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/mm.h>
32#include <linux/sched/signal.h>
33#include <linux/export.h>
34#include <linux/swap.h>
35#include <linux/uio.h>
36#include <linux/khugepaged.h>
37#include <linux/hugetlb.h>
38
39#include <asm/tlbflush.h> /* for arch/microblaze update_mmu_cache() */
40
41static struct vfsmount *shm_mnt;
42
43#ifdef CONFIG_SHMEM
44/*
45 * This virtual memory filesystem is heavily based on the ramfs. It
46 * extends ramfs by the ability to use swap and honor resource limits
47 * which makes it a completely usable filesystem.
48 */
49
50#include <linux/xattr.h>
51#include <linux/exportfs.h>
52#include <linux/posix_acl.h>
53#include <linux/posix_acl_xattr.h>
54#include <linux/mman.h>
55#include <linux/string.h>
56#include <linux/slab.h>
57#include <linux/backing-dev.h>
58#include <linux/shmem_fs.h>
59#include <linux/writeback.h>
60#include <linux/blkdev.h>
61#include <linux/pagevec.h>
62#include <linux/percpu_counter.h>
63#include <linux/falloc.h>
64#include <linux/splice.h>
65#include <linux/security.h>
66#include <linux/swapops.h>
67#include <linux/mempolicy.h>
68#include <linux/namei.h>
69#include <linux/ctype.h>
70#include <linux/migrate.h>
71#include <linux/highmem.h>
72#include <linux/seq_file.h>
73#include <linux/magic.h>
74#include <linux/syscalls.h>
75#include <linux/fcntl.h>
76#include <uapi/linux/memfd.h>
77#include <linux/userfaultfd_k.h>
78#include <linux/rmap.h>
79#include <linux/uuid.h>
80
81#include <linux/uaccess.h>
82#include <asm/pgtable.h>
83
84#include "internal.h"
85
86#define BLOCKS_PER_PAGE (PAGE_SIZE/512)
87#define VM_ACCT(size) (PAGE_ALIGN(size) >> PAGE_SHIFT)
88
89/* Pretend that each entry is of this size in directory's i_size */
90#define BOGO_DIRENT_SIZE 20
91
92/* Symlink up to this size is kmalloc'ed instead of using a swappable page */
93#define SHORT_SYMLINK_LEN 128
94
95/*
96 * shmem_fallocate communicates with shmem_fault or shmem_writepage via
97 * inode->i_private (with i_mutex making sure that it has only one user at
98 * a time): we would prefer not to enlarge the shmem inode just for that.
99 */
100struct shmem_falloc {
101 wait_queue_head_t *waitq; /* faults into hole wait for punch to end */
102 pgoff_t start; /* start of range currently being fallocated */
103 pgoff_t next; /* the next page offset to be fallocated */
104 pgoff_t nr_falloced; /* how many new pages have been fallocated */
105 pgoff_t nr_unswapped; /* how often writepage refused to swap out */
106};
107
108#ifdef CONFIG_TMPFS
109static unsigned long shmem_default_max_blocks(void)
110{
111 return totalram_pages / 2;
112}
113
114static unsigned long shmem_default_max_inodes(void)
115{
116 return min(totalram_pages - totalhigh_pages, totalram_pages / 2);
117}
118#endif
119
120static bool shmem_should_replace_page(struct page *page, gfp_t gfp);
121static int shmem_replace_page(struct page **pagep, gfp_t gfp,
122 struct shmem_inode_info *info, pgoff_t index);
123static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
124 struct page **pagep, enum sgp_type sgp,
125 gfp_t gfp, struct vm_area_struct *vma,
126 struct vm_fault *vmf, int *fault_type);
127
128int shmem_getpage(struct inode *inode, pgoff_t index,
129 struct page **pagep, enum sgp_type sgp)
130{
131 return shmem_getpage_gfp(inode, index, pagep, sgp,
132 mapping_gfp_mask(inode->i_mapping), NULL, NULL, NULL);
133}
134
135static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
136{
137 return sb->s_fs_info;
138}
139
140/*
141 * shmem_file_setup pre-accounts the whole fixed size of a VM object,
142 * for shared memory and for shared anonymous (/dev/zero) mappings
143 * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
144 * consistent with the pre-accounting of private mappings ...
145 */
146static inline int shmem_acct_size(unsigned long flags, loff_t size)
147{
148 return (flags & VM_NORESERVE) ?
149 0 : security_vm_enough_memory_mm(current->mm, VM_ACCT(size));
150}
151
152static inline void shmem_unacct_size(unsigned long flags, loff_t size)
153{
154 if (!(flags & VM_NORESERVE))
155 vm_unacct_memory(VM_ACCT(size));
156}
157
158static inline int shmem_reacct_size(unsigned long flags,
159 loff_t oldsize, loff_t newsize)
160{
161 if (!(flags & VM_NORESERVE)) {
162 if (VM_ACCT(newsize) > VM_ACCT(oldsize))
163 return security_vm_enough_memory_mm(current->mm,
164 VM_ACCT(newsize) - VM_ACCT(oldsize));
165 else if (VM_ACCT(newsize) < VM_ACCT(oldsize))
166 vm_unacct_memory(VM_ACCT(oldsize) - VM_ACCT(newsize));
167 }
168 return 0;
169}
170
171/*
172 * ... whereas tmpfs objects are accounted incrementally as
173 * pages are allocated, in order to allow large sparse files.
174 * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
175 * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
176 */
177static inline int shmem_acct_block(unsigned long flags, long pages)
178{
179 if (!(flags & VM_NORESERVE))
180 return 0;
181
182 return security_vm_enough_memory_mm(current->mm,
183 pages * VM_ACCT(PAGE_SIZE));
184}
185
186static inline void shmem_unacct_blocks(unsigned long flags, long pages)
187{
188 if (flags & VM_NORESERVE)
189 vm_unacct_memory(pages * VM_ACCT(PAGE_SIZE));
190}
191
192static inline bool shmem_inode_acct_block(struct inode *inode, long pages)
193{
194 struct shmem_inode_info *info = SHMEM_I(inode);
195 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
196
197 if (shmem_acct_block(info->flags, pages))
198 return false;
199
200 if (sbinfo->max_blocks) {
201 if (percpu_counter_compare(&sbinfo->used_blocks,
202 sbinfo->max_blocks - pages) > 0)
203 goto unacct;
204 percpu_counter_add(&sbinfo->used_blocks, pages);
205 }
206
207 return true;
208
209unacct:
210 shmem_unacct_blocks(info->flags, pages);
211 return false;
212}
213
214static inline void shmem_inode_unacct_blocks(struct inode *inode, long pages)
215{
216 struct shmem_inode_info *info = SHMEM_I(inode);
217 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
218
219 if (sbinfo->max_blocks)
220 percpu_counter_sub(&sbinfo->used_blocks, pages);
221 shmem_unacct_blocks(info->flags, pages);
222}
223
224static const struct super_operations shmem_ops;
225static const struct address_space_operations shmem_aops;
226static const struct file_operations shmem_file_operations;
227static const struct inode_operations shmem_inode_operations;
228static const struct inode_operations shmem_dir_inode_operations;
229static const struct inode_operations shmem_special_inode_operations;
230static const struct vm_operations_struct shmem_vm_ops;
231static struct file_system_type shmem_fs_type;
232
233bool vma_is_shmem(struct vm_area_struct *vma)
234{
235 return vma->vm_ops == &shmem_vm_ops;
236}
237
238static LIST_HEAD(shmem_swaplist);
239static DEFINE_MUTEX(shmem_swaplist_mutex);
240
241static int shmem_reserve_inode(struct super_block *sb)
242{
243 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
244 if (sbinfo->max_inodes) {
245 spin_lock(&sbinfo->stat_lock);
246 if (!sbinfo->free_inodes) {
247 spin_unlock(&sbinfo->stat_lock);
248 return -ENOSPC;
249 }
250 sbinfo->free_inodes--;
251 spin_unlock(&sbinfo->stat_lock);
252 }
253 return 0;
254}
255
256static void shmem_free_inode(struct super_block *sb)
257{
258 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
259 if (sbinfo->max_inodes) {
260 spin_lock(&sbinfo->stat_lock);
261 sbinfo->free_inodes++;
262 spin_unlock(&sbinfo->stat_lock);
263 }
264}
265
266/**
267 * shmem_recalc_inode - recalculate the block usage of an inode
268 * @inode: inode to recalc
269 *
270 * We have to calculate the free blocks since the mm can drop
271 * undirtied hole pages behind our back.
272 *
273 * But normally info->alloced == inode->i_mapping->nrpages + info->swapped
274 * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
275 *
276 * It has to be called with the spinlock held.
277 */
278static void shmem_recalc_inode(struct inode *inode)
279{
280 struct shmem_inode_info *info = SHMEM_I(inode);
281 long freed;
282
283 freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
284 if (freed > 0) {
285 info->alloced -= freed;
286 inode->i_blocks -= freed * BLOCKS_PER_PAGE;
287 shmem_inode_unacct_blocks(inode, freed);
288 }
289}
290
291bool shmem_charge(struct inode *inode, long pages)
292{
293 struct shmem_inode_info *info = SHMEM_I(inode);
294 unsigned long flags;
295
296 if (!shmem_inode_acct_block(inode, pages))
297 return false;
298
299 spin_lock_irqsave(&info->lock, flags);
300 info->alloced += pages;
301 inode->i_blocks += pages * BLOCKS_PER_PAGE;
302 shmem_recalc_inode(inode);
303 spin_unlock_irqrestore(&info->lock, flags);
304 inode->i_mapping->nrpages += pages;
305
306 return true;
307}
308
309void shmem_uncharge(struct inode *inode, long pages)
310{
311 struct shmem_inode_info *info = SHMEM_I(inode);
312 unsigned long flags;
313
314 spin_lock_irqsave(&info->lock, flags);
315 info->alloced -= pages;
316 inode->i_blocks -= pages * BLOCKS_PER_PAGE;
317 shmem_recalc_inode(inode);
318 spin_unlock_irqrestore(&info->lock, flags);
319
320 shmem_inode_unacct_blocks(inode, pages);
321}
322
323/*
324 * Replace item expected in radix tree by a new item, while holding tree lock.
325 */
326static int shmem_radix_tree_replace(struct address_space *mapping,
327 pgoff_t index, void *expected, void *replacement)
328{
329 struct radix_tree_node *node;
330 void **pslot;
331 void *item;
332
333 VM_BUG_ON(!expected);
334 VM_BUG_ON(!replacement);
335 item = __radix_tree_lookup(&mapping->i_pages, index, &node, &pslot);
336 if (!item)
337 return -ENOENT;
338 if (item != expected)
339 return -ENOENT;
340 __radix_tree_replace(&mapping->i_pages, node, pslot,
341 replacement, NULL);
342 return 0;
343}
344
345/*
346 * Sometimes, before we decide whether to proceed or to fail, we must check
347 * that an entry was not already brought back from swap by a racing thread.
348 *
349 * Checking page is not enough: by the time a SwapCache page is locked, it
350 * might be reused, and again be SwapCache, using the same swap as before.
351 */
352static bool shmem_confirm_swap(struct address_space *mapping,
353 pgoff_t index, swp_entry_t swap)
354{
355 void *item;
356
357 rcu_read_lock();
358 item = radix_tree_lookup(&mapping->i_pages, index);
359 rcu_read_unlock();
360 return item == swp_to_radix_entry(swap);
361}
362
363/*
364 * Definitions for "huge tmpfs": tmpfs mounted with the huge= option
365 *
366 * SHMEM_HUGE_NEVER:
367 * disables huge pages for the mount;
368 * SHMEM_HUGE_ALWAYS:
369 * enables huge pages for the mount;
370 * SHMEM_HUGE_WITHIN_SIZE:
371 * only allocate huge pages if the page will be fully within i_size,
372 * also respect fadvise()/madvise() hints;
373 * SHMEM_HUGE_ADVISE:
374 * only allocate huge pages if requested with fadvise()/madvise();
375 */
376
377#define SHMEM_HUGE_NEVER 0
378#define SHMEM_HUGE_ALWAYS 1
379#define SHMEM_HUGE_WITHIN_SIZE 2
380#define SHMEM_HUGE_ADVISE 3
381
382/*
383 * Special values.
384 * Only can be set via /sys/kernel/mm/transparent_hugepage/shmem_enabled:
385 *
386 * SHMEM_HUGE_DENY:
387 * disables huge on shm_mnt and all mounts, for emergency use;
388 * SHMEM_HUGE_FORCE:
389 * enables huge on shm_mnt and all mounts, w/o needing option, for testing;
390 *
391 */
392#define SHMEM_HUGE_DENY (-1)
393#define SHMEM_HUGE_FORCE (-2)
394
395#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
396/* ifdef here to avoid bloating shmem.o when not necessary */
397
398int shmem_huge __read_mostly;
399
400#if defined(CONFIG_SYSFS) || defined(CONFIG_TMPFS)
401static int shmem_parse_huge(const char *str)
402{
403 if (!strcmp(str, "never"))
404 return SHMEM_HUGE_NEVER;
405 if (!strcmp(str, "always"))
406 return SHMEM_HUGE_ALWAYS;
407 if (!strcmp(str, "within_size"))
408 return SHMEM_HUGE_WITHIN_SIZE;
409 if (!strcmp(str, "advise"))
410 return SHMEM_HUGE_ADVISE;
411 if (!strcmp(str, "deny"))
412 return SHMEM_HUGE_DENY;
413 if (!strcmp(str, "force"))
414 return SHMEM_HUGE_FORCE;
415 return -EINVAL;
416}
417
418static const char *shmem_format_huge(int huge)
419{
420 switch (huge) {
421 case SHMEM_HUGE_NEVER:
422 return "never";
423 case SHMEM_HUGE_ALWAYS:
424 return "always";
425 case SHMEM_HUGE_WITHIN_SIZE:
426 return "within_size";
427 case SHMEM_HUGE_ADVISE:
428 return "advise";
429 case SHMEM_HUGE_DENY:
430 return "deny";
431 case SHMEM_HUGE_FORCE:
432 return "force";
433 default:
434 VM_BUG_ON(1);
435 return "bad_val";
436 }
437}
438#endif
439
440static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
441 struct shrink_control *sc, unsigned long nr_to_split)
442{
443 LIST_HEAD(list), *pos, *next;
444 LIST_HEAD(to_remove);
445 struct inode *inode;
446 struct shmem_inode_info *info;
447 struct page *page;
448 unsigned long batch = sc ? sc->nr_to_scan : 128;
449 int removed = 0, split = 0;
450
451 if (list_empty(&sbinfo->shrinklist))
452 return SHRINK_STOP;
453
454 spin_lock(&sbinfo->shrinklist_lock);
455 list_for_each_safe(pos, next, &sbinfo->shrinklist) {
456 info = list_entry(pos, struct shmem_inode_info, shrinklist);
457
458 /* pin the inode */
459 inode = igrab(&info->vfs_inode);
460
461 /* inode is about to be evicted */
462 if (!inode) {
463 list_del_init(&info->shrinklist);
464 removed++;
465 goto next;
466 }
467
468 /* Check if there's anything to gain */
469 if (round_up(inode->i_size, PAGE_SIZE) ==
470 round_up(inode->i_size, HPAGE_PMD_SIZE)) {
471 list_move(&info->shrinklist, &to_remove);
472 removed++;
473 goto next;
474 }
475
476 list_move(&info->shrinklist, &list);
477next:
478 if (!--batch)
479 break;
480 }
481 spin_unlock(&sbinfo->shrinklist_lock);
482
483 list_for_each_safe(pos, next, &to_remove) {
484 info = list_entry(pos, struct shmem_inode_info, shrinklist);
485 inode = &info->vfs_inode;
486 list_del_init(&info->shrinklist);
487 iput(inode);
488 }
489
490 list_for_each_safe(pos, next, &list) {
491 int ret;
492
493 info = list_entry(pos, struct shmem_inode_info, shrinklist);
494 inode = &info->vfs_inode;
495
496 if (nr_to_split && split >= nr_to_split)
497 goto leave;
498
499 page = find_get_page(inode->i_mapping,
500 (inode->i_size & HPAGE_PMD_MASK) >> PAGE_SHIFT);
501 if (!page)
502 goto drop;
503
504 /* No huge page at the end of the file: nothing to split */
505 if (!PageTransHuge(page)) {
506 put_page(page);
507 goto drop;
508 }
509
510 /*
511 * Leave the inode on the list if we failed to lock
512 * the page at this time.
513 *
514 * Waiting for the lock may lead to deadlock in the
515 * reclaim path.
516 */
517 if (!trylock_page(page)) {
518 put_page(page);
519 goto leave;
520 }
521
522 ret = split_huge_page(page);
523 unlock_page(page);
524 put_page(page);
525
526 /* If split failed leave the inode on the list */
527 if (ret)
528 goto leave;
529
530 split++;
531drop:
532 list_del_init(&info->shrinklist);
533 removed++;
534leave:
535 iput(inode);
536 }
537
538 spin_lock(&sbinfo->shrinklist_lock);
539 list_splice_tail(&list, &sbinfo->shrinklist);
540 sbinfo->shrinklist_len -= removed;
541 spin_unlock(&sbinfo->shrinklist_lock);
542
543 return split;
544}
545
546static long shmem_unused_huge_scan(struct super_block *sb,
547 struct shrink_control *sc)
548{
549 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
550
551 if (!READ_ONCE(sbinfo->shrinklist_len))
552 return SHRINK_STOP;
553
554 return shmem_unused_huge_shrink(sbinfo, sc, 0);
555}
556
557static long shmem_unused_huge_count(struct super_block *sb,
558 struct shrink_control *sc)
559{
560 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
561 return READ_ONCE(sbinfo->shrinklist_len);
562}
563#else /* !CONFIG_TRANSPARENT_HUGE_PAGECACHE */
564
565#define shmem_huge SHMEM_HUGE_DENY
566
567static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
568 struct shrink_control *sc, unsigned long nr_to_split)
569{
570 return 0;
571}
572#endif /* CONFIG_TRANSPARENT_HUGE_PAGECACHE */
573
574/*
575 * Like add_to_page_cache_locked, but error if expected item has gone.
