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/sched/signal.h>
33#include <linux/export.h>
34#include <linux/swap.h>
35#include <linux/uio.h>
36#include <linux/khugepaged.h>
37#include <linux/hugetlb.h>
38
39#include <asm/tlbflush.h> /* for arch/microblaze update_mmu_cache() */
40
41static struct vfsmount *shm_mnt;
42
43#ifdef CONFIG_SHMEM
44/*
45 * This virtual memory filesystem is heavily based on the ramfs. It
46 * extends ramfs by the ability to use swap and honor resource limits
47 * which makes it a completely usable filesystem.
48 */
49
50#include <linux/xattr.h>
51#include <linux/exportfs.h>
52#include <linux/posix_acl.h>
53#include <linux/posix_acl_xattr.h>
54#include <linux/mman.h>
55#include <linux/string.h>
56#include <linux/slab.h>
57#include <linux/backing-dev.h>
58#include <linux/shmem_fs.h>
59#include <linux/writeback.h>
60#include <linux/blkdev.h>
61#include <linux/pagevec.h>
62#include <linux/percpu_counter.h>
63#include <linux/falloc.h>
64#include <linux/splice.h>
65#include <linux/security.h>
66#include <linux/swapops.h>
67#include <linux/mempolicy.h>
68#include <linux/namei.h>
69#include <linux/ctype.h>
70#include <linux/migrate.h>
71#include <linux/highmem.h>
72#include <linux/seq_file.h>
73#include <linux/magic.h>
74#include <linux/syscalls.h>
75#include <linux/fcntl.h>
76#include <uapi/linux/memfd.h>
77#include <linux/userfaultfd_k.h>
78#include <linux/rmap.h>
79#include <linux/uuid.h>
80
81#include <linux/uaccess.h>
82#include <asm/pgtable.h>
83
84#include "internal.h"
85
86#define BLOCKS_PER_PAGE (PAGE_SIZE/512)
87#define VM_ACCT(size) (PAGE_ALIGN(size) >> PAGE_SHIFT)
88
89/* Pretend that each entry is of this size in directory's i_size */
90#define BOGO_DIRENT_SIZE 20
91
92/* Symlink up to this size is kmalloc'ed instead of using a swappable page */
93#define SHORT_SYMLINK_LEN 128
94
95/*
96 * shmem_fallocate communicates with shmem_fault or shmem_writepage via
97 * inode->i_private (with i_mutex making sure that it has only one user at
98 * a time): we would prefer not to enlarge the shmem inode just for that.
99 */
100struct shmem_falloc {
101 wait_queue_head_t *waitq; /* faults into hole wait for punch to end */
102 pgoff_t start; /* start of range currently being fallocated */
103 pgoff_t next; /* the next page offset to be fallocated */
104 pgoff_t nr_falloced; /* how many new pages have been fallocated */
105 pgoff_t nr_unswapped; /* how often writepage refused to swap out */
106};
107
108#ifdef CONFIG_TMPFS
109static unsigned long shmem_default_max_blocks(void)
110{
111 return totalram_pages / 2;
112}
113
114static unsigned long shmem_default_max_inodes(void)
115{
116 return min(totalram_pages - totalhigh_pages, totalram_pages / 2);
117}
118#endif
119
120static bool shmem_should_replace_page(struct page *page, gfp_t gfp);
121static int shmem_replace_page(struct page **pagep, gfp_t gfp,
122 struct shmem_inode_info *info, pgoff_t index);
123static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
124 struct page **pagep, enum sgp_type sgp,
125 gfp_t gfp, struct vm_area_struct *vma,
126 struct vm_fault *vmf, int *fault_type);
127
128int shmem_getpage(struct inode *inode, pgoff_t index,
129 struct page **pagep, enum sgp_type sgp)
130{
131 return shmem_getpage_gfp(inode, index, pagep, sgp,
132 mapping_gfp_mask(inode->i_mapping), NULL, NULL, NULL);
133}
134
135static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
136{
137 return sb->s_fs_info;
138}
139
140/*
141 * shmem_file_setup pre-accounts the whole fixed size of a VM object,
142 * for shared memory and for shared anonymous (/dev/zero) mappings
143 * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
144 * consistent with the pre-accounting of private mappings ...
145 */
146static inline int shmem_acct_size(unsigned long flags, loff_t size)
147{
148 return (flags & VM_NORESERVE) ?
149 0 : security_vm_enough_memory_mm(current->mm, VM_ACCT(size));
150}
151
152static inline void shmem_unacct_size(unsigned long flags, loff_t size)
153{
154 if (!(flags & VM_NORESERVE))
155 vm_unacct_memory(VM_ACCT(size));
156}
157
158static inline int shmem_reacct_size(unsigned long flags,
159 loff_t oldsize, loff_t newsize)
160{
161 if (!(flags & VM_NORESERVE)) {
162 if (VM_ACCT(newsize) > VM_ACCT(oldsize))
163 return security_vm_enough_memory_mm(current->mm,
164 VM_ACCT(newsize) - VM_ACCT(oldsize));
165 else if (VM_ACCT(newsize) < VM_ACCT(oldsize))
166 vm_unacct_memory(VM_ACCT(oldsize) - VM_ACCT(newsize));
167 }
168 return 0;
169}
170
171/*
172 * ... whereas tmpfs objects are accounted incrementally as
173 * pages are allocated, in order to allow large sparse files.
174 * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
175 * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
176 */
177static inline int shmem_acct_block(unsigned long flags, long pages)
178{
179 if (!(flags & VM_NORESERVE))
180 return 0;
181
182 return security_vm_enough_memory_mm(current->mm,
183 pages * VM_ACCT(PAGE_SIZE));
184}
185
186static inline void shmem_unacct_blocks(unsigned long flags, long pages)
187{
188 if (flags & VM_NORESERVE)
189 vm_unacct_memory(pages * VM_ACCT(PAGE_SIZE));
190}
191
192static inline bool shmem_inode_acct_block(struct inode *inode, long pages)
193{
194 struct shmem_inode_info *info = SHMEM_I(inode);
195 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
196
197 if (shmem_acct_block(info->flags, pages))
198 return false;
199
200 if (sbinfo->max_blocks) {
201 if (percpu_counter_compare(&sbinfo->used_blocks,
202 sbinfo->max_blocks - pages) > 0)
203 goto unacct;
204 percpu_counter_add(&sbinfo->used_blocks, pages);
205 }
206
207 return true;
208
209unacct:
210 shmem_unacct_blocks(info->flags, pages);
211 return false;
212}
213
214static inline void shmem_inode_unacct_blocks(struct inode *inode, long pages)
215{
216 struct shmem_inode_info *info = SHMEM_I(inode);
217 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
218
219 if (sbinfo->max_blocks)
220 percpu_counter_sub(&sbinfo->used_blocks, pages);
221 shmem_unacct_blocks(info->flags, pages);
222}
223
224static const struct super_operations shmem_ops;
225static const struct address_space_operations shmem_aops;
226static const struct file_operations shmem_file_operations;
227static const struct inode_operations shmem_inode_operations;
228static const struct inode_operations shmem_dir_inode_operations;
229static const struct inode_operations shmem_special_inode_operations;
230static const struct vm_operations_struct shmem_vm_ops;
231static struct file_system_type shmem_fs_type;
232
233bool vma_is_shmem(struct vm_area_struct *vma)
234{
235 return vma->vm_ops == &shmem_vm_ops;
236}
237
238static LIST_HEAD(shmem_swaplist);
239static DEFINE_MUTEX(shmem_swaplist_mutex);
240
241static int shmem_reserve_inode(struct super_block *sb)
242{
243 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
244 if (sbinfo->max_inodes) {
245 spin_lock(&sbinfo->stat_lock);
246 if (!sbinfo->free_inodes) {
247 spin_unlock(&sbinfo->stat_lock);
248 return -ENOSPC;
249 }
250 sbinfo->free_inodes--;
251 spin_unlock(&sbinfo->stat_lock);
252 }
253 return 0;
254}
255
256static void shmem_free_inode(struct super_block *sb)
257{
258 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
259 if (sbinfo->max_inodes) {
260 spin_lock(&sbinfo->stat_lock);
261 sbinfo->free_inodes++;
262 spin_unlock(&sbinfo->stat_lock);
263 }
264}
265
266/**
267 * shmem_recalc_inode - recalculate the block usage of an inode
268 * @inode: inode to recalc
269 *
270 * We have to calculate the free blocks since the mm can drop
271 * undirtied hole pages behind our back.
272 *
273 * But normally info->alloced == inode->i_mapping->nrpages + info->swapped
274 * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
275 *
276 * It has to be called with the spinlock held.
277 */
278static void shmem_recalc_inode(struct inode *inode)
279{
280 struct shmem_inode_info *info = SHMEM_I(inode);
281 long freed;
282
283 freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
284 if (freed > 0) {
285 info->alloced -= freed;
286 inode->i_blocks -= freed * BLOCKS_PER_PAGE;
287 shmem_inode_unacct_blocks(inode, freed);
288 }
289}
290
291bool shmem_charge(struct inode *inode, long pages)
292{
293 struct shmem_inode_info *info = SHMEM_I(inode);
294 unsigned long flags;
295
296 if (!shmem_inode_acct_block(inode, pages))
297 return false;
298
299 spin_lock_irqsave(&info->lock, flags);
300 info->alloced += pages;
301 inode->i_blocks += pages * BLOCKS_PER_PAGE;
302 shmem_recalc_inode(inode);
303 spin_unlock_irqrestore(&info->lock, flags);
304 inode->i_mapping->nrpages += pages;
305
306 return true;
307}
308
309void shmem_uncharge(struct inode *inode, long pages)
310{
311 struct shmem_inode_info *info = SHMEM_I(inode);
312 unsigned long flags;
313
314 spin_lock_irqsave(&info->lock, flags);
315 info->alloced -= pages;
316 inode->i_blocks -= pages * BLOCKS_PER_PAGE;
317 shmem_recalc_inode(inode);
318 spin_unlock_irqrestore(&info->lock, flags);
319
320 shmem_inode_unacct_blocks(inode, pages);
321}
322
323/*
324 * Replace item expected in radix tree by a new item, while holding tree lock.
325 */
326static int shmem_radix_tree_replace(struct address_space *mapping,
327 pgoff_t index, void *expected, void *replacement)
328{
329 struct radix_tree_node *node;
330 void **pslot;
331 void *item;
332
333 VM_BUG_ON(!expected);
334 VM_BUG_ON(!replacement);
335 item = __radix_tree_lookup(&mapping->i_pages, index, &node, &pslot);
336 if (!item)
337 return -ENOENT;
338 if (item != expected)
339 return -ENOENT;
340 __radix_tree_replace(&mapping->i_pages, node, pslot,
341 replacement, NULL);
342 return 0;
343}
344
345/*
346 * Sometimes, before we decide whether to proceed or to fail, we must check
347 * that an entry was not already brought back from swap by a racing thread.
