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/export.h>
33#include <linux/swap.h>
34#include <linux/uio.h>
35
36static struct vfsmount *shm_mnt;
37
38#ifdef CONFIG_SHMEM
39/*
40 * This virtual memory filesystem is heavily based on the ramfs. It
41 * extends ramfs by the ability to use swap and honor resource limits
42 * which makes it a completely usable filesystem.
43 */
44
45#include <linux/xattr.h>
46#include <linux/exportfs.h>
47#include <linux/posix_acl.h>
48#include <linux/posix_acl_xattr.h>
49#include <linux/mman.h>
50#include <linux/string.h>
51#include <linux/slab.h>
52#include <linux/backing-dev.h>
53#include <linux/shmem_fs.h>
54#include <linux/writeback.h>
55#include <linux/blkdev.h>
56#include <linux/pagevec.h>
57#include <linux/percpu_counter.h>
58#include <linux/falloc.h>
59#include <linux/splice.h>
60#include <linux/security.h>
61#include <linux/swapops.h>
62#include <linux/mempolicy.h>
63#include <linux/namei.h>
64#include <linux/ctype.h>
65#include <linux/migrate.h>
66#include <linux/highmem.h>
67#include <linux/seq_file.h>
68#include <linux/magic.h>
69#include <linux/syscalls.h>
70#include <linux/fcntl.h>
71#include <uapi/linux/memfd.h>
72
73#include <asm/uaccess.h>
74#include <asm/pgtable.h>
75
76#include "internal.h"
77
78#define BLOCKS_PER_PAGE (PAGE_SIZE/512)
79#define VM_ACCT(size) (PAGE_ALIGN(size) >> PAGE_SHIFT)
80
81/* Pretend that each entry is of this size in directory's i_size */
82#define BOGO_DIRENT_SIZE 20
83
84/* Symlink up to this size is kmalloc'ed instead of using a swappable page */
85#define SHORT_SYMLINK_LEN 128
86
87/*
88 * shmem_fallocate communicates with shmem_fault or shmem_writepage via
89 * inode->i_private (with i_mutex making sure that it has only one user at
90 * a time): we would prefer not to enlarge the shmem inode just for that.
91 */
92struct shmem_falloc {
93 wait_queue_head_t *waitq; /* faults into hole wait for punch to end */
94 pgoff_t start; /* start of range currently being fallocated */
95 pgoff_t next; /* the next page offset to be fallocated */
96 pgoff_t nr_falloced; /* how many new pages have been fallocated */
97 pgoff_t nr_unswapped; /* how often writepage refused to swap out */
98};
99
100/* Flag allocation requirements to shmem_getpage */
101enum sgp_type {
102 SGP_READ, /* don't exceed i_size, don't allocate page */
103 SGP_CACHE, /* don't exceed i_size, may allocate page */
104 SGP_DIRTY, /* like SGP_CACHE, but set new page dirty */
105 SGP_WRITE, /* may exceed i_size, may allocate !Uptodate page */
106 SGP_FALLOC, /* like SGP_WRITE, but make existing page Uptodate */
107};
108
109#ifdef CONFIG_TMPFS
110static unsigned long shmem_default_max_blocks(void)
111{
112 return totalram_pages / 2;
113}
114
115static unsigned long shmem_default_max_inodes(void)
116{
117 return min(totalram_pages - totalhigh_pages, totalram_pages / 2);
118}
119#endif
120
121static bool shmem_should_replace_page(struct page *page, gfp_t gfp);
122static int shmem_replace_page(struct page **pagep, gfp_t gfp,
123 struct shmem_inode_info *info, pgoff_t index);
124static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
125 struct page **pagep, enum sgp_type sgp, gfp_t gfp, int *fault_type);
126
127static inline int shmem_getpage(struct inode *inode, pgoff_t index,
128 struct page **pagep, enum sgp_type sgp, int *fault_type)
129{
130 return shmem_getpage_gfp(inode, index, pagep, sgp,
131 mapping_gfp_mask(inode->i_mapping), fault_type);
132}
133
134static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
135{
136 return sb->s_fs_info;
137}
138
139/*
140 * shmem_file_setup pre-accounts the whole fixed size of a VM object,
141 * for shared memory and for shared anonymous (/dev/zero) mappings
142 * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
143 * consistent with the pre-accounting of private mappings ...
144 */
145static inline int shmem_acct_size(unsigned long flags, loff_t size)
146{
147 return (flags & VM_NORESERVE) ?
148 0 : security_vm_enough_memory_mm(current->mm, VM_ACCT(size));
149}
150
151static inline void shmem_unacct_size(unsigned long flags, loff_t size)
152{
153 if (!(flags & VM_NORESERVE))
154 vm_unacct_memory(VM_ACCT(size));
155}
156
157static inline int shmem_reacct_size(unsigned long flags,
158 loff_t oldsize, loff_t newsize)
159{
160 if (!(flags & VM_NORESERVE)) {
161 if (VM_ACCT(newsize) > VM_ACCT(oldsize))
162 return security_vm_enough_memory_mm(current->mm,
163 VM_ACCT(newsize) - VM_ACCT(oldsize));
164 else if (VM_ACCT(newsize) < VM_ACCT(oldsize))
165 vm_unacct_memory(VM_ACCT(oldsize) - VM_ACCT(newsize));
166 }
167 return 0;
168}
169
170/*
171 * ... whereas tmpfs objects are accounted incrementally as
172 * pages are allocated, in order to allow huge sparse files.
173 * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
174 * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
175 */
176static inline int shmem_acct_block(unsigned long flags)
177{
178 return (flags & VM_NORESERVE) ?
179 security_vm_enough_memory_mm(current->mm, VM_ACCT(PAGE_SIZE)) : 0;
180}
181
182static inline void shmem_unacct_blocks(unsigned long flags, long pages)
183{
184 if (flags & VM_NORESERVE)
185 vm_unacct_memory(pages * VM_ACCT(PAGE_SIZE));
186}
187
188static const struct super_operations shmem_ops;
189static const struct address_space_operations shmem_aops;
190static const struct file_operations shmem_file_operations;
191static const struct inode_operations shmem_inode_operations;
192static const struct inode_operations shmem_dir_inode_operations;
193static const struct inode_operations shmem_special_inode_operations;
194static const struct vm_operations_struct shmem_vm_ops;
195
196static LIST_HEAD(shmem_swaplist);
197static DEFINE_MUTEX(shmem_swaplist_mutex);
198
199static int shmem_reserve_inode(struct super_block *sb)
200{
201 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
202 if (sbinfo->max_inodes) {
203 spin_lock(&sbinfo->stat_lock);
204 if (!sbinfo->free_inodes) {
205 spin_unlock(&sbinfo->stat_lock);
206 return -ENOSPC;
207 }
208 sbinfo->free_inodes--;
209 spin_unlock(&sbinfo->stat_lock);
210 }
211 return 0;
212}
213
214static void shmem_free_inode(struct super_block *sb)
215{
216 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
217 if (sbinfo->max_inodes) {
218 spin_lock(&sbinfo->stat_lock);
219 sbinfo->free_inodes++;
220 spin_unlock(&sbinfo->stat_lock);
221 }
222}
223
224/**
225 * shmem_recalc_inode - recalculate the block usage of an inode
226 * @inode: inode to recalc
227 *
228 * We have to calculate the free blocks since the mm can drop
229 * undirtied hole pages behind our back.
230 *
231 * But normally info->alloced == inode->i_mapping->nrpages + info->swapped
232 * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
233 *
234 * It has to be called with the spinlock held.
235 */
236static void shmem_recalc_inode(struct inode *inode)
237{
238 struct shmem_inode_info *info = SHMEM_I(inode);
239 long freed;
240
241 freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
242 if (freed > 0) {
243 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
244 if (sbinfo->max_blocks)
245 percpu_counter_add(&sbinfo->used_blocks, -freed);
246 info->alloced -= freed;
247 inode->i_blocks -= freed * BLOCKS_PER_PAGE;
248 shmem_unacct_blocks(info->flags, freed);
249 }
250}
251
252/*
253 * Replace item expected in radix tree by a new item, while holding tree lock.
254 */
255static int shmem_radix_tree_replace(struct address_space *mapping,
256 pgoff_t index, void *expected, void *replacement)
257{
258 void **pslot;
259 void *item;
260
261 VM_BUG_ON(!expected);
262 VM_BUG_ON(!replacement);
263 pslot = radix_tree_lookup_slot(&mapping->page_tree, index);
264 if (!pslot)
265 return -ENOENT;
266 item = radix_tree_deref_slot_protected(pslot, &mapping->tree_lock);
267 if (item != expected)
268 return -ENOENT;
269 radix_tree_replace_slot(pslot, replacement);
270 return 0;
271}
272
273/*
274 * Sometimes, before we decide whether to proceed or to fail, we must check
275 * that an entry was not already brought back from swap by a racing thread.
276 *
277 * Checking page is not enough: by the time a SwapCache page is locked, it
278 * might be reused, and again be SwapCache, using the same swap as before.
279 */
280static bool shmem_confirm_swap(struct address_space *mapping,
281 pgoff_t index, swp_entry_t swap)
282{
283 void *item;
284
285 rcu_read_lock();
286 item = radix_tree_lookup(&mapping->page_tree, index);
287 rcu_read_unlock();
288 return item == swp_to_radix_entry(swap);
289}
290
291/*
292 * Like add_to_page_cache_locked, but error if expected item has gone.
293 */
294static int shmem_add_to_page_cache(struct page *page,
295 struct address_space *mapping,
296 pgoff_t index, void *expected)
297{
298 int error;
299
300 VM_BUG_ON_PAGE(!PageLocked(page), page);
301 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
302
303 get_page(page);
304 page->mapping = mapping;
305 page->index = index;
306
307 spin_lock_irq(&mapping->tree_lock);
308 if (!expected)
309 error = radix_tree_insert(&mapping->page_tree, index, page);
310 else
311 error = shmem_radix_tree_replace(mapping, index, expected,
312 page);
313 if (!error) {
314 mapping->nrpages++;
315 __inc_zone_page_state(page, NR_FILE_PAGES);
316 __inc_zone_page_state(page, NR_SHMEM);
317 spin_unlock_irq(&mapping->tree_lock);
318 } else {
319 page->mapping = NULL;
320 spin_unlock_irq(&mapping->tree_lock);
321 put_page(page);
322 }
323 return error;
324}
325
326/*
327 * Like delete_from_page_cache, but substitutes swap for page.
328 */
329static void shmem_delete_from_page_cache(struct page *page, void *radswap)
330{
331 struct address_space *mapping = page->mapping;
332 int error;
333
334 spin_lock_irq(&mapping->tree_lock);
335 error = shmem_radix_tree_replace(mapping, page->index, page, radswap);
336 page->mapping = NULL;
337 mapping->nrpages--;
338 __dec_zone_page_state(page, NR_FILE_PAGES);
339 __dec_zone_page_state(page, NR_SHMEM);
340 spin_unlock_irq(&mapping->tree_lock);
341 put_page(page);
342 BUG_ON(error);
343}
344
345/*
346 * Remove swap entry from radix tree, free the swap and its page cache.
347 */
348static int shmem_free_swap(struct address_space *mapping,
349 pgoff_t index, void *radswap)
350{
351 void *old;
352
353 spin_lock_irq(&mapping->tree_lock);
354 old = radix_tree_delete_item(&mapping->page_tree, index, radswap);
355 spin_unlock_irq(&mapping->tree_lock);
356 if (old != radswap)
357 return -ENOENT;
358 free_swap_and_cache(radix_to_swp_entry(radswap));
359 return 0;
360}
361
362/*
363 * Determine (in bytes) how many of the shmem object's pages mapped by the
364 * given offsets are swapped out.
365 *
366 * This is safe to call without i_mutex or mapping->tree_lock thanks to RCU,
367 * as long as the inode doesn't go away and racy results are not a problem.
368 */
369unsigned long shmem_partial_swap_usage(struct address_space *mapping,
370 pgoff_t start, pgoff_t end)
371{
372 struct radix_tree_iter iter;
373 void **slot;
374 struct page *page;
375 unsigned long swapped = 0;
376
377 rcu_read_lock();
378
379 radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
380 if (iter.index >= end)
381 break;
382
383 page = radix_tree_deref_slot(slot);
384
385 if (radix_tree_deref_retry(page)) {
386 slot = radix_tree_iter_retry(&iter);
387 continue;
388 }
389
390 if (radix_tree_exceptional_entry(page))
391 swapped++;
392
393 if (need_resched()) {
394 cond_resched_rcu();
395 slot = radix_tree_iter_next(&iter);
396 }
397 }
398
399 rcu_read_unlock();
400
401 return swapped << PAGE_SHIFT;
402}
403
404/*
405 * Determine (in bytes) how many of the shmem object's pages mapped by the
406 * given vma is swapped out.
407 *
408 * This is safe to call without i_mutex or mapping->tree_lock thanks to RCU,
409 * as long as the inode doesn't go away and racy results are not a problem.
410 */
411unsigned long shmem_swap_usage(struct vm_area_struct *vma)
412{
413 struct inode *inode = file_inode(vma->vm_file);
414 struct shmem_inode_info *info = SHMEM_I(inode);
415 struct address_space *mapping = inode->i_mapping;
416 unsigned long swapped;
417
418 /* Be careful as we don't hold info->lock */
419 swapped = READ_ONCE(info->swapped);
420
421 /*
422 * The easier cases are when the shmem object has nothing in swap, or
423 * the vma maps it whole. Then we can simply use the stats that we
424 * already track.
425 */
426 if (!swapped)
427 return 0;
428
429 if (!vma->vm_pgoff && vma->vm_end - vma->vm_start >= inode->i_size)
430 return swapped << PAGE_SHIFT;
431
432 /* Here comes the more involved part */
433 return shmem_partial_swap_usage(mapping,
434 linear_page_index(vma, vma->vm_start),
435 linear_page_index(vma, vma->vm_end));
436}
437
438/*
439 * SysV IPC SHM_UNLOCK restore Unevictable pages to their evictable lists.
440 */
441void shmem_unlock_mapping(struct address_space *mapping)
442{
443 struct pagevec pvec;
444 pgoff_t indices[PAGEVEC_SIZE];
445 pgoff_t index = 0;
446
447 pagevec_init(&pvec, 0);
448 /*
449 * Minor point, but we might as well stop if someone else SHM_LOCKs it.
450 */
451 while (!mapping_unevictable(mapping)) {
452 /*
453 * Avoid pagevec_lookup(): find_get_pages() returns 0 as if it
454 * has finished, if it hits a row of PAGEVEC_SIZE swap entries.