576 */
577static int shmem_add_to_page_cache(struct page *page,
578 struct address_space *mapping,
579 pgoff_t index, void *expected)
580{
581 int error, nr = hpage_nr_pages(page);
582
583 VM_BUG_ON_PAGE(PageTail(page), page);
584 VM_BUG_ON_PAGE(index != round_down(index, nr), page);
585 VM_BUG_ON_PAGE(!PageLocked(page), page);
586 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
587 VM_BUG_ON(expected && PageTransHuge(page));
588
589 page_ref_add(page, nr);
590 page->mapping = mapping;
591 page->index = index;
592
593 xa_lock_irq(&mapping->i_pages);
594 if (PageTransHuge(page)) {
595 void __rcu **results;
596 pgoff_t idx;
597 int i;
598
599 error = 0;
600 if (radix_tree_gang_lookup_slot(&mapping->i_pages,
601 &results, &idx, index, 1) &&
602 idx < index + HPAGE_PMD_NR) {
603 error = -EEXIST;
604 }
605
606 if (!error) {
607 for (i = 0; i < HPAGE_PMD_NR; i++) {
608 error = radix_tree_insert(&mapping->i_pages,
609 index + i, page + i);
610 VM_BUG_ON(error);
611 }
612 count_vm_event(THP_FILE_ALLOC);
613 }
614 } else if (!expected) {
615 error = radix_tree_insert(&mapping->i_pages, index, page);
616 } else {
617 error = shmem_radix_tree_replace(mapping, index, expected,
618 page);
619 }
620
621 if (!error) {
622 mapping->nrpages += nr;
623 if (PageTransHuge(page))
624 __inc_node_page_state(page, NR_SHMEM_THPS);
625 __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, nr);
626 __mod_node_page_state(page_pgdat(page), NR_SHMEM, nr);
627 xa_unlock_irq(&mapping->i_pages);
628 } else {
629 page->mapping = NULL;
630 xa_unlock_irq(&mapping->i_pages);
631 page_ref_sub(page, nr);
632 }
633 return error;
634}
635
636/*
637 * Like delete_from_page_cache, but substitutes swap for page.
638 */
639static void shmem_delete_from_page_cache(struct page *page, void *radswap)
640{
641 struct address_space *mapping = page->mapping;
642 int error;
643
644 VM_BUG_ON_PAGE(PageCompound(page), page);
645
646 xa_lock_irq(&mapping->i_pages);
647 error = shmem_radix_tree_replace(mapping, page->index, page, radswap);
648 page->mapping = NULL;
649 mapping->nrpages--;
650 __dec_node_page_state(page, NR_FILE_PAGES);
651 __dec_node_page_state(page, NR_SHMEM);
652 xa_unlock_irq(&mapping->i_pages);
653 put_page(page);
654 BUG_ON(error);
655}
656
657/*
658 * Remove swap entry from radix tree, free the swap and its page cache.
659 */
660static int shmem_free_swap(struct address_space *mapping,
661 pgoff_t index, void *radswap)
662{
663 void *old;
664
665 xa_lock_irq(&mapping->i_pages);
666 old = radix_tree_delete_item(&mapping->i_pages, index, radswap);
667 xa_unlock_irq(&mapping->i_pages);
668 if (old != radswap)
669 return -ENOENT;
670 free_swap_and_cache(radix_to_swp_entry(radswap));
671 return 0;
672}
673
674/*
675 * Determine (in bytes) how many of the shmem object's pages mapped by the
676 * given offsets are swapped out.
677 *
678 * This is safe to call without i_mutex or the i_pages lock thanks to RCU,
679 * as long as the inode doesn't go away and racy results are not a problem.
680 */
681unsigned long shmem_partial_swap_usage(struct address_space *mapping,
682 pgoff_t start, pgoff_t end)
683{
684 struct radix_tree_iter iter;
685 void **slot;
686 struct page *page;
687 unsigned long swapped = 0;
688
689 rcu_read_lock();
690
691 radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start) {
692 if (iter.index >= end)
693 break;
694
695 page = radix_tree_deref_slot(slot);
696
697 if (radix_tree_deref_retry(page)) {
698 slot = radix_tree_iter_retry(&iter);
699 continue;
700 }
701
702 if (radix_tree_exceptional_entry(page))
703 swapped++;
704
705 if (need_resched()) {
706 slot = radix_tree_iter_resume(slot, &iter);
707 cond_resched_rcu();
708 }
709 }
710
711 rcu_read_unlock();
712
713 return swapped << PAGE_SHIFT;
714}
715
716/*
717 * Determine (in bytes) how many of the shmem object's pages mapped by the
718 * given vma is swapped out.
719 *
720 * This is safe to call without i_mutex or the i_pages lock thanks to RCU,
721 * as long as the inode doesn't go away and racy results are not a problem.
722 */
723unsigned long shmem_swap_usage(struct vm_area_struct *vma)
724{
725 struct inode *inode = file_inode(vma->vm_file);
726 struct shmem_inode_info *info = SHMEM_I(inode);
727 struct address_space *mapping = inode->i_mapping;
728 unsigned long swapped;
729
730 /* Be careful as we don't hold info->lock */
731 swapped = READ_ONCE(info->swapped);
732
733 /*
734 * The easier cases are when the shmem object has nothing in swap, or
735 * the vma maps it whole. Then we can simply use the stats that we
736 * already track.
737 */
738 if (!swapped)
739 return 0;
740
741 if (!vma->vm_pgoff && vma->vm_end - vma->vm_start >= inode->i_size)
742 return swapped << PAGE_SHIFT;
743
744 /* Here comes the more involved part */
745 return shmem_partial_swap_usage(mapping,
746 linear_page_index(vma, vma->vm_start),
747 linear_page_index(vma, vma->vm_end));
748}
749
750/*
751 * SysV IPC SHM_UNLOCK restore Unevictable pages to their evictable lists.
752 */
753void shmem_unlock_mapping(struct address_space *mapping)
754{
755 struct pagevec pvec;
756 pgoff_t indices[PAGEVEC_SIZE];
757 pgoff_t index = 0;
758
759 pagevec_init(&pvec);
760 /*
761 * Minor point, but we might as well stop if someone else SHM_LOCKs it.
762 */
763 while (!mapping_unevictable(mapping)) {
764 /*
765 * Avoid pagevec_lookup(): find_get_pages() returns 0 as if it
766 * has finished, if it hits a row of PAGEVEC_SIZE swap entries.
767 */
768 pvec.nr = find_get_entries(mapping, index,
769 PAGEVEC_SIZE, pvec.pages, indices);
770 if (!pvec.nr)
771 break;
772 index = indices[pvec.nr - 1] + 1;
773 pagevec_remove_exceptionals(&pvec);
774 check_move_unevictable_pages(pvec.pages, pvec.nr);
775 pagevec_release(&pvec);
776 cond_resched();
777 }
778}
779
780/*
781 * Remove range of pages and swap entries from radix tree, and free them.
782 * If !unfalloc, truncate or punch hole; if unfalloc, undo failed fallocate.
783 */
784static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
785 bool unfalloc)
786{
787 struct address_space *mapping = inode->i_mapping;
788 struct shmem_inode_info *info = SHMEM_I(inode);
789 pgoff_t start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
790 pgoff_t end = (lend + 1) >> PAGE_SHIFT;
791 unsigned int partial_start = lstart & (PAGE_SIZE - 1);
792 unsigned int partial_end = (lend + 1) & (PAGE_SIZE - 1);
793 struct pagevec pvec;
794 pgoff_t indices[PAGEVEC_SIZE];
795 long nr_swaps_freed = 0;
796 pgoff_t index;
797 int i;
798
799 if (lend == -1)
800 end = -1; /* unsigned, so actually very big */
801
802 pagevec_init(&pvec);
803 index = start;
804 while (index < end) {
805 pvec.nr = find_get_entries(mapping, index,
806 min(end - index, (pgoff_t)PAGEVEC_SIZE),
807 pvec.pages, indices);
808 if (!pvec.nr)
809 break;
810 for (i = 0; i < pagevec_count(&pvec); i++) {
811 struct page *page = pvec.pages[i];
812
813 index = indices[i];
814 if (index >= end)
815 break;
816
817 if (radix_tree_exceptional_entry(page)) {
818 if (unfalloc)
819 continue;
820 nr_swaps_freed += !shmem_free_swap(mapping,
821 index, page);
822 continue;
823 }
824
825 VM_BUG_ON_PAGE(page_to_pgoff(page) != index, page);
826
827 if (!trylock_page(page))
828 continue;
829
830 if (PageTransTail(page)) {
831 /* Middle of THP: zero out the page */
832 clear_highpage(page);
833 unlock_page(page);
834 continue;
835 } else if (PageTransHuge(page)) {
836 if (index == round_down(end, HPAGE_PMD_NR)) {
837 /*
838 * Range ends in the middle of THP:
839 * zero out the page
840 */
841 clear_highpage(page);
842 unlock_page(page);
843 continue;
844 }
845 index += HPAGE_PMD_NR - 1;
846 i += HPAGE_PMD_NR - 1;
847 }
848
849 if (!unfalloc || !PageUptodate(page)) {
850 VM_BUG_ON_PAGE(PageTail(page), page);
851 if (page_mapping(page) == mapping) {
852 VM_BUG_ON_PAGE(PageWriteback(page), page);
853 truncate_inode_page(mapping, page);
854 }
855 }
856 unlock_page(page);
857 }
858 pagevec_remove_exceptionals(&pvec);
859 pagevec_release(&pvec);
860 cond_resched();
861 index++;
862 }
863
864 if (partial_start) {
865 struct page *page = NULL;
866 shmem_getpage(inode, start - 1, &page, SGP_READ);
867 if (page) {
868 unsigned int top = PAGE_SIZE;
869 if (start > end) {
870 top = partial_end;
871 partial_end = 0;
872 }
873 zero_user_segment(page, partial_start, top);
874 set_page_dirty(page);
875 unlock_page(page);
876 put_page(page);
877 }
878 }
879 if (partial_end) {
880 struct page *page = NULL;
881 shmem_getpage(inode, end, &page, SGP_READ);
882 if (page) {
883 zero_user_segment(page, 0, partial_end);
884 set_page_dirty(page);
885 unlock_page(page);
886 put_page(page);
887 }
888 }
889 if (start >= end)
890 return;
891
892 index = start;
893 while (index < end) {
894 cond_resched();
895
896 pvec.nr = find_get_entries(mapping, index,
897 min(end - index, (pgoff_t)PAGEVEC_SIZE),
898 pvec.pages, indices);
899 if (!pvec.nr) {
900 /* If all gone or hole-punch or unfalloc, we're done */
901 if (index == start || end != -1)
902 break;
903 /* But if truncating, restart to make sure all gone */
904 index = start;
905 continue;
906 }
907 for (i = 0; i < pagevec_count(&pvec); i++) {
908 struct page *page = pvec.pages[i];
909
910 index = indices[i];
911 if (index >= end)
912 break;
913
914 if (radix_tree_exceptional_entry(page)) {
915 if (unfalloc)
916 continue;
917 if (shmem_free_swap(mapping, index, page)) {
918 /* Swap was replaced by page: retry */
919 index--;
920 break;
921 }
922 nr_swaps_freed++;
923 continue;
924 }
925
926 lock_page(page);
927
928 if (PageTransTail(page)) {
929 /* Middle of THP: zero out the page */
930 clear_highpage(page);
931 unlock_page(page);
932 /*
933 * Partial thp truncate due 'start' in middle
934 * of THP: don't need to look on these pages
935 * again on !pvec.nr restart.
936 */
937 if (index != round_down(end, HPAGE_PMD_NR))
938 start++;
939 continue;
940 } else if (PageTransHuge(page)) {
941 if (index == round_down(end, HPAGE_PMD_NR)) {
942 /*
943 * Range ends in the middle of THP:
944 * zero out the page
945 */
946 clear_highpage(page);
947 unlock_page(page);
948 continue;
949 }
950 index += HPAGE_PMD_NR - 1;
951 i += HPAGE_PMD_NR - 1;
952 }
953
954 if (!unfalloc || !PageUptodate(page)) {
955 VM_BUG_ON_PAGE(PageTail(page), page);
956 if (page_mapping(page) == mapping) {
957 VM_BUG_ON_PAGE(PageWriteback(page), page);
958 truncate_inode_page(mapping, page);
959 } else {
960 /* Page was replaced by swap: retry */
961 unlock_page(page);
962 index--;
963 break;
964 }
965 }
966 unlock_page(page);
967 }
968 pagevec_remove_exceptionals(&pvec);
969 pagevec_release(&pvec);
970 index++;
971 }
972
973 spin_lock_irq(&info->lock);
974 info->swapped -= nr_swaps_freed;
975 shmem_recalc_inode(inode);
976 spin_unlock_irq(&info->lock);
977}
978
979void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
980{
981 shmem_undo_range(inode, lstart, lend, false);
982 inode->i_ctime = inode->i_mtime = current_time(inode);
983}
984EXPORT_SYMBOL_GPL(shmem_truncate_range);
985
986static int shmem_getattr(const struct path *path, struct kstat *stat,
987 u32 request_mask, unsigned int query_flags)
988{
989 struct inode *inode = path->dentry->d_inode;
990 struct shmem_inode_info *info = SHMEM_I(inode);
991
992 if (info->alloced - info->swapped != inode->i_mapping->nrpages) {
993 spin_lock_irq(&info->lock);
994 shmem_recalc_inode(inode);
995 spin_unlock_irq(&info->lock);
996 }
997 generic_fillattr(inode, stat);
998 return 0;
999}
1000
1001static int shmem_setattr(struct dentry *dentry, struct iattr *attr)
1002{
1003 struct inode *inode = d_inode(dentry);
1004 struct shmem_inode_info *info = SHMEM_I(inode);
1005 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1006 int error;
1007
1008 error = setattr_prepare(dentry, attr);
1009 if (error)
1010 return error;
1011
1012 if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
1013 loff_t oldsize = inode->i_size;
1014 loff_t newsize = attr->ia_size;
1015
1016 /* protected by i_mutex */
1017 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
1018 (newsize > oldsize && (info->seals & F_SEAL_GROW)))
1019 return -EPERM;
1020
1021 if (newsize != oldsize) {
1022 error = shmem_reacct_size(SHMEM_I(inode)->flags,
1023 oldsize, newsize);
1024 if (error)
1025 return error;
1026 i_size_write(inode, newsize);
1027 inode->i_ctime = inode->i_mtime = current_time(inode);
1028 }
1029 if (newsize <= oldsize) {
1030 loff_t holebegin = round_up(newsize, PAGE_SIZE);
1031 if (oldsize > holebegin)
1032 unmap_mapping_range(inode->i_mapping,
1033 holebegin, 0, 1);
1034 if (info->alloced)
1035 shmem_truncate_range(inode,
1036 newsize, (loff_t)-1);
1037 /* unmap again to remove racily COWed private pages */
1038 if (oldsize > holebegin)
1039 unmap_mapping_range(inode->i_mapping,
1040 holebegin, 0, 1);
1041
1042 /*
1043 * Part of the huge page can be beyond i_size: subject
1044 * to shrink under memory pressure.
1045 */
1046 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
1047 spin_lock(&sbinfo->shrinklist_lock);
1048 /*
1049 * _careful to defend against unlocked access to
1050 * ->shrink_list in shmem_unused_huge_shrink()
1051 */
1052 if (list_empty_careful(&info->shrinklist)) {
1053 list_add_tail(&info->shrinklist,
1054 &sbinfo->shrinklist);
1055 sbinfo->shrinklist_len++;
1056 }
1057 spin_unlock(&sbinfo->shrinklist_lock);
1058 }
1059 }
1060 }
1061
1062 setattr_copy(inode, attr);
1063 if (attr->ia_valid & ATTR_MODE)
1064 error = posix_acl_chmod(inode, inode->i_mode);
1065 return error;
1066}
1067
1068static void shmem_evict_inode(struct inode *inode)
1069{
1070 struct shmem_inode_info *info = SHMEM_I(inode);
1071 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1072
1073 if (inode->i_mapping->a_ops == &shmem_aops) {
1074 shmem_unacct_size(info->flags, inode->i_size);
1075 inode->i_size = 0;
1076 shmem_truncate_range(inode, 0, (loff_t)-1);
1077 if (!list_empty(&info->shrinklist)) {
1078 spin_lock(&sbinfo->shrinklist_lock);
1079 if (!list_empty(&info->shrinklist)) {
1080 list_del_init(&info->shrinklist);
1081 sbinfo->shrinklist_len--;
1082 }
1083 spin_unlock(&sbinfo->shrinklist_lock);
1084 }
1085 if (!list_empty(&info->swaplist)) {
1086 mutex_lock(&shmem_swaplist_mutex);
1087 list_del_init(&info->swaplist);
1088 mutex_unlock(&shmem_swaplist_mutex);
1089 }
1090 }
1091
1092 simple_xattrs_free(&info->xattrs);
1093 WARN_ON(inode->i_blocks);
1094 shmem_free_inode(inode->i_sb);
1095 clear_inode(inode);
1096}
1097
1098static unsigned long find_swap_entry(struct radix_tree_root *root, void *item)
1099{
1100 struct radix_tree_iter iter;
1101 void **slot;
1102 unsigned long found = -1;
1103 unsigned int checked = 0;
1104
1105 rcu_read_lock();
1106 radix_tree_for_each_slot(slot, root, &iter, 0) {
1107 if (*slot == item) {
1108 found = iter.index;
1109 break;
1110 }
1111 checked++;
1112 if ((checked % 4096) != 0)
1113 continue;
1114 slot = radix_tree_iter_resume(slot, &iter);
1115 cond_resched_rcu();
1116 }
1117
1118 rcu_read_unlock();
1119 return found;
1120}
1121
1122/*
1123 * If swap found in inode, free it and move page from swapcache to filecache.
1124 */
1125static int shmem_unuse_inode(struct shmem_inode_info *info,
1126 swp_entry_t swap, struct page **pagep)
1127{
1128 struct address_space *mapping = info->vfs_inode.i_mapping;
1129 void *radswap;
1130 pgoff_t index;
1131 gfp_t gfp;
1132 int error = 0;
1133
1134 radswap = swp_to_radix_entry(swap);
1135 index = find_swap_entry(&mapping->i_pages, radswap);
1136 if (index == -1)
1137 return -EAGAIN; /* tell shmem_unuse we found nothing */
1138
1139 /*
1140 * Move _head_ to start search for next from here.
1141 * But be careful: shmem_evict_inode checks list_empty without taking
1142 * mutex, and there's an instant in list_move_tail when info->swaplist
1143 * would appear empty, if it were the only one on shmem_swaplist.
1144 */
1145 if (shmem_swaplist.next != &info->swaplist)
1146 list_move_tail(&shmem_swaplist, &info->swaplist);
1147
1148 gfp = mapping_gfp_mask(mapping);
1149 if (shmem_should_replace_page(*pagep, gfp)) {
1150 mutex_unlock(&shmem_swaplist_mutex);
1151 error = shmem_replace_page(pagep, gfp, info, index);
1152 mutex_lock(&shmem_swaplist_mutex);
1153 /*
1154 * We needed to drop mutex to make that restrictive page
1155 * allocation, but the inode might have been freed while we
1156 * dropped it: although a racing shmem_evict_inode() cannot
1157 * complete without emptying the radix_tree, our page lock
1158 * on this swapcache page is not enough to prevent that -
1159 * free_swap_and_cache() of our swap entry will only
1160 * trylock_page(), removing swap from radix_tree whatever.
