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