455 */
456 pvec.nr = find_get_entries(mapping, index,
457 PAGEVEC_SIZE, pvec.pages, indices);
458 if (!pvec.nr)
459 break;
460 index = indices[pvec.nr - 1] + 1;
461 pagevec_remove_exceptionals(&pvec);
462 check_move_unevictable_pages(pvec.pages, pvec.nr);
463 pagevec_release(&pvec);
464 cond_resched();
465 }
466}
467
468/*
469 * Remove range of pages and swap entries from radix tree, and free them.
470 * If !unfalloc, truncate or punch hole; if unfalloc, undo failed fallocate.
471 */
472static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
473 bool unfalloc)
474{
475 struct address_space *mapping = inode->i_mapping;
476 struct shmem_inode_info *info = SHMEM_I(inode);
477 pgoff_t start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
478 pgoff_t end = (lend + 1) >> PAGE_SHIFT;
479 unsigned int partial_start = lstart & (PAGE_SIZE - 1);
480 unsigned int partial_end = (lend + 1) & (PAGE_SIZE - 1);
481 struct pagevec pvec;
482 pgoff_t indices[PAGEVEC_SIZE];
483 long nr_swaps_freed = 0;
484 pgoff_t index;
485 int i;
486
487 if (lend == -1)
488 end = -1; /* unsigned, so actually very big */
489
490 pagevec_init(&pvec, 0);
491 index = start;
492 while (index < end) {
493 pvec.nr = find_get_entries(mapping, index,
494 min(end - index, (pgoff_t)PAGEVEC_SIZE),
495 pvec.pages, indices);
496 if (!pvec.nr)
497 break;
498 for (i = 0; i < pagevec_count(&pvec); i++) {
499 struct page *page = pvec.pages[i];
500
501 index = indices[i];
502 if (index >= end)
503 break;
504
505 if (radix_tree_exceptional_entry(page)) {
506 if (unfalloc)
507 continue;
508 nr_swaps_freed += !shmem_free_swap(mapping,
509 index, page);
510 continue;
511 }
512
513 if (!trylock_page(page))
514 continue;
515 if (!unfalloc || !PageUptodate(page)) {
516 if (page->mapping == mapping) {
517 VM_BUG_ON_PAGE(PageWriteback(page), page);
518 truncate_inode_page(mapping, page);
519 }
520 }
521 unlock_page(page);
522 }
523 pagevec_remove_exceptionals(&pvec);
524 pagevec_release(&pvec);
525 cond_resched();
526 index++;
527 }
528
529 if (partial_start) {
530 struct page *page = NULL;
531 shmem_getpage(inode, start - 1, &page, SGP_READ, NULL);
532 if (page) {
533 unsigned int top = PAGE_SIZE;
534 if (start > end) {
535 top = partial_end;
536 partial_end = 0;
537 }
538 zero_user_segment(page, partial_start, top);
539 set_page_dirty(page);
540 unlock_page(page);
541 put_page(page);
542 }
543 }
544 if (partial_end) {
545 struct page *page = NULL;
546 shmem_getpage(inode, end, &page, SGP_READ, NULL);
547 if (page) {
548 zero_user_segment(page, 0, partial_end);
549 set_page_dirty(page);
550 unlock_page(page);
551 put_page(page);
552 }
553 }
554 if (start >= end)
555 return;
556
557 index = start;
558 while (index < end) {
559 cond_resched();
560
561 pvec.nr = find_get_entries(mapping, index,
562 min(end - index, (pgoff_t)PAGEVEC_SIZE),
563 pvec.pages, indices);
564 if (!pvec.nr) {
565 /* If all gone or hole-punch or unfalloc, we're done */
566 if (index == start || end != -1)
567 break;
568 /* But if truncating, restart to make sure all gone */
569 index = start;
570 continue;
571 }
572 for (i = 0; i < pagevec_count(&pvec); i++) {
573 struct page *page = pvec.pages[i];
574
575 index = indices[i];
576 if (index >= end)
577 break;
578
579 if (radix_tree_exceptional_entry(page)) {
580 if (unfalloc)
581 continue;
582 if (shmem_free_swap(mapping, index, page)) {
583 /* Swap was replaced by page: retry */
584 index--;
585 break;
586 }
587 nr_swaps_freed++;
588 continue;
589 }
590
591 lock_page(page);
592 if (!unfalloc || !PageUptodate(page)) {
593 if (page->mapping == mapping) {
594 VM_BUG_ON_PAGE(PageWriteback(page), page);
595 truncate_inode_page(mapping, page);
596 } else {
597 /* Page was replaced by swap: retry */
598 unlock_page(page);
599 index--;
600 break;
601 }
602 }
603 unlock_page(page);
604 }
605 pagevec_remove_exceptionals(&pvec);
606 pagevec_release(&pvec);
607 index++;
608 }
609
610 spin_lock(&info->lock);
611 info->swapped -= nr_swaps_freed;
612 shmem_recalc_inode(inode);
613 spin_unlock(&info->lock);
614}
615
616void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
617{
618 shmem_undo_range(inode, lstart, lend, false);
619 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
620}
621EXPORT_SYMBOL_GPL(shmem_truncate_range);
622
623static int shmem_getattr(struct vfsmount *mnt, struct dentry *dentry,
624 struct kstat *stat)
625{
626 struct inode *inode = dentry->d_inode;
627 struct shmem_inode_info *info = SHMEM_I(inode);
628
629 if (info->alloced - info->swapped != inode->i_mapping->nrpages) {
630 spin_lock(&info->lock);
631 shmem_recalc_inode(inode);
632 spin_unlock(&info->lock);
633 }
634 generic_fillattr(inode, stat);
635 return 0;
636}
637
638static int shmem_setattr(struct dentry *dentry, struct iattr *attr)
639{
640 struct inode *inode = d_inode(dentry);
641 struct shmem_inode_info *info = SHMEM_I(inode);
642 int error;
643
644 error = inode_change_ok(inode, attr);
645 if (error)
646 return error;
647
648 if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
649 loff_t oldsize = inode->i_size;
650 loff_t newsize = attr->ia_size;
651
652 /* protected by i_mutex */
653 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
654 (newsize > oldsize && (info->seals & F_SEAL_GROW)))
655 return -EPERM;
656
657 if (newsize != oldsize) {
658 error = shmem_reacct_size(SHMEM_I(inode)->flags,
659 oldsize, newsize);
660 if (error)
661 return error;
662 i_size_write(inode, newsize);
663 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
664 }
665 if (newsize <= oldsize) {
666 loff_t holebegin = round_up(newsize, PAGE_SIZE);
667 if (oldsize > holebegin)
668 unmap_mapping_range(inode->i_mapping,
669 holebegin, 0, 1);
670 if (info->alloced)
671 shmem_truncate_range(inode,
672 newsize, (loff_t)-1);
673 /* unmap again to remove racily COWed private pages */
674 if (oldsize > holebegin)
675 unmap_mapping_range(inode->i_mapping,
676 holebegin, 0, 1);
677 }
678 }
679
680 setattr_copy(inode, attr);
681 if (attr->ia_valid & ATTR_MODE)
682 error = posix_acl_chmod(inode, inode->i_mode);
683 return error;
684}
685
686static void shmem_evict_inode(struct inode *inode)
687{
688 struct shmem_inode_info *info = SHMEM_I(inode);
689
690 if (inode->i_mapping->a_ops == &shmem_aops) {
691 shmem_unacct_size(info->flags, inode->i_size);
692 inode->i_size = 0;
693 shmem_truncate_range(inode, 0, (loff_t)-1);
694 if (!list_empty(&info->swaplist)) {
695 mutex_lock(&shmem_swaplist_mutex);
696 list_del_init(&info->swaplist);
697 mutex_unlock(&shmem_swaplist_mutex);
698 }
699 }
700
701 simple_xattrs_free(&info->xattrs);
702 WARN_ON(inode->i_blocks);
703 shmem_free_inode(inode->i_sb);
704 clear_inode(inode);
705}
706
707/*
708 * If swap found in inode, free it and move page from swapcache to filecache.
709 */
710static int shmem_unuse_inode(struct shmem_inode_info *info,
711 swp_entry_t swap, struct page **pagep)
712{
713 struct address_space *mapping = info->vfs_inode.i_mapping;
714 void *radswap;
715 pgoff_t index;
716 gfp_t gfp;
717 int error = 0;
718
719 radswap = swp_to_radix_entry(swap);
720 index = radix_tree_locate_item(&mapping->page_tree, radswap);
721 if (index == -1)
722 return -EAGAIN; /* tell shmem_unuse we found nothing */
723
724 /*
725 * Move _head_ to start search for next from here.
726 * But be careful: shmem_evict_inode checks list_empty without taking
727 * mutex, and there's an instant in list_move_tail when info->swaplist
728 * would appear empty, if it were the only one on shmem_swaplist.
729 */
730 if (shmem_swaplist.next != &info->swaplist)
731 list_move_tail(&shmem_swaplist, &info->swaplist);
732
733 gfp = mapping_gfp_mask(mapping);
734 if (shmem_should_replace_page(*pagep, gfp)) {
735 mutex_unlock(&shmem_swaplist_mutex);
736 error = shmem_replace_page(pagep, gfp, info, index);
737 mutex_lock(&shmem_swaplist_mutex);
738 /*
739 * We needed to drop mutex to make that restrictive page
740 * allocation, but the inode might have been freed while we
741 * dropped it: although a racing shmem_evict_inode() cannot
742 * complete without emptying the radix_tree, our page lock
743 * on this swapcache page is not enough to prevent that -
744 * free_swap_and_cache() of our swap entry will only
745 * trylock_page(), removing swap from radix_tree whatever.
746 *
747 * We must not proceed to shmem_add_to_page_cache() if the
748 * inode has been freed, but of course we cannot rely on
749 * inode or mapping or info to check that. However, we can
750 * safely check if our swap entry is still in use (and here
751 * it can't have got reused for another page): if it's still
752 * in use, then the inode cannot have been freed yet, and we
753 * can safely proceed (if it's no longer in use, that tells
754 * nothing about the inode, but we don't need to unuse swap).
755 */
756 if (!page_swapcount(*pagep))
757 error = -ENOENT;
758 }
759
760 /*
761 * We rely on shmem_swaplist_mutex, not only to protect the swaplist,
762 * but also to hold up shmem_evict_inode(): so inode cannot be freed
763 * beneath us (pagelock doesn't help until the page is in pagecache).
764 */
765 if (!error)
766 error = shmem_add_to_page_cache(*pagep, mapping, index,
767 radswap);
768 if (error != -ENOMEM) {
769 /*
770 * Truncation and eviction use free_swap_and_cache(), which
771 * only does trylock page: if we raced, best clean up here.
772 */
773 delete_from_swap_cache(*pagep);
774 set_page_dirty(*pagep);
775 if (!error) {
776 spin_lock(&info->lock);
777 info->swapped--;
778 spin_unlock(&info->lock);
779 swap_free(swap);
780 }
781 }
782 return error;
783}
784
785/*
786 * Search through swapped inodes to find and replace swap by page.
787 */
788int shmem_unuse(swp_entry_t swap, struct page *page)
789{
790 struct list_head *this, *next;
791 struct shmem_inode_info *info;
792 struct mem_cgroup *memcg;
793 int error = 0;
794
795 /*
796 * There's a faint possibility that swap page was replaced before
797 * caller locked it: caller will come back later with the right page.
798 */
799 if (unlikely(!PageSwapCache(page) || page_private(page) != swap.val))
800 goto out;
801
802 /*
803 * Charge page using GFP_KERNEL while we can wait, before taking
804 * the shmem_swaplist_mutex which might hold up shmem_writepage().
805 * Charged back to the user (not to caller) when swap account is used.
806 */
807 error = mem_cgroup_try_charge(page, current->mm, GFP_KERNEL, &memcg,
808 false);
809 if (error)
810 goto out;
811 /* No radix_tree_preload: swap entry keeps a place for page in tree */
812 error = -EAGAIN;
813
814 mutex_lock(&shmem_swaplist_mutex);
815 list_for_each_safe(this, next, &shmem_swaplist) {
816 info = list_entry(this, struct shmem_inode_info, swaplist);
817 if (info->swapped)
818 error = shmem_unuse_inode(info, swap, &page);
819 else
820 list_del_init(&info->swaplist);
821 cond_resched();
822 if (error != -EAGAIN)
823 break;
824 /* found nothing in this: move on to search the next */
825 }
826 mutex_unlock(&shmem_swaplist_mutex);
827
828 if (error) {
829 if (error != -ENOMEM)
830 error = 0;
831 mem_cgroup_cancel_charge(page, memcg, false);
832 } else
833 mem_cgroup_commit_charge(page, memcg, true, false);
834out:
835 unlock_page(page);
836 put_page(page);
837 return error;
838}
839
840/*
841 * Move the page from the page cache to the swap cache.
842 */
843static int shmem_writepage(struct page *page, struct writeback_control *wbc)
844{
845 struct shmem_inode_info *info;
846 struct address_space *mapping;
847 struct inode *inode;
848 swp_entry_t swap;
849 pgoff_t index;
850
851 BUG_ON(!PageLocked(page));
852 mapping = page->mapping;
853 index = page->index;
854 inode = mapping->host;
855 info = SHMEM_I(inode);
856 if (info->flags & VM_LOCKED)
857 goto redirty;
858 if (!total_swap_pages)
859 goto redirty;
860
861 /*
862 * Our capabilities prevent regular writeback or sync from ever calling
863 * shmem_writepage; but a stacking filesystem might use ->writepage of
864 * its underlying filesystem, in which case tmpfs should write out to
865 * swap only in response to memory pressure, and not for the writeback
866 * threads or sync.
867 */
868 if (!wbc->for_reclaim) {
869 WARN_ON_ONCE(1); /* Still happens? Tell us about it! */
870 goto redirty;
871 }
872
873 /*
874 * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
875 * value into swapfile.c, the only way we can correctly account for a
876 * fallocated page arriving here is now to initialize it and write it.
877 *
878 * That's okay for a page already fallocated earlier, but if we have
879 * not yet completed the fallocation, then (a) we want to keep track
880 * of this page in case we have to undo it, and (b) it may not be a
881 * good idea to continue anyway, once we're pushing into swap. So
882 * reactivate the page, and let shmem_fallocate() quit when too many.