1161 *
1162 * We must not proceed to shmem_add_to_page_cache() if the
1163 * inode has been freed, but of course we cannot rely on
1164 * inode or mapping or info to check that. However, we can
1165 * safely check if our swap entry is still in use (and here
1166 * it can't have got reused for another page): if it's still
1167 * in use, then the inode cannot have been freed yet, and we
1168 * can safely proceed (if it's no longer in use, that tells
1169 * nothing about the inode, but we don't need to unuse swap).
1170 */
1171 if (!page_swapcount(*pagep))
1172 error = -ENOENT;
1173 }
1174
1175 /*
1176 * We rely on shmem_swaplist_mutex, not only to protect the swaplist,
1177 * but also to hold up shmem_evict_inode(): so inode cannot be freed
1178 * beneath us (pagelock doesn't help until the page is in pagecache).
1179 */
1180 if (!error)
1181 error = shmem_add_to_page_cache(*pagep, mapping, index,
1182 radswap);
1183 if (error != -ENOMEM) {
1184 /*
1185 * Truncation and eviction use free_swap_and_cache(), which
1186 * only does trylock page: if we raced, best clean up here.
1187 */
1188 delete_from_swap_cache(*pagep);
1189 set_page_dirty(*pagep);
1190 if (!error) {
1191 spin_lock_irq(&info->lock);
1192 info->swapped--;
1193 spin_unlock_irq(&info->lock);
1194 swap_free(swap);
1195 }
1196 }
1197 return error;
1198}
1199
1200/*
1201 * Search through swapped inodes to find and replace swap by page.
1202 */
1203int shmem_unuse(swp_entry_t swap, struct page *page)
1204{
1205 struct list_head *this, *next;
1206 struct shmem_inode_info *info;
1207 struct mem_cgroup *memcg;
1208 int error = 0;
1209
1210 /*
1211 * There's a faint possibility that swap page was replaced before
1212 * caller locked it: caller will come back later with the right page.
1213 */
1214 if (unlikely(!PageSwapCache(page) || page_private(page) != swap.val))
1215 goto out;
1216
1217 /*
1218 * Charge page using GFP_KERNEL while we can wait, before taking
1219 * the shmem_swaplist_mutex which might hold up shmem_writepage().
1220 * Charged back to the user (not to caller) when swap account is used.
1221 */
1222 error = mem_cgroup_try_charge(page, current->mm, GFP_KERNEL, &memcg,
1223 false);
1224 if (error)
1225 goto out;
1226 /* No radix_tree_preload: swap entry keeps a place for page in tree */
1227 error = -EAGAIN;
1228
1229 mutex_lock(&shmem_swaplist_mutex);
1230 list_for_each_safe(this, next, &shmem_swaplist) {
1231 info = list_entry(this, struct shmem_inode_info, swaplist);
1232 if (info->swapped)
1233 error = shmem_unuse_inode(info, swap, &page);
1234 else
1235 list_del_init(&info->swaplist);
1236 cond_resched();
1237 if (error != -EAGAIN)
1238 break;
1239 /* found nothing in this: move on to search the next */
1240 }
1241 mutex_unlock(&shmem_swaplist_mutex);
1242
1243 if (error) {
1244 if (error != -ENOMEM)
1245 error = 0;
1246 mem_cgroup_cancel_charge(page, memcg, false);
1247 } else
1248 mem_cgroup_commit_charge(page, memcg, true, false);
1249out:
1250 unlock_page(page);
1251 put_page(page);
1252 return error;
1253}
1254
1255/*
1256 * Move the page from the page cache to the swap cache.
1257 */
1258static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1259{
1260 struct shmem_inode_info *info;
1261 struct address_space *mapping;
1262 struct inode *inode;
1263 swp_entry_t swap;
1264 pgoff_t index;
1265
1266 VM_BUG_ON_PAGE(PageCompound(page), page);
1267 BUG_ON(!PageLocked(page));
1268 mapping = page->mapping;
1269 index = page->index;
1270 inode = mapping->host;
1271 info = SHMEM_I(inode);
1272 if (info->flags & VM_LOCKED)
1273 goto redirty;
1274 if (!total_swap_pages)
1275 goto redirty;
1276
1277 /*
1278 * Our capabilities prevent regular writeback or sync from ever calling
1279 * shmem_writepage; but a stacking filesystem might use ->writepage of
1280 * its underlying filesystem, in which case tmpfs should write out to
1281 * swap only in response to memory pressure, and not for the writeback
1282 * threads or sync.
1283 */
1284 if (!wbc->for_reclaim) {
1285 WARN_ON_ONCE(1); /* Still happens? Tell us about it! */
1286 goto redirty;
1287 }
1288
1289 /*
1290 * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
1291 * value into swapfile.c, the only way we can correctly account for a
1292 * fallocated page arriving here is now to initialize it and write it.
1293 *
1294 * That's okay for a page already fallocated earlier, but if we have
1295 * not yet completed the fallocation, then (a) we want to keep track
1296 * of this page in case we have to undo it, and (b) it may not be a
1297 * good idea to continue anyway, once we're pushing into swap. So
1298 * reactivate the page, and let shmem_fallocate() quit when too many.
1299 */
1300 if (!PageUptodate(page)) {
1301 if (inode->i_private) {
1302 struct shmem_falloc *shmem_falloc;
1303 spin_lock(&inode->i_lock);
1304 shmem_falloc = inode->i_private;
1305 if (shmem_falloc &&
1306 !shmem_falloc->waitq &&
1307 index >= shmem_falloc->start &&
1308 index < shmem_falloc->next)
1309 shmem_falloc->nr_unswapped++;
1310 else
1311 shmem_falloc = NULL;
1312 spin_unlock(&inode->i_lock);
1313 if (shmem_falloc)
1314 goto redirty;
1315 }
1316 clear_highpage(page);
1317 flush_dcache_page(page);
1318 SetPageUptodate(page);
1319 }
1320
1321 swap = get_swap_page(page);
1322 if (!swap.val)
1323 goto redirty;
1324
1325 if (mem_cgroup_try_charge_swap(page, swap))
1326 goto free_swap;
1327
1328 /*
1329 * Add inode to shmem_unuse()'s list of swapped-out inodes,
1330 * if it's not already there. Do it now before the page is
1331 * moved to swap cache, when its pagelock no longer protects
1332 * the inode from eviction. But don't unlock the mutex until
1333 * we've incremented swapped, because shmem_unuse_inode() will
1334 * prune a !swapped inode from the swaplist under this mutex.
1335 */
1336 mutex_lock(&shmem_swaplist_mutex);
1337 if (list_empty(&info->swaplist))
1338 list_add_tail(&info->swaplist, &shmem_swaplist);
1339
1340 if (add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
1341 spin_lock_irq(&info->lock);
1342 shmem_recalc_inode(inode);
1343 info->swapped++;
1344 spin_unlock_irq(&info->lock);
1345
1346 swap_shmem_alloc(swap);
1347 shmem_delete_from_page_cache(page, swp_to_radix_entry(swap));
1348
1349 mutex_unlock(&shmem_swaplist_mutex);
1350 BUG_ON(page_mapped(page));
1351 swap_writepage(page, wbc);
1352 return 0;
1353 }
1354
1355 mutex_unlock(&shmem_swaplist_mutex);
1356free_swap:
1357 put_swap_page(page, swap);
1358redirty:
1359 set_page_dirty(page);
1360 if (wbc->for_reclaim)
1361 return AOP_WRITEPAGE_ACTIVATE; /* Return with page locked */
1362 unlock_page(page);
1363 return 0;
1364}
1365
1366#if defined(CONFIG_NUMA) && defined(CONFIG_TMPFS)
1367static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1368{
1369 char buffer[64];
1370
1371 if (!mpol || mpol->mode == MPOL_DEFAULT)
1372 return; /* show nothing */
1373
1374 mpol_to_str(buffer, sizeof(buffer), mpol);
1375
1376 seq_printf(seq, ",mpol=%s", buffer);
1377}
1378
1379static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1380{
1381 struct mempolicy *mpol = NULL;
1382 if (sbinfo->mpol) {
1383 spin_lock(&sbinfo->stat_lock); /* prevent replace/use races */
1384 mpol = sbinfo->mpol;
1385 mpol_get(mpol);
1386 spin_unlock(&sbinfo->stat_lock);
1387 }
1388 return mpol;
1389}
1390#else /* !CONFIG_NUMA || !CONFIG_TMPFS */
1391static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1392{
1393}
1394static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1395{
1396 return NULL;
1397}
1398#endif /* CONFIG_NUMA && CONFIG_TMPFS */
1399#ifndef CONFIG_NUMA
1400#define vm_policy vm_private_data
1401#endif
1402
1403static void shmem_pseudo_vma_init(struct vm_area_struct *vma,
1404 struct shmem_inode_info *info, pgoff_t index)
1405{
1406 /* Create a pseudo vma that just contains the policy */
1407 vma->vm_start = 0;
1408 /* Bias interleave by inode number to distribute better across nodes */
1409 vma->vm_pgoff = index + info->vfs_inode.i_ino;
1410 vma->vm_ops = NULL;
1411 vma->vm_policy = mpol_shared_policy_lookup(&info->policy, index);
1412}
1413
1414static void shmem_pseudo_vma_destroy(struct vm_area_struct *vma)
1415{
1416 /* Drop reference taken by mpol_shared_policy_lookup() */
1417 mpol_cond_put(vma->vm_policy);
1418}
1419
1420static struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
1421 struct shmem_inode_info *info, pgoff_t index)
1422{
1423 struct vm_area_struct pvma;
1424 struct page *page;
1425 struct vm_fault vmf;
1426
1427 shmem_pseudo_vma_init(&pvma, info, index);
1428 vmf.vma = &pvma;
1429 vmf.address = 0;
1430 page = swap_cluster_readahead(swap, gfp, &vmf);
1431 shmem_pseudo_vma_destroy(&pvma);
1432
1433 return page;
1434}
1435
1436static struct page *shmem_alloc_hugepage(gfp_t gfp,
1437 struct shmem_inode_info *info, pgoff_t index)
1438{
1439 struct vm_area_struct pvma;
1440 struct inode *inode = &info->vfs_inode;
1441 struct address_space *mapping = inode->i_mapping;
1442 pgoff_t idx, hindex;
1443 void __rcu **results;
1444 struct page *page;
1445
1446 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
1447 return NULL;
1448
1449 hindex = round_down(index, HPAGE_PMD_NR);
1450 rcu_read_lock();
1451 if (radix_tree_gang_lookup_slot(&mapping->i_pages, &results, &idx,
1452 hindex, 1) && idx < hindex + HPAGE_PMD_NR) {
1453 rcu_read_unlock();
1454 return NULL;
1455 }
1456 rcu_read_unlock();
1457
1458 shmem_pseudo_vma_init(&pvma, info, hindex);
1459 page = alloc_pages_vma(gfp | __GFP_COMP | __GFP_NORETRY | __GFP_NOWARN,
1460 HPAGE_PMD_ORDER, &pvma, 0, numa_node_id(), true);
1461 shmem_pseudo_vma_destroy(&pvma);
1462 if (page)
1463 prep_transhuge_page(page);
1464 return page;
1465}
1466
1467static struct page *shmem_alloc_page(gfp_t gfp,
1468 struct shmem_inode_info *info, pgoff_t index)
1469{
1470 struct vm_area_struct pvma;
1471 struct page *page;
1472
1473 shmem_pseudo_vma_init(&pvma, info, index);
1474 page = alloc_page_vma(gfp, &pvma, 0);
1475 shmem_pseudo_vma_destroy(&pvma);
1476
1477 return page;
1478}
1479
1480static struct page *shmem_alloc_and_acct_page(gfp_t gfp,
1481 struct inode *inode,
1482 pgoff_t index, bool huge)
1483{
1484 struct shmem_inode_info *info = SHMEM_I(inode);
1485 struct page *page;
1486 int nr;
1487 int err = -ENOSPC;
1488
1489 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
1490 huge = false;
1491 nr = huge ? HPAGE_PMD_NR : 1;
1492
1493 if (!shmem_inode_acct_block(inode, nr))
1494 goto failed;
1495
1496 if (huge)
1497 page = shmem_alloc_hugepage(gfp, info, index);
1498 else
1499 page = shmem_alloc_page(gfp, info, index);
1500 if (page) {
1501 __SetPageLocked(page);
1502 __SetPageSwapBacked(page);
1503 return page;
1504 }
1505
1506 err = -ENOMEM;
1507 shmem_inode_unacct_blocks(inode, nr);
1508failed:
1509 return ERR_PTR(err);
1510}
1511
1512/*
1513 * When a page is moved from swapcache to shmem filecache (either by the
1514 * usual swapin of shmem_getpage_gfp(), or by the less common swapoff of
1515 * shmem_unuse_inode()), it may have been read in earlier from swap, in
1516 * ignorance of the mapping it belongs to. If that mapping has special
1517 * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
1518 * we may need to copy to a suitable page before moving to filecache.
1519 *
1520 * In a future release, this may well be extended to respect cpuset and
1521 * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
1522 * but for now it is a simple matter of zone.
1523 */
1524static bool shmem_should_replace_page(struct page *page, gfp_t gfp)
1525{
1526 return page_zonenum(page) > gfp_zone(gfp);
1527}
1528
1529static int shmem_replace_page(struct page **pagep, gfp_t gfp,
1530 struct shmem_inode_info *info, pgoff_t index)
1531{
1532 struct page *oldpage, *newpage;
1533 struct address_space *swap_mapping;
1534 pgoff_t swap_index;
1535 int error;
1536
1537 oldpage = *pagep;
1538 swap_index = page_private(oldpage);
1539 swap_mapping = page_mapping(oldpage);
1540
1541 /*
1542 * We have arrived here because our zones are constrained, so don't
1543 * limit chance of success by further cpuset and node constraints.
1544 */
1545 gfp &= ~GFP_CONSTRAINT_MASK;
1546 newpage = shmem_alloc_page(gfp, info, index);
1547 if (!newpage)
1548 return -ENOMEM;
1549
1550 get_page(newpage);
1551 copy_highpage(newpage, oldpage);
1552 flush_dcache_page(newpage);
1553
1554 __SetPageLocked(newpage);
1555 __SetPageSwapBacked(newpage);
1556 SetPageUptodate(newpage);
1557 set_page_private(newpage, swap_index);
1558 SetPageSwapCache(newpage);
1559
1560 /*
1561 * Our caller will very soon move newpage out of swapcache, but it's
1562 * a nice clean interface for us to replace oldpage by newpage there.
1563 */
1564 xa_lock_irq(&swap_mapping->i_pages);
1565 error = shmem_radix_tree_replace(swap_mapping, swap_index, oldpage,
1566 newpage);
1567 if (!error) {
1568 __inc_node_page_state(newpage, NR_FILE_PAGES);
1569 __dec_node_page_state(oldpage, NR_FILE_PAGES);
1570 }
1571 xa_unlock_irq(&swap_mapping->i_pages);
1572
1573 if (unlikely(error)) {
1574 /*
1575 * Is this possible? I think not, now that our callers check
1576 * both PageSwapCache and page_private after getting page lock;
1577 * but be defensive. Reverse old to newpage for clear and free.
1578 */
1579 oldpage = newpage;
1580 } else {
1581 mem_cgroup_migrate(oldpage, newpage);
1582 lru_cache_add_anon(newpage);
1583 *pagep = newpage;
1584 }
1585
1586 ClearPageSwapCache(oldpage);
1587 set_page_private(oldpage, 0);
1588
1589 unlock_page(oldpage);
1590 put_page(oldpage);
1591 put_page(oldpage);
1592 return error;
1593}
1594
1595/*
1596 * shmem_getpage_gfp - find page in cache, or get from swap, or allocate
1597 *
1598 * If we allocate a new one we do not mark it dirty. That's up to the
1599 * vm. If we swap it in we mark it dirty since we also free the swap
1600 * entry since a page cannot live in both the swap and page cache.
1601 *
1602 * fault_mm and fault_type are only supplied by shmem_fault:
1603 * otherwise they are NULL.
1604 */
1605static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
1606 struct page **pagep, enum sgp_type sgp, gfp_t gfp,
1607 struct vm_area_struct *vma, struct vm_fault *vmf, int *fault_type)
1608{
1609 struct address_space *mapping = inode->i_mapping;
1610 struct shmem_inode_info *info = SHMEM_I(inode);
1611 struct shmem_sb_info *sbinfo;
1612 struct mm_struct *charge_mm;
1613 struct mem_cgroup *memcg;
1614 struct page *page;
1615 swp_entry_t swap;
1616 enum sgp_type sgp_huge = sgp;
1617 pgoff_t hindex = index;
1618 int error;
1619 int once = 0;
1620 int alloced = 0;
1621
1622 if (index > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
1623 return -EFBIG;
1624 if (sgp == SGP_NOHUGE || sgp == SGP_HUGE)
1625 sgp = SGP_CACHE;
1626repeat:
1627 swap.val = 0;
1628 page = find_lock_entry(mapping, index);
1629 if (radix_tree_exceptional_entry(page)) {
1630 swap = radix_to_swp_entry(page);
1631 page = NULL;
1632 }
1633
1634 if (sgp <= SGP_CACHE &&
1635 ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1636 error = -EINVAL;
1637 goto unlock;
1638 }
1639
1640 if (page && sgp == SGP_WRITE)
1641 mark_page_accessed(page);
1642
1643 /* fallocated page? */
1644 if (page && !PageUptodate(page)) {
1645 if (sgp != SGP_READ)
1646 goto clear;
1647 unlock_page(page);
1648 put_page(page);
1649 page = NULL;
1650 }
1651 if (page || (sgp == SGP_READ && !swap.val)) {
1652 *pagep = page;
1653 return 0;
1654 }
1655
1656 /*
1657 * Fast cache lookup did not find it:
1658 * bring it back from swap or allocate.
