattr = shmem_setattr,
3941#ifdef CONFIG_TMPFS_XATTR
3942 .listxattr = shmem_listxattr,
3943 .set_acl = simple_set_acl,
3944 .fileattr_get = shmem_fileattr_get,
3945 .fileattr_set = shmem_fileattr_set,
3946#endif
3947};
3948
3949static const struct inode_operations shmem_dir_inode_operations = {
3950#ifdef CONFIG_TMPFS
3951 .getattr = shmem_getattr,
3952 .create = shmem_create,
3953 .lookup = simple_lookup,
3954 .link = shmem_link,
3955 .unlink = shmem_unlink,
3956 .symlink = shmem_symlink,
3957 .mkdir = shmem_mkdir,
3958 .rmdir = shmem_rmdir,
3959 .mknod = shmem_mknod,
3960 .rename = shmem_rename2,
3961 .tmpfile = shmem_tmpfile,
3962#endif
3963#ifdef CONFIG_TMPFS_XATTR
3964 .listxattr = shmem_listxattr,
3965 .fileattr_get = shmem_fileattr_get,
3966 .fileattr_set = shmem_fileattr_set,
3967#endif
3968#ifdef CONFIG_TMPFS_POSIX_ACL
3969 .setattr = shmem_setattr,
3970 .set_acl = simple_set_acl,
3971#endif
3972};
3973
3974static const struct inode_operations shmem_special_inode_operations = {
3975 .getattr = shmem_getattr,
3976#ifdef CONFIG_TMPFS_XATTR
3977 .listxattr = shmem_listxattr,
3978#endif
3979#ifdef CONFIG_TMPFS_POSIX_ACL
3980 .setattr = shmem_setattr,
3981 .set_acl = simple_set_acl,
3982#endif
3983};
3984
3985static const struct super_operations shmem_ops = {
3986 .alloc_inode = shmem_alloc_inode,
3987 .free_inode = shmem_free_in_core_inode,
3988 .destroy_inode = shmem_destroy_inode,
3989#ifdef CONFIG_TMPFS
3990 .statfs = shmem_statfs,
3991 .show_options = shmem_show_options,
3992#endif
3993 .evict_inode = shmem_evict_inode,
3994 .drop_inode = generic_delete_inode,
3995 .put_super = shmem_put_super,
3996#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3997 .nr_cached_objects = shmem_unused_huge_count,
3998 .free_cached_objects = shmem_unused_huge_scan,
3999#endif
4000};
4001
4002static const struct vm_operations_struct shmem_vm_ops = {
4003 .fault = shmem_fault,
4004 .map_pages = filemap_map_pages,
4005#ifdef CONFIG_NUMA
4006 .set_policy = shmem_set_policy,
4007 .get_policy = shmem_get_policy,
4008#endif
4009};
4010
4011static const struct vm_operations_struct shmem_anon_vm_ops = {
4012 .fault = shmem_fault,
4013 .map_pages = filemap_map_pages,
4014#ifdef CONFIG_NUMA
4015 .set_policy = shmem_set_policy,
4016 .get_policy = shmem_get_policy,
4017#endif
4018};
4019
4020int shmem_init_fs_context(struct fs_context *fc)
4021{
4022 struct shmem_options *ctx;
4023
4024 ctx = kzalloc(sizeof(struct shmem_options), GFP_KERNEL);
4025 if (!ctx)
4026 return -ENOMEM;
4027
4028 ctx->mode = 0777 | S_ISVTX;
4029 ctx->uid = current_fsuid();
4030 ctx->gid = current_fsgid();
4031
4032 fc->fs_private = ctx;
4033 fc->ops = &shmem_fs_context_ops;
4034 return 0;
4035}
4036
4037static struct file_system_type shmem_fs_type = {
4038 .owner = THIS_MODULE,
4039 .name = "tmpfs",
4040 .init_fs_context = shmem_init_fs_context,
4041#ifdef CONFIG_TMPFS
4042 .parameters = shmem_fs_parameters,
4043#endif
4044 .kill_sb = kill_litter_super,
4045 .fs_flags = FS_USERNS_MOUNT,
4046};
4047
4048void __init shmem_init(void)
4049{
4050 int error;
4051
4052 shmem_init_inodecache();
4053
4054 error = register_filesystem(&shmem_fs_type);
4055 if (error) {
4056 pr_err("Could not register tmpfs\n");
4057 goto out2;
4058 }
4059
4060 shm_mnt = kern_mount(&shmem_fs_type);
4061 if (IS_ERR(shm_mnt)) {
4062 error = PTR_ERR(shm_mnt);
4063 pr_err("Could not kern_mount tmpfs\n");
4064 goto out1;
4065 }
4066
4067#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4068 if (has_transparent_hugepage() && shmem_huge > SHMEM_HUGE_DENY)
4069 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
4070 else
4071 shmem_huge = SHMEM_HUGE_NEVER; /* just in case it was patched */
4072#endif
4073 return;
4074
4075out1:
4076 unregister_filesystem(&shmem_fs_type);
4077out2:
4078 shmem_destroy_inodecache();
4079 shm_mnt = ERR_PTR(error);
4080}
4081
4082#if defined(CONFIG_TRANSPARENT_HUGEPAGE) && defined(CONFIG_SYSFS)
4083static ssize_t shmem_enabled_show(struct kobject *kobj,
4084 struct kobj_attribute *attr, char *buf)
4085{
4086 static const int values[] = {
4087 SHMEM_HUGE_ALWAYS,
4088 SHMEM_HUGE_WITHIN_SIZE,
4089 SHMEM_HUGE_ADVISE,
4090 SHMEM_HUGE_NEVER,
4091 SHMEM_HUGE_DENY,
4092 SHMEM_HUGE_FORCE,
4093 };
4094 int len = 0;
4095 int i;
4096
4097 for (i = 0; i < ARRAY_SIZE(values); i++) {
4098 len += sysfs_emit_at(buf, len,