883 */
884 if (!PageUptodate(page)) {
885 if (inode->i_private) {
886 struct shmem_falloc *shmem_falloc;
887 spin_lock(&inode->i_lock);
888 shmem_falloc = inode->i_private;
889 if (shmem_falloc &&
890 !shmem_falloc->waitq &&
891 index >= shmem_falloc->start &&
892 index < shmem_falloc->next)
893 shmem_falloc->nr_unswapped++;
894 else
895 shmem_falloc = NULL;
896 spin_unlock(&inode->i_lock);
897 if (shmem_falloc)
898 goto redirty;
899 }
900 clear_highpage(page);
901 flush_dcache_page(page);
902 SetPageUptodate(page);
903 }
904
905 swap = get_swap_page();
906 if (!swap.val)
907 goto redirty;
908
909 if (mem_cgroup_try_charge_swap(page, swap))
910 goto free_swap;
911
912 /*
913 * Add inode to shmem_unuse()'s list of swapped-out inodes,
914 * if it's not already there. Do it now before the page is
915 * moved to swap cache, when its pagelock no longer protects
916 * the inode from eviction. But don't unlock the mutex until
917 * we've incremented swapped, because shmem_unuse_inode() will
918 * prune a !swapped inode from the swaplist under this mutex.
919 */
920 mutex_lock(&shmem_swaplist_mutex);
921 if (list_empty(&info->swaplist))
922 list_add_tail(&info->swaplist, &shmem_swaplist);
923
924 if (add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
925 spin_lock(&info->lock);
926 shmem_recalc_inode(inode);
927 info->swapped++;
928 spin_unlock(&info->lock);
929
930 swap_shmem_alloc(swap);
931 shmem_delete_from_page_cache(page, swp_to_radix_entry(swap));
932
933 mutex_unlock(&shmem_swaplist_mutex);
934 BUG_ON(page_mapped(page));
935 swap_writepage(page, wbc);
936 return 0;
937 }
938
939 mutex_unlock(&shmem_swaplist_mutex);
940free_swap:
941 swapcache_free(swap);
942redirty:
943 set_page_dirty(page);
944 if (wbc->for_reclaim)
945 return AOP_WRITEPAGE_ACTIVATE; /* Return with page locked */
946 unlock_page(page);
947 return 0;
948}
949
950#ifdef CONFIG_NUMA
951#ifdef CONFIG_TMPFS
952static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
953{
954 char buffer[64];
955
956 if (!mpol || mpol->mode == MPOL_DEFAULT)
957 return; /* show nothing */
958
959 mpol_to_str(buffer, sizeof(buffer), mpol);
960
961 seq_printf(seq, ",mpol=%s", buffer);
962}
963
964static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
965{
966 struct mempolicy *mpol = NULL;
967 if (sbinfo->mpol) {
968 spin_lock(&sbinfo->stat_lock); /* prevent replace/use races */
969 mpol = sbinfo->mpol;
970 mpol_get(mpol);
971 spin_unlock(&sbinfo->stat_lock);
972 }
973 return mpol;
974}
975#endif /* CONFIG_TMPFS */
976
977static struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
978 struct shmem_inode_info *info, pgoff_t index)
979{
980 struct vm_area_struct pvma;
981 struct page *page;
982
983 /* Create a pseudo vma that just contains the policy */
984 pvma.vm_start = 0;
985 /* Bias interleave by inode number to distribute better across nodes */
986 pvma.vm_pgoff = index + info->vfs_inode.i_ino;
987 pvma.vm_ops = NULL;
988 pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, index);
989
990 page = swapin_readahead(swap, gfp, &pvma, 0);
991
992 /* Drop reference taken by mpol_shared_policy_lookup() */
993 mpol_cond_put(pvma.vm_policy);
994
995 return page;
996}
997
998static struct page *shmem_alloc_page(gfp_t gfp,
999 struct shmem_inode_info *info, pgoff_t index)
1000{
1001 struct vm_area_struct pvma;
1002 struct page *page;
1003
1004 /* Create a pseudo vma that just contains the policy */
1005 pvma.vm_start = 0;
1006 /* Bias interleave by inode number to distribute better across nodes */
1007 pvma.vm_pgoff = index + info->vfs_inode.i_ino;
1008 pvma.vm_ops = NULL;
1009 pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, index);
1010
1011 page = alloc_page_vma(gfp, &pvma, 0);
1012
1013 /* Drop reference taken by mpol_shared_policy_lookup() */
1014 mpol_cond_put(pvma.vm_policy);
1015
1016 return page;
1017}
1018#else /* !CONFIG_NUMA */
1019#ifdef CONFIG_TMPFS
1020static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1021{
1022}
1023#endif /* CONFIG_TMPFS */
1024
1025static inline struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
1026 struct shmem_inode_info *info, pgoff_t index)
1027{
1028 return swapin_readahead(swap, gfp, NULL, 0);
1029}
1030
1031static inline struct page *shmem_alloc_page(gfp_t gfp,
1032 struct shmem_inode_info *info, pgoff_t index)
1033{
1034 return alloc_page(gfp);
1035}
1036#endif /* CONFIG_NUMA */
1037
1038#if !defined(CONFIG_NUMA) || !defined(CONFIG_TMPFS)
1039static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1040{
1041 return NULL;
1042}
1043#endif
1044
1045/*
1046 * When a page is moved from swapcache to shmem filecache (either by the
1047 * usual swapin of shmem_getpage_gfp(), or by the less common swapoff of
1048 * shmem_unuse_inode()), it may have been read in earlier from swap, in
1049 * ignorance of the mapping it belongs to. If that mapping has special
1050 * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
1051 * we may need to copy to a suitable page before moving to filecache.
1052 *
1053 * In a future release, this may well be extended to respect cpuset and
1054 * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
1055 * but for now it is a simple matter of zone.
1056 */
1057static bool shmem_should_replace_page(struct page *page, gfp_t gfp)
1058{
1059 return page_zonenum(page) > gfp_zone(gfp);
1060}
1061
1062static int shmem_replace_page(struct page **pagep, gfp_t gfp,
1063 struct shmem_inode_info *info, pgoff_t index)
1064{
1065 struct page *oldpage, *newpage;
1066 struct address_space *swap_mapping;
1067 pgoff_t swap_index;
1068 int error;
1069
1070 oldpage = *pagep;
1071 swap_index = page_private(oldpage);
1072 swap_mapping = page_mapping(oldpage);
1073
1074 /*
1075 * We have arrived here because our zones are constrained, so don't
1076 * limit chance of success by further cpuset and node constraints.
1077 */
1078 gfp &= ~GFP_CONSTRAINT_MASK;
1079 newpage = shmem_alloc_page(gfp, info, index);
1080 if (!newpage)
1081 return -ENOMEM;
1082
1083 get_page(newpage);
1084 copy_highpage(newpage, oldpage);
1085 flush_dcache_page(newpage);
1086
1087 __SetPageLocked(newpage);
1088 SetPageUptodate(newpage);
1089 SetPageSwapBacked(newpage);
1090 set_page_private(newpage, swap_index);
1091 SetPageSwapCache(newpage);
1092
1093 /*
1094 * Our caller will very soon move newpage out of swapcache, but it's
1095 * a nice clean interface for us to replace oldpage by newpage there.
1096 */
1097 spin_lock_irq(&swap_mapping->tree_lock);
1098 error = shmem_radix_tree_replace(swap_mapping, swap_index, oldpage,
1099 newpage);
1100 if (!error) {
1101 __inc_zone_page_state(newpage, NR_FILE_PAGES);
1102 __dec_zone_page_state(oldpage, NR_FILE_PAGES);
1103 }
1104 spin_unlock_irq(&swap_mapping->tree_lock);
1105
1106 if (unlikely(error)) {
1107 /*
1108 * Is this possible? I think not, now that our callers check
1109 * both PageSwapCache and page_private after getting page lock;
1110 * but be defensive. Reverse old to newpage for clear and free.
1111 */
1112 oldpage = newpage;
1113 } else {
1114 mem_cgroup_migrate(oldpage, newpage);
1115 lru_cache_add_anon(newpage);
1116 *pagep = newpage;
1117 }
1118
1119 ClearPageSwapCache(oldpage);
1120 set_page_private(oldpage, 0);
1121
1122 unlock_page(oldpage);
1123 put_page(oldpage);
1124 put_page(oldpage);
1125 return error;
1126}
1127
1128/*
1129 * shmem_getpage_gfp - find page in cache, or get from swap, or allocate
1130 *
1131 * If we allocate a new one we do not mark it dirty. That's up to the
1132 * vm. If we swap it in we mark it dirty since we also free the swap
1133 * entry since a page cannot live in both the swap and page cache
1134 */
1135static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
1136 struct page **pagep, enum sgp_type sgp, gfp_t gfp, int *fault_type)
1137{
1138 struct address_space *mapping = inode->i_mapping;
1139 struct shmem_inode_info *info;
1140 struct shmem_sb_info *sbinfo;
1141 struct mem_cgroup *memcg;
1142 struct page *page;
1143 swp_entry_t swap;
1144 int error;
1145 int once = 0;
1146 int alloced = 0;
1147
1148 if (index > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
1149 return -EFBIG;
1150repeat:
1151 swap.val = 0;
1152 page = find_lock_entry(mapping, index);
1153 if (radix_tree_exceptional_entry(page)) {
1154 swap = radix_to_swp_entry(page);
1155 page = NULL;
1156 }
1157
1158 if (sgp != SGP_WRITE && sgp != SGP_FALLOC &&
1159 ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1160 error = -EINVAL;
1161 goto unlock;
1162 }
1163
1164 if (page && sgp == SGP_WRITE)
1165 mark_page_accessed(page);
1166
1167 /* fallocated page? */
1168 if (page && !PageUptodate(page)) {
1169 if (sgp != SGP_READ)
1170 goto clear;
1171 unlock_page(page);
1172 put_page(page);
1173 page = NULL;
1174 }
1175 if (page || (sgp == SGP_READ && !swap.val)) {
1176 *pagep = page;
1177 return 0;
1178 }
1179
1180 /*
1181 * Fast cache lookup did not find it:
1182 * bring it back from swap or allocate.
1183 */
1184 info = SHMEM_I(inode);
1185 sbinfo = SHMEM_SB(inode->i_sb);
1186
1187 if (swap.val) {
1188 /* Look it up and read it in.. */
1189 page = lookup_swap_cache(swap);
1190 if (!page) {
1191 /* here we actually do the io */
1192 if (fault_type)
1193 *fault_type |= VM_FAULT_MAJOR;
1194 page = shmem_swapin(swap, gfp, info, index);
1195 if (!page) {
1196 error = -ENOMEM;
1197 goto failed;
1198 }
1199 }
1200
1201 /* We have to do this with page locked to prevent races */
1202 lock_page(page);
1203 if (!PageSwapCache(page) || page_private(page) != swap.val ||
1204 !shmem_confirm_swap(mapping, index, swap)) {
1205 error = -EEXIST; /* try again */
1206 goto unlock;
1207 }
1208 if (!PageUptodate(page)) {
1209 error = -EIO;
1210 goto failed;
1211 }
1212 wait_on_page_writeback(page);
1213
1214 if (shmem_should_replace_page(page, gfp)) {
1215 error = shmem_replace_page(&page, gfp, info, index);
1216 if (error)
1217 goto failed;
1218 }
1219
1220 error = mem_cgroup_try_charge(page, current->mm, gfp, &memcg,
1221 false);
1222 if (!error) {
1223 error = shmem_add_to_page_cache(page, mapping, index,
1224 swp_to_radix_entry(swap));
1225 /*
1226 * We already confirmed swap under page lock, and make
1227 * no memory allocation here, so usually no possibility
1228 * of error; but free_swap_and_cache() only trylocks a
1229 * page, so it is just possible that the entry has been
1230 * truncated or holepunched since swap was confirmed.
1231 * shmem_undo_range() will have done some of the
1232 * unaccounting, now delete_from_swap_cache() will do
1233 * the rest.
1234 * Reset swap.val? No, leave it so "failed" goes back to
1235 * "repeat": reading a hole and writing should succeed.
1236 */
1237 if (error) {
1238 mem_cgroup_cancel_charge(page, memcg, false);
1239 delete_from_swap_cache(page);
1240 }
1241 }
1242 if (error)
1243 goto failed;
1244
1245 mem_cgroup_commit_charge(page, memcg, true, false);
1246
1247 spin_lock(&info->lock);
1248 info->swapped--;
1249 shmem_recalc_inode(inode);
1250 spin_unlock(&info->lock);
1251
1252 if (sgp == SGP_WRITE)
1253 mark_page_accessed(page);
1254
1255 delete_from_swap_cache(page);
1256 set_page_dirty(page);
1257 swap_free(swap);
1258
1259 } else {
1260 if (shmem_acct_block(info->flags)) {
1261 error = -ENOSPC;
1262 goto failed;
1263 }
1264 if (sbinfo->max_blocks) {
1265 if (percpu_counter_compare(&sbinfo->used_blocks,
1266 sbinfo->max_blocks) >= 0) {
1267 error = -ENOSPC;
1268 goto unacct;
1269 }
1270 percpu_counter_inc(&sbinfo->used_blocks);
1271 }
1272
1273 page = shmem_alloc_page(gfp, info, index);
1274 if (!page) {
1275 error = -ENOMEM;
1276 goto decused;
1277 }
1278
1279 __SetPageSwapBacked(page);
1280 __SetPageLocked(page);
1281 if (sgp == SGP_WRITE)
1282 __SetPageReferenced(page);
1283
1284 error = mem_cgroup_try_charge(page, current->mm, gfp, &memcg,
1285 false);
1286 if (error)
1287 goto decused;
1288 error = radix_tree_maybe_preload(gfp & GFP_RECLAIM_MASK);
1289 if (!error) {
1290 error = shmem_add_to_page_cache(page, mapping, index,
1291 NULL);
1292 radix_tree_preload_end();
1293 }
1294 if (error) {
1295 mem_cgroup_cancel_charge(page, memcg, false);
1296 goto decused;
1297 }
1298 mem_cgroup_commit_charge(page, memcg, false, false);
1299 lru_cache_add_anon(page);
1300
1301 spin_lock(&info->lock);
1302 info->alloced++;
1303 inode->i_blocks += BLOCKS_PER_PAGE;
1304 shmem_recalc_inode(inode);
1305 spin_unlock(&info->lock);
1306 alloced = true;
1307
1308 /*
1309 * Let SGP_FALLOC use the SGP_WRITE optimization on a new page.
1310 */
1311 if (sgp == SGP_FALLOC)
1312 sgp = SGP_WRITE;
1313clear:
1314 /*
1315 * Let SGP_WRITE caller clear ends if write does not fill page;
1316 * but SGP_FALLOC on a page fallocated earlier must initialize
1317 * it now, lest undo on failure cancel our earlier guarantee.