1659 */
1660 sbinfo = SHMEM_SB(inode->i_sb);
1661 charge_mm = vma ? vma->vm_mm : current->mm;
1662
1663 if (swap.val) {
1664 /* Look it up and read it in.. */
1665 page = lookup_swap_cache(swap, NULL, 0);
1666 if (!page) {
1667 /* Or update major stats only when swapin succeeds?? */
1668 if (fault_type) {
1669 *fault_type |= VM_FAULT_MAJOR;
1670 count_vm_event(PGMAJFAULT);
1671 count_memcg_event_mm(charge_mm, PGMAJFAULT);
1672 }
1673 /* Here we actually start the io */
1674 page = shmem_swapin(swap, gfp, info, index);
1675 if (!page) {
1676 error = -ENOMEM;
1677 goto failed;
1678 }
1679 }
1680
1681 /* We have to do this with page locked to prevent races */
1682 lock_page(page);
1683 if (!PageSwapCache(page) || page_private(page) != swap.val ||
1684 !shmem_confirm_swap(mapping, index, swap)) {
1685 error = -EEXIST; /* try again */
1686 goto unlock;
1687 }
1688 if (!PageUptodate(page)) {
1689 error = -EIO;
1690 goto failed;
1691 }
1692 wait_on_page_writeback(page);
1693
1694 if (shmem_should_replace_page(page, gfp)) {
1695 error = shmem_replace_page(&page, gfp, info, index);
1696 if (error)
1697 goto failed;
1698 }
1699
1700 error = mem_cgroup_try_charge(page, charge_mm, gfp, &memcg,
1701 false);
1702 if (!error) {
1703 error = shmem_add_to_page_cache(page, mapping, index,
1704 swp_to_radix_entry(swap));
1705 /*
1706 * We already confirmed swap under page lock, and make
1707 * no memory allocation here, so usually no possibility
1708 * of error; but free_swap_and_cache() only trylocks a
1709 * page, so it is just possible that the entry has been
1710 * truncated or holepunched since swap was confirmed.
1711 * shmem_undo_range() will have done some of the
1712 * unaccounting, now delete_from_swap_cache() will do
1713 * the rest.
1714 * Reset swap.val? No, leave it so "failed" goes back to
1715 * "repeat": reading a hole and writing should succeed.
1716 */
1717 if (error) {
1718 mem_cgroup_cancel_charge(page, memcg, false);
1719 delete_from_swap_cache(page);
1720 }
1721 }
1722 if (error)
1723 goto failed;
1724
1725 mem_cgroup_commit_charge(page, memcg, true, false);
1726
1727 spin_lock_irq(&info->lock);
1728 info->swapped--;
1729 shmem_recalc_inode(inode);
1730 spin_unlock_irq(&info->lock);
1731
1732 if (sgp == SGP_WRITE)
1733 mark_page_accessed(page);
1734
1735 delete_from_swap_cache(page);
1736 set_page_dirty(page);
1737 swap_free(swap);
1738
1739 } else {
1740 if (vma && userfaultfd_missing(vma)) {
1741 *fault_type = handle_userfault(vmf, VM_UFFD_MISSING);
1742 return 0;
1743 }
1744
1745 /* shmem_symlink() */
1746 if (mapping->a_ops != &shmem_aops)
1747 goto alloc_nohuge;
1748 if (shmem_huge == SHMEM_HUGE_DENY || sgp_huge == SGP_NOHUGE)
1749 goto alloc_nohuge;
1750 if (shmem_huge == SHMEM_HUGE_FORCE)
1751 goto alloc_huge;
1752 switch (sbinfo->huge) {
1753 loff_t i_size;
1754 pgoff_t off;
1755 case SHMEM_HUGE_NEVER:
1756 goto alloc_nohuge;
1757 case SHMEM_HUGE_WITHIN_SIZE:
1758 off = round_up(index, HPAGE_PMD_NR);
1759 i_size = round_up(i_size_read(inode), PAGE_SIZE);
1760 if (i_size >= HPAGE_PMD_SIZE &&
1761 i_size >> PAGE_SHIFT >= off)
1762 goto alloc_huge;
1763 /* fallthrough */
1764 case SHMEM_HUGE_ADVISE:
1765 if (sgp_huge == SGP_HUGE)
1766 goto alloc_huge;
1767 /* TODO: implement fadvise() hints */
1768 goto alloc_nohuge;
1769 }
1770
1771alloc_huge:
1772 page = shmem_alloc_and_acct_page(gfp, inode, index, true);
1773 if (IS_ERR(page)) {
1774alloc_nohuge: page = shmem_alloc_and_acct_page(gfp, inode,
1775 index, false);
1776 }
1777 if (IS_ERR(page)) {
1778 int retry = 5;
1779 error = PTR_ERR(page);
1780 page = NULL;
1781 if (error != -ENOSPC)
1782 goto failed;
1783 /*
1784 * Try to reclaim some spece by splitting a huge page
1785 * beyond i_size on the filesystem.
1786 */
1787 while (retry--) {
1788 int ret;
1789 ret = shmem_unused_huge_shrink(sbinfo, NULL, 1);
1790 if (ret == SHRINK_STOP)
1791 break;
1792 if (ret)
1793 goto alloc_nohuge;
1794 }
1795 goto failed;
1796 }
1797
1798 if (PageTransHuge(page))
1799 hindex = round_down(index, HPAGE_PMD_NR);
1800 else
1801 hindex = index;
1802
1803 if (sgp == SGP_WRITE)
1804 __SetPageReferenced(page);
1805
1806 error = mem_cgroup_try_charge(page, charge_mm, gfp, &memcg,
1807 PageTransHuge(page));
1808 if (error)
1809 goto unacct;
1810 error = radix_tree_maybe_preload_order(gfp & GFP_RECLAIM_MASK,
1811 compound_order(page));
1812 if (!error) {
1813 error = shmem_add_to_page_cache(page, mapping, hindex,
1814 NULL);
1815 radix_tree_preload_end();
1816 }
1817 if (error) {
1818 mem_cgroup_cancel_charge(page, memcg,
1819 PageTransHuge(page));
1820 goto unacct;
1821 }
1822 mem_cgroup_commit_charge(page, memcg, false,
1823 PageTransHuge(page));
1824 lru_cache_add_anon(page);
1825
1826 spin_lock_irq(&info->lock);
1827 info->alloced += 1 << compound_order(page);
1828 inode->i_blocks += BLOCKS_PER_PAGE << compound_order(page);
1829 shmem_recalc_inode(inode);
1830 spin_unlock_irq(&info->lock);
1831 alloced = true;
1832
1833 if (PageTransHuge(page) &&
1834 DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE) <
1835 hindex + HPAGE_PMD_NR - 1) {
1836 /*
1837 * Part of the huge page is beyond i_size: subject
1838 * to shrink under memory pressure.
1839 */
1840 spin_lock(&sbinfo->shrinklist_lock);
1841 /*
1842 * _careful to defend against unlocked access to
1843 * ->shrink_list in shmem_unused_huge_shrink()
1844 */
1845 if (list_empty_careful(&info->shrinklist)) {
1846 list_add_tail(&info->shrinklist,
1847 &sbinfo->shrinklist);
1848 sbinfo->shrinklist_len++;
1849 }
1850 spin_unlock(&sbinfo->shrinklist_lock);
1851 }
1852
1853 /*
1854 * Let SGP_FALLOC use the SGP_WRITE optimization on a new page.
1855 */
1856 if (sgp == SGP_FALLOC)
1857 sgp = SGP_WRITE;
1858clear:
1859 /*
1860 * Let SGP_WRITE caller clear ends if write does not fill page;
1861 * but SGP_FALLOC on a page fallocated earlier must initialize
1862 * it now, lest undo on failure cancel our earlier guarantee.
1863 */
1864 if (sgp != SGP_WRITE && !PageUptodate(page)) {
1865 struct page *head = compound_head(page);
1866 int i;
1867
1868 for (i = 0; i < (1 << compound_order(head)); i++) {
1869 clear_highpage(head + i);
1870 flush_dcache_page(head + i);
1871 }
1872 SetPageUptodate(head);
1873 }
1874 }
1875
1876 /* Perhaps the file has been truncated since we checked */
1877 if (sgp <= SGP_CACHE &&
1878 ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1879 if (alloced) {
1880 ClearPageDirty(page);
1881 delete_from_page_cache(page);
1882 spin_lock_irq(&info->lock);
1883 shmem_recalc_inode(inode);
1884 spin_unlock_irq(&info->lock);
1885 }
1886 error = -EINVAL;
1887 goto unlock;
1888 }
1889 *pagep = page + index - hindex;
1890 return 0;
1891
1892 /*
1893 * Error recovery.
1894 */
1895unacct:
1896 shmem_inode_unacct_blocks(inode, 1 << compound_order(page));
1897
1898 if (PageTransHuge(page)) {
1899 unlock_page(page);
1900 put_page(page);
1901 goto alloc_nohuge;
1902 }
1903failed:
1904 if (swap.val && !shmem_confirm_swap(mapping, index, swap))
1905 error = -EEXIST;
1906unlock:
1907 if (page) {
1908 unlock_page(page);
1909 put_page(page);
1910 }
1911 if (error == -ENOSPC && !once++) {
1912 spin_lock_irq(&info->lock);
1913 shmem_recalc_inode(inode);
1914 spin_unlock_irq(&info->lock);
1915 goto repeat;
1916 }
1917 if (error == -EEXIST) /* from above or from radix_tree_insert */
1918 goto repeat;
1919 return error;
1920}
1921
1922/*
1923 * This is like autoremove_wake_function, but it removes the wait queue
1924 * entry unconditionally - even if something else had already woken the
1925 * target.
1926 */
1927static int synchronous_wake_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
1928{
1929 int ret = default_wake_function(wait, mode, sync, key);
1930 list_del_init(&wait->entry);
1931 return ret;
1932}
1933
1934static int shmem_fault(struct vm_fault *vmf)
1935{
1936 struct vm_area_struct *vma = vmf->vma;
1937 struct inode *inode = file_inode(vma->vm_file);
1938 gfp_t gfp = mapping_gfp_mask(inode->i_mapping);
1939 enum sgp_type sgp;
1940 int error;
1941 int ret = VM_FAULT_LOCKED;
1942
1943 /*
1944 * Trinity finds that probing a hole which tmpfs is punching can
1945 * prevent the hole-punch from ever completing: which in turn
1946 * locks writers out with its hold on i_mutex. So refrain from
1947 * faulting pages into the hole while it's being punched. Although
1948 * shmem_undo_range() does remove the additions, it may be unable to
1949 * keep up, as each new page needs its own unmap_mapping_range() call,
1950 * and the i_mmap tree grows ever slower to scan if new vmas are added.
1951 *
1952 * It does not matter if we sometimes reach this check just before the
1953 * hole-punch begins, so that one fault then races with the punch:
1954 * we just need to make racing faults a rare case.
1955 *
1956 * The implementation below would be much simpler if we just used a
1957 * standard mutex or completion: but we cannot take i_mutex in fault,
1958 * and bloating every shmem inode for this unlikely case would be sad.
1959 */
1960 if (unlikely(inode->i_private)) {
1961 struct shmem_falloc *shmem_falloc;
1962
1963 spin_lock(&inode->i_lock);
1964 shmem_falloc = inode->i_private;
1965 if (shmem_falloc &&
1966 shmem_falloc->waitq &&
1967 vmf->pgoff >= shmem_falloc->start &&
1968 vmf->pgoff < shmem_falloc->next) {
1969 wait_queue_head_t *shmem_falloc_waitq;
1970 DEFINE_WAIT_FUNC(shmem_fault_wait, synchronous_wake_function);
1971
1972 ret = VM_FAULT_NOPAGE;
1973 if ((vmf->flags & FAULT_FLAG_ALLOW_RETRY) &&
1974 !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
1975 /* It's polite to up mmap_sem if we can */
1976 up_read(&vma->vm_mm->mmap_sem);
1977 ret = VM_FAULT_RETRY;
1978 }
1979
1980 shmem_falloc_waitq = shmem_falloc->waitq;
1981 prepare_to_wait(shmem_falloc_waitq, &shmem_fault_wait,
1982 TASK_UNINTERRUPTIBLE);
1983 spin_unlock(&inode->i_lock);
1984 schedule();
1985
1986 /*
1987 * shmem_falloc_waitq points into the shmem_fallocate()
1988 * stack of the hole-punching task: shmem_falloc_waitq
1989 * is usually invalid by the time we reach here, but
1990 * finish_wait() does not dereference it in that case;
1991 * though i_lock needed lest racing with wake_up_all().
1992 */
1993 spin_lock(&inode->i_lock);
1994 finish_wait(shmem_falloc_waitq, &shmem_fault_wait);
1995 spin_unlock(&inode->i_lock);
1996 return ret;
1997 }
1998 spin_unlock(&inode->i_lock);
1999 }
2000
2001 sgp = SGP_CACHE;
2002
2003 if ((vma->vm_flags & VM_NOHUGEPAGE) ||
2004 test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
2005 sgp = SGP_NOHUGE;
2006 else if (vma->vm_flags & VM_HUGEPAGE)
2007 sgp = SGP_HUGE;
2008
2009 error = shmem_getpage_gfp(inode, vmf->pgoff, &vmf->page, sgp,
2010 gfp, vma, vmf, &ret);
2011 if (error)
2012 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
2013 return ret;
2014}
2015
2016unsigned long shmem_get_unmapped_area(struct file *file,
2017 unsigned long uaddr, unsigned long len,
2018 unsigned long pgoff, unsigned long flags)
2019{
2020 unsigned long (*get_area)(struct file *,
2021 unsigned long, unsigned long, unsigned long, unsigned long);
2022 unsigned long addr;
2023 unsigned long offset;
2024 unsigned long inflated_len;
2025 unsigned long inflated_addr;
2026 unsigned long inflated_offset;
2027
2028 if (len > TASK_SIZE)
2029 return -ENOMEM;
2030
2031 get_area = current->mm->get_unmapped_area;
2032 addr = get_area(file, uaddr, len, pgoff, flags);
2033
2034 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
2035 return addr;
2036 if (IS_ERR_VALUE(addr))
2037 return addr;
2038 if (addr & ~PAGE_MASK)
2039 return addr;
2040 if (addr > TASK_SIZE - len)
2041 return addr;
2042
2043 if (shmem_huge == SHMEM_HUGE_DENY)
2044 return addr;
2045 if (len < HPAGE_PMD_SIZE)
2046 return addr;
2047 if (flags & MAP_FIXED)
2048 return addr;
2049 /*
2050 * Our priority is to support MAP_SHARED mapped hugely;
2051 * and support MAP_PRIVATE mapped hugely too, until it is COWed.
2052 * But if caller specified an address hint, respect that as before.
2053 */
2054 if (uaddr)
2055 return addr;
2056
2057 if (shmem_huge != SHMEM_HUGE_FORCE) {
2058 struct super_block *sb;
2059
2060 if (file) {
2061 VM_BUG_ON(file->f_op != &shmem_file_operations);
2062 sb = file_inode(file)->i_sb;
2063 } else {
2064 /*
2065 * Called directly from mm/mmap.c, or drivers/char/mem.c
2066 * for "/dev/zero", to create a shared anonymous object.
2067 */
2068 if (IS_ERR(shm_mnt))
2069 return addr;
2070 sb = shm_mnt->mnt_sb;
2071 }
2072 if (SHMEM_SB(sb)->huge == SHMEM_HUGE_NEVER)
2073 return addr;
2074 }
2075
2076 offset = (pgoff << PAGE_SHIFT) & (HPAGE_PMD_SIZE-1);
2077 if (offset && offset + len < 2 * HPAGE_PMD_SIZE)
2078 return addr;
2079 if ((addr & (HPAGE_PMD_SIZE-1)) == offset)
2080 return addr;
2081
2082 inflated_len = len + HPAGE_PMD_SIZE - PAGE_SIZE;
2083 if (inflated_len > TASK_SIZE)
2084 return addr;
2085 if (inflated_len < len)
2086 return addr;
2087
2088 inflated_addr = get_area(NULL, 0, inflated_len, 0, flags);
2089 if (IS_ERR_VALUE(inflated_addr))
2090 return addr;
2091 if (inflated_addr & ~PAGE_MASK)
2092 return addr;
2093
2094 inflated_offset = inflated_addr & (HPAGE_PMD_SIZE-1);
2095 inflated_addr += offset - inflated_offset;
2096 if (inflated_offset > offset)
2097 inflated_addr += HPAGE_PMD_SIZE;
2098
2099 if (inflated_addr > TASK_SIZE - len)
2100 return addr;
2101 return inflated_addr;
2102}
2103
2104#ifdef CONFIG_NUMA
2105static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
2106{
2107 struct inode *inode = file_inode(vma->vm_file);
2108 return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
2109}
2110
2111static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
2112 unsigned long addr)
2113{
2114 struct inode *inode = file_inode(vma->vm_file);
2115 pgoff_t index;
2116
2117 index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2118 return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
2119}
2120#endif
2121
2122int shmem_lock(struct file *file, int lock, struct user_struct *user)
2123{
2124 struct inode *inode = file_inode(file);
2125 struct shmem_inode_info *info = SHMEM_I(inode);
2126 int retval = -ENOMEM;
2127
2128 spin_lock_irq(&info->lock);
2129 if (lock && !(info->flags & VM_LOCKED)) {
2130 if (!user_shm_lock(inode->i_size, user))
2131 goto out_nomem;
2132 info->flags |= VM_LOCKED;
2133 mapping_set_unevictable(file->f_mapping);
2134 }
2135 if (!lock && (info->flags & VM_LOCKED) && user) {
2136 user_shm_unlock(inode->i_size, user);
2137 info->flags &= ~VM_LOCKED;
2138 mapping_clear_unevictable(file->f_mapping);
2139 }
2140 retval = 0;
2141
2142out_nomem:
2143 spin_unlock_irq(&info->lock);
2144 return retval;
2145}
2146
2147static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
2148{
2149 file_accessed(file);
2150 vma->vm_ops = &shmem_vm_ops;
2151 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE) &&
2152 ((vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK) <
2153 (vma->vm_end & HPAGE_PMD_MASK)) {
2154 khugepaged_enter(vma, vma->vm_flags);
2155 }
2156 return 0;
2157}
2158
2159static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
2160 umode_t mode, dev_t dev, unsigned long flags)
2161{
2162 struct inode *inode;
2163 struct shmem_inode_info *info;
2164 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2165
2166 if (shmem_reserve_inode(sb))
2167 return NULL;
2168
2169 inode = new_inode(sb);
2170 if (inode) {
2171 inode->i_ino = get_next_ino();
2172 inode_init_owner(inode, dir, mode);
2173 inode->i_blocks = 0;
2174 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
2175 inode->i_generation = get_seconds();
2176 info = SHMEM_I(inode);
2177 memset(info, 0, (char *)inode - (char *)info);
2178 spin_lock_init(&info->lock);
2179 info->seals = F_SEAL_SEAL;
2180 info->flags = flags & VM_NORESERVE;
2181 INIT_LIST_HEAD(&info->shrinklist);
2182 INIT_LIST_HEAD(&info->swaplist);
2183 simple_xattrs_init(&info->xattrs);
2184 cache_no_acl(inode);
2185
2186 switch (mode & S_IFMT) {
2187 default:
2188 inode->i_op = &shmem_special_inode_operations;
2189 init_special_inode(inode, mode, dev);
2190 break;
2191 case S_IFREG:
2192 inode->i_mapping->a_ops = &shmem_aops;
2193 inode->i_op = &shmem_inode_operations;
2194 inode->i_fop = &shmem_file_operations;
2195 mpol_shared_policy_init(&info->policy,
2196 shmem_get_sbmpol(sbinfo));
2197 break;
2198 case S_IFDIR:
2199 inc_nlink(inode);
2200 /* Some things misbehave if size == 0 on a directory */
2201 inode->i_size = 2 * BOGO_DIRENT_SIZE;
2202 inode->i_op = &shmem_dir_inode_operations;
2203 inode->i_fop = &simple_dir_operations;
2204 break;
2205 case S_IFLNK:
2206 /*
2207 * Must not load anything in the rbtree,
2208 * mpol_free_shared_policy will not be called.