4099 shmem_huge == values[i] ? "%s[%s]" : "%s%s",
4100 i ? " " : "",
4101 shmem_format_huge(values[i]));
4102 }
4103
4104 len += sysfs_emit_at(buf, len, "\n");
4105
4106 return len;
4107}
4108
4109static ssize_t shmem_enabled_store(struct kobject *kobj,
4110 struct kobj_attribute *attr, const char *buf, size_t count)
4111{
4112 char tmp[16];
4113 int huge;
4114
4115 if (count + 1 > sizeof(tmp))
4116 return -EINVAL;
4117 memcpy(tmp, buf, count);
4118 tmp[count] = '\0';
4119 if (count && tmp[count - 1] == '\n')
4120 tmp[count - 1] = '\0';
4121
4122 huge = shmem_parse_huge(tmp);
4123 if (huge == -EINVAL)
4124 return -EINVAL;
4125 if (!has_transparent_hugepage() &&
4126 huge != SHMEM_HUGE_NEVER && huge != SHMEM_HUGE_DENY)
4127 return -EINVAL;
4128
4129 shmem_huge = huge;
4130 if (shmem_huge > SHMEM_HUGE_DENY)
4131 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
4132 return count;
4133}
4134
4135struct kobj_attribute shmem_enabled_attr = __ATTR_RW(shmem_enabled);
4136#endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_SYSFS */
4137
4138#else /* !CONFIG_SHMEM */
4139
4140/*
4141 * tiny-shmem: simple shmemfs and tmpfs using ramfs code
4142 *
4143 * This is intended for small system where the benefits of the full
4144 * shmem code (swap-backed and resource-limited) are outweighed by
4145 * their complexity. On systems without swap this code should be
4146 * effectively equivalent, but much lighter weight.
4147 */
4148
4149static struct file_system_type shmem_fs_type = {
4150 .name = "tmpfs",
4151 .init_fs_context = ramfs_init_fs_context,
4152 .parameters = ramfs_fs_parameters,
4153 .kill_sb = kill_litter_super,
4154 .fs_flags = FS_USERNS_MOUNT,
4155};
4156
4157void __init shmem_init(void)
4158{
4159 BUG_ON(register_filesystem(&shmem_fs_type) != 0);
4160
4161 shm_mnt = kern_mount(&shmem_fs_type);
4162 BUG_ON(IS_ERR(shm_mnt));
4163}
4164
4165int shmem_unuse(unsigned int type)
4166{
4167 return 0;
4168}
4169
4170int shmem_lock(struct file *file, int lock, struct ucounts *ucounts)
4171{
4172 return 0;
4173}
4174
4175void shmem_unlock_mapping(struct address_space *mapping)
4176{
4177}
4178
4179#ifdef CONFIG_MMU
4180unsigned long shmem_get_unmapped_area(struct file *file,
4181 unsigned long addr, unsigned long len,
4182 unsigned long pgoff, unsigned long flags)
4183{
4184 return current->mm->get_unmapped_area(file, addr, len, pgoff, flags);
4185}
4186#endif
4187
4188void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
4189{
4190 truncate_inode_pages_range(inode->i_mapping, lstart, lend);
4191}
4192EXPORT_SYMBOL_GPL(shmem_truncate_range);
4193
4194#define shmem_vm_ops generic_file_vm_ops
4195#define shmem_anon_vm_ops generic_file_vm_ops
4196#define shmem_file_operations ramfs_file_operations
4197#define shmem_get_inode(sb, dir, mode, dev, flags) ramfs_get_inode(sb, dir, mode, dev)
4198#define shmem_acct_size(flags, size) 0
4199#define shmem_unacct_size(flags, size) do {} while (0)
4200
4201#endif /* CONFIG_SHMEM */
4202
4203/* common code */
4204
4205static struct file *__shmem_file_setup(struct vfsmount *mnt, const char *name, loff_t size,
4206 unsigned long flags, unsigned int i_flags)
4207{
4208 struct inode *inode;
4209 struct file *res;
4210
4211 if (IS_ERR(mnt))
4212 return ERR_CAST(mnt);
4213
4214 if (size < 0 || size > MAX_LFS_FILESIZE)
4215 return ERR_PTR(-EINVAL);
4216
4217 if (shmem_acct_size(flags, size))
4218 return ERR_PTR(-ENOMEM);
4219
4220 inode = shmem_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0,
4221 flags);
4222 if (unlikely(!inode)) {
4223 shmem_unacct_size(flags, size);
4224 return ERR_PTR(-ENOSPC);
4225 }
4226 inode->i_flags |= i_flags;
4227 inode->i_size = size;
4228 clear_nlink(inode); /* It is unlinked */
4229 res = ERR_PTR(ramfs_nommu_expand_for_mapping(inode, size));
4230 if (!IS_ERR(res))
4231 res = alloc_file_pseudo(inode, mnt, name, O_RDWR,
4232 &shmem_file_operations);
4233 if (IS_ERR(res))
4234 iput(inode);
4235 return res;
4236}
4237
4238/**
4239 * shmem_kernel_file_setup - get an unlinked file living in tmpfs which must be
4240 * kernel internal. There will be NO LSM permission checks against the
4241 * underlying inode. So users of this interface must do LSM checks at a
4242 * higher layer. The users are the big_key and shm implementations. LSM
4243 * checks are provided at the key or shm level rather than the inode.