1318 */
1319 if (sgp != SGP_WRITE) {
1320 clear_highpage(page);
1321 flush_dcache_page(page);
1322 SetPageUptodate(page);
1323 }
1324 if (sgp == SGP_DIRTY)
1325 set_page_dirty(page);
1326 }
1327
1328 /* Perhaps the file has been truncated since we checked */
1329 if (sgp != SGP_WRITE && sgp != SGP_FALLOC &&
1330 ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1331 if (alloced) {
1332 ClearPageDirty(page);
1333 delete_from_page_cache(page);
1334 spin_lock(&info->lock);
1335 shmem_recalc_inode(inode);
1336 spin_unlock(&info->lock);
1337 }
1338 error = -EINVAL;
1339 goto unlock;
1340 }
1341 *pagep = page;
1342 return 0;
1343
1344 /*
1345 * Error recovery.
1346 */
1347decused:
1348 if (sbinfo->max_blocks)
1349 percpu_counter_add(&sbinfo->used_blocks, -1);
1350unacct:
1351 shmem_unacct_blocks(info->flags, 1);
1352failed:
1353 if (swap.val && !shmem_confirm_swap(mapping, index, swap))
1354 error = -EEXIST;
1355unlock:
1356 if (page) {
1357 unlock_page(page);
1358 put_page(page);
1359 }
1360 if (error == -ENOSPC && !once++) {
1361 info = SHMEM_I(inode);
1362 spin_lock(&info->lock);
1363 shmem_recalc_inode(inode);
1364 spin_unlock(&info->lock);
1365 goto repeat;
1366 }
1367 if (error == -EEXIST) /* from above or from radix_tree_insert */
1368 goto repeat;
1369 return error;
1370}
1371
1372static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1373{
1374 struct inode *inode = file_inode(vma->vm_file);
1375 int error;
1376 int ret = VM_FAULT_LOCKED;
1377
1378 /*
1379 * Trinity finds that probing a hole which tmpfs is punching can
1380 * prevent the hole-punch from ever completing: which in turn
1381 * locks writers out with its hold on i_mutex. So refrain from
1382 * faulting pages into the hole while it's being punched. Although
1383 * shmem_undo_range() does remove the additions, it may be unable to
1384 * keep up, as each new page needs its own unmap_mapping_range() call,
1385 * and the i_mmap tree grows ever slower to scan if new vmas are added.
1386 *
1387 * It does not matter if we sometimes reach this check just before the
1388 * hole-punch begins, so that one fault then races with the punch:
1389 * we just need to make racing faults a rare case.
1390 *
1391 * The implementation below would be much simpler if we just used a
1392 * standard mutex or completion: but we cannot take i_mutex in fault,
1393 * and bloating every shmem inode for this unlikely case would be sad.
1394 */
1395 if (unlikely(inode->i_private)) {
1396 struct shmem_falloc *shmem_falloc;
1397
1398 spin_lock(&inode->i_lock);
1399 shmem_falloc = inode->i_private;
1400 if (shmem_falloc &&
1401 shmem_falloc->waitq &&
1402 vmf->pgoff >= shmem_falloc->start &&
1403 vmf->pgoff < shmem_falloc->next) {
1404 wait_queue_head_t *shmem_falloc_waitq;
1405 DEFINE_WAIT(shmem_fault_wait);
1406
1407 ret = VM_FAULT_NOPAGE;
1408 if ((vmf->flags & FAULT_FLAG_ALLOW_RETRY) &&
1409 !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
1410 /* It's polite to up mmap_sem if we can */
1411 up_read(&vma->vm_mm->mmap_sem);
1412 ret = VM_FAULT_RETRY;
1413 }
1414
1415 shmem_falloc_waitq = shmem_falloc->waitq;
1416 prepare_to_wait(shmem_falloc_waitq, &shmem_fault_wait,
1417 TASK_UNINTERRUPTIBLE);
1418 spin_unlock(&inode->i_lock);
1419 schedule();
1420
1421 /*
1422 * shmem_falloc_waitq points into the shmem_fallocate()
1423 * stack of the hole-punching task: shmem_falloc_waitq
1424 * is usually invalid by the time we reach here, but
1425 * finish_wait() does not dereference it in that case;
1426 * though i_lock needed lest racing with wake_up_all().
1427 */
1428 spin_lock(&inode->i_lock);
1429 finish_wait(shmem_falloc_waitq, &shmem_fault_wait);
1430 spin_unlock(&inode->i_lock);
1431 return ret;
1432 }
1433 spin_unlock(&inode->i_lock);
1434 }
1435
1436 error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_CACHE, &ret);
1437 if (error)
1438 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1439
1440 if (ret & VM_FAULT_MAJOR) {
1441 count_vm_event(PGMAJFAULT);
1442 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1443 }
1444 return ret;
1445}
1446
1447#ifdef CONFIG_NUMA
1448static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
1449{
1450 struct inode *inode = file_inode(vma->vm_file);
1451 return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
1452}
1453
1454static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
1455 unsigned long addr)
1456{
1457 struct inode *inode = file_inode(vma->vm_file);
1458 pgoff_t index;
1459
1460 index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1461 return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
1462}
1463#endif
1464
1465int shmem_lock(struct file *file, int lock, struct user_struct *user)
1466{
1467 struct inode *inode = file_inode(file);
1468 struct shmem_inode_info *info = SHMEM_I(inode);
1469 int retval = -ENOMEM;
1470
1471 spin_lock(&info->lock);
1472 if (lock && !(info->flags & VM_LOCKED)) {
1473 if (!user_shm_lock(inode->i_size, user))
1474 goto out_nomem;
1475 info->flags |= VM_LOCKED;
1476 mapping_set_unevictable(file->f_mapping);
1477 }
1478 if (!lock && (info->flags & VM_LOCKED) && user) {
1479 user_shm_unlock(inode->i_size, user);
1480 info->flags &= ~VM_LOCKED;
1481 mapping_clear_unevictable(file->f_mapping);
1482 }
1483 retval = 0;
1484
1485out_nomem:
1486 spin_unlock(&info->lock);
1487 return retval;
1488}
1489
1490static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1491{
1492 file_accessed(file);
1493 vma->vm_ops = &shmem_vm_ops;
1494 return 0;
1495}
1496
1497static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
1498 umode_t mode, dev_t dev, unsigned long flags)
1499{
1500 struct inode *inode;
1501 struct shmem_inode_info *info;
1502 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1503
1504 if (shmem_reserve_inode(sb))
1505 return NULL;
1506
1507 inode = new_inode(sb);
1508 if (inode) {
1509 inode->i_ino = get_next_ino();
1510 inode_init_owner(inode, dir, mode);
1511 inode->i_blocks = 0;
1512 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1513 inode->i_generation = get_seconds();
1514 info = SHMEM_I(inode);
1515 memset(info, 0, (char *)inode - (char *)info);
1516 spin_lock_init(&info->lock);
1517 info->seals = F_SEAL_SEAL;
1518 info->flags = flags & VM_NORESERVE;
1519 INIT_LIST_HEAD(&info->swaplist);
1520 simple_xattrs_init(&info->xattrs);
1521 cache_no_acl(inode);
1522
1523 switch (mode & S_IFMT) {
1524 default:
1525 inode->i_op = &shmem_special_inode_operations;
1526 init_special_inode(inode, mode, dev);
1527 break;
1528 case S_IFREG:
1529 inode->i_mapping->a_ops = &shmem_aops;
1530 inode->i_op = &shmem_inode_operations;
1531 inode->i_fop = &shmem_file_operations;
1532 mpol_shared_policy_init(&info->policy,
1533 shmem_get_sbmpol(sbinfo));
1534 break;
1535 case S_IFDIR:
1536 inc_nlink(inode);
1537 /* Some things misbehave if size == 0 on a directory */
1538 inode->i_size = 2 * BOGO_DIRENT_SIZE;
1539 inode->i_op = &shmem_dir_inode_operations;
1540 inode->i_fop = &simple_dir_operations;
1541 break;
1542 case S_IFLNK:
1543 /*
1544 * Must not load anything in the rbtree,
1545 * mpol_free_shared_policy will not be called.
1546 */
1547 mpol_shared_policy_init(&info->policy, NULL);
1548 break;
1549 }
1550 } else
1551 shmem_free_inode(sb);
1552 return inode;
1553}
1554
1555bool shmem_mapping(struct address_space *mapping)
1556{
1557 if (!mapping->host)
1558 return false;
1559
1560 return mapping->host->i_sb->s_op == &shmem_ops;
1561}
1562
1563#ifdef CONFIG_TMPFS
1564static const struct inode_operations shmem_symlink_inode_operations;
1565static const struct inode_operations shmem_short_symlink_operations;
1566
1567#ifdef CONFIG_TMPFS_XATTR
1568static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
1569#else
1570#define shmem_initxattrs NULL
1571#endif
1572
1573static int
1574shmem_write_begin(struct file *file, struct address_space *mapping,
1575 loff_t pos, unsigned len, unsigned flags,
1576 struct page **pagep, void **fsdata)
1577{
1578 struct inode *inode = mapping->host;
1579 struct shmem_inode_info *info = SHMEM_I(inode);
1580 pgoff_t index = pos >> PAGE_SHIFT;
1581
1582 /* i_mutex is held by caller */
1583 if (unlikely(info->seals)) {
1584 if (info->seals & F_SEAL_WRITE)
1585 return -EPERM;
1586 if ((info->seals & F_SEAL_GROW) && pos + len > inode->i_size)
1587 return -EPERM;
1588 }
1589
1590 return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1591}
1592
1593static int
1594shmem_write_end(struct file *file, struct address_space *mapping,
1595 loff_t pos, unsigned len, unsigned copied,
1596 struct page *page, void *fsdata)
1597{
1598 struct inode *inode = mapping->host;
1599
1600 if (pos + copied > inode->i_size)
1601 i_size_write(inode, pos + copied);
1602
1603 if (!PageUptodate(page)) {
1604 if (copied < PAGE_SIZE) {
1605 unsigned from = pos & (PAGE_SIZE - 1);
1606 zero_user_segments(page, 0, from,
1607 from + copied, PAGE_SIZE);
1608 }
1609 SetPageUptodate(page);
1610 }
1611 set_page_dirty(page);
1612 unlock_page(page);
1613 put_page(page);
1614
1615 return copied;
1616}
1617
1618static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
1619{
1620 struct file *file = iocb->ki_filp;
1621 struct inode *inode = file_inode(file);
1622 struct address_space *mapping = inode->i_mapping;
1623 pgoff_t index;
1624 unsigned long offset;
1625 enum sgp_type sgp = SGP_READ;
1626 int error = 0;
1627 ssize_t retval = 0;
1628 loff_t *ppos = &iocb->ki_pos;
1629
1630 /*
1631 * Might this read be for a stacking filesystem? Then when reading
1632 * holes of a sparse file, we actually need to allocate those pages,
1633 * and even mark them dirty, so it cannot exceed the max_blocks limit.
1634 */
1635 if (!iter_is_iovec(to))
1636 sgp = SGP_DIRTY;
1637
1638 index = *ppos >> PAGE_SHIFT;
1639 offset = *ppos & ~PAGE_MASK;
1640
1641 for (;;) {
1642 struct page *page = NULL;
1643 pgoff_t end_index;
1644 unsigned long nr, ret;
1645 loff_t i_size = i_size_read(inode);
1646
1647 end_index = i_size >> PAGE_SHIFT;
1648 if (index > end_index)
1649 break;
1650 if (index == end_index) {
1651 nr = i_size & ~PAGE_MASK;
1652 if (nr <= offset)
1653 break;
1654 }
1655
1656 error = shmem_getpage(inode, index, &page, sgp, NULL);
1657 if (error) {
1658 if (error == -EINVAL)
1659 error = 0;
1660 break;
1661 }
1662 if (page)
1663 unlock_page(page);
1664
1665 /*
1666 * We must evaluate after, since reads (unlike writes)
1667 * are called without i_mutex protection against truncate
1668 */
1669 nr = PAGE_SIZE;
1670 i_size = i_size_read(inode);
1671 end_index = i_size >> PAGE_SHIFT;
1672 if (index == end_index) {
1673 nr = i_size & ~PAGE_MASK;
1674 if (nr <= offset) {
1675 if (page)
1676 put_page(page);
1677 break;
1678 }
1679 }
1680 nr -= offset;
1681
1682 if (page) {
1683 /*
1684 * If users can be writing to this page using arbitrary
1685 * virtual addresses, take care about potential aliasing
1686 * before reading the page on the kernel side.
1687 */
1688 if (mapping_writably_mapped(mapping))
1689 flush_dcache_page(page);
1690 /*
1691 * Mark the page accessed if we read the beginning.
1692 */
1693 if (!offset)
1694 mark_page_accessed(page);
1695 } else {
1696 page = ZERO_PAGE(0);
1697 get_page(page);
1698 }
1699
1700 /*
1701 * Ok, we have the page, and it's up-to-date, so
1702 * now we can copy it to user space...