2209 */
2210 mpol_shared_policy_init(&info->policy, NULL);
2211 break;
2212 }
2213 } else
2214 shmem_free_inode(sb);
2215 return inode;
2216}
2217
2218bool shmem_mapping(struct address_space *mapping)
2219{
2220 return mapping->a_ops == &shmem_aops;
2221}
2222
2223static int shmem_mfill_atomic_pte(struct mm_struct *dst_mm,
2224 pmd_t *dst_pmd,
2225 struct vm_area_struct *dst_vma,
2226 unsigned long dst_addr,
2227 unsigned long src_addr,
2228 bool zeropage,
2229 struct page **pagep)
2230{
2231 struct inode *inode = file_inode(dst_vma->vm_file);
2232 struct shmem_inode_info *info = SHMEM_I(inode);
2233 struct address_space *mapping = inode->i_mapping;
2234 gfp_t gfp = mapping_gfp_mask(mapping);
2235 pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
2236 struct mem_cgroup *memcg;
2237 spinlock_t *ptl;
2238 void *page_kaddr;
2239 struct page *page;
2240 pte_t _dst_pte, *dst_pte;
2241 int ret;
2242
2243 ret = -ENOMEM;
2244 if (!shmem_inode_acct_block(inode, 1))
2245 goto out;
2246
2247 if (!*pagep) {
2248 page = shmem_alloc_page(gfp, info, pgoff);
2249 if (!page)
2250 goto out_unacct_blocks;
2251
2252 if (!zeropage) { /* mcopy_atomic */
2253 page_kaddr = kmap_atomic(page);
2254 ret = copy_from_user(page_kaddr,
2255 (const void __user *)src_addr,
2256 PAGE_SIZE);
2257 kunmap_atomic(page_kaddr);
2258
2259 /* fallback to copy_from_user outside mmap_sem */
2260 if (unlikely(ret)) {
2261 *pagep = page;
2262 shmem_inode_unacct_blocks(inode, 1);
2263 /* don't free the page */
2264 return -EFAULT;
2265 }
2266 } else { /* mfill_zeropage_atomic */
2267 clear_highpage(page);
2268 }
2269 } else {
2270 page = *pagep;
2271 *pagep = NULL;
2272 }
2273
2274 VM_BUG_ON(PageLocked(page) || PageSwapBacked(page));
2275 __SetPageLocked(page);
2276 __SetPageSwapBacked(page);
2277 __SetPageUptodate(page);
2278
2279 ret = mem_cgroup_try_charge(page, dst_mm, gfp, &memcg, false);
2280 if (ret)
2281 goto out_release;
2282
2283 ret = radix_tree_maybe_preload(gfp & GFP_RECLAIM_MASK);
2284 if (!ret) {
2285 ret = shmem_add_to_page_cache(page, mapping, pgoff, NULL);
2286 radix_tree_preload_end();
2287 }
2288 if (ret)
2289 goto out_release_uncharge;
2290
2291 mem_cgroup_commit_charge(page, memcg, false, false);
2292
2293 _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
2294 if (dst_vma->vm_flags & VM_WRITE)
2295 _dst_pte = pte_mkwrite(pte_mkdirty(_dst_pte));
2296
2297 ret = -EEXIST;
2298 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
2299 if (!pte_none(*dst_pte))
2300 goto out_release_uncharge_unlock;
2301
2302 lru_cache_add_anon(page);
2303
2304 spin_lock(&info->lock);
2305 info->alloced++;
2306 inode->i_blocks += BLOCKS_PER_PAGE;
2307 shmem_recalc_inode(inode);
2308 spin_unlock(&info->lock);
2309
2310 inc_mm_counter(dst_mm, mm_counter_file(page));
2311 page_add_file_rmap(page, false);
2312 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
2313
2314 /* No need to invalidate - it was non-present before */
2315 update_mmu_cache(dst_vma, dst_addr, dst_pte);
2316 unlock_page(page);
2317 pte_unmap_unlock(dst_pte, ptl);
2318 ret = 0;
2319out:
2320 return ret;
2321out_release_uncharge_unlock:
2322 pte_unmap_unlock(dst_pte, ptl);
2323out_release_uncharge:
2324 mem_cgroup_cancel_charge(page, memcg, false);
2325out_release:
2326 unlock_page(page);
2327 put_page(page);
2328out_unacct_blocks:
2329 shmem_inode_unacct_blocks(inode, 1);
2330 goto out;
2331}
2332
2333int shmem_mcopy_atomic_pte(struct mm_struct *dst_mm,
2334 pmd_t *dst_pmd,
2335 struct vm_area_struct *dst_vma,
2336 unsigned long dst_addr,
2337 unsigned long src_addr,
2338 struct page **pagep)
2339{
2340 return shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
2341 dst_addr, src_addr, false, pagep);
2342}
2343
2344int shmem_mfill_zeropage_pte(struct mm_struct *dst_mm,
2345 pmd_t *dst_pmd,
2346 struct vm_area_struct *dst_vma,
2347 unsigned long dst_addr)
2348{
2349 struct page *page = NULL;
2350
2351 return shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
2352 dst_addr, 0, true, &page);
2353}
2354
2355#ifdef CONFIG_TMPFS
2356static const struct inode_operations shmem_symlink_inode_operations;
2357static const struct inode_operations shmem_short_symlink_operations;
2358
2359#ifdef CONFIG_TMPFS_XATTR
2360static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
2361#else
2362#define shmem_initxattrs NULL
2363#endif
2364
2365static int
2366shmem_write_begin(struct file *file, struct address_space *mapping,
2367 loff_t pos, unsigned len, unsigned flags,
2368 struct page **pagep, void **fsdata)
2369{
2370 struct inode *inode = mapping->host;
2371 struct shmem_inode_info *info = SHMEM_I(inode);
2372 pgoff_t index = pos >> PAGE_SHIFT;
2373
2374 /* i_mutex is held by caller */
2375 if (unlikely(info->seals & (F_SEAL_WRITE | F_SEAL_GROW))) {
2376 if (info->seals & F_SEAL_WRITE)
2377 return -EPERM;
2378 if ((info->seals & F_SEAL_GROW) && pos + len > inode->i_size)
2379 return -EPERM;
2380 }
2381
2382 return shmem_getpage(inode, index, pagep, SGP_WRITE);
2383}
2384
2385static int
2386shmem_write_end(struct file *file, struct address_space *mapping,
2387 loff_t pos, unsigned len, unsigned copied,
2388 struct page *page, void *fsdata)
2389{
2390 struct inode *inode = mapping->host;
2391
2392 if (pos + copied > inode->i_size)
2393 i_size_write(inode, pos + copied);
2394
2395 if (!PageUptodate(page)) {
2396 struct page *head = compound_head(page);
2397 if (PageTransCompound(page)) {
2398 int i;
2399
2400 for (i = 0; i < HPAGE_PMD_NR; i++) {
2401 if (head + i == page)
2402 continue;
2403 clear_highpage(head + i);
2404 flush_dcache_page(head + i);
2405 }
2406 }
2407 if (copied < PAGE_SIZE) {
2408 unsigned from = pos & (PAGE_SIZE - 1);
2409 zero_user_segments(page, 0, from,
2410 from + copied, PAGE_SIZE);
2411 }
2412 SetPageUptodate(head);
2413 }
2414 set_page_dirty(page);
2415 unlock_page(page);
2416 put_page(page);
2417
2418 return copied;
2419}
2420
2421static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
2422{
2423 struct file *file = iocb->ki_filp;
2424 struct inode *inode = file_inode(file);
2425 struct address_space *mapping = inode->i_mapping;
2426 pgoff_t index;
2427 unsigned long offset;
2428 enum sgp_type sgp = SGP_READ;
2429 int error = 0;
2430 ssize_t retval = 0;
2431 loff_t *ppos = &iocb->ki_pos;
2432
2433 /*
2434 * Might this read be for a stacking filesystem? Then when reading
2435 * holes of a sparse file, we actually need to allocate those pages,
2436 * and even mark them dirty, so it cannot exceed the max_blocks limit.
2437 */
2438 if (!iter_is_iovec(to))
2439 sgp = SGP_CACHE;
2440
2441 index = *ppos >> PAGE_SHIFT;
2442 offset = *ppos & ~PAGE_MASK;
2443
2444 for (;;) {
2445 struct page *page = NULL;
2446 pgoff_t end_index;
2447 unsigned long nr, ret;
2448 loff_t i_size = i_size_read(inode);
2449
2450 end_index = i_size >> PAGE_SHIFT;
2451 if (index > end_index)
2452 break;
2453 if (index == end_index) {
2454 nr = i_size & ~PAGE_MASK;
2455 if (nr <= offset)
2456 break;
2457 }
2458
2459 error = shmem_getpage(inode, index, &page, sgp);
2460 if (error) {
2461 if (error == -EINVAL)
2462 error = 0;
2463 break;
2464 }
2465 if (page) {
2466 if (sgp == SGP_CACHE)
2467 set_page_dirty(page);
2468 unlock_page(page);
2469 }
2470
2471 /*
2472 * We must evaluate after, since reads (unlike writes)
2473 * are called without i_mutex protection against truncate
2474 */
2475 nr = PAGE_SIZE;
2476 i_size = i_size_read(inode);
2477 end_index = i_size >> PAGE_SHIFT;
2478 if (index == end_index) {
2479 nr = i_size & ~PAGE_MASK;
2480 if (nr <= offset) {
2481 if (page)
2482 put_page(page);
2483 break;
2484 }
2485 }
2486 nr -= offset;
2487
2488 if (page) {
2489 /*
2490 * If users can be writing to this page using arbitrary
2491 * virtual addresses, take care about potential aliasing
2492 * before reading the page on the kernel side.
2493 */
2494 if (mapping_writably_mapped(mapping))
2495 flush_dcache_page(page);
2496 /*
2497 * Mark the page accessed if we read the beginning.
2498 */
2499 if (!offset)
2500 mark_page_accessed(page);
2501 } else {
2502 page = ZERO_PAGE(0);
2503 get_page(page);
2504 }
2505
2506 /*
2507 * Ok, we have the page, and it's up-to-date, so
2508 * now we can copy it to user space...
2509 */
2510 ret = copy_page_to_iter(page, offset, nr, to);
2511 retval += ret;
2512 offset += ret;
2513 index += offset >> PAGE_SHIFT;
2514 offset &= ~PAGE_MASK;
2515
2516 put_page(page);
2517 if (!iov_iter_count(to))
2518 break;
2519 if (ret < nr) {
2520 error = -EFAULT;
2521 break;
2522 }
2523 cond_resched();
2524 }
2525
2526 *ppos = ((loff_t) index << PAGE_SHIFT) + offset;
2527 file_accessed(file);
2528 return retval ? retval : error;
2529}
2530
2531/*
2532 * llseek SEEK_DATA or SEEK_HOLE through the radix_tree.
2533 */
2534static pgoff_t shmem_seek_hole_data(struct address_space *mapping,
2535 pgoff_t index, pgoff_t end, int whence)
2536{
2537 struct page *page;
2538 struct pagevec pvec;
2539 pgoff_t indices[PAGEVEC_SIZE];
2540 bool done = false;
2541 int i;
2542
2543 pagevec_init(&pvec);
2544 pvec.nr = 1; /* start small: we may be there already */
2545 while (!done) {
2546 pvec.nr = find_get_entries(mapping, index,
2547 pvec.nr, pvec.pages, indices);
2548 if (!pvec.nr) {
2549 if (whence == SEEK_DATA)
2550 index = end;
2551 break;
2552 }
2553 for (i = 0; i < pvec.nr; i++, index++) {
2554 if (index < indices[i]) {
2555 if (whence == SEEK_HOLE) {
2556 done = true;
2557 break;
2558 }
2559 index = indices[i];
2560 }
2561 page = pvec.pages[i];
2562 if (page && !radix_tree_exceptional_entry(page)) {
2563 if (!PageUptodate(page))
2564 page = NULL;
2565 }
2566 if (index >= end ||
2567 (page && whence == SEEK_DATA) ||
2568 (!page && whence == SEEK_HOLE)) {
2569 done = true;
2570 break;
2571 }
2572 }
2573 pagevec_remove_exceptionals(&pvec);
2574 pagevec_release(&pvec);
2575 pvec.nr = PAGEVEC_SIZE;
2576 cond_resched();
2577 }
2578 return index;
2579}
2580
2581static loff_t shmem_file_llseek(struct file *file, loff_t offset, int whence)
2582{
2583 struct address_space *mapping = file->f_mapping;
2584 struct inode *inode = mapping->host;
2585 pgoff_t start, end;
2586 loff_t new_offset;
2587
2588 if (whence != SEEK_DATA && whence != SEEK_HOLE)
2589 return generic_file_llseek_size(file, offset, whence,
2590 MAX_LFS_FILESIZE, i_size_read(inode));
2591 inode_lock(inode);
2592 /* We're holding i_mutex so we can access i_size directly */
2593
2594 if (offset < 0)
2595 offset = -EINVAL;
2596 else if (offset >= inode->i_size)
2597 offset = -ENXIO;
2598 else {
2599 start = offset >> PAGE_SHIFT;
2600 end = (inode->i_size + PAGE_SIZE - 1) >> PAGE_SHIFT;
2601 new_offset = shmem_seek_hole_data(mapping, start, end, whence);
2602 new_offset <<= PAGE_SHIFT;
2603 if (new_offset > offset) {
2604 if (new_offset < inode->i_size)
2605 offset = new_offset;
2606 else if (whence == SEEK_DATA)
2607 offset = -ENXIO;
2608 else
2609 offset = inode->i_size;
2610 }
2611 }
2612
2613 if (offset >= 0)
2614 offset = vfs_setpos(file, offset, MAX_LFS_FILESIZE);
2615 inode_unlock(inode);
2616 return offset;
2617}
2618
2619/*
2620 * We need a tag: a new tag would expand every radix_tree_node by 8 bytes,
2621 * so reuse a tag which we firmly believe is never set or cleared on shmem.
2622 */
2623#define SHMEM_TAG_PINNED PAGECACHE_TAG_TOWRITE
2624#define LAST_SCAN 4 /* about 150ms max */
2625
2626static void shmem_tag_pins(struct address_space *mapping)
2627{
2628 struct radix_tree_iter iter;
2629 void **slot;
2630 pgoff_t start;
2631 struct page *page;
2632
2633 lru_add_drain();
2634 start = 0;
2635 rcu_read_lock();
2636
2637 radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start) {
2638 page = radix_tree_deref_slot(slot);
2639 if (!page || radix_tree_exception(page)) {
2640 if (radix_tree_deref_retry(page)) {
2641 slot = radix_tree_iter_retry(&iter);
2642 continue;
2643 }
2644 } else if (page_count(page) - page_mapcount(page) > 1) {
2645 xa_lock_irq(&mapping->i_pages);
2646 radix_tree_tag_set(&mapping->i_pages, iter.index,
2647 SHMEM_TAG_PINNED);
2648 xa_unlock_irq(&mapping->i_pages);
2649 }
2650
2651 if (need_resched()) {
2652 slot = radix_tree_iter_resume(slot, &iter);
2653 cond_resched_rcu();
2654 }
2655 }
2656 rcu_read_unlock();
2657}
2658
2659/*
2660 * Setting SEAL_WRITE requires us to verify there's no pending writer. However,
2661 * via get_user_pages(), drivers might have some pending I/O without any active
2662 * user-space mappings (eg., direct-IO, AIO). Therefore, we look at all pages
2663 * and see whether it has an elevated ref-count. If so, we tag them and wait for
2664 * them to be dropped.
2665 * The caller must guarantee that no new user will acquire writable references
2666 * to those pages to avoid races.
2667 */
2668static int shmem_wait_for_pins(struct address_space *mapping)
2669{
2670 struct radix_tree_iter iter;
2671 void **slot;
2672 pgoff_t start;
2673 struct page *page;
2674 int error, scan;
2675
2676 shmem_tag_pins(mapping);
2677
2678 error = 0;
2679 for (scan = 0; scan <= LAST_SCAN; scan++) {
2680 if (!radix_tree_tagged(&mapping->i_pages, SHMEM_TAG_PINNED))
2681 break;
2682
2683 if (!scan)
2684 lru_add_drain_all();
2685 else if (schedule_timeout_killable((HZ << scan) / 200))
2686 scan = LAST_SCAN;
2687
2688 start = 0;
2689 rcu_read_lock();
2690 radix_tree_for_each_tagged(slot, &mapping->i_pages, &iter,
2691 start, SHMEM_TAG_PINNED) {
2692
2693 page = radix_tree_deref_slot(slot);
2694 if (radix_tree_exception(page)) {
2695 if (radix_tree_deref_retry(page)) {
2696 slot = radix_tree_iter_retry(&iter);
2697 continue;
2698 }
2699
2700 page = NULL;
2701 }
2702
2703 if (page &&
2704 page_count(page) - page_mapcount(page) != 1) {
2705 if (scan < LAST_SCAN)
2706 goto continue_resched;
2707
2708 /*
2709 * On the last scan, we clean up all those tags
2710 * we inserted; but make a note that we still
2711 * found pages pinned.