348 *
349 * Checking page is not enough: by the time a SwapCache page is locked, it
350 * might be reused, and again be SwapCache, using the same swap as before.
351 */
352static bool shmem_confirm_swap(struct address_space *mapping,
353 pgoff_t index, swp_entry_t swap)
354{
355 void *item;
356
357 rcu_read_lock();
358 item = radix_tree_lookup(&mapping->i_pages, index);
359 rcu_read_unlock();
360 return item == swp_to_radix_entry(swap);
361}
362
363/*
364 * Definitions for "huge tmpfs": tmpfs mounted with the huge= option
365 *
366 * SHMEM_HUGE_NEVER:
367 * disables huge pages for the mount;
368 * SHMEM_HUGE_ALWAYS:
369 * enables huge pages for the mount;
370 * SHMEM_HUGE_WITHIN_SIZE:
371 * only allocate huge pages if the page will be fully within i_size,
372 * also respect fadvise()/madvise() hints;
373 * SHMEM_HUGE_ADVISE:
374 * only allocate huge pages if requested with fadvise()/madvise();
375 */
376
377#define SHMEM_HUGE_NEVER 0
378#define SHMEM_HUGE_ALWAYS 1
379#define SHMEM_HUGE_WITHIN_SIZE 2
380#define SHMEM_HUGE_ADVISE 3
381
382/*
383 * Special values.
384 * Only can be set via /sys/kernel/mm/transparent_hugepage/shmem_enabled:
385 *
386 * SHMEM_HUGE_DENY:
387 * disables huge on shm_mnt and all mounts, for emergency use;
388 * SHMEM_HUGE_FORCE:
389 * enables huge on shm_mnt and all mounts, w/o needing option, for testing;
390 *
391 */
392#define SHMEM_HUGE_DENY (-1)
393#define SHMEM_HUGE_FORCE (-2)
394
395#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
396/* ifdef here to avoid bloating shmem.o when not necessary */
397
398int shmem_huge __read_mostly;
399
400#if defined(CONFIG_SYSFS) || defined(CONFIG_TMPFS)
401static int shmem_parse_huge(const char *str)
402{
403 if (!strcmp(str, "never"))
404 return SHMEM_HUGE_NEVER;
405 if (!strcmp(str, "always"))
406 return SHMEM_HUGE_ALWAYS;
407 if (!strcmp(str, "within_size"))
408 return SHMEM_HUGE_WITHIN_SIZE;
409 if (!strcmp(str, "advise"))
410 return SHMEM_HUGE_ADVISE;
411 if (!strcmp(str, "deny"))
412 return SHMEM_HUGE_DENY;
413 if (!strcmp(str, "force"))
414 return SHMEM_HUGE_FORCE;
415 return -EINVAL;
416}
417
418static const char *shmem_format_huge(int huge)
419{
420 switch (huge) {
421 case SHMEM_HUGE_NEVER:
422 return "never";
423 case SHMEM_HUGE_ALWAYS:
424 return "always";
425 case SHMEM_HUGE_WITHIN_SIZE:
426 return "within_size";
427 case SHMEM_HUGE_ADVISE:
428 return "advise";
429 case SHMEM_HUGE_DENY:
430 return "deny";
431 case SHMEM_HUGE_FORCE:
432 return "force";
433 default:
434 VM_BUG_ON(1);
435 return "bad_val";
436 }
437}
438#endif
439
440static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
441 struct shrink_control *sc, unsigned long nr_to_split)
442{
443 LIST_HEAD(list), *pos, *next;
444 LIST_HEAD(to_remove);
445 struct inode *inode;
446 struct shmem_inode_info *info;
447 struct page *page;
448 unsigned long batch = sc ? sc->nr_to_scan : 128;
449 int removed = 0, split = 0;
450
451 if (list_empty(&sbinfo->shrinklist))
452 return SHRINK_STOP;
453
454 spin_lock(&sbinfo->shrinklist_lock);
455 list_for_each_safe(pos, next, &sbinfo->shrinklist) {
456 info = list_entry(pos, struct shmem_inode_info, shrinklist);
457
458 /* pin the inode */
459 inode = igrab(&info->vfs_inode);
460
461 /* inode is about to be evicted */
462 if (!inode) {
463 list_del_init(&info->shrinklist);
464 removed++;
465 goto next;
466 }
467
468 /* Check if there's anything to gain */
469 if (round_up(inode->i_size, PAGE_SIZE) ==
470 round_up(inode->i_size, HPAGE_PMD_SIZE)) {
471 list_move(&info->shrinklist, &to_remove);
472 removed++;
473 goto next;
474 }
475
476 list_move(&info->shrinklist, &list);
477next:
478 if (!--batch)
479 break;
480 }
481 spin_unlock(&sbinfo->shrinklist_lock);
482
483 list_for_each_safe(pos, next, &to_remove) {
484 info = list_entry(pos, struct shmem_inode_info, shrinklist);
485 inode = &info->vfs_inode;
486 list_del_init(&info->shrinklist);
487 iput(inode);
488 }
489
490 list_for_each_safe(pos, next, &list) {
491 int ret;
492
493 info = list_entry(pos, struct shmem_inode_info, shrinklist);
494 inode = &info->vfs_inode;
495
496 if (nr_to_split && split >= nr_to_split)
497 goto leave;
498
499 page = find_get_page(inode->i_mapping,
500 (inode->i_size & HPAGE_PMD_MASK) >> PAGE_SHIFT);
501 if (!page)
502 goto drop;
503
504 /* No huge page at the end of the file: nothing to split */
505 if (!PageTransHuge(page)) {
506 put_page(page);
507 goto drop;
508 }
509
510 /*
511 * Leave the inode on the list if we failed to lock
512 * the page at this time.
513 *
514 * Waiting for the lock may lead to deadlock in the
515 * reclaim path.
516 */
517 if (!trylock_page(page)) {
518 put_page(page);
519 goto leave;
520 }
521
522 ret = split_huge_page(page);
523 unlock_page(page);
524 put_page(page);
525
526 /* If split failed leave the inode on the list */
527 if (ret)
528 goto leave;
529
530 split++;
531drop:
532 list_del_init(&info->shrinklist);
533 removed++;
534leave:
535 iput(inode);
536 }
537
538 spin_lock(&sbinfo->shrinklist_lock);
539 list_splice_tail(&list, &sbinfo->shrinklist);
540 sbinfo->shrinklist_len -= removed;
541 spin_unlock(&sbinfo->shrinklist_lock);
542
543 return split;
544}
545
546static long shmem_unused_huge_scan(struct super_block *sb,
547 struct shrink_control *sc)
548{
549 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
550
551 if (!READ_ONCE(sbinfo->shrinklist_len))
552 return SHRINK_STOP;
553
554 return shmem_unused_huge_shrink(sbinfo, sc, 0);
555}
556
557static long shmem_unused_huge_count(struct super_block *sb,
558 struct shrink_control *sc)
559{
560 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
561 return READ_ONCE(sbinfo->shrinklist_len);
562}
563#else /* !CONFIG_TRANSPARENT_HUGE_PAGECACHE */
564
565#define shmem_huge SHMEM_HUGE_DENY
566
567static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
568 struct shrink_control *sc, unsigned long nr_to_split)
569{
570 return 0;
571}
572#endif /* CONFIG_TRANSPARENT_HUGE_PAGECACHE */
573
574/*
575 * Like add_to_page_cache_locked, but error if expected item has gone.
576 */
577static int shmem_add_to_page_cache(struct page *page,
578 struct address_space *mapping,
579 pgoff_t index, void *expected)
580{
581 int error, nr = hpage_nr_pages(page);
582
583 VM_BUG_ON_PAGE(PageTail(page), page);
584 VM_BUG_ON_PAGE(index != round_down(index, nr), page);
585 VM_BUG_ON_PAGE(!PageLocked(page), page);
586 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
587 VM_BUG_ON(expected && PageTransHuge(page));
588
589 page_ref_add(page, nr);
590 page->mapping = mapping;
591 page->index = index;
592
593 xa_lock_irq(&mapping->i_pages);
594 if (PageTransHuge(page)) {
595 void __rcu **results;
596 pgoff_t idx;
597 int i;
598
599 error = 0;
600 if (radix_tree_gang_lookup_slot(&mapping->i_pages,
601 &results, &idx, index, 1) &&
602 idx < index + HPAGE_PMD_NR) {
603 error = -EEXIST;
604 }
605
606 if (!error) {
607 for (i = 0; i < HPAGE_PMD_NR; i++) {
608 error = radix_tree_insert(&mapping->i_pages,
609 index + i, page + i);
610 VM_BUG_ON(error);
611 }
612 count_vm_event(THP_FILE_ALLOC);
613 }
614 } else if (!expected) {
615 error = radix_tree_insert(&mapping->i_pages, index, page);
616 } else {
617 error = shmem_radix_tree_replace(mapping, index, expected,
618 page);
619 }
620
621 if (!error) {
622 mapping->nrpages += nr;
623 if (PageTransHuge(page))
624 __inc_node_page_state(page, NR_SHMEM_THPS);
625 __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, nr);
626 __mod_node_page_state(page_pgdat(page), NR_SHMEM, nr);
627 xa_unlock_irq(&mapping->i_pages);
628 } else {
629 page->mapping = NULL;
630 xa_unlock_irq(&mapping->i_pages);
631 page_ref_sub(page, nr);
632 }
633 return error;
634}
635
636/*
637 * Like delete_from_page_cache, but substitutes swap for page.
638 */
639static void shmem_delete_from_page_cache(struct page *page, void *radswap)
640{
641 struct address_space *mapping = page->mapping;
642 int error;
643
644 VM_BUG_ON_PAGE(PageCompound(page), page);
645
646 xa_lock_irq(&mapping->i_pages);
647 error = shmem_radix_tree_replace(mapping, page->index, page, radswap);
648 page->mapping = NULL;
649 mapping->nrpages--;
650 __dec_node_page_state(page, NR_FILE_PAGES);
651 __dec_node_page_state(page, NR_SHMEM);
652 xa_unlock_irq(&mapping->i_pages);
653 put_page(page);
654 BUG_ON(error);
655}
656
657/*
658 * Remove swap entry from radix tree, free the swap and its page cache.
659 */
660static int shmem_free_swap(struct address_space *mapping,
661 pgoff_t index, void *radswap)
662{
663 void *old;
664
665 xa_lock_irq(&mapping->i_pages);
666 old = radix_tree_delete_item(&mapping->i_pages, index, radswap);
667 xa_unlock_irq(&mapping->i_pages);
668 if (old != radswap)
669 return -ENOENT;
670 free_swap_and_cache(radix_to_swp_entry(radswap));
671 return 0;
672}
673
674/*
675 * Determine (in bytes) how many of the shmem object's pages mapped by the
676 * given offsets are swapped out.
677 *
678 * This is safe to call without i_mutex or the i_pages lock thanks to RCU,
679 * as long as the inode doesn't go away and racy results are not a problem.