1703 */
1704 ret = copy_page_to_iter(page, offset, nr, to);
1705 retval += ret;
1706 offset += ret;
1707 index += offset >> PAGE_SHIFT;
1708 offset &= ~PAGE_MASK;
1709
1710 put_page(page);
1711 if (!iov_iter_count(to))
1712 break;
1713 if (ret < nr) {
1714 error = -EFAULT;
1715 break;
1716 }
1717 cond_resched();
1718 }
1719
1720 *ppos = ((loff_t) index << PAGE_SHIFT) + offset;
1721 file_accessed(file);
1722 return retval ? retval : error;
1723}
1724
1725static ssize_t shmem_file_splice_read(struct file *in, loff_t *ppos,
1726 struct pipe_inode_info *pipe, size_t len,
1727 unsigned int flags)
1728{
1729 struct address_space *mapping = in->f_mapping;
1730 struct inode *inode = mapping->host;
1731 unsigned int loff, nr_pages, req_pages;
1732 struct page *pages[PIPE_DEF_BUFFERS];
1733 struct partial_page partial[PIPE_DEF_BUFFERS];
1734 struct page *page;
1735 pgoff_t index, end_index;
1736 loff_t isize, left;
1737 int error, page_nr;
1738 struct splice_pipe_desc spd = {
1739 .pages = pages,
1740 .partial = partial,
1741 .nr_pages_max = PIPE_DEF_BUFFERS,
1742 .flags = flags,
1743 .ops = &page_cache_pipe_buf_ops,
1744 .spd_release = spd_release_page,
1745 };
1746
1747 isize = i_size_read(inode);
1748 if (unlikely(*ppos >= isize))
1749 return 0;
1750
1751 left = isize - *ppos;
1752 if (unlikely(left < len))
1753 len = left;
1754
1755 if (splice_grow_spd(pipe, &spd))
1756 return -ENOMEM;
1757
1758 index = *ppos >> PAGE_SHIFT;
1759 loff = *ppos & ~PAGE_MASK;
1760 req_pages = (len + loff + PAGE_SIZE - 1) >> PAGE_SHIFT;
1761 nr_pages = min(req_pages, spd.nr_pages_max);
1762
1763 spd.nr_pages = find_get_pages_contig(mapping, index,
1764 nr_pages, spd.pages);
1765 index += spd.nr_pages;
1766 error = 0;
1767
1768 while (spd.nr_pages < nr_pages) {
1769 error = shmem_getpage(inode, index, &page, SGP_CACHE, NULL);
1770 if (error)
1771 break;
1772 unlock_page(page);
1773 spd.pages[spd.nr_pages++] = page;
1774 index++;
1775 }
1776
1777 index = *ppos >> PAGE_SHIFT;
1778 nr_pages = spd.nr_pages;
1779 spd.nr_pages = 0;
1780
1781 for (page_nr = 0; page_nr < nr_pages; page_nr++) {
1782 unsigned int this_len;
1783
1784 if (!len)
1785 break;
1786
1787 this_len = min_t(unsigned long, len, PAGE_SIZE - loff);
1788 page = spd.pages[page_nr];
1789
1790 if (!PageUptodate(page) || page->mapping != mapping) {
1791 error = shmem_getpage(inode, index, &page,
1792 SGP_CACHE, NULL);
1793 if (error)
1794 break;
1795 unlock_page(page);
1796 put_page(spd.pages[page_nr]);
1797 spd.pages[page_nr] = page;
1798 }
1799
1800 isize = i_size_read(inode);
1801 end_index = (isize - 1) >> PAGE_SHIFT;
1802 if (unlikely(!isize || index > end_index))
1803 break;
1804
1805 if (end_index == index) {
1806 unsigned int plen;
1807
1808 plen = ((isize - 1) & ~PAGE_MASK) + 1;
1809 if (plen <= loff)
1810 break;
1811
1812 this_len = min(this_len, plen - loff);
1813 len = this_len;
1814 }
1815
1816 spd.partial[page_nr].offset = loff;
1817 spd.partial[page_nr].len = this_len;
1818 len -= this_len;
1819 loff = 0;
1820 spd.nr_pages++;
1821 index++;
1822 }
1823
1824 while (page_nr < nr_pages)
1825 put_page(spd.pages[page_nr++]);
1826
1827 if (spd.nr_pages)
1828 error = splice_to_pipe(pipe, &spd);
1829
1830 splice_shrink_spd(&spd);
1831
1832 if (error > 0) {
1833 *ppos += error;
1834 file_accessed(in);
1835 }
1836 return error;
1837}
1838
1839/*
1840 * llseek SEEK_DATA or SEEK_HOLE through the radix_tree.
1841 */
1842static pgoff_t shmem_seek_hole_data(struct address_space *mapping,
1843 pgoff_t index, pgoff_t end, int whence)
1844{
1845 struct page *page;
1846 struct pagevec pvec;
1847 pgoff_t indices[PAGEVEC_SIZE];
1848 bool done = false;
1849 int i;
1850
1851 pagevec_init(&pvec, 0);
1852 pvec.nr = 1; /* start small: we may be there already */
1853 while (!done) {
1854 pvec.nr = find_get_entries(mapping, index,
1855 pvec.nr, pvec.pages, indices);
1856 if (!pvec.nr) {
1857 if (whence == SEEK_DATA)
1858 index = end;
1859 break;
1860 }
1861 for (i = 0; i < pvec.nr; i++, index++) {
1862 if (index < indices[i]) {
1863 if (whence == SEEK_HOLE) {
1864 done = true;
1865 break;
1866 }
1867 index = indices[i];
1868 }
1869 page = pvec.pages[i];
1870 if (page && !radix_tree_exceptional_entry(page)) {
1871 if (!PageUptodate(page))
1872 page = NULL;
1873 }
1874 if (index >= end ||
1875 (page && whence == SEEK_DATA) ||
1876 (!page && whence == SEEK_HOLE)) {
1877 done = true;
1878 break;
1879 }
1880 }
1881 pagevec_remove_exceptionals(&pvec);
1882 pagevec_release(&pvec);
1883 pvec.nr = PAGEVEC_SIZE;
1884 cond_resched();
1885 }
1886 return index;
1887}
1888
1889static loff_t shmem_file_llseek(struct file *file, loff_t offset, int whence)
1890{
1891 struct address_space *mapping = file->f_mapping;
1892 struct inode *inode = mapping->host;
1893 pgoff_t start, end;
1894 loff_t new_offset;
1895
1896 if (whence != SEEK_DATA && whence != SEEK_HOLE)
1897 return generic_file_llseek_size(file, offset, whence,
1898 MAX_LFS_FILESIZE, i_size_read(inode));
1899 inode_lock(inode);
1900 /* We're holding i_mutex so we can access i_size directly */
1901
1902 if (offset < 0)
1903 offset = -EINVAL;
1904 else if (offset >= inode->i_size)
1905 offset = -ENXIO;
1906 else {
1907 start = offset >> PAGE_SHIFT;
1908 end = (inode->i_size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1909 new_offset = shmem_seek_hole_data(mapping, start, end, whence);
1910 new_offset <<= PAGE_SHIFT;
1911 if (new_offset > offset) {
1912 if (new_offset < inode->i_size)
1913 offset = new_offset;
1914 else if (whence == SEEK_DATA)
1915 offset = -ENXIO;
1916 else
1917 offset = inode->i_size;
1918 }
1919 }
1920
1921 if (offset >= 0)
1922 offset = vfs_setpos(file, offset, MAX_LFS_FILESIZE);
1923 inode_unlock(inode);
1924 return offset;
1925}
1926
1927/*
1928 * We need a tag: a new tag would expand every radix_tree_node by 8 bytes,
1929 * so reuse a tag which we firmly believe is never set or cleared on shmem.
1930 */
1931#define SHMEM_TAG_PINNED PAGECACHE_TAG_TOWRITE
1932#define LAST_SCAN 4 /* about 150ms max */
1933
1934static void shmem_tag_pins(struct address_space *mapping)
1935{
1936 struct radix_tree_iter iter;
1937 void **slot;
1938 pgoff_t start;
1939 struct page *page;
1940
1941 lru_add_drain();
1942 start = 0;
1943 rcu_read_lock();
1944
1945 radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1946 page = radix_tree_deref_slot(slot);
1947 if (!page || radix_tree_exception(page)) {
1948 if (radix_tree_deref_retry(page)) {
1949 slot = radix_tree_iter_retry(&iter);
1950 continue;
1951 }
1952 } else if (page_count(page) - page_mapcount(page) > 1) {
1953 spin_lock_irq(&mapping->tree_lock);
1954 radix_tree_tag_set(&mapping->page_tree, iter.index,
1955 SHMEM_TAG_PINNED);
1956 spin_unlock_irq(&mapping->tree_lock);
1957 }
1958
1959 if (need_resched()) {
1960 cond_resched_rcu();
1961 slot = radix_tree_iter_next(&iter);
1962 }
1963 }
1964 rcu_read_unlock();
1965}
1966
1967/*
1968 * Setting SEAL_WRITE requires us to verify there's no pending writer. However,
1969 * via get_user_pages(), drivers might have some pending I/O without any active
1970 * user-space mappings (eg., direct-IO, AIO). Therefore, we look at all pages
1971 * and see whether it has an elevated ref-count. If so, we tag them and wait for
1972 * them to be dropped.
1973 * The caller must guarantee that no new user will acquire writable references
1974 * to those pages to avoid races.
1975 */
1976static int shmem_wait_for_pins(struct address_space *mapping)
1977{
1978 struct radix_tree_iter iter;
1979 void **slot;
1980 pgoff_t start;
1981 struct page *page;
1982 int error, scan;
1983
1984 shmem_tag_pins(mapping);
1985
1986 error = 0;
1987 for (scan = 0; scan <= LAST_SCAN; scan++) {
1988 if (!radix_tree_tagged(&mapping->page_tree, SHMEM_TAG_PINNED))
1989 break;
1990
1991 if (!scan)
1992 lru_add_drain_all();
1993 else if (schedule_timeout_killable((HZ << scan) / 200))
1994 scan = LAST_SCAN;
1995
1996 start = 0;
1997 rcu_read_lock();
1998 radix_tree_for_each_tagged(slot, &mapping->page_tree, &iter,
1999 start, SHMEM_TAG_PINNED) {
2000
2001 page = radix_tree_deref_slot(slot);
2002 if (radix_tree_exception(page)) {
2003 if (radix_tree_deref_retry(page)) {
2004 slot = radix_tree_iter_retry(&iter);
2005 continue;
2006 }
2007
2008 page = NULL;
2009 }
2010
2011 if (page &&
2012 page_count(page) - page_mapcount(page) != 1) {
2013 if (scan < LAST_SCAN)
2014 goto continue_resched;
2015
2016 /*
2017 * On the last scan, we clean up all those tags
2018 * we inserted; but make a note that we still
2019 * found pages pinned.
2020 */
2021 error = -EBUSY;
2022 }
2023
2024 spin_lock_irq(&mapping->tree_lock);
2025 radix_tree_tag_clear(&mapping->page_tree,
2026 iter.index, SHMEM_TAG_PINNED);
2027 spin_unlock_irq(&mapping->tree_lock);
2028continue_resched:
2029 if (need_resched()) {
2030 cond_resched_rcu();
2031 slot = radix_tree_iter_next(&iter);
2032 }
2033 }
2034 rcu_read_unlock();
2035 }
2036
2037 return error;
2038}
2039
2040#define F_ALL_SEALS (F_SEAL_SEAL | \
2041 F_SEAL_SHRINK | \
2042 F_SEAL_GROW | \
2043 F_SEAL_WRITE)
2044
2045int shmem_add_seals(struct file *file, unsigned int seals)
2046{
2047 struct inode *inode = file_inode(file);
2048 struct shmem_inode_info *info = SHMEM_I(inode);
2049 int error;
2050
2051 /*
2052 * SEALING
2053 * Sealing allows multiple parties to share a shmem-file but restrict
2054 * access to a specific subset of file operations. Seals can only be
2055 * added, but never removed. This way, mutually untrusted parties can
2056 * share common memory regions with a well-defined policy. A malicious
2057 * peer can thus never perform unwanted operations on a shared object.
2058 *
2059 * Seals are only supported on special shmem-files and always affect
2060 * the whole underlying inode. Once a seal is set, it may prevent some
2061 * kinds of access to the file. Currently, the following seals are
2062 * defined:
2063 * SEAL_SEAL: Prevent further seals from being set on this file
2064 * SEAL_SHRINK: Prevent the file from shrinking
2065 * SEAL_GROW: Prevent the file from growing
2066 * SEAL_WRITE: Prevent write access to the file
2067 *
2068 * As we don't require any trust relationship between two parties, we
2069 * must prevent seals from being removed. Therefore, sealing a file
2070 * only adds a given set of seals to the file, it never touches
2071 * existing seals. Furthermore, the "setting seals"-operation can be
2072 * sealed itself, which basically prevents any further seal from being
2073 * added.
2074 *
2075 * Semantics of sealing are only defined on volatile files. Only
2076 * anonymous shmem files support sealing. More importantly, seals are
2077 * never written to disk. Therefore, there's no plan to support it on
2078 * other file types.
2079 */
2080
2081 if (file->f_op != &shmem_file_operations)
2082 return -EINVAL;
2083 if (!(file->f_mode & FMODE_WRITE))
2084 return -EPERM;
2085 if (seals & ~(unsigned int)F_ALL_SEALS)
2086 return -EINVAL;
2087
2088 inode_lock(inode);
2089
2090 if (info->seals & F_SEAL_SEAL) {
2091 error = -EPERM;
2092 goto unlock;
2093 }
2094
2095 if ((seals & F_SEAL_WRITE) && !(info->seals & F_SEAL_WRITE)) {
2096 error = mapping_deny_writable(file->f_mapping);
2097 if (error)
2098 goto unlock;
2099
2100 error = shmem_wait_for_pins(file->f_mapping);
2101 if (error) {
2102 mapping_allow_writable(file->f_mapping);
2103 goto unlock;
2104 }
2105 }
2106
2107 info->seals |= seals;
2108 error = 0;
2109
2110unlock:
2111 inode_unlock(inode);
2112 return error;
2113}
2114EXPORT_SYMBOL_GPL(shmem_add_seals);
2115
2116int shmem_get_seals(struct file *file)
2117{
2118 if (file->f_op != &shmem_file_operations)
2119 return -EINVAL;
2120
2121 return SHMEM_I(file_inode(file))->seals;
2122}
2123EXPORT_SYMBOL_GPL(shmem_get_seals);
2124
2125long shmem_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
2126{
2127 long error;
2128
2129 switch (cmd) {
2130 case F_ADD_SEALS:
2131 /* disallow upper 32bit */
2132 if (arg > UINT_MAX)
2133 return -EINVAL;
2134
2135 error = shmem_add_seals(file, arg);
2136 break;
2137 case F_GET_SEALS:
2138 error = shmem_get_seals(file);
2139 break;
2140 default:
2141 error = -EINVAL;
2142 break;
2143 }
2144
2145 return error;
2146}
2147
2148static long shmem_fallocate(struct file *file, int mode, loff_t offset,
2149 loff_t len)
2150{
2151 struct inode *inode = file_inode(file);
2152 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
2153 struct shmem_inode_info *info = SHMEM_I(inode);
2154 struct shmem_falloc shmem_falloc;
2155 pgoff_t start, index, end;
2156 int error;
2157
2158 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2159 return -EOPNOTSUPP;
2160
2161 inode_lock(inode);
2162
2163 if (mode & FALLOC_FL_PUNCH_HOLE) {
2164 struct address_space *mapping = file->f_mapping;
2165 loff_t unmap_start = round_up(offset, PAGE_SIZE);
2166 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
2167 DECLARE_WAIT_QUEUE_HEAD_ONSTACK(shmem_falloc_waitq);
2168
2169 /* protected by i_mutex */
2170 if (info->seals & F_SEAL_WRITE) {
2171 error = -EPERM;
2172 goto out;
2173 }
2174
2175 shmem_falloc.waitq = &shmem_falloc_waitq;
2176 shmem_falloc.start = unmap_start >> PAGE_SHIFT;
2177 shmem_falloc.next = (unmap_end + 1) >> PAGE_SHIFT;
2178 spin_lock(&inode->i_lock);
2179 inode->i_private = &shmem_falloc;
2180 spin_unlock(&inode->i_lock);
2181
2182 if ((u64)unmap_end > (u64)unmap_start)
2183 unmap_mapping_range(mapping, unmap_start,
2184 1 + unmap_end - unmap_start, 0);
2185 shmem_truncate_range(inode, offset, offset + len - 1);
2186 /* No need to unmap again: hole-punching leaves COWed pages */
2187
2188 spin_lock(&inode->i_lock);
2189 inode->i_private = NULL;
2190 wake_up_all(&shmem_falloc_waitq);
2191 spin_unlock(&inode->i_lock);
2192 error = 0;
2193 goto out;
2194 }
2195
2196 /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
2197 error = inode_newsize_ok(inode, offset + len);
2198 if (error)
2199 goto out;
2200
2201 if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
2202 error = -EPERM;
2203 goto out;
2204 }
2205
2206 start = offset >> PAGE_SHIFT;
2207 end = (offset + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
2208 /* Try to avoid a swapstorm if len is impossible to satisfy */
2209 if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
2210 error = -ENOSPC;
2211 goto out;
2212 }
2213
2214 shmem_falloc.waitq = NULL;
2215 shmem_falloc.start = start;
2216 shmem_falloc.next = start;
2217 shmem_falloc.nr_falloced = 0;
2218 shmem_falloc.nr_unswapped = 0;
2219 spin_lock(&inode->i_lock);
2220 inode->i_private = &shmem_falloc;
2221 spin_unlock(&inode->i_lock);
2222
2223 for (index = start; index < end; index++) {
2224 struct page *page;
2225
2226 /*
2227 * Good, the fallocate(2) manpage permits EINTR: we may have
2228 * been interrupted because we are using up too much memory.