2712 */
2713 error = -EBUSY;
2714 }
2715
2716 xa_lock_irq(&mapping->i_pages);
2717 radix_tree_tag_clear(&mapping->i_pages,
2718 iter.index, SHMEM_TAG_PINNED);
2719 xa_unlock_irq(&mapping->i_pages);
2720continue_resched:
2721 if (need_resched()) {
2722 slot = radix_tree_iter_resume(slot, &iter);
2723 cond_resched_rcu();
2724 }
2725 }
2726 rcu_read_unlock();
2727 }
2728
2729 return error;
2730}
2731
2732static unsigned int *memfd_file_seals_ptr(struct file *file)
2733{
2734 if (file->f_op == &shmem_file_operations)
2735 return &SHMEM_I(file_inode(file))->seals;
2736
2737#ifdef CONFIG_HUGETLBFS
2738 if (file->f_op == &hugetlbfs_file_operations)
2739 return &HUGETLBFS_I(file_inode(file))->seals;
2740#endif
2741
2742 return NULL;
2743}
2744
2745#define F_ALL_SEALS (F_SEAL_SEAL | \
2746 F_SEAL_SHRINK | \
2747 F_SEAL_GROW | \
2748 F_SEAL_WRITE)
2749
2750static int memfd_add_seals(struct file *file, unsigned int seals)
2751{
2752 struct inode *inode = file_inode(file);
2753 unsigned int *file_seals;
2754 int error;
2755
2756 /*
2757 * SEALING
2758 * Sealing allows multiple parties to share a shmem-file but restrict
2759 * access to a specific subset of file operations. Seals can only be
2760 * added, but never removed. This way, mutually untrusted parties can
2761 * share common memory regions with a well-defined policy. A malicious
2762 * peer can thus never perform unwanted operations on a shared object.
2763 *
2764 * Seals are only supported on special shmem-files and always affect
2765 * the whole underlying inode. Once a seal is set, it may prevent some
2766 * kinds of access to the file. Currently, the following seals are
2767 * defined:
2768 * SEAL_SEAL: Prevent further seals from being set on this file
2769 * SEAL_SHRINK: Prevent the file from shrinking
2770 * SEAL_GROW: Prevent the file from growing
2771 * SEAL_WRITE: Prevent write access to the file
2772 *
2773 * As we don't require any trust relationship between two parties, we
2774 * must prevent seals from being removed. Therefore, sealing a file
2775 * only adds a given set of seals to the file, it never touches
2776 * existing seals. Furthermore, the "setting seals"-operation can be
2777 * sealed itself, which basically prevents any further seal from being
2778 * added.
2779 *
2780 * Semantics of sealing are only defined on volatile files. Only
2781 * anonymous shmem files support sealing. More importantly, seals are
2782 * never written to disk. Therefore, there's no plan to support it on
2783 * other file types.
2784 */
2785
2786 if (!(file->f_mode & FMODE_WRITE))
2787 return -EPERM;
2788 if (seals & ~(unsigned int)F_ALL_SEALS)
2789 return -EINVAL;
2790
2791 inode_lock(inode);
2792
2793 file_seals = memfd_file_seals_ptr(file);
2794 if (!file_seals) {
2795 error = -EINVAL;
2796 goto unlock;
2797 }
2798
2799 if (*file_seals & F_SEAL_SEAL) {
2800 error = -EPERM;
2801 goto unlock;
2802 }
2803
2804 if ((seals & F_SEAL_WRITE) && !(*file_seals & F_SEAL_WRITE)) {
2805 error = mapping_deny_writable(file->f_mapping);
2806 if (error)
2807 goto unlock;
2808
2809 error = shmem_wait_for_pins(file->f_mapping);
2810 if (error) {
2811 mapping_allow_writable(file->f_mapping);
2812 goto unlock;
2813 }
2814 }
2815
2816 *file_seals |= seals;
2817 error = 0;
2818
2819unlock:
2820 inode_unlock(inode);
2821 return error;
2822}
2823
2824static int memfd_get_seals(struct file *file)
2825{
2826 unsigned int *seals = memfd_file_seals_ptr(file);
2827
2828 return seals ? *seals : -EINVAL;
2829}
2830
2831long memfd_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
2832{
2833 long error;
2834
2835 switch (cmd) {
2836 case F_ADD_SEALS:
2837 /* disallow upper 32bit */
2838 if (arg > UINT_MAX)
2839 return -EINVAL;
2840
2841 error = memfd_add_seals(file, arg);
2842 break;
2843 case F_GET_SEALS:
2844 error = memfd_get_seals(file);
2845 break;
2846 default:
2847 error = -EINVAL;
2848 break;
2849 }
2850
2851 return error;
2852}
2853
2854static long shmem_fallocate(struct file *file, int mode, loff_t offset,
2855 loff_t len)
2856{
2857 struct inode *inode = file_inode(file);
2858 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
2859 struct shmem_inode_info *info = SHMEM_I(inode);
2860 struct shmem_falloc shmem_falloc;
2861 pgoff_t start, index, end;
2862 int error;
2863
2864 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2865 return -EOPNOTSUPP;
2866
2867 inode_lock(inode);
2868
2869 if (mode & FALLOC_FL_PUNCH_HOLE) {
2870 struct address_space *mapping = file->f_mapping;
2871 loff_t unmap_start = round_up(offset, PAGE_SIZE);
2872 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
2873 DECLARE_WAIT_QUEUE_HEAD_ONSTACK(shmem_falloc_waitq);
2874
2875 /* protected by i_mutex */
2876 if (info->seals & F_SEAL_WRITE) {
2877 error = -EPERM;
2878 goto out;
2879 }
2880
2881 shmem_falloc.waitq = &shmem_falloc_waitq;
2882 shmem_falloc.start = unmap_start >> PAGE_SHIFT;
2883 shmem_falloc.next = (unmap_end + 1) >> PAGE_SHIFT;
2884 spin_lock(&inode->i_lock);
2885 inode->i_private = &shmem_falloc;
2886 spin_unlock(&inode->i_lock);
2887
2888 if ((u64)unmap_end > (u64)unmap_start)
2889 unmap_mapping_range(mapping, unmap_start,
2890 1 + unmap_end - unmap_start, 0);
2891 shmem_truncate_range(inode, offset, offset + len - 1);
2892 /* No need to unmap again: hole-punching leaves COWed pages */
2893
2894 spin_lock(&inode->i_lock);
2895 inode->i_private = NULL;
2896 wake_up_all(&shmem_falloc_waitq);
2897 WARN_ON_ONCE(!list_empty(&shmem_falloc_waitq.head));
2898 spin_unlock(&inode->i_lock);
2899 error = 0;
2900 goto out;
2901 }
2902
2903 /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
2904 error = inode_newsize_ok(inode, offset + len);
2905 if (error)
2906 goto out;
2907
2908 if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
2909 error = -EPERM;
2910 goto out;
2911 }
2912
2913 start = offset >> PAGE_SHIFT;
2914 end = (offset + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
2915 /* Try to avoid a swapstorm if len is impossible to satisfy */
2916 if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
2917 error = -ENOSPC;
2918 goto out;
2919 }
2920
2921 shmem_falloc.waitq = NULL;
2922 shmem_falloc.start = start;
2923 shmem_falloc.next = start;
2924 shmem_falloc.nr_falloced = 0;
2925 shmem_falloc.nr_unswapped = 0;
2926 spin_lock(&inode->i_lock);
2927 inode->i_private = &shmem_falloc;
2928 spin_unlock(&inode->i_lock);
2929
2930 for (index = start; index < end; index++) {
2931 struct page *page;
2932
2933 /*
2934 * Good, the fallocate(2) manpage permits EINTR: we may have
2935 * been interrupted because we are using up too much memory.
2936 */
2937 if (signal_pending(current))
2938 error = -EINTR;
2939 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
2940 error = -ENOMEM;
2941 else
2942 error = shmem_getpage(inode, index, &page, SGP_FALLOC);
2943 if (error) {
2944 /* Remove the !PageUptodate pages we added */
2945 if (index > start) {
2946 shmem_undo_range(inode,
2947 (loff_t)start << PAGE_SHIFT,
2948 ((loff_t)index << PAGE_SHIFT) - 1, true);
2949 }
2950 goto undone;
2951 }
2952
2953 /*
2954 * Inform shmem_writepage() how far we have reached.
2955 * No need for lock or barrier: we have the page lock.
2956 */
2957 shmem_falloc.next++;
2958 if (!PageUptodate(page))
2959 shmem_falloc.nr_falloced++;
2960
2961 /*
2962 * If !PageUptodate, leave it that way so that freeable pages
2963 * can be recognized if we need to rollback on error later.
2964 * But set_page_dirty so that memory pressure will swap rather
2965 * than free the pages we are allocating (and SGP_CACHE pages
2966 * might still be clean: we now need to mark those dirty too).
2967 */
2968 set_page_dirty(page);
2969 unlock_page(page);
2970 put_page(page);
2971 cond_resched();
2972 }
2973
2974 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
2975 i_size_write(inode, offset + len);
2976 inode->i_ctime = current_time(inode);
2977undone:
2978 spin_lock(&inode->i_lock);
2979 inode->i_private = NULL;
2980 spin_unlock(&inode->i_lock);
2981out:
2982 inode_unlock(inode);
2983 return error;
2984}
2985
2986static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
2987{
2988 struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
2989
2990 buf->f_type = TMPFS_MAGIC;
2991 buf->f_bsize = PAGE_SIZE;
2992 buf->f_namelen = NAME_MAX;
2993 if (sbinfo->max_blocks) {
2994 buf->f_blocks = sbinfo->max_blocks;
2995 buf->f_bavail =
2996 buf->f_bfree = sbinfo->max_blocks -
2997 percpu_counter_sum(&sbinfo->used_blocks);
2998 }
2999 if (sbinfo->max_inodes) {
3000 buf->f_files = sbinfo->max_inodes;
3001 buf->f_ffree = sbinfo->free_inodes;
3002 }
3003 /* else leave those fields 0 like simple_statfs */
3004 return 0;
3005}
3006
3007/*
3008 * File creation. Allocate an inode, and we're done..
3009 */
3010static int
3011shmem_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
3012{
3013 struct inode *inode;
3014 int error = -ENOSPC;
3015
3016 inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
3017 if (inode) {
3018 error = simple_acl_create(dir, inode);
3019 if (error)
3020 goto out_iput;
3021 error = security_inode_init_security(inode, dir,
3022 &dentry->d_name,
3023 shmem_initxattrs, NULL);
3024 if (error && error != -EOPNOTSUPP)
3025 goto out_iput;
3026
3027 error = 0;
3028 dir->i_size += BOGO_DIRENT_SIZE;
3029 dir->i_ctime = dir->i_mtime = current_time(dir);
3030 d_instantiate(dentry, inode);
3031 dget(dentry); /* Extra count - pin the dentry in core */
3032 }
3033 return error;
3034out_iput:
3035 iput(inode);
3036 return error;
3037}
3038
3039static int
3040shmem_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
3041{
3042 struct inode *inode;
3043 int error = -ENOSPC;
3044
3045 inode = shmem_get_inode(dir->i_sb, dir, mode, 0, VM_NORESERVE);
3046 if (inode) {
3047 error = security_inode_init_security(inode, dir,
3048 NULL,
3049 shmem_initxattrs, NULL);
3050 if (error && error != -EOPNOTSUPP)
3051 goto out_iput;
3052 error = simple_acl_create(dir, inode);
3053 if (error)
3054 goto out_iput;
3055 d_tmpfile(dentry, inode);
3056 }
3057 return error;
3058out_iput:
3059 iput(inode);
3060 return error;
3061}
3062
3063static int shmem_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
3064{
3065 int error;
3066
3067 if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
3068 return error;
3069 inc_nlink(dir);
3070 return 0;
3071}
3072
3073static int shmem_create(struct inode *dir, struct dentry *dentry, umode_t mode,
3074 bool excl)
3075{
3076 return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
3077}
3078
3079/*
3080 * Link a file..
3081 */
3082static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
3083{
3084 struct inode *inode = d_inode(old_dentry);
3085 int ret;
3086
3087 /*
3088 * No ordinary (disk based) filesystem counts links as inodes;
3089 * but each new link needs a new dentry, pinning lowmem, and
3090 * tmpfs dentries cannot be pruned until they are unlinked.
3091 */
3092 ret = shmem_reserve_inode(inode->i_sb);
3093 if (ret)
3094 goto out;
3095
3096 dir->i_size += BOGO_DIRENT_SIZE;
3097 inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
3098 inc_nlink(inode);
3099 ihold(inode); /* New dentry reference */
3100 dget(dentry); /* Extra pinning count for the created dentry */
3101 d_instantiate(dentry, inode);
3102out:
3103 return ret;
3104}
3105
3106static int shmem_unlink(struct inode *dir, struct dentry *dentry)
3107{
3108 struct inode *inode = d_inode(dentry);
3109
3110 if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
3111 shmem_free_inode(inode->i_sb);
3112
3113 dir->i_size -= BOGO_DIRENT_SIZE;
3114 inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
3115 drop_nlink(inode);
3116 dput(dentry); /* Undo the count from "create" - this does all the work */
3117 return 0;
3118}
3119
3120static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
3121{
3122 if (!simple_empty(dentry))
3123 return -ENOTEMPTY;
3124
3125 drop_nlink(d_inode(dentry));
3126 drop_nlink(dir);
3127 return shmem_unlink(dir, dentry);
3128}
3129
3130static int shmem_exchange(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
3131{
3132 bool old_is_dir = d_is_dir(old_dentry);
3133 bool new_is_dir = d_is_dir(new_dentry);
3134
3135 if (old_dir != new_dir && old_is_dir != new_is_dir) {
3136 if (old_is_dir) {
3137 drop_nlink(old_dir);
3138 inc_nlink(new_dir);
3139 } else {
3140 drop_nlink(new_dir);
3141 inc_nlink(old_dir);
3142 }
3143 }
3144 old_dir->i_ctime = old_dir->i_mtime =
3145 new_dir->i_ctime = new_dir->i_mtime =
3146 d_inode(old_dentry)->i_ctime =
3147 d_inode(new_dentry)->i_ctime = current_time(old_dir);
3148
3149 return 0;
3150}
3151
3152static int shmem_whiteout(struct inode *old_dir, struct dentry *old_dentry)
3153{
3154 struct dentry *whiteout;
3155 int error;
3156
3157 whiteout = d_alloc(old_dentry->d_parent, &old_dentry->d_name);
3158 if (!whiteout)
3159 return -ENOMEM;
3160
3161 error = shmem_mknod(old_dir, whiteout,
3162 S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
3163 dput(whiteout);
3164 if (error)
3165 return error;
3166
3167 /*
3168 * Cheat and hash the whiteout while the old dentry is still in
3169 * place, instead of playing games with FS_RENAME_DOES_D_MOVE.
3170 *
3171 * d_lookup() will consistently find one of them at this point,
3172 * not sure which one, but that isn't even important.
3173 */
3174 d_rehash(whiteout);
3175 return 0;
3176}
3177
3178/*
3179 * The VFS layer already does all the dentry stuff for rename,
3180 * we just have to decrement the usage count for the target if
3181 * it exists so that the VFS layer correctly free's it when it
3182 * gets overwritten.
3183 */
3184static int shmem_rename2(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry, unsigned int flags)
3185{
3186 struct inode *inode = d_inode(old_dentry);
3187 int they_are_dirs = S_ISDIR(inode->i_mode);
3188
3189 if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
3190 return -EINVAL;
3191
3192 if (flags & RENAME_EXCHANGE)
3193 return shmem_exchange(old_dir, old_dentry, new_dir, new_dentry);
3194
3195 if (!simple_empty(new_dentry))
3196 return -ENOTEMPTY;
3197
3198 if (flags & RENAME_WHITEOUT) {
3199 int error;
3200
3201 error = shmem_whiteout(old_dir, old_dentry);
3202 if (error)
3203 return error;
3204 }
3205
3206 if (d_really_is_positive(new_dentry)) {
3207 (void) shmem_unlink(new_dir, new_dentry);
3208 if (they_are_dirs) {
3209 drop_nlink(d_inode(new_dentry));
3210 drop_nlink(old_dir);
3211 }
3212 } else if (they_are_dirs) {
3213 drop_nlink(old_dir);
3214 inc_nlink(new_dir);
3215 }
3216
3217 old_dir->i_size -= BOGO_DIRENT_SIZE;
3218 new_dir->i_size += BOGO_DIRENT_SIZE;
3219 old_dir->i_ctime = old_dir->i_mtime =
3220 new_dir->i_ctime = new_dir->i_mtime =
3221 inode->i_ctime = current_time(old_dir);
3222 return 0;
3223}
3224
3225static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
3226{
3227 int error;
3228 int len;
3229 struct inode *inode;
3230 struct page *page;
3231
3232 len = strlen(symname) + 1;
3233 if (len > PAGE_SIZE)
3234 return -ENAMETOOLONG;
3235
3236 inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0, VM_NORESERVE);
3237 if (!inode)
3238 return -ENOSPC;
3239
3240 error = security_inode_init_security(inode, dir, &dentry->d_name,
3241 shmem_initxattrs, NULL);
3242 if (error) {
3243 if (error != -EOPNOTSUPP) {
3244 iput(inode);
3245 return error;
3246 }
3247 error = 0;
3248 }
3249
3250 inode->i_size = len-1;
3251 if (len <= SHORT_SYMLINK_LEN) {
3252 inode->i_link = kmemdup(symname, len, GFP_KERNEL);
3253 if (!inode->i_link) {
3254 iput(inode);
3255 return -ENOMEM;
3256 }
3257 inode->i_op = &shmem_short_symlink_operations;
3258 } else {
3259 inode_nohighmem(inode);
3260 error = shmem_getpage(inode, 0, &page, SGP_WRITE);
3261 if (error) {
3262 iput(inode);
3263 return error;
3264 }
3265 inode->i_mapping->a_ops = &shmem_aops;
3266 inode->i_op = &shmem_symlink_inode_operations;
3267 memcpy(page_address(page), symname, len);
3268 SetPageUptodate(page);
3269 set_page_dirty(page);
3270 unlock_page(page);
3271 put_page(page);
3272 }
3273 dir->i_size += BOGO_DIRENT_SIZE;
3274 dir->i_ctime = dir->i_mtime = current_time(dir);
3275 d_instantiate(dentry, inode);
3276 dget(dentry);
3277 return 0;
3278}
3279
3280static void shmem_put_link(void *arg)
3281{
3282 mark_page_accessed(arg);
3283 put_page(arg);
3284}
3285
3286static const char *shmem_get_link(struct dentry *dentry,
3287 struct inode *inode,
3288 struct delayed_call *done)
3289{
3290 struct page *page = NULL;
3291 int error;
3292 if (!dentry) {
3293 page = find_get_page(inode->i_mapping, 0);
3294 if (!page)
3295 return ERR_PTR(-ECHILD);
3296 if (!PageUptodate(page)) {
3297 put_page(page);
3298 return ERR_PTR(-ECHILD);
3299 }
3300 } else {
3301 error = shmem_getpage(inode, 0, &page, SGP_READ);
3302 if (error)
3303 return ERR_PTR(error);
3304 unlock_page(page);
3305 }
3306 set_delayed_call(done, shmem_put_link, page);
3307 return page_address(page);
3308}
3309
3310#ifdef CONFIG_TMPFS_XATTR
3311/*
3312 * Superblocks without xattr inode operations may get some security.* xattr
3313 * support from the LSM "for free". As soon as we have any other xattrs
3314 * like ACLs, we also need to implement the security.* handlers at
3315 * filesystem level, though.