680 */
681unsigned long shmem_partial_swap_usage(struct address_space *mapping,
682 pgoff_t start, pgoff_t end)
683{
684 struct radix_tree_iter iter;
685 void **slot;
686 struct page *page;
687 unsigned long swapped = 0;
688
689 rcu_read_lock();
690
691 radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start) {
692 if (iter.index >= end)
693 break;
694
695 page = radix_tree_deref_slot(slot);
696
697 if (radix_tree_deref_retry(page)) {
698 slot = radix_tree_iter_retry(&iter);
699 continue;
700 }
701
702 if (radix_tree_exceptional_entry(page))
703 swapped++;
704
705 if (need_resched()) {
706 slot = radix_tree_iter_resume(slot, &iter);
707 cond_resched_rcu();
708 }
709 }
710
711 rcu_read_unlock();
712
713 return swapped << PAGE_SHIFT;
714}
715
716/*
717 * Determine (in bytes) how many of the shmem object's pages mapped by the
718 * given vma is swapped out.
719 *
720 * This is safe to call without i_mutex or the i_pages lock thanks to RCU,
721 * as long as the inode doesn't go away and racy results are not a problem.
722 */
723unsigned long shmem_swap_usage(struct vm_area_struct *vma)
724{
725 struct inode *inode = file_inode(vma->vm_file);
726 struct shmem_inode_info *info = SHMEM_I(inode);
727 struct address_space *mapping = inode->i_mapping;
728 unsigned long swapped;
729
730 /* Be careful as we don't hold info->lock */
731 swapped = READ_ONCE(info->swapped);
732
733 /*
734 * The easier cases are when the shmem object has nothing in swap, or
735 * the vma maps it whole. Then we can simply use the stats that we
736 * already track.
737 */
738 if (!swapped)
739 return 0;
740
741 if (!vma->vm_pgoff && vma->vm_end - vma->vm_start >= inode->i_size)
742 return swapped << PAGE_SHIFT;
743
744 /* Here comes the more involved part */
745 return shmem_partial_swap_usage(mapping,
746 linear_page_index(vma, vma->vm_start),
747 linear_page_index(vma, vma->vm_end));
748}
749
750/*
751 * SysV IPC SHM_UNLOCK restore Unevictable pages to their evictable lists.
752 */
753void shmem_unlock_mapping(struct address_space *mapping)
754{
755 struct pagevec pvec;
756 pgoff_t indices[PAGEVEC_SIZE];
757 pgoff_t index = 0;
758
759 pagevec_init(&pvec);
760 /*
761 * Minor point, but we might as well stop if someone else SHM_LOCKs it.
762 */
763 while (!mapping_unevictable(mapping)) {
764 /*
765 * Avoid pagevec_lookup(): find_get_pages() returns 0 as if it
766 * has finished, if it hits a row of PAGEVEC_SIZE swap entries.
767 */
768 pvec.nr = find_get_entries(mapping, index,
769 PAGEVEC_SIZE, pvec.pages, indices);
770 if (!pvec.nr)
771 break;
772 index = indices[pvec.nr - 1] + 1;
773 pagevec_remove_exceptionals(&pvec);
774 check_move_unevictable_pages(pvec.pages, pvec.nr);
775 pagevec_release(&pvec);
776 cond_resched();
777 }
778}
779
780/*
781 * Remove range of pages and swap entries from radix tree, and free them.
782 * If !unfalloc, truncate or punch hole; if unfalloc, undo failed fallocate.
783 */
784static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
785 bool unfalloc)
786{
787 struct address_space *mapping = inode->i_mapping;
788 struct shmem_inode_info *info = SHMEM_I(inode);
789 pgoff_t start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
790 pgoff_t end = (lend + 1) >> PAGE_SHIFT;
791 unsigned int partial_start = lstart & (PAGE_SIZE - 1);
792 unsigned int partial_end = (lend + 1) & (PAGE_SIZE - 1);
793 struct pagevec pvec;
794 pgoff_t indices[PAGEVEC_SIZE];
795 long nr_swaps_freed = 0;
796 pgoff_t index;
797 int i;
798
799 if (lend == -1)
800 end = -1; /* unsigned, so actually very big */
801
802 pagevec_init(&pvec);
803 index = start;
804 while (index < end) {
805 pvec.nr = find_get_entries(mapping, index,
806 min(end - index, (pgoff_t)PAGEVEC_SIZE),
807 pvec.pages, indices);
808 if (!pvec.nr)
809 break;
810 for (i = 0; i < pagevec_count(&pvec); i++) {
811 struct page *page = pvec.pages[i];
812
813 index = indices[i];
814 if (index >= end)
815 break;
816
817 if (radix_tree_exceptional_entry(page)) {
818 if (unfalloc)
819 continue;
820 nr_swaps_freed += !shmem_free_swap(mapping,
821 index, page);
822 continue;
823 }
824
825 VM_BUG_ON_PAGE(page_to_pgoff(page) != index, page);
826
827 if (!trylock_page(page))
828 continue;
829
830 if (PageTransTail(page)) {
831 /* Middle of THP: zero out the page */
832 clear_highpage(page);
833 unlock_page(page);
834 continue;
835 } else if (PageTransHuge(page)) {
836 if (index == round_down(end, HPAGE_PMD_NR)) {
837 /*
838 * Range ends in the middle of THP:
839 * zero out the page
840 */
841 clear_highpage(page);
842 unlock_page(page);
843 continue;
844 }
845 index += HPAGE_PMD_NR - 1;
846 i += HPAGE_PMD_NR - 1;
847 }
848
849 if (!unfalloc || !PageUptodate(page)) {
850 VM_BUG_ON_PAGE(PageTail(page), page);
851 if (page_mapping(page) == mapping) {
852 VM_BUG_ON_PAGE(PageWriteback(page), page);
853 truncate_inode_page(mapping, page);
854 }
855 }
856 unlock_page(page);
857 }
858 pagevec_remove_exceptionals(&pvec);
859 pagevec_release(&pvec);
860 cond_resched();
861 index++;
862 }
863
864 if (partial_start) {
865 struct page *page = NULL;
866 shmem_getpage(inode, start - 1, &page, SGP_READ);
867 if (page) {
868 unsigned int top = PAGE_SIZE;
869 if (start > end) {
870 top = partial_end;
871 partial_end = 0;
872 }
873 zero_user_segment(page, partial_start, top);
874 set_page_dirty(page);
875 unlock_page(page);
876 put_page(page);
877 }
878 }
879 if (partial_end) {
880 struct page *page = NULL;
881 shmem_getpage(inode, end, &page, SGP_READ);
882 if (page) {
883 zero_user_segment(page, 0, partial_end);
884 set_page_dirty(page);
885 unlock_page(page);
886 put_page(page);
887 }
888 }
889 if (start >= end)
890 return;
891
892 index = start;
893 while (index < end) {
894 cond_resched();
895
896 pvec.nr = find_get_entries(mapping, index,
897 min(end - index, (pgoff_t)PAGEVEC_SIZE),
898 pvec.pages, indices);
899 if (!pvec.nr) {
900 /* If all gone or hole-punch or unfalloc, we're done */
901 if (index == start || end != -1)
902 break;
903 /* But if truncating, restart to make sure all gone */
904 index = start;
905 continue;
906 }
907 for (i = 0; i < pagevec_count(&pvec); i++) {
908 struct page *page = pvec.pages[i];
909
910 index = indices[i];
911 if (index >= end)
912 break;
913
914 if (radix_tree_exceptional_entry(page)) {
915 if (unfalloc)
916 continue;
917 if (shmem_free_swap(mapping, index, page)) {
918 /* Swap was replaced by page: retry */
919 index--;
920 break;
921 }
922 nr_swaps_freed++;
923 continue;
924 }
925
926 lock_page(page);
927
928 if (PageTransTail(page)) {
929 /* Middle of THP: zero out the page */
930 clear_highpage(page);
931 unlock_page(page);
932 /*
933 * Partial thp truncate due 'start' in middle
934 * of THP: don't need to look on these pages
935 * again on !pvec.nr restart.
936 */
937 if (index != round_down(end, HPAGE_PMD_NR))
938 start++;
939 continue;
940 } else if (PageTransHuge(page)) {
941 if (index == round_down(end, HPAGE_PMD_NR)) {
942 /*
943 * Range ends in the middle of THP:
944 * zero out the page
945 */
946 clear_highpage(page);
947 unlock_page(page);
948 continue;
949 }
950 index += HPAGE_PMD_NR - 1;
951 i += HPAGE_PMD_NR - 1;
952 }
953
954 if (!unfalloc || !PageUptodate(page)) {
955 VM_BUG_ON_PAGE(PageTail(page), page);
956 if (page_mapping(page) == mapping) {
957 VM_BUG_ON_PAGE(PageWriteback(page), page);
958 truncate_inode_page(mapping, page);
959 } else {
960 /* Page was replaced by swap: retry */
961 unlock_page(page);
962 index--;
963 break;
964 }
965 }
966 unlock_page(page);
967 }
968 pagevec_remove_exceptionals(&pvec);
969 pagevec_release(&pvec);
970 index++;
971 }
972
973 spin_lock_irq(&info->lock);
974 info->swapped -= nr_swaps_freed;
975 shmem_recalc_inode(inode);
976 spin_unlock_irq(&info->lock);
977}
978
979void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
980{
981 shmem_undo_range(inode, lstart, lend, false);
982 inode->i_ctime = inode->i_mtime = current_time(inode);
983}
984EXPORT_SYMBOL_GPL(shmem_truncate_range);
985
986static int shmem_getattr(const struct path *path, struct kstat *stat,
987 u32 request_mask, unsigned int query_flags)
988{
989 struct inode *inode = path->dentry->d_inode;
990 struct shmem_inode_info *info = SHMEM_I(inode);
991
992 if (info->alloced - info->swapped != inode->i_mapping->nrpages) {
993 spin_lock_irq(&info->lock);
994 shmem_recalc_inode(inode);
995 spin_unlock_irq(&info->lock);
996 }
997 generic_fillattr(inode, stat);
998 return 0;
999}
1000
1001static int shmem_setattr(struct dentry *dentry, struct iattr *attr)
1002{
1003 struct inode *inode = d_inode(dentry);
1004 struct shmem_inode_info *info = SHMEM_I(inode);
1005 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1006 int error;
1007
1008 error = setattr_prepare(dentry, attr);
1009 if (error)
1010 return error;
1011
1012 if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
1013 loff_t oldsize = inode->i_size;
1014 loff_t newsize = attr->ia_size;
1015
1016 /* protected by i_mutex */
1017 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
1018 (newsize > oldsize && (info->seals & F_SEAL_GROW)))
1019 return -EPERM;
1020
1021 if (newsize != oldsize) {
1022 error = shmem_reacct_size(SHMEM_I(inode)->flags,
1023 oldsize, newsize);
1024 if (error)
1025 return error;
1026 i_size_write(inode, newsize);
1027 inode->i_ctime = inode->i_mtime = current_time(inode);
1028 }
1029 if (newsize <= oldsize) {
1030 loff_t holebegin = round_up(newsize, PAGE_SIZE);
1031 if (oldsize > holebegin)
1032 unmap_mapping_range(inode->i_mapping,
1033 holebegin, 0, 1);
1034 if (info->alloced)
1035 shmem_truncate_range(inode,
1036 newsize, (loff_t)-1);
1037 /* unmap again to remove racily COWed private pages */
1038 if (oldsize > holebegin)
1039 unmap_mapping_range(inode->i_mapping,
1040 holebegin, 0, 1);
1041
1042 /*
1043 * Part of the huge page can be beyond i_size: subject
1044 * to shrink under memory pressure.