2229 */
2230 if (signal_pending(current))
2231 error = -EINTR;
2232 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
2233 error = -ENOMEM;
2234 else
2235 error = shmem_getpage(inode, index, &page, SGP_FALLOC,
2236 NULL);
2237 if (error) {
2238 /* Remove the !PageUptodate pages we added */
2239 shmem_undo_range(inode,
2240 (loff_t)start << PAGE_SHIFT,
2241 (loff_t)index << PAGE_SHIFT, true);
2242 goto undone;
2243 }
2244
2245 /*
2246 * Inform shmem_writepage() how far we have reached.
2247 * No need for lock or barrier: we have the page lock.
2248 */
2249 shmem_falloc.next++;
2250 if (!PageUptodate(page))
2251 shmem_falloc.nr_falloced++;
2252
2253 /*
2254 * If !PageUptodate, leave it that way so that freeable pages
2255 * can be recognized if we need to rollback on error later.
2256 * But set_page_dirty so that memory pressure will swap rather
2257 * than free the pages we are allocating (and SGP_CACHE pages
2258 * might still be clean: we now need to mark those dirty too).
2259 */
2260 set_page_dirty(page);
2261 unlock_page(page);
2262 put_page(page);
2263 cond_resched();
2264 }
2265
2266 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
2267 i_size_write(inode, offset + len);
2268 inode->i_ctime = CURRENT_TIME;
2269undone:
2270 spin_lock(&inode->i_lock);
2271 inode->i_private = NULL;
2272 spin_unlock(&inode->i_lock);
2273out:
2274 inode_unlock(inode);
2275 return error;
2276}
2277
2278static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
2279{
2280 struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
2281
2282 buf->f_type = TMPFS_MAGIC;
2283 buf->f_bsize = PAGE_SIZE;
2284 buf->f_namelen = NAME_MAX;
2285 if (sbinfo->max_blocks) {
2286 buf->f_blocks = sbinfo->max_blocks;
2287 buf->f_bavail =
2288 buf->f_bfree = sbinfo->max_blocks -
2289 percpu_counter_sum(&sbinfo->used_blocks);
2290 }
2291 if (sbinfo->max_inodes) {
2292 buf->f_files = sbinfo->max_inodes;
2293 buf->f_ffree = sbinfo->free_inodes;
2294 }
2295 /* else leave those fields 0 like simple_statfs */
2296 return 0;
2297}
2298
2299/*
2300 * File creation. Allocate an inode, and we're done..
2301 */
2302static int
2303shmem_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
2304{
2305 struct inode *inode;
2306 int error = -ENOSPC;
2307
2308 inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
2309 if (inode) {
2310 error = simple_acl_create(dir, inode);
2311 if (error)
2312 goto out_iput;
2313 error = security_inode_init_security(inode, dir,
2314 &dentry->d_name,
2315 shmem_initxattrs, NULL);
2316 if (error && error != -EOPNOTSUPP)
2317 goto out_iput;
2318
2319 error = 0;
2320 dir->i_size += BOGO_DIRENT_SIZE;
2321 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2322 d_instantiate(dentry, inode);
2323 dget(dentry); /* Extra count - pin the dentry in core */
2324 }
2325 return error;
2326out_iput:
2327 iput(inode);
2328 return error;
2329}
2330
2331static int
2332shmem_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
2333{
2334 struct inode *inode;
2335 int error = -ENOSPC;
2336
2337 inode = shmem_get_inode(dir->i_sb, dir, mode, 0, VM_NORESERVE);
2338 if (inode) {
2339 error = security_inode_init_security(inode, dir,
2340 NULL,
2341 shmem_initxattrs, NULL);
2342 if (error && error != -EOPNOTSUPP)
2343 goto out_iput;
2344 error = simple_acl_create(dir, inode);
2345 if (error)
2346 goto out_iput;
2347 d_tmpfile(dentry, inode);
2348 }
2349 return error;
2350out_iput:
2351 iput(inode);
2352 return error;
2353}
2354
2355static int shmem_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
2356{
2357 int error;
2358
2359 if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
2360 return error;
2361 inc_nlink(dir);
2362 return 0;
2363}
2364
2365static int shmem_create(struct inode *dir, struct dentry *dentry, umode_t mode,
2366 bool excl)
2367{
2368 return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
2369}
2370
2371/*
2372 * Link a file..
2373 */
2374static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
2375{
2376 struct inode *inode = d_inode(old_dentry);
2377 int ret;
2378
2379 /*
2380 * No ordinary (disk based) filesystem counts links as inodes;
2381 * but each new link needs a new dentry, pinning lowmem, and
2382 * tmpfs dentries cannot be pruned until they are unlinked.
2383 */
2384 ret = shmem_reserve_inode(inode->i_sb);
2385 if (ret)
2386 goto out;
2387
2388 dir->i_size += BOGO_DIRENT_SIZE;
2389 inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2390 inc_nlink(inode);
2391 ihold(inode); /* New dentry reference */
2392 dget(dentry); /* Extra pinning count for the created dentry */
2393 d_instantiate(dentry, inode);
2394out:
2395 return ret;
2396}
2397
2398static int shmem_unlink(struct inode *dir, struct dentry *dentry)
2399{
2400 struct inode *inode = d_inode(dentry);
2401
2402 if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
2403 shmem_free_inode(inode->i_sb);
2404
2405 dir->i_size -= BOGO_DIRENT_SIZE;
2406 inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2407 drop_nlink(inode);
2408 dput(dentry); /* Undo the count from "create" - this does all the work */
2409 return 0;
2410}
2411
2412static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
2413{
2414 if (!simple_empty(dentry))
2415 return -ENOTEMPTY;
2416
2417 drop_nlink(d_inode(dentry));
2418 drop_nlink(dir);
2419 return shmem_unlink(dir, dentry);
2420}
2421
2422static int shmem_exchange(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
2423{
2424 bool old_is_dir = d_is_dir(old_dentry);
2425 bool new_is_dir = d_is_dir(new_dentry);
2426
2427 if (old_dir != new_dir && old_is_dir != new_is_dir) {
2428 if (old_is_dir) {
2429 drop_nlink(old_dir);
2430 inc_nlink(new_dir);
2431 } else {
2432 drop_nlink(new_dir);
2433 inc_nlink(old_dir);
2434 }
2435 }
2436 old_dir->i_ctime = old_dir->i_mtime =
2437 new_dir->i_ctime = new_dir->i_mtime =
2438 d_inode(old_dentry)->i_ctime =
2439 d_inode(new_dentry)->i_ctime = CURRENT_TIME;
2440
2441 return 0;
2442}
2443
2444static int shmem_whiteout(struct inode *old_dir, struct dentry *old_dentry)
2445{
2446 struct dentry *whiteout;
2447 int error;
2448
2449 whiteout = d_alloc(old_dentry->d_parent, &old_dentry->d_name);
2450 if (!whiteout)
2451 return -ENOMEM;
2452
2453 error = shmem_mknod(old_dir, whiteout,
2454 S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
2455 dput(whiteout);
2456 if (error)
2457 return error;
2458
2459 /*
2460 * Cheat and hash the whiteout while the old dentry is still in
2461 * place, instead of playing games with FS_RENAME_DOES_D_MOVE.
2462 *
2463 * d_lookup() will consistently find one of them at this point,
2464 * not sure which one, but that isn't even important.
2465 */
2466 d_rehash(whiteout);
2467 return 0;
2468}
2469
2470/*
2471 * The VFS layer already does all the dentry stuff for rename,
2472 * we just have to decrement the usage count for the target if
2473 * it exists so that the VFS layer correctly free's it when it
2474 * gets overwritten.
2475 */
2476static int shmem_rename2(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry, unsigned int flags)
2477{
2478 struct inode *inode = d_inode(old_dentry);
2479 int they_are_dirs = S_ISDIR(inode->i_mode);
2480
2481 if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
2482 return -EINVAL;
2483
2484 if (flags & RENAME_EXCHANGE)
2485 return shmem_exchange(old_dir, old_dentry, new_dir, new_dentry);
2486
2487 if (!simple_empty(new_dentry))
2488 return -ENOTEMPTY;
2489
2490 if (flags & RENAME_WHITEOUT) {
2491 int error;
2492
2493 error = shmem_whiteout(old_dir, old_dentry);
2494 if (error)
2495 return error;
2496 }
2497
2498 if (d_really_is_positive(new_dentry)) {
2499 (void) shmem_unlink(new_dir, new_dentry);
2500 if (they_are_dirs) {
2501 drop_nlink(d_inode(new_dentry));
2502 drop_nlink(old_dir);
2503 }
2504 } else if (they_are_dirs) {
2505 drop_nlink(old_dir);
2506 inc_nlink(new_dir);
2507 }
2508
2509 old_dir->i_size -= BOGO_DIRENT_SIZE;
2510 new_dir->i_size += BOGO_DIRENT_SIZE;
2511 old_dir->i_ctime = old_dir->i_mtime =
2512 new_dir->i_ctime = new_dir->i_mtime =
2513 inode->i_ctime = CURRENT_TIME;
2514 return 0;
2515}
2516
2517static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
2518{
2519 int error;
2520 int len;
2521 struct inode *inode;
2522 struct page *page;
2523 struct shmem_inode_info *info;
2524
2525 len = strlen(symname) + 1;
2526 if (len > PAGE_SIZE)
2527 return -ENAMETOOLONG;
2528
2529 inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0, VM_NORESERVE);
2530 if (!inode)
2531 return -ENOSPC;
2532
2533 error = security_inode_init_security(inode, dir, &dentry->d_name,
2534 shmem_initxattrs, NULL);
2535 if (error) {
2536 if (error != -EOPNOTSUPP) {
2537 iput(inode);
2538 return error;
2539 }
2540 error = 0;
2541 }
2542
2543 info = SHMEM_I(inode);
2544 inode->i_size = len-1;
2545 if (len <= SHORT_SYMLINK_LEN) {
2546 inode->i_link = kmemdup(symname, len, GFP_KERNEL);
2547 if (!inode->i_link) {
2548 iput(inode);
2549 return -ENOMEM;
2550 }
2551 inode->i_op = &shmem_short_symlink_operations;
2552 } else {
2553 inode_nohighmem(inode);
2554 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
2555 if (error) {
2556 iput(inode);
2557 return error;
2558 }
2559 inode->i_mapping->a_ops = &shmem_aops;
2560 inode->i_op = &shmem_symlink_inode_operations;
2561 memcpy(page_address(page), symname, len);
2562 SetPageUptodate(page);
2563 set_page_dirty(page);
2564 unlock_page(page);
2565 put_page(page);
2566 }
2567 dir->i_size += BOGO_DIRENT_SIZE;
2568 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2569 d_instantiate(dentry, inode);
2570 dget(dentry);
2571 return 0;
2572}
2573
2574static void shmem_put_link(void *arg)
2575{
2576 mark_page_accessed(arg);
2577 put_page(arg);
2578}
2579
2580static const char *shmem_get_link(struct dentry *dentry,
2581 struct inode *inode,
2582 struct delayed_call *done)
2583{
2584 struct page *page = NULL;
2585 int error;
2586 if (!dentry) {
2587 page = find_get_page(inode->i_mapping, 0);
2588 if (!page)
2589 return ERR_PTR(-ECHILD);
2590 if (!PageUptodate(page)) {
2591 put_page(page);
2592 return ERR_PTR(-ECHILD);
2593 }
2594 } else {
2595 error = shmem_getpage(inode, 0, &page, SGP_READ, NULL);
2596 if (error)
2597 return ERR_PTR(error);
2598 unlock_page(page);
2599 }
2600 set_delayed_call(done, shmem_put_link, page);
2601 return page_address(page);
2602}
2603
2604#ifdef CONFIG_TMPFS_XATTR
2605/*
2606 * Superblocks without xattr inode operations may get some security.* xattr
2607 * support from the LSM "for free". As soon as we have any other xattrs
2608 * like ACLs, we also need to implement the security.* handlers at
2609 * filesystem level, though.
2610 */
2611
2612/*
2613 * Callback for security_inode_init_security() for acquiring xattrs.