3316 */
3317
3318/*
3319 * Callback for security_inode_init_security() for acquiring xattrs.
3320 */
3321static int shmem_initxattrs(struct inode *inode,
3322 const struct xattr *xattr_array,
3323 void *fs_info)
3324{
3325 struct shmem_inode_info *info = SHMEM_I(inode);
3326 const struct xattr *xattr;
3327 struct simple_xattr *new_xattr;
3328 size_t len;
3329
3330 for (xattr = xattr_array; xattr->name != NULL; xattr++) {
3331 new_xattr = simple_xattr_alloc(xattr->value, xattr->value_len);
3332 if (!new_xattr)
3333 return -ENOMEM;
3334
3335 len = strlen(xattr->name) + 1;
3336 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
3337 GFP_KERNEL);
3338 if (!new_xattr->name) {
3339 kfree(new_xattr);
3340 return -ENOMEM;
3341 }
3342
3343 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
3344 XATTR_SECURITY_PREFIX_LEN);
3345 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
3346 xattr->name, len);
3347
3348 simple_xattr_list_add(&info->xattrs, new_xattr);
3349 }
3350
3351 return 0;
3352}
3353
3354static int shmem_xattr_handler_get(const struct xattr_handler *handler,
3355 struct dentry *unused, struct inode *inode,
3356 const char *name, void *buffer, size_t size)
3357{
3358 struct shmem_inode_info *info = SHMEM_I(inode);
3359
3360 name = xattr_full_name(handler, name);
3361 return simple_xattr_get(&info->xattrs, name, buffer, size);
3362}
3363
3364static int shmem_xattr_handler_set(const struct xattr_handler *handler,
3365 struct dentry *unused, struct inode *inode,
3366 const char *name, const void *value,
3367 size_t size, int flags)
3368{
3369 struct shmem_inode_info *info = SHMEM_I(inode);
3370
3371 name = xattr_full_name(handler, name);
3372 return simple_xattr_set(&info->xattrs, name, value, size, flags);
3373}
3374
3375static const struct xattr_handler shmem_security_xattr_handler = {
3376 .prefix = XATTR_SECURITY_PREFIX,
3377 .get = shmem_xattr_handler_get,
3378 .set = shmem_xattr_handler_set,
3379};
3380
3381static const struct xattr_handler shmem_trusted_xattr_handler = {
3382 .prefix = XATTR_TRUSTED_PREFIX,
3383 .get = shmem_xattr_handler_get,
3384 .set = shmem_xattr_handler_set,
3385};
3386
3387static const struct xattr_handler *shmem_xattr_handlers[] = {
3388#ifdef CONFIG_TMPFS_POSIX_ACL
3389 &posix_acl_access_xattr_handler,
3390 &posix_acl_default_xattr_handler,
3391#endif
3392 &shmem_security_xattr_handler,
3393 &shmem_trusted_xattr_handler,
3394 NULL
3395};
3396
3397static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
3398{
3399 struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
3400 return simple_xattr_list(d_inode(dentry), &info->xattrs, buffer, size);
3401}
3402#endif /* CONFIG_TMPFS_XATTR */
3403
3404static const struct inode_operations shmem_short_symlink_operations = {
3405 .get_link = simple_get_link,
3406#ifdef CONFIG_TMPFS_XATTR
3407 .listxattr = shmem_listxattr,
3408#endif
3409};
3410
3411static const struct inode_operations shmem_symlink_inode_operations = {
3412 .get_link = shmem_get_link,
3413#ifdef CONFIG_TMPFS_XATTR
3414 .listxattr = shmem_listxattr,
3415#endif
3416};
3417
3418static struct dentry *shmem_get_parent(struct dentry *child)
3419{
3420 return ERR_PTR(-ESTALE);
3421}
3422
3423static int shmem_match(struct inode *ino, void *vfh)
3424{
3425 __u32 *fh = vfh;
3426 __u64 inum = fh[2];
3427 inum = (inum << 32) | fh[1];
3428 return ino->i_ino == inum && fh[0] == ino->i_generation;
3429}
3430
3431static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
3432 struct fid *fid, int fh_len, int fh_type)
3433{
3434 struct inode *inode;
3435 struct dentry *dentry = NULL;
3436 u64 inum;
3437
3438 if (fh_len < 3)
3439 return NULL;
3440
3441 inum = fid->raw[2];
3442 inum = (inum << 32) | fid->raw[1];
3443
3444 inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
3445 shmem_match, fid->raw);
3446 if (inode) {
3447 dentry = d_find_alias(inode);
3448 iput(inode);
3449 }
3450
3451 return dentry;
3452}
3453
3454static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
3455 struct inode *parent)
3456{
3457 if (*len < 3) {
3458 *len = 3;
3459 return FILEID_INVALID;
3460 }
3461
3462 if (inode_unhashed(inode)) {
3463 /* Unfortunately insert_inode_hash is not idempotent,
3464 * so as we hash inodes here rather than at creation
3465 * time, we need a lock to ensure we only try
3466 * to do it once
3467 */
3468 static DEFINE_SPINLOCK(lock);
3469 spin_lock(&lock);
3470 if (inode_unhashed(inode))
3471 __insert_inode_hash(inode,
3472 inode->i_ino + inode->i_generation);
3473 spin_unlock(&lock);
3474 }
3475
3476 fh[0] = inode->i_generation;
3477 fh[1] = inode->i_ino;
3478 fh[2] = ((__u64)inode->i_ino) >> 32;
3479
3480 *len = 3;
3481 return 1;
3482}
3483
3484static const struct export_operations shmem_export_ops = {
3485 .get_parent = shmem_get_parent,
3486 .encode_fh = shmem_encode_fh,
3487 .fh_to_dentry = shmem_fh_to_dentry,
3488};
3489
3490static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
3491 bool remount)
3492{
3493 char *this_char, *value, *rest;
3494 struct mempolicy *mpol = NULL;
3495 uid_t uid;
3496 gid_t gid;
3497
3498 while (options != NULL) {
3499 this_char = options;
3500 for (;;) {
3501 /*
3502 * NUL-terminate this option: unfortunately,
3503 * mount options form a comma-separated list,
3504 * but mpol's nodelist may also contain commas.
3505 */
3506 options = strchr(options, ',');
3507 if (options == NULL)
3508 break;
3509 options++;
3510 if (!isdigit(*options)) {
3511 options[-1] = '\0';
3512 break;
3513 }
3514 }
3515 if (!*this_char)
3516 continue;
3517 if ((value = strchr(this_char,'=')) != NULL) {
3518 *value++ = 0;
3519 } else {
3520 pr_err("tmpfs: No value for mount option '%s'\n",
3521 this_char);
3522 goto error;
3523 }
3524
3525 if (!strcmp(this_char,"size")) {
3526 unsigned long long size;
3527 size = memparse(value,&rest);
3528 if (*rest == '%') {
3529 size <<= PAGE_SHIFT;
3530 size *= totalram_pages;
3531 do_div(size, 100);
3532 rest++;
3533 }
3534 if (*rest)
3535 goto bad_val;
3536 sbinfo->max_blocks =
3537 DIV_ROUND_UP(size, PAGE_SIZE);
3538 } else if (!strcmp(this_char,"nr_blocks")) {
3539 sbinfo->max_blocks = memparse(value, &rest);
3540 if (*rest)
3541 goto bad_val;
3542 } else if (!strcmp(this_char,"nr_inodes")) {
3543 sbinfo->max_inodes = memparse(value, &rest);
3544 if (*rest)
3545 goto bad_val;
3546 } else if (!strcmp(this_char,"mode")) {
3547 if (remount)
3548 continue;
3549 sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
3550 if (*rest)
3551 goto bad_val;
3552 } else if (!strcmp(this_char,"uid")) {
3553 if (remount)
3554 continue;
3555 uid = simple_strtoul(value, &rest, 0);
3556 if (*rest)
3557 goto bad_val;
3558 sbinfo->uid = make_kuid(current_user_ns(), uid);
3559 if (!uid_valid(sbinfo->uid))
3560 goto bad_val;
3561 } else if (!strcmp(this_char,"gid")) {
3562 if (remount)
3563 continue;
3564 gid = simple_strtoul(value, &rest, 0);
3565 if (*rest)
3566 goto bad_val;
3567 sbinfo->gid = make_kgid(current_user_ns(), gid);
3568 if (!gid_valid(sbinfo->gid))
3569 goto bad_val;
3570#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3571 } else if (!strcmp(this_char, "huge")) {
3572 int huge;
3573 huge = shmem_parse_huge(value);
3574 if (huge < 0)
3575 goto bad_val;
3576 if (!has_transparent_hugepage() &&
3577 huge != SHMEM_HUGE_NEVER)
3578 goto bad_val;
3579 sbinfo->huge = huge;
3580#endif
3581#ifdef CONFIG_NUMA
3582 } else if (!strcmp(this_char,"mpol")) {
3583 mpol_put(mpol);
3584 mpol = NULL;
3585 if (mpol_parse_str(value, &mpol))
3586 goto bad_val;
3587#endif
3588 } else {
3589 pr_err("tmpfs: Bad mount option %s\n", this_char);
3590 goto error;
3591 }
3592 }
3593 sbinfo->mpol = mpol;
3594 return 0;
3595
3596bad_val:
3597 pr_err("tmpfs: Bad value '%s' for mount option '%s'\n",
3598 value, this_char);
3599error:
3600 mpol_put(mpol);
3601 return 1;
3602
3603}
3604
3605static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
3606{
3607 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
3608 struct shmem_sb_info config = *sbinfo;
3609 unsigned long inodes;
3610 int error = -EINVAL;
3611
3612 config.mpol = NULL;
3613 if (shmem_parse_options(data, &config, true))
3614 return error;
3615
3616 spin_lock(&sbinfo->stat_lock);
3617 inodes = sbinfo->max_inodes - sbinfo->free_inodes;
3618 if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
3619 goto out;
3620 if (config.max_inodes < inodes)
3621 goto out;
3622 /*
3623 * Those tests disallow limited->unlimited while any are in use;
3624 * but we must separately disallow unlimited->limited, because
3625 * in that case we have no record of how much is already in use.
3626 */
3627 if (config.max_blocks && !sbinfo->max_blocks)
3628 goto out;
3629 if (config.max_inodes && !sbinfo->max_inodes)
3630 goto out;
3631
3632 error = 0;
3633 sbinfo->huge = config.huge;
3634 sbinfo->max_blocks = config.max_blocks;
3635 sbinfo->max_inodes = config.max_inodes;
3636 sbinfo->free_inodes = config.max_inodes - inodes;
3637
3638 /*
3639 * Preserve previous mempolicy unless mpol remount option was specified.
3640 */
3641 if (config.mpol) {
3642 mpol_put(sbinfo->mpol);
3643 sbinfo->mpol = config.mpol; /* transfers initial ref */
3644 }
3645out:
3646 spin_unlock(&sbinfo->stat_lock);
3647 return error;
3648}
3649
3650static int shmem_show_options(struct seq_file *seq, struct dentry *root)
3651{
3652 struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
3653
3654 if (sbinfo->max_blocks != shmem_default_max_blocks())
3655 seq_printf(seq, ",size=%luk",
3656 sbinfo->max_blocks << (PAGE_SHIFT - 10));
3657 if (sbinfo->max_inodes != shmem_default_max_inodes())
3658 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
3659 if (sbinfo->mode != (S_IRWXUGO | S_ISVTX))
3660 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
3661 if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
3662 seq_printf(seq, ",uid=%u",
3663 from_kuid_munged(&init_user_ns, sbinfo->uid));
3664 if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
3665 seq_printf(seq, ",gid=%u",
3666 from_kgid_munged(&init_user_ns, sbinfo->gid));
3667#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3668 /* Rightly or wrongly, show huge mount option unmasked by shmem_huge */
3669 if (sbinfo->huge)
3670 seq_printf(seq, ",huge=%s", shmem_format_huge(sbinfo->huge));
3671#endif
3672 shmem_show_mpol(seq, sbinfo->mpol);
3673 return 0;
3674}
3675
3676#define MFD_NAME_PREFIX "memfd:"
3677#define MFD_NAME_PREFIX_LEN (sizeof(MFD_NAME_PREFIX) - 1)
3678#define MFD_NAME_MAX_LEN (NAME_MAX - MFD_NAME_PREFIX_LEN)
3679
3680#define MFD_ALL_FLAGS (MFD_CLOEXEC | MFD_ALLOW_SEALING | MFD_HUGETLB)
3681
3682SYSCALL_DEFINE2(memfd_create,
3683 const char __user *, uname,
3684 unsigned int, flags)
3685{
3686 unsigned int *file_seals;
3687 struct file *file;
3688 int fd, error;
3689 char *name;
3690 long len;
3691
3692 if (!(flags & MFD_HUGETLB)) {
3693 if (flags & ~(unsigned int)MFD_ALL_FLAGS)
3694 return -EINVAL;
3695 } else {
3696 /* Allow huge page size encoding in flags. */
3697 if (flags & ~(unsigned int)(MFD_ALL_FLAGS |
3698 (MFD_HUGE_MASK << MFD_HUGE_SHIFT)))
3699 return -EINVAL;
3700 }
3701
3702 /* length includes terminating zero */
3703 len = strnlen_user(uname, MFD_NAME_MAX_LEN + 1);
3704 if (len <= 0)
3705 return -EFAULT;
3706 if (len > MFD_NAME_MAX_LEN + 1)
3707 return -EINVAL;
3708
3709 name = kmalloc(len + MFD_NAME_PREFIX_LEN, GFP_KERNEL);
3710 if (!name)
3711 return -ENOMEM;
3712
3713 strcpy(name, MFD_NAME_PREFIX);
3714 if (copy_from_user(&name[MFD_NAME_PREFIX_LEN], uname, len)) {
3715 error = -EFAULT;
3716 goto err_name;
3717 }
3718
3719 /* terminating-zero may have changed after strnlen_user() returned */
3720 if (name[len + MFD_NAME_PREFIX_LEN - 1]) {
3721 error = -EFAULT;
3722 goto err_name;
3723 }
3724
3725 fd = get_unused_fd_flags((flags & MFD_CLOEXEC) ? O_CLOEXEC : 0);
3726 if (fd < 0) {
3727 error = fd;
3728 goto err_name;
3729 }
3730
3731 if (flags & MFD_HUGETLB) {
3732 struct user_struct *user = NULL;
3733
3734 file = hugetlb_file_setup(name, 0, VM_NORESERVE, &user,
3735 HUGETLB_ANONHUGE_INODE,
3736 (flags >> MFD_HUGE_SHIFT) &
3737 MFD_HUGE_MASK);
3738 } else
3739 file = shmem_file_setup(name, 0, VM_NORESERVE);
3740 if (IS_ERR(file)) {
3741 error = PTR_ERR(file);
3742 goto err_fd;
3743 }
3744 file->f_mode |= FMODE_LSEEK | FMODE_PREAD | FMODE_PWRITE;
3745 file->f_flags |= O_RDWR | O_LARGEFILE;
3746
3747 if (flags & MFD_ALLOW_SEALING) {
3748 file_seals = memfd_file_seals_ptr(file);
3749 *file_seals &= ~F_SEAL_SEAL;
3750 }
3751
3752 fd_install(fd, file);
3753 kfree(name);
3754 return fd;
3755
3756err_fd:
3757 put_unused_fd(fd);
3758err_name:
3759 kfree(name);
3760 return error;
3761}
3762
3763#endif /* CONFIG_TMPFS */
3764
3765static void shmem_put_super(struct super_block *sb)
3766{
3767 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
3768
3769 percpu_counter_destroy(&sbinfo->used_blocks);
3770 mpol_put(sbinfo->mpol);
3771 kfree(sbinfo);
3772 sb->s_fs_info = NULL;
3773}
3774
3775int shmem_fill_super(struct super_block *sb, void *data, int silent)
3776{
3777 struct inode *inode;
3778 struct shmem_sb_info *sbinfo;
3779 int err = -ENOMEM;
3780
3781 /* Round up to L1_CACHE_BYTES to resist false sharing */
3782 sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
3783 L1_CACHE_BYTES), GFP_KERNEL);
3784 if (!sbinfo)
3785 return -ENOMEM;
3786
3787 sbinfo->mode = S_IRWXUGO | S_ISVTX;
3788 sbinfo->uid = current_fsuid();
3789 sbinfo->gid = current_fsgid();
3790 sb->s_fs_info = sbinfo;
3791
3792#ifdef CONFIG_TMPFS
3793 /*
3794 * Per default we only allow half of the physical ram per
3795 * tmpfs instance, limiting inodes to one per page of lowmem;
3796 * but the internal instance is left unlimited.