4244 * @name: name for dentry (to be seen in /proc/<pid>/maps
4245 * @size: size to be set for the file
4246 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4247 */
4248struct file *shmem_kernel_file_setup(const char *name, loff_t size, unsigned long flags)
4249{
4250 return __shmem_file_setup(shm_mnt, name, size, flags, S_PRIVATE);
4251}
4252
4253/**
4254 * shmem_file_setup - get an unlinked file living in tmpfs
4255 * @name: name for dentry (to be seen in /proc/<pid>/maps
4256 * @size: size to be set for the file
4257 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4258 */
4259struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
4260{
4261 return __shmem_file_setup(shm_mnt, name, size, flags, 0);
4262}
4263EXPORT_SYMBOL_GPL(shmem_file_setup);
4264
4265/**
4266 * shmem_file_setup_with_mnt - get an unlinked file living in tmpfs
4267 * @mnt: the tmpfs mount where the file will be created
4268 * @name: name for dentry (to be seen in /proc/<pid>/maps
4269 * @size: size to be set for the file
4270 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4271 */
4272struct file *shmem_file_setup_with_mnt(struct vfsmount *mnt, const char *name,
4273 loff_t size, unsigned long flags)
4274{
4275 return __shmem_file_setup(mnt, name, size, flags, 0);
4276}
4277EXPORT_SYMBOL_GPL(shmem_file_setup_with_mnt);
4278
4279/**
4280 * shmem_zero_setup - setup a shared anonymous mapping
4281 * @vma: the vma to be mmapped is prepared by do_mmap
4282 */
4283int shmem_zero_setup(struct vm_area_struct *vma)
4284{
4285 struct file *file;
4286 loff_t size = vma->vm_end - vma->vm_start;
4287
4288 /*
4289 * Cloning a new file under mmap_lock leads to a lock ordering conflict
4290 * between XFS directory reading and selinux: since this file is only
4291 * accessible to the user through its mapping, use S_PRIVATE flag to
4292 * bypass file security, in the same way as shmem_kernel_file_setup().
4293 */
4294 file = shmem_kernel_file_setup("dev/zero", size, vma->vm_flags);
4295 if (IS_ERR(file))
4296 return PTR_ERR(file);
4297
4298 if (vma->vm_file)
4299 fput(vma->vm_file);
4300 vma->vm_file = file;
4301 vma->vm_ops = &shmem_anon_vm_ops;
4302
4303 return 0;
4304}
4305
4306/**
4307 * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
4308 * @mapping: the page's address_space
4309 * @index: the page index
4310 * @gfp: the page allocator flags to use if allocating
4311 *
4312 * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
4313 * with any new page allocations done using the specified allocation flags.
4314 * But read_cache_page_gfp() uses the ->read_folio() method: which does not
4315 * suit tmpfs, since it may have pages in swapcache, and needs to find those
4316 * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
4317 *
4318 * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
4319 * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
4320 */
4321struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
4322 pgoff_t index, gfp_t gfp)
4323{
4324#ifdef CONFIG_SHMEM
4325 struct inode *inode = mapping->host;
4326 struct folio *folio;
4327 struct page *page;
4328 int error;
4329
4330 BUG_ON(!shmem_mapping(mapping));
4331 error = shmem_get_folio_gfp(inode, index, &folio, SGP_CACHE,
4332 gfp, NULL, NULL, NULL);
4333 if (error)
4334 return ERR_PTR(error);
4335
4336 folio_unlock(folio);
4337 page = folio_file_page(folio, index);
4338 if (PageHWPoison(page)) {
4339 folio_put(folio);
4340 return ERR_PTR(-EIO);
4341 }
4342
4343 return page;
4344#else
4345 /*
4346 * The tiny !SHMEM case uses ramfs without swap
4347 */
4348 return read_cache_page_gfp(mapping, index, gfp);
4349#endif
4350}
4351EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);