2614 */
2615static int shmem_initxattrs(struct inode *inode,
2616 const struct xattr *xattr_array,
2617 void *fs_info)
2618{
2619 struct shmem_inode_info *info = SHMEM_I(inode);
2620 const struct xattr *xattr;
2621 struct simple_xattr *new_xattr;
2622 size_t len;
2623
2624 for (xattr = xattr_array; xattr->name != NULL; xattr++) {
2625 new_xattr = simple_xattr_alloc(xattr->value, xattr->value_len);
2626 if (!new_xattr)
2627 return -ENOMEM;
2628
2629 len = strlen(xattr->name) + 1;
2630 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
2631 GFP_KERNEL);
2632 if (!new_xattr->name) {
2633 kfree(new_xattr);
2634 return -ENOMEM;
2635 }
2636
2637 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
2638 XATTR_SECURITY_PREFIX_LEN);
2639 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
2640 xattr->name, len);
2641
2642 simple_xattr_list_add(&info->xattrs, new_xattr);
2643 }
2644
2645 return 0;
2646}
2647
2648static int shmem_xattr_handler_get(const struct xattr_handler *handler,
2649 struct dentry *dentry, const char *name,
2650 void *buffer, size_t size)
2651{
2652 struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
2653
2654 name = xattr_full_name(handler, name);
2655 return simple_xattr_get(&info->xattrs, name, buffer, size);
2656}
2657
2658static int shmem_xattr_handler_set(const struct xattr_handler *handler,
2659 struct dentry *dentry, const char *name,
2660 const void *value, size_t size, int flags)
2661{
2662 struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
2663
2664 name = xattr_full_name(handler, name);
2665 return simple_xattr_set(&info->xattrs, name, value, size, flags);
2666}
2667
2668static const struct xattr_handler shmem_security_xattr_handler = {
2669 .prefix = XATTR_SECURITY_PREFIX,
2670 .get = shmem_xattr_handler_get,
2671 .set = shmem_xattr_handler_set,
2672};
2673
2674static const struct xattr_handler shmem_trusted_xattr_handler = {
2675 .prefix = XATTR_TRUSTED_PREFIX,
2676 .get = shmem_xattr_handler_get,
2677 .set = shmem_xattr_handler_set,
2678};
2679
2680static const struct xattr_handler *shmem_xattr_handlers[] = {
2681#ifdef CONFIG_TMPFS_POSIX_ACL
2682 &posix_acl_access_xattr_handler,
2683 &posix_acl_default_xattr_handler,
2684#endif
2685 &shmem_security_xattr_handler,
2686 &shmem_trusted_xattr_handler,
2687 NULL
2688};
2689
2690static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
2691{
2692 struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
2693 return simple_xattr_list(d_inode(dentry), &info->xattrs, buffer, size);
2694}
2695#endif /* CONFIG_TMPFS_XATTR */
2696
2697static const struct inode_operations shmem_short_symlink_operations = {
2698 .readlink = generic_readlink,
2699 .get_link = simple_get_link,
2700#ifdef CONFIG_TMPFS_XATTR
2701 .setxattr = generic_setxattr,
2702 .getxattr = generic_getxattr,
2703 .listxattr = shmem_listxattr,
2704 .removexattr = generic_removexattr,
2705#endif
2706};
2707
2708static const struct inode_operations shmem_symlink_inode_operations = {
2709 .readlink = generic_readlink,
2710 .get_link = shmem_get_link,
2711#ifdef CONFIG_TMPFS_XATTR
2712 .setxattr = generic_setxattr,
2713 .getxattr = generic_getxattr,
2714 .listxattr = shmem_listxattr,
2715 .removexattr = generic_removexattr,
2716#endif
2717};
2718
2719static struct dentry *shmem_get_parent(struct dentry *child)
2720{
2721 return ERR_PTR(-ESTALE);
2722}
2723
2724static int shmem_match(struct inode *ino, void *vfh)
2725{
2726 __u32 *fh = vfh;
2727 __u64 inum = fh[2];
2728 inum = (inum << 32) | fh[1];
2729 return ino->i_ino == inum && fh[0] == ino->i_generation;
2730}
2731
2732static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
2733 struct fid *fid, int fh_len, int fh_type)
2734{
2735 struct inode *inode;
2736 struct dentry *dentry = NULL;
2737 u64 inum;
2738
2739 if (fh_len < 3)
2740 return NULL;
2741
2742 inum = fid->raw[2];
2743 inum = (inum << 32) | fid->raw[1];
2744
2745 inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
2746 shmem_match, fid->raw);
2747 if (inode) {
2748 dentry = d_find_alias(inode);
2749 iput(inode);
2750 }
2751
2752 return dentry;
2753}
2754
2755static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
2756 struct inode *parent)
2757{
2758 if (*len < 3) {
2759 *len = 3;
2760 return FILEID_INVALID;
2761 }
2762
2763 if (inode_unhashed(inode)) {
2764 /* Unfortunately insert_inode_hash is not idempotent,
2765 * so as we hash inodes here rather than at creation
2766 * time, we need a lock to ensure we only try
2767 * to do it once
2768 */
2769 static DEFINE_SPINLOCK(lock);
2770 spin_lock(&lock);
2771 if (inode_unhashed(inode))
2772 __insert_inode_hash(inode,
2773 inode->i_ino + inode->i_generation);
2774 spin_unlock(&lock);
2775 }
2776
2777 fh[0] = inode->i_generation;
2778 fh[1] = inode->i_ino;
2779 fh[2] = ((__u64)inode->i_ino) >> 32;
2780
2781 *len = 3;
2782 return 1;
2783}
2784
2785static const struct export_operations shmem_export_ops = {
2786 .get_parent = shmem_get_parent,
2787 .encode_fh = shmem_encode_fh,
2788 .fh_to_dentry = shmem_fh_to_dentry,
2789};
2790
2791static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
2792 bool remount)
2793{
2794 char *this_char, *value, *rest;
2795 struct mempolicy *mpol = NULL;
2796 uid_t uid;
2797 gid_t gid;
2798
2799 while (options != NULL) {
2800 this_char = options;
2801 for (;;) {
2802 /*
2803 * NUL-terminate this option: unfortunately,
2804 * mount options form a comma-separated list,
2805 * but mpol's nodelist may also contain commas.
2806 */
2807 options = strchr(options, ',');
2808 if (options == NULL)
2809 break;
2810 options++;
2811 if (!isdigit(*options)) {
2812 options[-1] = '\0';
2813 break;
2814 }
2815 }
2816 if (!*this_char)
2817 continue;
2818 if ((value = strchr(this_char,'=')) != NULL) {
2819 *value++ = 0;
2820 } else {
2821 pr_err("tmpfs: No value for mount option '%s'\n",
2822 this_char);
2823 goto error;
2824 }
2825
2826 if (!strcmp(this_char,"size")) {
2827 unsigned long long size;
2828 size = memparse(value,&rest);
2829 if (*rest == '%') {
2830 size <<= PAGE_SHIFT;
2831 size *= totalram_pages;
2832 do_div(size, 100);
2833 rest++;
2834 }
2835 if (*rest)
2836 goto bad_val;
2837 sbinfo->max_blocks =
2838 DIV_ROUND_UP(size, PAGE_SIZE);
2839 } else if (!strcmp(this_char,"nr_blocks")) {
2840 sbinfo->max_blocks = memparse(value, &rest);
2841 if (*rest)
2842 goto bad_val;
2843 } else if (!strcmp(this_char,"nr_inodes")) {
2844 sbinfo->max_inodes = memparse(value, &rest);
2845 if (*rest)
2846 goto bad_val;
2847 } else if (!strcmp(this_char,"mode")) {
2848 if (remount)
2849 continue;
2850 sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
2851 if (*rest)
2852 goto bad_val;
2853 } else if (!strcmp(this_char,"uid")) {
2854 if (remount)
2855 continue;
2856 uid = simple_strtoul(value, &rest, 0);
2857 if (*rest)
2858 goto bad_val;
2859 sbinfo->uid = make_kuid(current_user_ns(), uid);
2860 if (!uid_valid(sbinfo->uid))
2861 goto bad_val;
2862 } else if (!strcmp(this_char,"gid")) {
2863 if (remount)
2864 continue;
2865 gid = simple_strtoul(value, &rest, 0);
2866 if (*rest)
2867 goto bad_val;
2868 sbinfo->gid = make_kgid(current_user_ns(), gid);
2869 if (!gid_valid(sbinfo->gid))
2870 goto bad_val;
2871 } else if (!strcmp(this_char,"mpol")) {
2872 mpol_put(mpol);
2873 mpol = NULL;
2874 if (mpol_parse_str(value, &mpol))
2875 goto bad_val;
2876 } else {
2877 pr_err("tmpfs: Bad mount option %s\n", this_char);
2878 goto error;
2879 }
2880 }
2881 sbinfo->mpol = mpol;
2882 return 0;
2883
2884bad_val:
2885 pr_err("tmpfs: Bad value '%s' for mount option '%s'\n",
2886 value, this_char);
2887error:
2888 mpol_put(mpol);
2889 return 1;
2890
2891}
2892
2893static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2894{
2895 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2896 struct shmem_sb_info config = *sbinfo;
2897 unsigned long inodes;
2898 int error = -EINVAL;
2899
2900 config.mpol = NULL;
2901 if (shmem_parse_options(data, &config, true))
2902 return error;
2903
2904 spin_lock(&sbinfo->stat_lock);
2905 inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2906 if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
2907 goto out;
2908 if (config.max_inodes < inodes)
2909 goto out;
2910 /*
2911 * Those tests disallow limited->unlimited while any are in use;
2912 * but we must separately disallow unlimited->limited, because
2913 * in that case we have no record of how much is already in use.
2914 */
2915 if (config.max_blocks && !sbinfo->max_blocks)
2916 goto out;
2917 if (config.max_inodes && !sbinfo->max_inodes)
2918 goto out;
2919
2920 error = 0;
2921 sbinfo->max_blocks = config.max_blocks;
2922 sbinfo->max_inodes = config.max_inodes;
2923 sbinfo->free_inodes = config.max_inodes - inodes;
2924
2925 /*
2926 * Preserve previous mempolicy unless mpol remount option was specified.
2927 */
2928 if (config.mpol) {
2929 mpol_put(sbinfo->mpol);
2930 sbinfo->mpol = config.mpol; /* transfers initial ref */
2931 }
2932out:
2933 spin_unlock(&sbinfo->stat_lock);
2934 return error;
2935}
2936
2937static int shmem_show_options(struct seq_file *seq, struct dentry *root)
2938{
2939 struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
2940
2941 if (sbinfo->max_blocks != shmem_default_max_blocks())
2942 seq_printf(seq, ",size=%luk",
2943 sbinfo->max_blocks << (PAGE_SHIFT - 10));
2944 if (sbinfo->max_inodes != shmem_default_max_inodes())
2945 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
2946 if (sbinfo->mode != (S_IRWXUGO | S_ISVTX))
2947 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
2948 if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
2949 seq_printf(seq, ",uid=%u",
2950 from_kuid_munged(&init_user_ns, sbinfo->uid));
2951 if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
2952 seq_printf(seq, ",gid=%u",
2953 from_kgid_munged(&init_user_ns, sbinfo->gid));
2954 shmem_show_mpol(seq, sbinfo->mpol);
2955 return 0;
2956}
2957
2958#define MFD_NAME_PREFIX "memfd:"
2959#define MFD_NAME_PREFIX_LEN (sizeof(MFD_NAME_PREFIX) - 1)
2960#define MFD_NAME_MAX_LEN (NAME_MAX - MFD_NAME_PREFIX_LEN)
2961
2962#define MFD_ALL_FLAGS (MFD_CLOEXEC | MFD_ALLOW_SEALING)
2963
2964SYSCALL_DEFINE2(memfd_create,
2965 const char __user *, uname,
2966 unsigned int, flags)
2967{
2968 struct shmem_inode_info *info;
2969 struct file *file;
2970 int fd, error;
2971 char *name;
2972 long len;
2973
2974 if (flags & ~(unsigned int)MFD_ALL_FLAGS)
2975 return -EINVAL;
2976
2977 /* length includes terminating zero */
2978 len = strnlen_user(uname, MFD_NAME_MAX_LEN + 1);
2979 if (len <= 0)
2980 return -EFAULT;
2981 if (len > MFD_NAME_MAX_LEN + 1)
2982 return -EINVAL;
2983
2984 name = kmalloc(len + MFD_NAME_PREFIX_LEN, GFP_TEMPORARY);
2985 if (!name)
2986 return -ENOMEM;
2987
2988 strcpy(name, MFD_NAME_PREFIX);
2989 if (copy_from_user(&name[MFD_NAME_PREFIX_LEN], uname, len)) {
2990 error = -EFAULT;
2991 goto err_name;
2992 }
2993
2994 /* terminating-zero may have changed after strnlen_user() returned */
2995 if (name[len + MFD_NAME_PREFIX_LEN - 1]) {
2996 error = -EFAULT;
2997 goto err_name;
2998 }
2999
3000 fd = get_unused_fd_flags((flags & MFD_CLOEXEC) ? O_CLOEXEC : 0);
3001 if (fd < 0) {
3002 error = fd;
3003 goto err_name;
3004 }
3005
3006 file = shmem_file_setup(name, 0, VM_NORESERVE);
3007 if (IS_ERR(file)) {
3008 error = PTR_ERR(file);
3009 goto err_fd;
3010 }
3011 info = SHMEM_I(file_inode(file));
3012 file->f_mode |= FMODE_LSEEK | FMODE_PREAD | FMODE_PWRITE;
3013 file->f_flags |= O_RDWR | O_LARGEFILE;
3014 if (flags & MFD_ALLOW_SEALING)
3015 info->seals &= ~F_SEAL_SEAL;
3016
3017 fd_install(fd, file);
3018 kfree(name);
3019 return fd;
3020
3021err_fd:
3022 put_unused_fd(fd);
3023err_name:
3024 kfree(name);
3025 return error;
3026}
3027
3028#endif /* CONFIG_TMPFS */
3029
3030static void shmem_put_super(struct super_block *sb)
3031{
3032 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
3033
3034 percpu_counter_destroy(&sbinfo->used_blocks);
3035 mpol_put(sbinfo->mpol);
3036 kfree(sbinfo);
3037 sb->s_fs_info = NULL;
3038}
3039
3040int shmem_fill_super(struct super_block *sb, void *data, int silent)
3041{
3042 struct inode *inode;
3043 struct shmem_sb_info *sbinfo;
3044 int err = -ENOMEM;
3045
3046 /* Round up to L1_CACHE_BYTES to resist false sharing */
3047 sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
3048 L1_CACHE_BYTES), GFP_KERNEL);
3049 if (!sbinfo)
3050 return -ENOMEM;
3051
3052 sbinfo->mode = S_IRWXUGO | S_ISVTX;
3053 sbinfo->uid = current_fsuid();
3054 sbinfo->gid = current_fsgid();
3055 sb->s_fs_info = sbinfo;
3056
3057#ifdef CONFIG_TMPFS
3058 /*
3059 * Per default we only allow half of the physical ram per
3060 * tmpfs instance, limiting inodes to one per page of lowmem;
3061 * but the internal instance is left unlimited.