3797 */
3798 if (!(sb->s_flags & SB_KERNMOUNT)) {
3799 sbinfo->max_blocks = shmem_default_max_blocks();
3800 sbinfo->max_inodes = shmem_default_max_inodes();
3801 if (shmem_parse_options(data, sbinfo, false)) {
3802 err = -EINVAL;
3803 goto failed;
3804 }
3805 } else {
3806 sb->s_flags |= SB_NOUSER;
3807 }
3808 sb->s_export_op = &shmem_export_ops;
3809 sb->s_flags |= SB_NOSEC;
3810#else
3811 sb->s_flags |= SB_NOUSER;
3812#endif
3813
3814 spin_lock_init(&sbinfo->stat_lock);
3815 if (percpu_counter_init(&sbinfo->used_blocks, 0, GFP_KERNEL))
3816 goto failed;
3817 sbinfo->free_inodes = sbinfo->max_inodes;
3818 spin_lock_init(&sbinfo->shrinklist_lock);
3819 INIT_LIST_HEAD(&sbinfo->shrinklist);
3820
3821 sb->s_maxbytes = MAX_LFS_FILESIZE;
3822 sb->s_blocksize = PAGE_SIZE;
3823 sb->s_blocksize_bits = PAGE_SHIFT;
3824 sb->s_magic = TMPFS_MAGIC;
3825 sb->s_op = &shmem_ops;
3826 sb->s_time_gran = 1;
3827#ifdef CONFIG_TMPFS_XATTR
3828 sb->s_xattr = shmem_xattr_handlers;
3829#endif
3830#ifdef CONFIG_TMPFS_POSIX_ACL
3831 sb->s_flags |= SB_POSIXACL;
3832#endif
3833 uuid_gen(&sb->s_uuid);
3834
3835 inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
3836 if (!inode)
3837 goto failed;
3838 inode->i_uid = sbinfo->uid;
3839 inode->i_gid = sbinfo->gid;
3840 sb->s_root = d_make_root(inode);
3841 if (!sb->s_root)
3842 goto failed;
3843 return 0;
3844
3845failed:
3846 shmem_put_super(sb);
3847 return err;
3848}
3849
3850static struct kmem_cache *shmem_inode_cachep;
3851
3852static struct inode *shmem_alloc_inode(struct super_block *sb)
3853{
3854 struct shmem_inode_info *info;
3855 info = kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
3856 if (!info)
3857 return NULL;
3858 return &info->vfs_inode;
3859}
3860
3861static void shmem_destroy_callback(struct rcu_head *head)
3862{
3863 struct inode *inode = container_of(head, struct inode, i_rcu);
3864 if (S_ISLNK(inode->i_mode))
3865 kfree(inode->i_link);
3866 kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
3867}
3868
3869static void shmem_destroy_inode(struct inode *inode)
3870{
3871 if (S_ISREG(inode->i_mode))
3872 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
3873 call_rcu(&inode->i_rcu, shmem_destroy_callback);
3874}
3875
3876static void shmem_init_inode(void *foo)
3877{
3878 struct shmem_inode_info *info = foo;
3879 inode_init_once(&info->vfs_inode);
3880}
3881
3882static void shmem_init_inodecache(void)
3883{
3884 shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
3885 sizeof(struct shmem_inode_info),
3886 0, SLAB_PANIC|SLAB_ACCOUNT, shmem_init_inode);
3887}
3888
3889static void shmem_destroy_inodecache(void)
3890{
3891 kmem_cache_destroy(shmem_inode_cachep);
3892}
3893
3894static const struct address_space_operations shmem_aops = {
3895 .writepage = shmem_writepage,
3896 .set_page_dirty = __set_page_dirty_no_writeback,
3897#ifdef CONFIG_TMPFS
3898 .write_begin = shmem_write_begin,
3899 .write_end = shmem_write_end,
3900#endif
3901#ifdef CONFIG_MIGRATION
3902 .migratepage = migrate_page,
3903#endif
3904 .error_remove_page = generic_error_remove_page,
3905};
3906
3907static const struct file_operations shmem_file_operations = {
3908 .mmap = shmem_mmap,
3909 .get_unmapped_area = shmem_get_unmapped_area,
3910#ifdef CONFIG_TMPFS
3911 .llseek = shmem_file_llseek,
3912 .read_iter = shmem_file_read_iter,
3913 .write_iter = generic_file_write_iter,
3914 .fsync = noop_fsync,
3915 .splice_read = generic_file_splice_read,
3916 .splice_write = iter_file_splice_write,
3917 .fallocate = shmem_fallocate,
3918#endif
3919};
3920
3921static const struct inode_operations shmem_inode_operations = {
3922 .getattr = shmem_getattr,
3923 .setattr = shmem_setattr,
3924#ifdef CONFIG_TMPFS_XATTR
3925 .listxattr = shmem_listxattr,
3926 .set_acl = simple_set_acl,
3927#endif
3928};
3929
3930static const struct inode_operations shmem_dir_inode_operations = {
3931#ifdef CONFIG_TMPFS
3932 .create = shmem_create,
3933 .lookup = simple_lookup,
3934 .link = shmem_link,
3935 .unlink = shmem_unlink,
3936 .symlink = shmem_symlink,
3937 .mkdir = shmem_mkdir,
3938 .rmdir = shmem_rmdir,
3939 .mknod = shmem_mknod,
3940 .rename = shmem_rename2,
3941 .tmpfile = shmem_tmpfile,
3942#endif
3943#ifdef CONFIG_TMPFS_XATTR
3944 .listxattr = shmem_listxattr,
3945#endif
3946#ifdef CONFIG_TMPFS_POSIX_ACL
3947 .setattr = shmem_setattr,
3948 .set_acl = simple_set_acl,
3949#endif
3950};
3951
3952static const struct inode_operations shmem_special_inode_operations = {
3953#ifdef CONFIG_TMPFS_XATTR
3954 .listxattr = shmem_listxattr,
3955#endif
3956#ifdef CONFIG_TMPFS_POSIX_ACL
3957 .setattr = shmem_setattr,
3958 .set_acl = simple_set_acl,
3959#endif
3960};
3961
3962static const struct super_operations shmem_ops = {
3963 .alloc_inode = shmem_alloc_inode,
3964 .destroy_inode = shmem_destroy_inode,
3965#ifdef CONFIG_TMPFS
3966 .statfs = shmem_statfs,
3967 .remount_fs = shmem_remount_fs,
3968 .show_options = shmem_show_options,
3969#endif
3970 .evict_inode = shmem_evict_inode,
3971 .drop_inode = generic_delete_inode,
3972 .put_super = shmem_put_super,
3973#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3974 .nr_cached_objects = shmem_unused_huge_count,
3975 .free_cached_objects = shmem_unused_huge_scan,
3976#endif
3977};
3978
3979static const struct vm_operations_struct shmem_vm_ops = {
3980 .fault = shmem_fault,
3981 .map_pages = filemap_map_pages,
3982#ifdef CONFIG_NUMA
3983 .set_policy = shmem_set_policy,
3984 .get_policy = shmem_get_policy,
3985#endif
3986};
3987
3988static struct dentry *shmem_mount(struct file_system_type *fs_type,
3989 int flags, const char *dev_name, void *data)
3990{
3991 return mount_nodev(fs_type, flags, data, shmem_fill_super);
3992}
3993
3994static struct file_system_type shmem_fs_type = {
3995 .owner = THIS_MODULE,
3996 .name = "tmpfs",
3997 .mount = shmem_mount,
3998 .kill_sb = kill_litter_super,
3999 .fs_flags = FS_USERNS_MOUNT,
4000};
4001
4002int __init shmem_init(void)
4003{
4004 int error;
4005
4006 /* If rootfs called this, don't re-init */
4007 if (shmem_inode_cachep)
4008 return 0;
4009
4010 shmem_init_inodecache();
4011
4012 error = register_filesystem(&shmem_fs_type);
4013 if (error) {
4014 pr_err("Could not register tmpfs\n");
4015 goto out2;
4016 }
4017
4018 shm_mnt = kern_mount(&shmem_fs_type);
4019 if (IS_ERR(shm_mnt)) {
4020 error = PTR_ERR(shm_mnt);
4021 pr_err("Could not kern_mount tmpfs\n");
4022 goto out1;
4023 }
4024
4025#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
4026 if (has_transparent_hugepage() && shmem_huge > SHMEM_HUGE_DENY)
4027 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
4028 else
4029 shmem_huge = 0; /* just in case it was patched */
4030#endif
4031 return 0;
4032
4033out1:
4034 unregister_filesystem(&shmem_fs_type);
4035out2:
4036 shmem_destroy_inodecache();
4037 shm_mnt = ERR_PTR(error);
4038 return error;
4039}
4040
4041#if defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE) && defined(CONFIG_SYSFS)
4042static ssize_t shmem_enabled_show(struct kobject *kobj,
4043 struct kobj_attribute *attr, char *buf)
4044{
4045 int values[] = {
4046 SHMEM_HUGE_ALWAYS,
4047 SHMEM_HUGE_WITHIN_SIZE,
4048 SHMEM_HUGE_ADVISE,
4049 SHMEM_HUGE_NEVER,
4050 SHMEM_HUGE_DENY,
4051 SHMEM_HUGE_FORCE,
4052 };
4053 int i, count;
4054
4055 for (i = 0, count = 0; i < ARRAY_SIZE(values); i++) {
4056 const char *fmt = shmem_huge == values[i] ? "[%s] " : "%s ";
4057
4058 count += sprintf(buf + count, fmt,
4059 shmem_format_huge(values[i]));
4060 }
4061 buf[count - 1] = '\n';
4062 return count;
4063}
4064
4065static ssize_t shmem_enabled_store(struct kobject *kobj,
4066 struct kobj_attribute *attr, const char *buf, size_t count)
4067{
4068 char tmp[16];
4069 int huge;
4070
4071 if (count + 1 > sizeof(tmp))
4072 return -EINVAL;
4073 memcpy(tmp, buf, count);
4074 tmp[count] = '\0';
4075 if (count && tmp[count - 1] == '\n')
4076 tmp[count - 1] = '\0';
4077
4078 huge = shmem_parse_huge(tmp);
4079 if (huge == -EINVAL)
4080 return -EINVAL;
4081 if (!has_transparent_hugepage() &&
4082 huge != SHMEM_HUGE_NEVER && huge != SHMEM_HUGE_DENY)
4083 return -EINVAL;
4084
4085 shmem_huge = huge;
4086 if (shmem_huge > SHMEM_HUGE_DENY)
4087 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
4088 return count;
4089}
4090
4091struct kobj_attribute shmem_enabled_attr =
4092 __ATTR(shmem_enabled, 0644, shmem_enabled_show, shmem_enabled_store);
4093#endif /* CONFIG_TRANSPARENT_HUGE_PAGECACHE && CONFIG_SYSFS */
4094
4095#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
4096bool shmem_huge_enabled(struct vm_area_struct *vma)
4097{
4098 struct inode *inode = file_inode(vma->vm_file);
4099 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
4100 loff_t i_size;
4101 pgoff_t off;
4102
4103 if (shmem_huge == SHMEM_HUGE_FORCE)
4104 return true;
4105 if (shmem_huge == SHMEM_HUGE_DENY)
4106 return false;
4107 switch (sbinfo->huge) {
4108 case SHMEM_HUGE_NEVER:
4109 return false;
4110 case SHMEM_HUGE_ALWAYS:
4111 return true;
4112 case SHMEM_HUGE_WITHIN_SIZE:
4113 off = round_up(vma->vm_pgoff, HPAGE_PMD_NR);
4114 i_size = round_up(i_size_read(inode), PAGE_SIZE);
4115 if (i_size >= HPAGE_PMD_SIZE &&
4116 i_size >> PAGE_SHIFT >= off)
4117 return true;
4118 /* fall through */
4119 case SHMEM_HUGE_ADVISE:
4120 /* TODO: implement fadvise() hints */
4121 return (vma->vm_flags & VM_HUGEPAGE);
4122 default:
4123 VM_BUG_ON(1);
4124 return false;
4125 }
4126}
4127#endif /* CONFIG_TRANSPARENT_HUGE_PAGECACHE */
4128
4129#else /* !CONFIG_SHMEM */
4130
4131/*
4132 * tiny-shmem: simple shmemfs and tmpfs using ramfs code
4133 *
4134 * This is intended for small system where the benefits of the full
4135 * shmem code (swap-backed and resource-limited) are outweighed by
4136 * their complexity. On systems without swap this code should be
4137 * effectively equivalent, but much lighter weight.
4138 */
4139
4140static struct file_system_type shmem_fs_type = {
4141 .name = "tmpfs",
4142 .mount = ramfs_mount,
4143 .kill_sb = kill_litter_super,
4144 .fs_flags = FS_USERNS_MOUNT,
4145};
4146
4147int __init shmem_init(void)
4148{
4149 BUG_ON(register_filesystem(&shmem_fs_type) != 0);
4150
4151 shm_mnt = kern_mount(&shmem_fs_type);
4152 BUG_ON(IS_ERR(shm_mnt));
4153
4154 return 0;
4155}
4156
4157int shmem_unuse(swp_entry_t swap, struct page *page)
4158{
4159 return 0;
4160}
4161
4162int shmem_lock(struct file *file, int lock, struct user_struct *user)
4163{
4164 return 0;
4165}
4166
4167void shmem_unlock_mapping(struct address_space *mapping)
4168{
4169}
4170
4171#ifdef CONFIG_MMU
4172unsigned long shmem_get_unmapped_area(struct file *file,
4173 unsigned long addr, unsigned long len,
4174 unsigned long pgoff, unsigned long flags)
4175{
4176 return current->mm->get_unmapped_area(file, addr, len, pgoff, flags);
4177}
4178#endif
4179
4180void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
4181{
4182 truncate_inode_pages_range(inode->i_mapping, lstart, lend);
4183}
4184EXPORT_SYMBOL_GPL(shmem_truncate_range);
4185
4186#define shmem_vm_ops generic_file_vm_ops
4187#define shmem_file_operations ramfs_file_operations
4188#define shmem_get_inode(sb, dir, mode, dev, flags) ramfs_get_inode(sb, dir, mode, dev)
4189#define shmem_acct_size(flags, size) 0
4190#define shmem_unacct_size(flags, size) do {} while (0)
4191
4192#endif /* CONFIG_SHMEM */
4193
4194/* common code */
4195
4196static const struct dentry_operations anon_ops = {
4197 .d_dname = simple_dname
4198};
4199
4200static struct file *__shmem_file_setup(struct vfsmount *mnt, const char *name, loff_t size,
4201 unsigned long flags, unsigned int i_flags)
4202{
4203 struct file *res;
4204 struct inode *inode;
4205 struct path path;
4206 struct super_block *sb;
4207 struct qstr this;
4208
4209 if (IS_ERR(mnt))
4210 return ERR_CAST(mnt);
4211
4212 if (size < 0 || size > MAX_LFS_FILESIZE)
4213 return ERR_PTR(-EINVAL);
4214
4215 if (shmem_acct_size(flags, size))
4216 return ERR_PTR(-ENOMEM);
4217
4218 res = ERR_PTR(-ENOMEM);
4219 this.name = name;
4220 this.len = strlen(name);
4221 this.hash = 0; /* will go */
4222 sb = mnt->mnt_sb;
4223 path.mnt = mntget(mnt);
4224 path.dentry = d_alloc_pseudo(sb, &this);
4225 if (!path.dentry)
4226 goto put_memory;
4227 d_set_d_op(path.dentry, &anon_ops);
4228
4229 res = ERR_PTR(-ENOSPC);
4230 inode = shmem_get_inode(sb, NULL, S_IFREG | S_IRWXUGO, 0, flags);
4231 if (!inode)
4232 goto put_memory;
4233
4234 inode->i_flags |= i_flags;
4235 d_instantiate(path.dentry, inode);
4236 inode->i_size = size;
4237 clear_nlink(inode); /* It is unlinked */
4238 res = ERR_PTR(ramfs_nommu_expand_for_mapping(inode, size));
4239 if (IS_ERR(res))
4240 goto put_path;
4241
4242 res = alloc_file(&path, FMODE_WRITE | FMODE_READ,
4243 &shmem_file_operations);
4244 if (IS_ERR(res))
4245 goto put_path;
4246
4247 return res;
4248
4249put_memory:
4250 shmem_unacct_size(flags, size);
4251put_path:
4252 path_put(&path);
4253 return res;
4254}
4255
4256/**
4257 * shmem_kernel_file_setup - get an unlinked file living in tmpfs which must be
4258 * kernel internal. There will be NO LSM permission checks against the
4259 * underlying inode. So users of this interface must do LSM checks at a
4260 * higher layer. The users are the big_key and shm implementations. LSM
4261 * checks are provided at the key or shm level rather than the inode.
4262 * @name: name for dentry (to be seen in /proc/<pid>/maps
4263 * @size: size to be set for the file
4264 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4265 */
4266struct file *shmem_kernel_file_setup(const char *name, loff_t size, unsigned long flags)
4267{
4268 return __shmem_file_setup(shm_mnt, name, size, flags, S_PRIVATE);
4269}
4270
4271/**
4272 * shmem_file_setup - get an unlinked file living in tmpfs
4273 * @name: name for dentry (to be seen in /proc/<pid>/maps
4274 * @size: size to be set for the file
4275 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4276 */
4277struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
4278{
4279 return __shmem_file_setup(shm_mnt, name, size, flags, 0);
4280}
4281EXPORT_SYMBOL_GPL(shmem_file_setup);
4282
4283/**
4284 * shmem_file_setup_with_mnt - get an unlinked file living in tmpfs
4285 * @mnt: the tmpfs mount where the file will be created
4286 * @name: name for dentry (to be seen in /proc/<pid>/maps
4287 * @size: size to be set for the file
4288 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4289 */
4290struct file *shmem_file_setup_with_mnt(struct vfsmount *mnt, const char *name,
4291 loff_t size, unsigned long flags)
4292{
4293 return __shmem_file_setup(mnt, name, size, flags, 0);
4294}
4295EXPORT_SYMBOL_GPL(shmem_file_setup_with_mnt);
4296
4297/**
4298 * shmem_zero_setup - setup a shared anonymous mapping
4299 * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
4300 */
4301int shmem_zero_setup(struct vm_area_struct *vma)
4302{
4303 struct file *file;
4304 loff_t size = vma->vm_end - vma->vm_start;
4305
4306 /*
4307 * Cloning a new file under mmap_sem leads to a lock ordering conflict
4308 * between XFS directory reading and selinux: since this file is only
4309 * accessible to the user through its mapping, use S_PRIVATE flag to
4310 * bypass file security, in the same way as shmem_kernel_file_setup().
4311 */
4312 file = shmem_kernel_file_setup("dev/zero", size, vma->vm_flags);
4313 if (IS_ERR(file))
4314 return PTR_ERR(file);
4315
4316 if (vma->vm_file)
4317 fput(vma->vm_file);
4318 vma->vm_file = file;
4319 vma->vm_ops = &shmem_vm_ops;
4320
4321 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE) &&
4322 ((vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK) <
4323 (vma->vm_end & HPAGE_PMD_MASK)) {
4324 khugepaged_enter(vma, vma->vm_flags);
4325 }
4326
4327 return 0;
4328}
4329
4330/**
4331 * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
4332 * @mapping: the page's address_space
4333 * @index: the page index
4334 * @gfp: the page allocator flags to use if allocating
4335 *
4336 * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
4337 * with any new page allocations done using the specified allocation flags.
4338 * But read_cache_page_gfp() uses the ->readpage() method: which does not
4339 * suit tmpfs, since it may have pages in swapcache, and needs to find those
4340 * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
4341 *
4342 * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
4343 * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
4344 */
4345struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
4346 pgoff_t index, gfp_t gfp)
4347{
4348#ifdef CONFIG_SHMEM
4349 struct inode *inode = mapping->host;
4350 struct page *page;
4351 int error;
4352
4353 BUG_ON(mapping->a_ops != &shmem_aops);
4354 error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE,
4355 gfp, NULL, NULL, NULL);
4356 if (error)
4357 page = ERR_PTR(error);
4358 else
4359 unlock_page(page);
4360 return page;
4361#else
4362 /*
4363 * The tiny !SHMEM case uses ramfs without swap
4364 */
4365 return read_cache_page_gfp(mapping, index, gfp);
4366#endif
4367}
4368EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);