1045 */
1046 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
1047 spin_lock(&sbinfo->shrinklist_lock);
1048 /*
1049 * _careful to defend against unlocked access to
1050 * ->shrink_list in shmem_unused_huge_shrink()
1051 */
1052 if (list_empty_careful(&info->shrinklist)) {
1053 list_add_tail(&info->shrinklist,
1054 &sbinfo->shrinklist);
1055 sbinfo->shrinklist_len++;
1056 }
1057 spin_unlock(&sbinfo->shrinklist_lock);
1058 }
1059 }
1060 }
1061
1062 setattr_copy(inode, attr);
1063 if (attr->ia_valid & ATTR_MODE)
1064 error = posix_acl_chmod(inode, inode->i_mode);
1065 return error;
1066}
1067
1068static void shmem_evict_inode(struct inode *inode)
1069{
1070 struct shmem_inode_info *info = SHMEM_I(inode);
1071 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1072
1073 if (inode->i_mapping->a_ops == &shmem_aops) {
1074 shmem_unacct_size(info->flags, inode->i_size);
1075 inode->i_size = 0;
1076 shmem_truncate_range(inode, 0, (loff_t)-1);
1077 if (!list_empty(&info->shrinklist)) {
1078 spin_lock(&sbinfo->shrinklist_lock);
1079 if (!list_empty(&info->shrinklist)) {
1080 list_del_init(&info->shrinklist);
1081 sbinfo->shrinklist_len--;
1082 }
1083 spin_unlock(&sbinfo->shrinklist_lock);
1084 }
1085 if (!list_empty(&info->swaplist)) {
1086 mutex_lock(&shmem_swaplist_mutex);
1087 list_del_init(&info->swaplist);
1088 mutex_unlock(&shmem_swaplist_mutex);
1089 }
1090 }
1091
1092 simple_xattrs_free(&info->xattrs);
1093 WARN_ON(inode->i_blocks);
1094 shmem_free_inode(inode->i_sb);
1095 clear_inode(inode);
1096}
1097
1098static unsigned long find_swap_entry(struct radix_tree_root *root, void *item)
1099{
1100 struct radix_tree_iter iter;
1101 void **slot;
1102 unsigned long found = -1;
1103 unsigned int checked = 0;
1104
1105 rcu_read_lock();
1106 radix_tree_for_each_slot(slot, root, &iter, 0) {
1107 if (*slot == item) {
1108 found = iter.index;
1109 break;
1110 }
1111 checked++;
1112 if ((checked % 4096) != 0)
1113 continue;
1114 slot = radix_tree_iter_resume(slot, &iter);
1115 cond_resched_rcu();
1116 }
1117
1118 rcu_read_unlock();
1119 return found;
1120}
1121
1122/*
1123 * If swap found in inode, free it and move page from swapcache to filecache.
1124 */
1125static int shmem_unuse_inode(struct shmem_inode_info *info,
1126 swp_entry_t swap, struct page **pagep)
1127{
1128 struct address_space *mapping = info->vfs_inode.i_mapping;
1129 void *radswap;
1130 pgoff_t index;
1131 gfp_t gfp;
1132 int error = 0;
1133
1134 radswap = swp_to_radix_entry(swap);
1135 index = find_swap_entry(&mapping->i_pages, radswap);
1136 if (index == -1)
1137 return -EAGAIN; /* tell shmem_unuse we found nothing */
1138
1139 /*
1140 * Move _head_ to start search for next from here.
1141 * But be careful: shmem_evict_inode checks list_empty without taking
1142 * mutex, and there's an instant in list_move_tail when info->swaplist
1143 * would appear empty, if it were the only one on shmem_swaplist.
1144 */
1145 if (shmem_swaplist.next != &info->swaplist)
1146 list_move_tail(&shmem_swaplist, &info->swaplist);
1147
1148 gfp = mapping_gfp_mask(mapping);
1149 if (shmem_should_replace_page(*pagep, gfp)) {
1150 mutex_unlock(&shmem_swaplist_mutex);
1151 error = shmem_replace_page(pagep, gfp, info, index);
1152 mutex_lock(&shmem_swaplist_mutex);
1153 /*
1154 * We needed to drop mutex to make that restrictive page
1155 * allocation, but the inode might have been freed while we
1156 * dropped it: although a racing shmem_evict_inode() cannot
1157 * complete without emptying the radix_tree, our page lock
1158 * on this swapcache page is not enough to prevent that -
1159 * free_swap_and_cache() of our swap entry will only
1160 * trylock_page(), removing swap from radix_tree whatever.
1161 *
1162 * We must not proceed to shmem_add_to_page_cache() if the
1163 * inode has been freed, but of course we cannot rely on
1164 * inode or mapping or info to check that. However, we can
1165 * safely check if our swap entry is still in use (and here
1166 * it can't have got reused for another page): if it's still
1167 * in use, then the inode cannot have been freed yet, and we
1168 * can safely proceed (if it's no longer in use, that tells
1169 * nothing about the inode, but we don't need to unuse swap).
1170 */
1171 if (!page_swapcount(*pagep))
1172 error = -ENOENT;
1173 }
1174
1175 /*
1176 * We rely on shmem_swaplist_mutex, not only to protect the swaplist,
1177 * but also to hold up shmem_evict_inode(): so inode cannot be freed
1178 * beneath us (pagelock doesn't help until the page is in pagecache).
1179 */
1180 if (!error)
1181 error = shmem_add_to_page_cache(*pagep, mapping, index,
1182 radswap);
1183 if (error != -ENOMEM) {
1184 /*
1185 * Truncation and eviction use free_swap_and_cache(), which
1186 * only does trylock page: if we raced, best clean up here.
1187 */
1188 delete_from_swap_cache(*pagep);
1189 set_page_dirty(*pagep);
1190 if (!error) {
1191 spin_lock_irq(&info->lock);
1192 info->swapped--;
1193 spin_unlock_irq(&info->lock);
1194 swap_free(swap);
1195 }
1196 }
1197 return error;
1198}
1199
1200/*
1201 * Search through swapped inodes to find and replace swap by page.
1202 */
1203int shmem_unuse(swp_entry_t swap, struct page *page)
1204{
1205 struct list_head *this, *next;
1206 struct shmem_inode_info *info;
1207 struct mem_cgroup *memcg;
1208 int error = 0;
1209
1210 /*
1211 * There's a faint possibility that swap page was replaced before
1212 * caller locked it: caller will come back later with the right page.
1213 */
1214 if (unlikely(!PageSwapCache(page) || page_private(page) != swap.val))
1215 goto out;
1216
1217 /*
1218 * Charge page using GFP_KERNEL while we can wait, before taking
1219 * the shmem_swaplist_mutex which might hold up shmem_writepage().
1220 * Charged back to the user (not to caller) when swap account is used.
1221 */
1222 error = mem_cgroup_try_charge(page, current->mm, GFP_KERNEL, &memcg,
1223 false);
1224 if (error)
1225 goto out;
1226 /* No radix_tree_preload: swap entry keeps a place for page in tree */
1227 error = -EAGAIN;
1228
1229 mutex_lock(&shmem_swaplist_mutex);
1230 list_for_each_safe(this, next, &shmem_swaplist) {
1231 info = list_entry(this, struct shmem_inode_info, swaplist);
1232 if (info->swapped)
1233 error = shmem_unuse_inode(info, swap, &page);
1234 else
1235 list_del_init(&info->swaplist);
1236 cond_resched();
1237 if (error != -EAGAIN)
1238 break;
1239 /* found nothing in this: move on to search the next */
1240 }
1241 mutex_unlock(&shmem_swaplist_mutex);
1242
1243 if (error) {
1244 if (error != -ENOMEM)
1245 error = 0;
1246 mem_cgroup_cancel_charge(page, memcg, false);
1247 } else
1248 mem_cgroup_commit_charge(page, memcg, true, false);
1249out:
1250 unlock_page(page);
1251 put_page(page);
1252 return error;
1253}
1254
1255/*
1256 * Move the page from the page cache to the swap cache.
1257 */
1258static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1259{
1260 struct shmem_inode_info *info;
1261 struct address_space *mapping;
1262 struct inode *inode;
1263 swp_entry_t swap;
1264 pgoff_t index;
1265
1266 VM_BUG_ON_PAGE(PageCompound(page), page);
1267 BUG_ON(!PageLocked(page));
1268 mapping = page->mapping;
1269 index = page->index;
1270 inode = mapping->host;
1271 info = SHMEM_I(inode);
1272 if (info->flags & VM_LOCKED)
1273 goto redirty;
1274 if (!total_swap_pages)
1275 goto redirty;
1276
1277 /*
1278 * Our capabilities prevent regular writeback or sync from ever calling
1279 * shmem_writepage; but a stacking filesystem might use ->writepage of
1280 * its underlying filesystem, in which case tmpfs should write out to
1281 * swap only in response to memory pressure, and not for the writeback
1282 * threads or sync.
1283 */
1284 if (!wbc->for_reclaim) {
1285 WARN_ON_ONCE(1); /* Still happens? Tell us about it! */
1286 goto redirty;
1287 }
1288
1289 /*
1290 * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
1291 * value into swapfile.c, the only way we can correctly account for a
1292 * fallocated page arriving here is now to initialize it and write it.
1293 *
1294 * That's okay for a page already fallocated earlier, but if we have
1295 * not yet completed the fallocation, then (a) we want to keep track
1296 * of this page in case we have to undo it, and (b) it may not be a
1297 * good idea to continue anyway, once we're pushing into swap. So
1298 * reactivate the page, and let shmem_fallocate() quit when too many.
1299 */
1300 if (!PageUptodate(page)) {
1301 if (inode->i_private) {
1302 struct shmem_falloc *shmem_falloc;
1303 spin_lock(&inode->i_lock);
1304 shmem_falloc = inode->i_private;
1305 if (shmem_falloc &&
1306 !shmem_falloc->waitq &&
1307 index >= shmem_falloc->start &&
1308 index < shmem_falloc->next)
1309 shmem_falloc->nr_unswapped++;
1310 else
1311 shmem_falloc = NULL;
1312 spin_unlock(&inode->i_lock);
1313 if (shmem_falloc)
1314 goto redirty;
1315 }
1316 clear_highpage(page);
1317 flush_dcache_page(page);
1318 SetPageUptodate(page);
1319 }
1320
1321 swap = get_swap_page(page);
1322 if (!swap.val)
1323 goto redirty;
1324
1325 if (mem_cgroup_try_charge_swap(page, swap))
1326 goto free_swap;
1327
1328 /*
1329 * Add inode to shmem_unuse()'s list of swapped-out inodes,
1330 * if it's not already there. Do it now before the page is
1331 * moved to swap cache, when its pagelock no longer protects
1332 * the inode from eviction. But don't unlock the mutex until
1333 * we've incremented swapped, because shmem_unuse_inode() will
1334 * prune a !swapped inode from the swaplist under this mutex.