3062 */
3063 if (!(sb->s_flags & MS_KERNMOUNT)) {
3064 sbinfo->max_blocks = shmem_default_max_blocks();
3065 sbinfo->max_inodes = shmem_default_max_inodes();
3066 if (shmem_parse_options(data, sbinfo, false)) {
3067 err = -EINVAL;
3068 goto failed;
3069 }
3070 } else {
3071 sb->s_flags |= MS_NOUSER;
3072 }
3073 sb->s_export_op = &shmem_export_ops;
3074 sb->s_flags |= MS_NOSEC;
3075#else
3076 sb->s_flags |= MS_NOUSER;
3077#endif
3078
3079 spin_lock_init(&sbinfo->stat_lock);
3080 if (percpu_counter_init(&sbinfo->used_blocks, 0, GFP_KERNEL))
3081 goto failed;
3082 sbinfo->free_inodes = sbinfo->max_inodes;
3083
3084 sb->s_maxbytes = MAX_LFS_FILESIZE;
3085 sb->s_blocksize = PAGE_SIZE;
3086 sb->s_blocksize_bits = PAGE_SHIFT;
3087 sb->s_magic = TMPFS_MAGIC;
3088 sb->s_op = &shmem_ops;
3089 sb->s_time_gran = 1;
3090#ifdef CONFIG_TMPFS_XATTR
3091 sb->s_xattr = shmem_xattr_handlers;
3092#endif
3093#ifdef CONFIG_TMPFS_POSIX_ACL
3094 sb->s_flags |= MS_POSIXACL;
3095#endif
3096
3097 inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
3098 if (!inode)
3099 goto failed;
3100 inode->i_uid = sbinfo->uid;
3101 inode->i_gid = sbinfo->gid;
3102 sb->s_root = d_make_root(inode);
3103 if (!sb->s_root)
3104 goto failed;
3105 return 0;
3106
3107failed:
3108 shmem_put_super(sb);
3109 return err;
3110}
3111
3112static struct kmem_cache *shmem_inode_cachep;
3113
3114static struct inode *shmem_alloc_inode(struct super_block *sb)
3115{
3116 struct shmem_inode_info *info;
3117 info = kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
3118 if (!info)
3119 return NULL;
3120 return &info->vfs_inode;
3121}
3122
3123static void shmem_destroy_callback(struct rcu_head *head)
3124{
3125 struct inode *inode = container_of(head, struct inode, i_rcu);
3126 kfree(inode->i_link);
3127 kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
3128}
3129
3130static void shmem_destroy_inode(struct inode *inode)
3131{
3132 if (S_ISREG(inode->i_mode))
3133 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
3134 call_rcu(&inode->i_rcu, shmem_destroy_callback);
3135}
3136
3137static void shmem_init_inode(void *foo)
3138{
3139 struct shmem_inode_info *info = foo;
3140 inode_init_once(&info->vfs_inode);
3141}
3142
3143static int shmem_init_inodecache(void)
3144{
3145 shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
3146 sizeof(struct shmem_inode_info),
3147 0, SLAB_PANIC|SLAB_ACCOUNT, shmem_init_inode);
3148 return 0;
3149}
3150
3151static void shmem_destroy_inodecache(void)
3152{
3153 kmem_cache_destroy(shmem_inode_cachep);
3154}
3155
3156static const struct address_space_operations shmem_aops = {
3157 .writepage = shmem_writepage,
3158 .set_page_dirty = __set_page_dirty_no_writeback,
3159#ifdef CONFIG_TMPFS
3160 .write_begin = shmem_write_begin,
3161 .write_end = shmem_write_end,
3162#endif
3163#ifdef CONFIG_MIGRATION
3164 .migratepage = migrate_page,
3165#endif
3166 .error_remove_page = generic_error_remove_page,
3167};
3168
3169static const struct file_operations shmem_file_operations = {
3170 .mmap = shmem_mmap,
3171#ifdef CONFIG_TMPFS
3172 .llseek = shmem_file_llseek,
3173 .read_iter = shmem_file_read_iter,
3174 .write_iter = generic_file_write_iter,
3175 .fsync = noop_fsync,
3176 .splice_read = shmem_file_splice_read,
3177 .splice_write = iter_file_splice_write,
3178 .fallocate = shmem_fallocate,
3179#endif
3180};
3181
3182static const struct inode_operations shmem_inode_operations = {
3183 .getattr = shmem_getattr,
3184 .setattr = shmem_setattr,
3185#ifdef CONFIG_TMPFS_XATTR
3186 .setxattr = generic_setxattr,
3187 .getxattr = generic_getxattr,
3188 .listxattr = shmem_listxattr,
3189 .removexattr = generic_removexattr,
3190 .set_acl = simple_set_acl,
3191#endif
3192};
3193
3194static const struct inode_operations shmem_dir_inode_operations = {
3195#ifdef CONFIG_TMPFS
3196 .create = shmem_create,
3197 .lookup = simple_lookup,
3198 .link = shmem_link,
3199 .unlink = shmem_unlink,
3200 .symlink = shmem_symlink,
3201 .mkdir = shmem_mkdir,
3202 .rmdir = shmem_rmdir,
3203 .mknod = shmem_mknod,
3204 .rename2 = shmem_rename2,
3205 .tmpfile = shmem_tmpfile,
3206#endif
3207#ifdef CONFIG_TMPFS_XATTR
3208 .setxattr = generic_setxattr,
3209 .getxattr = generic_getxattr,
3210 .listxattr = shmem_listxattr,
3211 .removexattr = generic_removexattr,
3212#endif
3213#ifdef CONFIG_TMPFS_POSIX_ACL
3214 .setattr = shmem_setattr,
3215 .set_acl = simple_set_acl,
3216#endif
3217};
3218
3219static const struct inode_operations shmem_special_inode_operations = {
3220#ifdef CONFIG_TMPFS_XATTR
3221 .setxattr = generic_setxattr,
3222 .getxattr = generic_getxattr,
3223 .listxattr = shmem_listxattr,
3224 .removexattr = generic_removexattr,
3225#endif
3226#ifdef CONFIG_TMPFS_POSIX_ACL
3227 .setattr = shmem_setattr,
3228 .set_acl = simple_set_acl,
3229#endif
3230};
3231
3232static const struct super_operations shmem_ops = {
3233 .alloc_inode = shmem_alloc_inode,
3234 .destroy_inode = shmem_destroy_inode,
3235#ifdef CONFIG_TMPFS
3236 .statfs = shmem_statfs,
3237 .remount_fs = shmem_remount_fs,
3238 .show_options = shmem_show_options,
3239#endif
3240 .evict_inode = shmem_evict_inode,
3241 .drop_inode = generic_delete_inode,
3242 .put_super = shmem_put_super,
3243};
3244
3245static const struct vm_operations_struct shmem_vm_ops = {
3246 .fault = shmem_fault,
3247 .map_pages = filemap_map_pages,
3248#ifdef CONFIG_NUMA
3249 .set_policy = shmem_set_policy,
3250 .get_policy = shmem_get_policy,
3251#endif
3252};
3253
3254static struct dentry *shmem_mount(struct file_system_type *fs_type,
3255 int flags, const char *dev_name, void *data)
3256{
3257 return mount_nodev(fs_type, flags, data, shmem_fill_super);
3258}
3259
3260static struct file_system_type shmem_fs_type = {
3261 .owner = THIS_MODULE,
3262 .name = "tmpfs",
3263 .mount = shmem_mount,
3264 .kill_sb = kill_litter_super,
3265 .fs_flags = FS_USERNS_MOUNT,
3266};
3267
3268int __init shmem_init(void)
3269{
3270 int error;
3271
3272 /* If rootfs called this, don't re-init */
3273 if (shmem_inode_cachep)
3274 return 0;
3275
3276 error = shmem_init_inodecache();
3277 if (error)
3278 goto out3;
3279
3280 error = register_filesystem(&shmem_fs_type);
3281 if (error) {
3282 pr_err("Could not register tmpfs\n");
3283 goto out2;
3284 }
3285
3286 shm_mnt = kern_mount(&shmem_fs_type);
3287 if (IS_ERR(shm_mnt)) {
3288 error = PTR_ERR(shm_mnt);
3289 pr_err("Could not kern_mount tmpfs\n");
3290 goto out1;
3291 }
3292 return 0;
3293
3294out1:
3295 unregister_filesystem(&shmem_fs_type);
3296out2:
3297 shmem_destroy_inodecache();
3298out3:
3299 shm_mnt = ERR_PTR(error);
3300 return error;
3301}
3302
3303#else /* !CONFIG_SHMEM */
3304
3305/*
3306 * tiny-shmem: simple shmemfs and tmpfs using ramfs code
3307 *
3308 * This is intended for small system where the benefits of the full
3309 * shmem code (swap-backed and resource-limited) are outweighed by
3310 * their complexity. On systems without swap this code should be
3311 * effectively equivalent, but much lighter weight.
3312 */
3313
3314static struct file_system_type shmem_fs_type = {
3315 .name = "tmpfs",
3316 .mount = ramfs_mount,
3317 .kill_sb = kill_litter_super,
3318 .fs_flags = FS_USERNS_MOUNT,
3319};
3320
3321int __init shmem_init(void)
3322{
3323 BUG_ON(register_filesystem(&shmem_fs_type) != 0);
3324
3325 shm_mnt = kern_mount(&shmem_fs_type);
3326 BUG_ON(IS_ERR(shm_mnt));
3327
3328 return 0;
3329}
3330
3331int shmem_unuse(swp_entry_t swap, struct page *page)
3332{
3333 return 0;
3334}
3335
3336int shmem_lock(struct file *file, int lock, struct user_struct *user)
3337{
3338 return 0;
3339}
3340
3341void shmem_unlock_mapping(struct address_space *mapping)
3342{
3343}
3344
3345void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
3346{
3347 truncate_inode_pages_range(inode->i_mapping, lstart, lend);
3348}
3349EXPORT_SYMBOL_GPL(shmem_truncate_range);
3350
3351#define shmem_vm_ops generic_file_vm_ops
3352#define shmem_file_operations ramfs_file_operations
3353#define shmem_get_inode(sb, dir, mode, dev, flags) ramfs_get_inode(sb, dir, mode, dev)
3354#define shmem_acct_size(flags, size) 0
3355#define shmem_unacct_size(flags, size) do {} while (0)
3356
3357#endif /* CONFIG_SHMEM */
3358
3359/* common code */
3360
3361static struct dentry_operations anon_ops = {
3362 .d_dname = simple_dname
3363};
3364
3365static struct file *__shmem_file_setup(const char *name, loff_t size,
3366 unsigned long flags, unsigned int i_flags)
3367{
3368 struct file *res;
3369 struct inode *inode;
3370 struct path path;
3371 struct super_block *sb;
3372 struct qstr this;
3373
3374 if (IS_ERR(shm_mnt))
3375 return ERR_CAST(shm_mnt);
3376
3377 if (size < 0 || size > MAX_LFS_FILESIZE)
3378 return ERR_PTR(-EINVAL);
3379
3380 if (shmem_acct_size(flags, size))
3381 return ERR_PTR(-ENOMEM);
3382
3383 res = ERR_PTR(-ENOMEM);
3384 this.name = name;
3385 this.len = strlen(name);
3386 this.hash = 0; /* will go */
3387 sb = shm_mnt->mnt_sb;
3388 path.mnt = mntget(shm_mnt);
3389 path.dentry = d_alloc_pseudo(sb, &this);
3390 if (!path.dentry)
3391 goto put_memory;
3392 d_set_d_op(path.dentry, &anon_ops);
3393
3394 res = ERR_PTR(-ENOSPC);
3395 inode = shmem_get_inode(sb, NULL, S_IFREG | S_IRWXUGO, 0, flags);
3396 if (!inode)
3397 goto put_memory;
3398
3399 inode->i_flags |= i_flags;
3400 d_instantiate(path.dentry, inode);
3401 inode->i_size = size;
3402 clear_nlink(inode); /* It is unlinked */
3403 res = ERR_PTR(ramfs_nommu_expand_for_mapping(inode, size));
3404 if (IS_ERR(res))
3405 goto put_path;
3406
3407 res = alloc_file(&path, FMODE_WRITE | FMODE_READ,
3408 &shmem_file_operations);
3409 if (IS_ERR(res))
3410 goto put_path;
3411
3412 return res;
3413
3414put_memory:
3415 shmem_unacct_size(flags, size);
3416put_path:
3417 path_put(&path);
3418 return res;
3419}
3420
3421/**
3422 * shmem_kernel_file_setup - get an unlinked file living in tmpfs which must be
3423 * kernel internal. There will be NO LSM permission checks against the
3424 * underlying inode. So users of this interface must do LSM checks at a
3425 * higher layer. The users are the big_key and shm implementations. LSM
3426 * checks are provided at the key or shm level rather than the inode.
3427 * @name: name for dentry (to be seen in /proc/<pid>/maps
3428 * @size: size to be set for the file
3429 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
3430 */
3431struct file *shmem_kernel_file_setup(const char *name, loff_t size, unsigned long flags)
3432{
3433 return __shmem_file_setup(name, size, flags, S_PRIVATE);
3434}
3435
3436/**
3437 * shmem_file_setup - get an unlinked file living in tmpfs
3438 * @name: name for dentry (to be seen in /proc/<pid>/maps
3439 * @size: size to be set for the file
3440 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
3441 */
3442struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
3443{
3444 return __shmem_file_setup(name, size, flags, 0);
3445}
3446EXPORT_SYMBOL_GPL(shmem_file_setup);
3447
3448/**
3449 * shmem_zero_setup - setup a shared anonymous mapping
3450 * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
3451 */
3452int shmem_zero_setup(struct vm_area_struct *vma)
3453{
3454 struct file *file;
3455 loff_t size = vma->vm_end - vma->vm_start;
3456
3457 /*
3458 * Cloning a new file under mmap_sem leads to a lock ordering conflict
3459 * between XFS directory reading and selinux: since this file is only
3460 * accessible to the user through its mapping, use S_PRIVATE flag to
3461 * bypass file security, in the same way as shmem_kernel_file_setup().
3462 */
3463 file = __shmem_file_setup("dev/zero", size, vma->vm_flags, S_PRIVATE);
3464 if (IS_ERR(file))
3465 return PTR_ERR(file);
3466
3467 if (vma->vm_file)
3468 fput(vma->vm_file);
3469 vma->vm_file = file;
3470 vma->vm_ops = &shmem_vm_ops;
3471 return 0;
3472}
3473
3474/**
3475 * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
3476 * @mapping: the page's address_space
3477 * @index: the page index
3478 * @gfp: the page allocator flags to use if allocating
3479 *
3480 * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
3481 * with any new page allocations done using the specified allocation flags.
3482 * But read_cache_page_gfp() uses the ->readpage() method: which does not
3483 * suit tmpfs, since it may have pages in swapcache, and needs to find those
3484 * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
3485 *
3486 * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
3487 * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
3488 */
3489struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
3490 pgoff_t index, gfp_t gfp)
3491{
3492#ifdef CONFIG_SHMEM
3493 struct inode *inode = mapping->host;
3494 struct page *page;
3495 int error;
3496
3497 BUG_ON(mapping->a_ops != &shmem_aops);
3498 error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE, gfp, NULL);
3499 if (error)
3500 page = ERR_PTR(error);
3501 else
3502 unlock_page(page);
3503 return page;
3504#else
3505 /*
3506 * The tiny !SHMEM case uses ramfs without swap
3507 */
3508 return read_cache_page_gfp(mapping, index, gfp);
3509#endif
3510}
3511EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);