1335 */
1336 mutex_lock(&shmem_swaplist_mutex);
1337 if (list_empty(&info->swaplist))
1338 list_add_tail(&info->swaplist, &shmem_swaplist);
1339
1340 if (add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
1341 spin_lock_irq(&info->lock);
1342 shmem_recalc_inode(inode);
1343 info->swapped++;
1344 spin_unlock_irq(&info->lock);
1345
1346 swap_shmem_alloc(swap);
1347 shmem_delete_from_page_cache(page, swp_to_radix_entry(swap));
1348
1349 mutex_unlock(&shmem_swaplist_mutex);
1350 BUG_ON(page_mapped(page));
1351 swap_writepage(page, wbc);
1352 return 0;
1353 }
1354
1355 mutex_unlock(&shmem_swaplist_mutex);
1356free_swap:
1357 put_swap_page(page, swap);
1358redirty:
1359 set_page_dirty(page);
1360 if (wbc->for_reclaim)
1361 return AOP_WRITEPAGE_ACTIVATE; /* Return with page locked */
1362 unlock_page(page);
1363 return 0;
1364}
1365
1366#if defined(CONFIG_NUMA) && defined(CONFIG_TMPFS)
1367static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1368{
1369 char buffer[64];
1370
1371 if (!mpol || mpol->mode == MPOL_DEFAULT)
1372 return; /* show nothing */
1373
1374 mpol_to_str(buffer, sizeof(buffer), mpol);
1375
1376 seq_printf(seq, ",mpol=%s", buffer);
1377}
1378
1379static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1380{
1381 struct mempolicy *mpol = NULL;
1382 if (sbinfo->mpol) {
1383 spin_lock(&sbinfo->stat_lock); /* prevent replace/use races */
1384 mpol = sbinfo->mpol;
1385 mpol_get(mpol);
1386 spin_unlock(&sbinfo->stat_lock);
1387 }
1388 return mpol;
1389}
1390#else /* !CONFIG_NUMA || !CONFIG_TMPFS */
1391static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1392{
1393}
1394static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1395{
1396 return NULL;
1397}
1398#endif /* CONFIG_NUMA && CONFIG_TMPFS */
1399#ifndef CONFIG_NUMA
1400#define vm_policy vm_private_data
1401#endif
1402
1403static void shmem_pseudo_vma_init(struct vm_area_struct *vma,
1404 struct shmem_inode_info *info, pgoff_t index)
1405{
1406 /* Create a pseudo vma that just contains the policy */
1407 vma->vm_start = 0;
1408 /* Bias interleave by inode number to distribute better across nodes */
1409 vma->vm_pgoff = index + info->vfs_inode.i_ino;
1410 vma->vm_ops = NULL;
1411 vma->vm_policy = mpol_shared_policy_lookup(&info->policy, index);
1412}
1413
1414static void shmem_pseudo_vma_destroy(struct vm_area_struct *vma)
1415{
1416 /* Drop reference taken by mpol_shared_policy_lookup() */
1417 mpol_cond_put(vma->vm_policy);
1418}
1419
1420static struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
1421 struct shmem_inode_info *info, pgoff_t index)
1422{
1423 struct vm_area_struct pvma;
1424 struct page *page;
1425 struct vm_fault vmf;
1426
1427 shmem_pseudo_vma_init(&pvma, info, index);
1428 vmf.vma = &pvma;
1429 vmf.address = 0;
1430 page = swap_cluster_readahead(swap, gfp, &vmf);
1431 shmem_pseudo_vma_destroy(&pvma);
1432
1433 return page;
1434}
1435
1436static struct page *shmem_alloc_hugepage(gfp_t gfp,
1437 struct shmem_inode_info *info, pgoff_t index)
1438{
1439 struct vm_area_struct pvma;
1440 struct inode *inode = &info->vfs_inode;
1441 struct address_space *mapping = inode->i_mapping;
1442 pgoff_t idx, hindex;
1443 void __rcu **results;
1444 struct page *page;
1445
1446 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
1447 return NULL;
1448
1449 hindex = round_down(index, HPAGE_PMD_NR);
1450 rcu_read_lock();
1451 if (radix_tree_gang_lookup_slot(&mapping->i_pages, &results, &idx,
1452 hindex, 1) && idx < hindex + HPAGE_PMD_NR) {
1453 rcu_read_unlock();
1454 return NULL;
1455 }
1456 rcu_read_unlock();
1457
1458 shmem_pseudo_vma_init(&pvma, info, hindex);
1459 page = alloc_pages_vma(gfp | __GFP_COMP | __GFP_NORETRY | __GFP_NOWARN,
1460 HPAGE_PMD_ORDER, &pvma, 0, numa_node_id(), true);
1461 shmem_pseudo_vma_destroy(&pvma);
1462 if (page)
1463 prep_transhuge_page(page);
1464 return page;
1465}
1466
1467static struct page *shmem_alloc_page(gfp_t gfp,
1468 struct shmem_inode_info *info, pgoff_t index)
1469{
1470 struct vm_area_struct pvma;
1471 struct page *page;
1472
1473 shmem_pseudo_vma_init(&pvma, info, index);
1474 page = alloc_page_vma(gfp, &pvma, 0);
1475 shmem_pseudo_vma_destroy(&pvma);
1476
1477 return page;
1478}
1479
1480static struct page *shmem_alloc_and_acct_page(gfp_t gfp,
1481 struct inode *inode,
1482 pgoff_t index, bool huge)
1483{
1484 struct shmem_inode_info *info = SHMEM_I(inode);
1485 struct page *page;
1486 int nr;
1487 int err = -ENOSPC;
1488
1489 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
1490 huge = false;
1491 nr = huge ? HPAGE_PMD_NR : 1;
1492
1493 if (!shmem_inode_acct_block(inode, nr))
1494 goto failed;
1495
1496 if (huge)
1497 page = shmem_alloc_hugepage(gfp, info, index);
1498 else
1499 page = shmem_alloc_page(gfp, info, index);
1500 if (page) {
1501 __SetPageLocked(page);
1502 __SetPageSwapBacked(page);
1503 return page;
1504 }
1505
1506 err = -ENOMEM;
1507 shmem_inode_unacct_blocks(inode, nr);
1508failed:
1509 return ERR_PTR(err);
1510}
1511
1512/*
1513 * When a page is moved from swapcache to shmem filecache (either by the
1514 * usual swapin of shmem_getpage_gfp(), or by the less common swapoff of
1515 * shmem_unuse_inode()), it may have been read in earlier from swap, in
1516 * ignorance of the mapping it belongs to. If that mapping has special
1517 * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
1518 * we may need to copy to a suitable page before moving to filecache.
1519 *
1520 * In a future release, this may well be extended to respect cpuset and
1521 * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
1522 * but for now it is a simple matter of zone.
1523 */
1524static bool shmem_should_replace_page(struct page *page, gfp_t gfp)
1525{
1526 return page_zonenum(page) > gfp_zone(gfp);
1527}
1528
1529static int shmem_replace_page(struct page **pagep, gfp_t gfp,
1530 struct shmem_inode_info *info, pgoff_t index)
1531{
1532 struct page *oldpage, *newpage;
1533 struct address_space *swap_mapping;
1534 pgoff_t swap_index;
1535 int error;
1536
1537 oldpage = *pagep;
1538 swap_index = page_private(oldpage);
1539 swap_mapping = page_mapping(oldpage);
1540
1541 /*
1542 * We have arrived here because our zones are constrained, so don't
1543 * limit chance of success by further cpuset and node constraints.
1544 */
1545 gfp &= ~GFP_CONSTRAINT_MASK;
1546 newpage = shmem_alloc_page(gfp, info, index);
1547 if (!newpage)
1548 return -ENOMEM;
1549
1550 get_page(newpage);
1551 copy_highpage(newpage, oldpage);
1552 flush_dcache_page(newpage);
1553
1554 __SetPageLocked(newpage);
1555 __SetPageSwapBacked(newpage);
1556 SetPageUptodate(newpage);
1557 set_page_private(newpage, swap_index);
1558 SetPageSwapCache(newpage);
1559
1560 /*
1561 * Our caller will very soon move newpage out of swapcache, but it's
1562 * a nice clean interface for us to replace oldpage by newpage there.
1563 */
1564 xa_lock_irq(&swap_mapping->i_pages);
1565 error = shmem_radix_tree_replace(swap_mapping, swap_index, oldpage,
1566 newpage);
1567 if (!error) {
1568 __inc_node_page_state(newpage, NR_FILE_PAGES);
1569 __dec_node_page_state(oldpage, NR_FILE_PAGES);
1570 }
1571 xa_unlock_irq(&swap_mapping->i_pages);
1572
1573 if (unlikely(error)) {
1574 /*
1575 * Is this possible? I think not, now that our callers check
1576 * both PageSwapCache and page_private after getting page lock;
1577 * but be defensive. Reverse old to newpage for clear and free.
1578 */
1579 oldpage = newpage;
1580 } else {
1581 mem_cgroup_migrate(oldpage, newpage);
1582 lru_cache_add_anon(newpage);
1583 *pagep = newpage;
1584 }
1585
1586 ClearPageSwapCache(oldpage);
1587 set_page_private(oldpage, 0);
1588
1589 unlock_page(oldpage);
1590 put_page(oldpage);
1591 put_page(oldpage);
1592 return error;
1593}
1594
1595/*
1596 * shmem_getpage_gfp - find page in cache, or get from swap, or allocate
1597 *
1598 * If we allocate a new one we do not mark it dirty. That's up to the
1599 * vm. If we swap it in we mark it dirty since we also free the swap
1600 * entry since a page cannot live in both the swap and page cache.
1601 *
1602 * fault_mm and fault_type are only supplied by shmem_fault:
1603 * otherwise they are NULL.
1604 */
1605static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
1606 struct page **pagep, enum sgp_type sgp, gfp_t gfp,
1607 struct vm_area_struct *vma, struct vm_fault *vmf, int *fault_type)
1608{
1609 struct address_space *mapping = inode->i_mapping;
1610 struct shmem_inode_info *info = SHMEM_I(inode);
1611 struct shmem_sb_info *sbinfo;
1612 struct mm_struct *charge_mm;
1613 struct mem_cgroup *memcg;
1614 struct page *page;
1615 swp_entry_t swap;
1616 enum sgp_type sgp_huge = sgp;
1617 pgoff_t hindex = index;
1618 int error;
1619 int once = 0;
1620 int alloced = 0;
1621
1622 if (index > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
1623 return -EFBIG;
1624 if (sgp == SGP_NOHUGE || sgp == SGP_HUGE)
1625 sgp = SGP_CACHE;
1626repeat:
1627 swap.val = 0;
1628 page = find_lock_entry(mapping, index);
1629 if (radix_tree_exceptional_entry(page)) {
1630 swap = radix_to_swp_entry(page);
1631 page = NULL;
1632 }
1633
1634 if (sgp <= SGP_CACHE &&
1635 ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1636 error = -EINVAL;
1637 goto unlock;
1638 }
1639
1640 if (page && sgp == SGP_WRITE)
1641 mark_page_accessed(page);
1642
1643 /* fallocated page? */
1644 if (page && !PageUptodate(page)) {
1645 if (sgp != SGP_READ)
1646 goto clear;
1647 unlock_page(page);
1648 put_page(page);
1649 page = NULL;
1650 }
1651 if (page || (sgp == SGP_READ && !swap.val)) {
1652 *pagep = page;
1653 return 0;
1654 }
1655
1656 /*
1657 * Fast cache lookup did not find it:
1658 * bring it back from swap or allocate.
1659 */
1660 sbinfo = SHMEM_SB(inode->i_sb);
1661 charge_mm = vma ? vma->vm_mm : current->mm;
1662
1663 if (swap.val) {
1664 /* Look it up and read it in.. */
1665 page = lookup_swap_cache(swap, NULL, 0);
1666 if (!page) {
1667 /* Or update major stats only when swapin succeeds?? */
1668 if (fault_type) {
1669 *fault_type |= VM_FAULT_MAJOR;
1670 count_vm_event(PGMAJFAULT);
1671 count_memcg_event_mm(charge_mm, PGMAJFAULT);
1672 }
1673 /* Here we actually start the io */
1674 page = shmem_swapin(swap, gfp, info, index);
1675 if (!page) {
1676 error = -ENOMEM;
1677 goto failed;
1678 }
1679 }
1680
1681 /* We have to do this with page locked to prevent races */
1682 lock_page(page);
1683 if (!PageSwapCache(page) || page_private(page) != swap.val ||
1684 !shmem_confirm_swap(mapping, index, swap)) {
1685 error = -EEXIST; /* try again */
1686 goto unlock;
1687 }
1688 if (!PageUptodate(page)) {
1689 error = -EIO;
1690 goto failed;
1691 }
1692 wait_on_page_writeback(page);
1693
1694 if (shmem_should_replace_page(page, gfp)) {
1695 error = shmem_replace_page(&page, gfp, info, index);
1696 if (error)
1697 goto failed;
1698 }
1699
1700 error = mem_cgroup_try_charge(page, charge_mm, gfp, &memcg,
1701 false);
1702 if (!error) {
1703 error = shmem_add_to_page_cache(page, mapping, index,
1704 swp_to_radix_entry(swap));
1705 /*
1706 * We already confirmed swap under page lock, and make
1707 * no memory allocation here, so usually no possibility
1708 * of error; but free_swap_and_cache() only trylocks a
1709 * page, so it is just possible that the entry has been
1710 * truncated or holepunched since swap was confirmed.
1711 * shmem_undo_range() will have done some of the
1712 * unaccounting, now delete_from_swap_cache() will do
1713 * the rest.
1714 * Reset swap.val? No, leave it so "failed" goes back to
1715 * "repeat": reading a hole and writing should succeed.
1716 */
1717 if (error) {
1718 mem_cgroup_cancel_charge(page, memcg, false);
1719 delete_from_swap_cache(page);
1720 }
1721 }
1722 if (error)
1723 goto failed;
1724
1725 mem_cgroup_commit_charge(page, memcg, true, false);
1726
1727 spin_lock_irq(&info->lock);
1728 info->swapped--;
1729 shmem_recalc_inode(inode);
1730 spin_unlock_irq(&info->lock);
1731
1732 if (sgp == SGP_WRITE)
1733 mark_page_accessed(page);
1734
1735 delete_from_swap_cache(page);
1736 set_page_dirty(page);
1737 swap_free(swap);
1738
1739 } else {
1740 if (vma && userfaultfd_missing(vma)) {
1741 *fault_type = handle_userfault(vmf, VM_UFFD_MISSING);
1742 return 0;
1743 }
1744
1745 /* shmem_symlink() */
1746 if (mapping->a_ops != &shmem_aops)
1747 goto alloc_nohuge;
1748 if (shmem_huge == SHMEM_HUGE_DENY || sgp_huge == SGP_NOHUGE)
1749 goto alloc_nohuge;
1750 if (shmem_huge == SHMEM_HUGE_FORCE)
1751 goto alloc_huge;
1752 switch (sbinfo->huge) {
1753 loff_t i_size;
1754 pgoff_t off;
1755 case SHMEM_HUGE_NEVER:
1756 goto alloc_nohuge;
1757 case SHMEM_HUGE_WITHIN_SIZE:
1758 off = round_up(index, HPAGE_PMD_NR);
1759 i_size = round_up(i_size_read(inode), PAGE_SIZE);
1760 if (i_size >= HPAGE_PMD_SIZE &&
1761 i_size >> PAGE_SHIFT >= off)
1762 goto alloc_huge;
1763 /* fallthrough */
1764 case SHMEM_HUGE_ADVISE:
1765 if (sgp_huge == SGP_HUGE)
1766 goto alloc_huge;
1767 /* TODO: implement fadvise() hints */
1768 goto alloc_nohuge;
1769 }
1770
1771alloc_huge:
1772 page = shmem_alloc_and_acct_page(gfp, inode, index, true);
1773 if (IS_ERR(page)) {
1774alloc_nohuge: page = shmem_alloc_and_acct_page(gfp, inode,
1775 index, false);
1776 }
1777 if (IS_ERR(page)) {
1778 int retry = 5;
1779 error = PTR_ERR(page);
1780 page = NULL;
1781 if (error != -ENOSPC)
1782 goto failed;
1783 /*
1784 * Try to reclaim some spece by splitting a huge page
1785 * beyond i_size on the filesystem.
1786 */
1787 while (retry--) {
1788 int ret;
1789 ret = shmem_unused_huge_shrink(sbinfo, NULL, 1);
1790 if (ret == SHRINK_STOP)
1791 break;
1792 if (ret)
1793 goto alloc_nohuge;
1794 }
1795 goto failed;
1796 }
1797
1798 if (PageTransHuge(page))
1799 hindex = round_down(index, HPAGE_PMD_NR);
1800 else
1801 hindex = index;
1802
1803 if (sgp == SGP_WRITE)
1804 __SetPageReferenced(page);
1805
1806 error = mem_cgroup_try_charge(page, charge_mm, gfp, &memcg,
1807 PageTransHuge(page));
1808 if (error)
1809 goto unacct;
1810 error = radix_tree_maybe_preload_order(gfp & GFP_RECLAIM_MASK,
1811 compound_order(page));
1812 if (!error) {
1813 error = shmem_add_to_page_cache(page, mapping, hindex,
1814 NULL);
1815 radix_tree_preload_end();
1816 }
1817 if (error) {
1818 mem_cgroup_cancel_charge(page, memcg,
1819 PageTransHuge(page));
1820 goto unacct;
1821 }
1822 mem_cgroup_commit_charge(page, memcg, false,
1823 PageTransHuge(page));
1824 lru_cache_add_anon(page);
1825
1826 spin_lock_irq(&info->lock);
1827 info->alloced += 1 << compound_order(page);
1828 inode->i_blocks += BLOCKS_PER_PAGE << compound_order(page);
1829 shmem_recalc_inode(inode);
1830 spin_unlock_irq(&info->lock);
1831 alloced = true;
1832
1833 if (PageTransHuge(page) &&
1834 DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE) <
1835 hindex + HPAGE_PMD_NR - 1) {
1836 /*
1837 * Part of the huge page is beyond i_size: subject
1838 * to shrink under memory pressure.
1839 */
1840 spin_lock(&sbinfo->shrinklist_lock);
1841 /*
1842 * _careful to defend against unlocked access to
1843 * ->shrink_list in shmem_unused_huge_shrink()
1844 */
1845 if (list_empty_careful(&info->shrinklist)) {
1846 list_add_tail(&info->shrinklist,
1847 &sbinfo->shrinklist);
1848 sbinfo->shrinklist_len++;
1849 }
1850 spin_unlock(&sbinfo->shrinklist_lock);
1851 }
1852
1853 /*
1854 * Let SGP_FALLOC use the SGP_WRITE optimization on a new page.
1855 */
1856 if (sgp == SGP_FALLOC)
1857 sgp = SGP_WRITE;
1858clear:
1859 /*
1860 * Let SGP_WRITE caller clear ends if write does not fill page;
1861 * but SGP_FALLOC on a page fallocated earlier must initialize
1862 * it now, lest undo on failure cancel our earlier guarantee.
1863 */
1864 if (sgp != SGP_WRITE && !PageUptodate(page)) {
1865 struct page *head = compound_head(page);
1866 int i;
1867
1868 for (i = 0; i < (1 << compound_order(head)); i++) {
1869 clear_highpage(head + i);
1870 flush_dcache_page(head + i);
1871 }
1872 SetPageUptodate(head);
1873 }
1874 }
1875
1876 /* Perhaps the file has been truncated since we checked */
1877 if (sgp <= SGP_CACHE &&
1878 ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1879 if (alloced) {
1880 ClearPageDirty(page);
1881 delete_from_page_cache(page);
1882 spin_lock_irq(&info->lock);
1883 shmem_recalc_inode(inode);
1884 spin_unlock_irq(&info->lock);
1885 }
1886 error = -EINVAL;
1887 goto unlock;
1888 }
1889 *pagep = page + index - hindex;
1890 return 0;
1891
1892 /*
1893 * Error recovery.
1894 */
1895unacct:
1896 shmem_inode_unacct_blocks(inode, 1 << compound_order(page));
1897
1898 if (PageTransHuge(page)) {
1899 unlock_page(page);
1900 put_page(page);
1901 goto alloc_nohuge;
1902 }
1903failed:
1904 if (swap.val && !shmem_confirm_swap(mapping, index, swap))
1905 error = -EEXIST;
1906unlock:
1907 if (page) {
1908 unlock_page(page);
1909 put_page(page);
1910 }
1911 if (error == -ENOSPC && !once++) {
1912 spin_lock_irq(&info->lock);
1913 shmem_recalc_inode(inode);
1914 spin_unlock_irq(&info->lock);
1915 goto repeat;
1916 }
1917 if (error == -EEXIST) /* from above or from radix_tree_insert */
1918 goto repeat;
1919 return error;
1920}
1921
1922/*
1923 * This is like autoremove_wake_function, but it removes the wait queue
1924 * entry unconditionally - even if something else had already woken the
1925 * target.
1926 */
1927static int synchronous_wake_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
1928{
1929 int ret = default_wake_function(wait, mode, sync, key);
1930 list_del_init(&wait->entry);
1931 return ret;
1932}
1933
1934static int shmem_fault(struct vm_fault *vmf)
1935{
1936 struct vm_area_struct *vma = vmf->vma;
1937 struct inode *inode = file_inode(vma->vm_file);
1938 gfp_t gfp = mapping_gfp_mask(inode->i_mapping);
1939 enum sgp_type sgp;
1940 int error;
1941 int ret = VM_FAULT_LOCKED;
1942
1943 /*
1944 * Trinity finds that probing a hole which tmpfs is punching can
1945 * prevent the hole-punch from ever completing: which in turn
1946 * locks writers out with its hold on i_mutex. So refrain from
1947 * faulting pages into the hole while it's being punched. Although
1948 * shmem_undo_range() does remove the additions, it may be unable to
1949 * keep up, as each new page needs its own unmap_mapping_range() call,
1950 * and the i_mmap tree grows ever slower to scan if new vmas are added.
1951 *
1952 * It does not matter if we sometimes reach this check just before the
1953 * hole-punch begins, so that one fault then races with the punch:
1954 * we just need to make racing faults a rare case.
1955 *
1956 * The implementation below would be much simpler if we just used a
1957 * standard mutex or completion: but we cannot take i_mutex in fault,
1958 * and bloating every shmem inode for this unlikely case would be sad.
1959 */
1960 if (unlikely(inode->i_private)) {
1961 struct shmem_falloc *shmem_falloc;
1962
1963 spin_lock(&inode->i_lock);
1964 shmem_falloc = inode->i_private;
1965 if (shmem_falloc &&
1966 shmem_falloc->waitq &&
1967 vmf->pgoff >= shmem_falloc->start &&
1968 vmf->pgoff < shmem_falloc->next) {
1969 wait_queue_head_t *shmem_falloc_waitq;
1970 DEFINE_WAIT_FUNC(shmem_fault_wait, synchronous_wake_function);
1971
1972 ret = VM_FAULT_NOPAGE;
1973 if ((vmf->flags & FAULT_FLAG_ALLOW_RETRY) &&
1974 !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
1975 /* It's polite to up mmap_sem if we can */
1976 up_read(&vma->vm_mm->mmap_sem);
1977 ret = VM_FAULT_RETRY;
1978 }
1979
1980 shmem_falloc_waitq = shmem_falloc->waitq;
1981 prepare_to_wait(shmem_falloc_waitq, &shmem_fault_wait,
1982 TASK_UNINTERRUPTIBLE);
1983 spin_unlock(&inode->i_lock);
1984 schedule();
1985
1986 /*
1987 * shmem_falloc_waitq points into the shmem_fallocate()
1988 * stack of the hole-punching task: shmem_falloc_waitq
1989 * is usually invalid by the time we reach here, but
1990 * finish_wait() does not dereference it in that case;
1991 * though i_lock needed lest racing with wake_up_all().
1992 */
1993 spin_lock(&inode->i_lock);
1994 finish_wait(shmem_falloc_waitq, &shmem_fault_wait);
1995 spin_unlock(&inode->i_lock);
1996 return ret;
1997 }
1998 spin_unlock(&inode->i_lock);
1999 }
2000
2001 sgp = SGP_CACHE;
2002
2003 if ((vma->vm_flags & VM_NOHUGEPAGE) ||
2004 test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
2005 sgp = SGP_NOHUGE;
2006 else if (vma->vm_flags & VM_HUGEPAGE)
2007 sgp = SGP_HUGE;
2008
2009 error = shmem_getpage_gfp(inode, vmf->pgoff, &vmf->page, sgp,
2010 gfp, vma, vmf, &ret);
2011 if (error)
2012 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
2013 return ret;
2014}
2015
2016unsigned long shmem_get_unmapped_area(struct file *file,
2017 unsigned long uaddr, unsigned long len,
2018 unsigned long pgoff, unsigned long flags)
2019{
2020 unsigned long (*get_area)(struct file *,
2021 unsigned long, unsigned long, unsigned long, unsigned long);
2022 unsigned long addr;
2023 unsigned long offset;
2024 unsigned long inflated_len;
2025 unsigned long inflated_addr;
2026 unsigned long inflated_offset;
2027
2028 if (len > TASK_SIZE)
2029 return -ENOMEM;
2030
2031 get_area = current->mm->get_unmapped_area;
2032 addr = get_area(file, uaddr, len, pgoff, flags);
2033
2034 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
2035 return addr;
2036 if (IS_ERR_VALUE(addr))
2037 return addr;
2038 if (addr & ~PAGE_MASK)
2039 return addr;
2040 if (addr > TASK_SIZE - len)
2041 return addr;
2042
2043 if (shmem_huge == SHMEM_HUGE_DENY)
2044 return addr;
2045 if (len < HPAGE_PMD_SIZE)
2046 return addr;
2047 if (flags & MAP_FIXED)
2048 return addr;
2049 /*
2050 * Our priority is to support MAP_SHARED mapped hugely;
2051 * and support MAP_PRIVATE mapped hugely too, until it is COWed.
2052 * But if caller specified an address hint, respect that as before.
2053 */
2054 if (uaddr)
2055 return addr;
2056
2057 if (shmem_huge != SHMEM_HUGE_FORCE) {
2058 struct super_block *sb;
2059
2060 if (file) {
2061 VM_BUG_ON(file->f_op != &shmem_file_operations);
2062 sb = file_inode(file)->i_sb;
2063 } else {
2064 /*
2065 * Called directly from mm/mmap.c, or drivers/char/mem.c
2066 * for "/dev/zero", to create a shared anonymous object.
2067 */
2068 if (IS_ERR(shm_mnt))
2069 return addr;
2070 sb = shm_mnt->mnt_sb;
2071 }
2072 if (SHMEM_SB(sb)->huge == SHMEM_HUGE_NEVER)
2073 return addr;
2074 }
2075
2076 offset = (pgoff << PAGE_SHIFT) & (HPAGE_PMD_SIZE-1);
2077 if (offset && offset + len < 2 * HPAGE_PMD_SIZE)
2078 return addr;
2079 if ((addr & (HPAGE_PMD_SIZE-1)) == offset)
2080 return addr;
2081
2082 inflated_len = len + HPAGE_PMD_SIZE - PAGE_SIZE;
2083 if (inflated_len > TASK_SIZE)
2084 return addr;
2085 if (inflated_len < len)
2086 return addr;
2087
2088 inflated_addr = get_area(NULL, 0, inflated_len, 0, flags);
2089 if (IS_ERR_VALUE(inflated_addr))
2090 return addr;
2091 if (inflated_addr & ~PAGE_MASK)
2092 return addr;
2093
2094 inflated_offset = inflated_addr & (HPAGE_PMD_SIZE-1);
2095 inflated_addr += offset - inflated_offset;
2096 if (inflated_offset > offset)
2097 inflated_addr += HPAGE_PMD_SIZE;
2098
2099 if (inflated_addr > TASK_SIZE - len)
2100 return addr;
2101 return inflated_addr;
2102}
2103
2104#ifdef CONFIG_NUMA
2105static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
2106{
2107 struct inode *inode = file_inode(vma->vm_file);
2108 return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
2109}
2110
2111static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
2112 unsigned long addr)
2113{
2114 struct inode *inode = file_inode(vma->vm_file);
2115 pgoff_t index;
2116
2117 index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2118 return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
2119}
2120#endif
2121
2122int shmem_lock(struct file *file, int lock, struct user_struct *user)
2123{
2124 struct inode *inode = file_inode(file);
2125 struct shmem_inode_info *info = SHMEM_I(inode);
2126 int retval = -ENOMEM;
2127
2128 spin_lock_irq(&info->lock);
2129 if (lock && !(info->flags & VM_LOCKED)) {
2130 if (!user_shm_lock(inode->i_size, user))
2131 goto out_nomem;
2132 info->flags |= VM_LOCKED;
2133 mapping_set_unevictable(file->f_mapping);
2134 }
2135 if (!lock && (info->flags & VM_LOCKED) && user) {
2136 user_shm_unlock(inode->i_size, user);
2137 info->flags &= ~VM_LOCKED;
2138 mapping_clear_unevictable(file->f_mapping);
2139 }
2140 retval = 0;
2141
2142out_nomem:
2143 spin_unlock_irq(&info->lock);
2144 return retval;
2145}
2146
2147static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
2148{
2149 file_accessed(file);
2150 vma->vm_ops = &shmem_vm_ops;
2151 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE) &&
2152 ((vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK) <
2153 (vma->vm_end & HPAGE_PMD_MASK)) {
2154 khugepaged_enter(vma, vma->vm_flags);
2155 }
2156 return 0;
2157}
2158
2159static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
2160 umode_t mode, dev_t dev, unsigned long flags)
2161{
2162 struct inode *inode;
2163 struct shmem_inode_info *info;
2164 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2165
2166 if (shmem_reserve_inode(sb))
2167 return NULL;
2168
2169 inode = new_inode(sb);
2170 if (inode) {
2171 inode->i_ino = get_next_ino();
2172 inode_init_owner(inode, dir, mode);
2173 inode->i_blocks = 0;
2174 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
2175 inode->i_generation = get_seconds();
2176 info = SHMEM_I(inode);
2177 memset(info, 0, (char *)inode - (char *)info);
2178 spin_lock_init(&info->lock);
2179 info->seals = F_SEAL_SEAL;
2180 info->flags = flags & VM_NORESERVE;
2181 INIT_LIST_HEAD(&info->shrinklist);
2182 INIT_LIST_HEAD(&info->swaplist);
2183 simple_xattrs_init(&info->xattrs);
2184 cache_no_acl(inode);
2185
2186 switch (mode & S_IFMT) {
2187 default:
2188 inode->i_op = &shmem_special_inode_operations;
2189 init_special_inode(inode, mode, dev);
2190 break;
2191 case S_IFREG:
2192 inode->i_mapping->a_ops = &shmem_aops;
2193 inode->i_op = &shmem_inode_operations;
2194 inode->i_fop = &shmem_file_operations;
2195 mpol_shared_policy_init(&info->policy,
2196 shmem_get_sbmpol(sbinfo));
2197 break;
2198 case S_IFDIR:
2199 inc_nlink(inode);
2200 /* Some things misbehave if size == 0 on a directory */
2201 inode->i_size = 2 * BOGO_DIRENT_SIZE;
2202 inode->i_op = &shmem_dir_inode_operations;
2203 inode->i_fop = &simple_dir_operations;
2204 break;
2205 case S_IFLNK:
2206 /*
2207 * Must not load anything in the rbtree,
2208 * mpol_free_shared_policy will not be called.
2209 */
2210 mpol_shared_policy_init(&info->policy, NULL);
2211 break;
2212 }
2213 } else
2214 shmem_free_inode(sb);
2215 return inode;
2216}
2217
2218bool shmem_mapping(struct address_space *mapping)
2219{
2220 return mapping->a_ops == &shmem_aops;
2221}
2222
2223static int shmem_mfill_atomic_pte(struct mm_struct *dst_mm,
2224 pmd_t *dst_pmd,
2225 struct vm_area_struct *dst_vma,
2226 unsigned long dst_addr,
2227 unsigned long src_addr,
2228 bool zeropage,
2229 struct page **pagep)
2230{
2231 struct inode *inode = file_inode(dst_vma->vm_file);
2232 struct shmem_inode_info *info = SHMEM_I(inode);
2233 struct address_space *mapping = inode->i_mapping;
2234 gfp_t gfp = mapping_gfp_mask(mapping);
2235 pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
2236 struct mem_cgroup *memcg;
2237 spinlock_t *ptl;
2238 void *page_kaddr;
2239 struct page *page;
2240 pte_t _dst_pte, *dst_pte;
2241 int ret;
2242
2243 ret = -ENOMEM;
2244 if (!shmem_inode_acct_block(inode, 1))
2245 goto out;
2246
2247 if (!*pagep) {
2248 page = shmem_alloc_page(gfp, info, pgoff);
2249 if (!page)
2250 goto out_unacct_blocks;
2251
2252 if (!zeropage) { /* mcopy_atomic */
2253 page_kaddr = kmap_atomic(page);
2254 ret = copy_from_user(page_kaddr,
2255 (const void __user *)src_addr,
2256 PAGE_SIZE);
2257 kunmap_atomic(page_kaddr);
2258
2259 /* fallback to copy_from_user outside mmap_sem */
2260 if (unlikely(ret)) {
2261 *pagep = page;
2262 shmem_inode_unacct_blocks(inode, 1);
2263 /* don't free the page */
2264 return -EFAULT;
2265 }
2266 } else { /* mfill_zeropage_atomic */
2267 clear_highpage(page);
2268 }
2269 } else {
2270 page = *pagep;
2271 *pagep = NULL;
2272 }
2273
2274 VM_BUG_ON(PageLocked(page) || PageSwapBacked(page));
2275 __SetPageLocked(page);
2276 __SetPageSwapBacked(page);
2277 __SetPageUptodate(page);
2278
2279 ret = mem_cgroup_try_charge(page, dst_mm, gfp, &memcg, false);
2280 if (ret)
2281 goto out_release;
2282
2283 ret = radix_tree_maybe_preload(gfp & GFP_RECLAIM_MASK);
2284 if (!ret) {
2285 ret = shmem_add_to_page_cache(page, mapping, pgoff, NULL);
2286 radix_tree_preload_end();
2287 }
2288 if (ret)
2289 goto out_release_uncharge;
2290
2291 mem_cgroup_commit_charge(page, memcg, false, false);
2292
2293 _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
2294 if (dst_vma->vm_flags & VM_WRITE)
2295 _dst_pte = pte_mkwrite(pte_mkdirty(_dst_pte));
2296
2297 ret = -EEXIST;
2298 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
2299 if (!pte_none(*dst_pte))
2300 goto out_release_uncharge_unlock;
2301
2302 lru_cache_add_anon(page);
2303
2304 spin_lock(&info->lock);
2305 info->alloced++;
2306 inode->i_blocks += BLOCKS_PER_PAGE;
2307 shmem_recalc_inode(inode);
2308 spin_unlock(&info->lock);
2309
2310 inc_mm_counter(dst_mm, mm_counter_file(page));
2311 page_add_file_rmap(page, false);
2312 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
2313
2314 /* No need to invalidate - it was non-present before */
2315 update_mmu_cache(dst_vma, dst_addr, dst_pte);
2316 unlock_page(page);
2317 pte_unmap_unlock(dst_pte, ptl);
2318 ret = 0;
2319out:
2320 return ret;
2321out_release_uncharge_unlock:
2322 pte_unmap_unlock(dst_pte, ptl);
2323out_release_uncharge:
2324 mem_cgroup_cancel_charge(page, memcg, false);
2325out_release:
2326 unlock_page(page);
2327 put_page(page);
2328out_unacct_blocks:
2329 shmem_inode_unacct_blocks(inode, 1);
2330 goto out;
2331}
2332
2333int shmem_mcopy_atomic_pte(struct mm_struct *dst_mm,
2334 pmd_t *dst_pmd,
2335 struct vm_area_struct *dst_vma,
2336 unsigned long dst_addr,
2337 unsigned long src_addr,
2338 struct page **pagep)
2339{
2340 return shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
2341 dst_addr, src_addr, false, pagep);
2342}
2343
2344int shmem_mfill_zeropage_pte(struct mm_struct *dst_mm,
2345 pmd_t *dst_pmd,
2346 struct vm_area_struct *dst_vma,
2347 unsigned long dst_addr)
2348{
2349 struct page *page = NULL;
2350
2351 return shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
2352 dst_addr, 0, true, &page);
2353}
2354
2355#ifdef CONFIG_TMPFS
2356static const struct inode_operations shmem_symlink_inode_operations;
2357static const struct inode_operations shmem_short_symlink_operations;
2358
2359#ifdef CONFIG_TMPFS_XATTR
2360static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
2361#else
2362#define shmem_initxattrs NULL
2363#endif
2364
2365static int
2366shmem_write_begin(struct file *file, struct address_space *mapping,
2367 loff_t pos, unsigned len, unsigned flags,
2368 struct page **pagep, void **fsdata)
2369{
2370 struct inode *inode = mapping->host;
2371 struct shmem_inode_info *info = SHMEM_I(inode);
2372 pgoff_t index = pos >> PAGE_SHIFT;
2373
2374 /* i_mutex is held by caller */
2375 if (unlikely(info->seals & (F_SEAL_WRITE | F_SEAL_GROW))) {
2376 if (info->seals & F_SEAL_WRITE)
2377 return -EPERM;
2378 if ((info->seals & F_SEAL_GROW) && pos + len > inode->i_size)
2379 return -EPERM;
2380 }
2381
2382 return shmem_getpage(inode, index, pagep, SGP_WRITE);
2383}
2384
2385static int
2386shmem_write_end(struct file *file, struct address_space *mapping,
2387 loff_t pos, unsigned len, unsigned copied,
2388 struct page *page, void *fsdata)
2389{
2390 struct inode *inode = mapping->host;
2391
2392 if (pos + copied > inode->i_size)
2393 i_size_write(inode, pos + copied);
2394
2395 if (!PageUptodate(page)) {
2396 struct page *head = compound_head(page);
2397 if (PageTransCompound(page)) {
2398 int i;
2399
2400 for (i = 0; i < HPAGE_PMD_NR; i++) {
2401 if (head + i == page)
2402 continue;
2403 clear_highpage(head + i);
2404 flush_dcache_page(head + i);
2405 }
2406 }
2407 if (copied < PAGE_SIZE) {
2408 unsigned from = pos & (PAGE_SIZE - 1);
2409 zero_user_segments(page, 0, from,
2410 from + copied, PAGE_SIZE);
2411 }
2412 SetPageUptodate(head);
2413 }
2414 set_page_dirty(page);
2415 unlock_page(page);
2416 put_page(page);
2417
2418 return copied;
2419}
2420
2421static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
2422{
2423 struct file *file = iocb->ki_filp;
2424 struct inode *inode = file_inode(file);
2425 struct address_space *mapping = inode->i_mapping;
2426 pgoff_t index;
2427 unsigned long offset;
2428 enum sgp_type sgp = SGP_READ;
2429 int error = 0;
2430 ssize_t retval = 0;
2431 loff_t *ppos = &iocb->ki_pos;
2432
2433 /*
2434 * Might this read be for a stacking filesystem? Then when reading
2435 * holes of a sparse file, we actually need to allocate those pages,
2436 * and even mark them dirty, so it cannot exceed the max_blocks limit.
2437 */
2438 if (!iter_is_iovec(to))
2439 sgp = SGP_CACHE;
2440
2441 index = *ppos >> PAGE_SHIFT;
2442 offset = *ppos & ~PAGE_MASK;
2443
2444 for (;;) {
2445 struct page *page = NULL;
2446 pgoff_t end_index;
2447 unsigned long nr, ret;
2448 loff_t i_size = i_size_read(inode);
2449
2450 end_index = i_size >> PAGE_SHIFT;
2451 if (index > end_index)
2452 break;
2453 if (index == end_index) {
2454 nr = i_size & ~PAGE_MASK;
2455 if (nr <= offset)
2456 break;
2457 }
2458
2459 error = shmem_getpage(inode, index, &page, sgp);
2460 if (error) {
2461 if (error == -EINVAL)
2462 error = 0;
2463 break;
2464 }
2465 if (page) {
2466 if (sgp == SGP_CACHE)
2467 set_page_dirty(page);
2468 unlock_page(page);
2469 }
2470
2471 /*
2472 * We must evaluate after, since reads (unlike writes)
2473 * are called without i_mutex protection against truncate
2474 */
2475 nr = PAGE_SIZE;
2476 i_size = i_size_read(inode);
2477 end_index = i_size >> PAGE_SHIFT;
2478 if (index == end_index) {
2479 nr = i_size & ~PAGE_MASK;
2480 if (nr <= offset) {
2481 if (page)
2482 put_page(page);
2483 break;
2484 }
2485 }
2486 nr -= offset;
2487
2488 if (page) {
2489 /*
2490 * If users can be writing to this page using arbitrary
2491 * virtual addresses, take care about potential aliasing
2492 * before reading the page on the kernel side.
2493 */
2494 if (mapping_writably_mapped(mapping))
2495 flush_dcache_page(page);
2496 /*
2497 * Mark the page accessed if we read the beginning.
2498 */
2499 if (!offset)
2500 mark_page_accessed(page);
2501 } else {
2502 page = ZERO_PAGE(0);
2503 get_page(page);
2504 }
2505
2506 /*
2507 * Ok, we have the page, and it's up-to-date, so
2508 * now we can copy it to user space...
2509 */
2510 ret = copy_page_to_iter(page, offset, nr, to);
2511 retval += ret;
2512 offset += ret;
2513 index += offset >> PAGE_SHIFT;
2514 offset &= ~PAGE_MASK;
2515
2516 put_page(page);
2517 if (!iov_iter_count(to))
2518 break;
2519 if (ret < nr) {
2520 error = -EFAULT;
2521 break;
2522 }
2523 cond_resched();
2524 }
2525
2526 *ppos = ((loff_t) index << PAGE_SHIFT) + offset;
2527 file_accessed(file);
2528 return retval ? retval : error;
2529}
2530
2531/*
2532 * llseek SEEK_DATA or SEEK_HOLE through the radix_tree.
2533 */
2534static pgoff_t shmem_seek_hole_data(struct address_space *mapping,
2535 pgoff_t index, pgoff_t end, int whence)
2536{
2537 struct page *page;
2538 struct pagevec pvec;
2539 pgoff_t indices[PAGEVEC_SIZE];
2540 bool done = false;
2541 int i;
2542
2543 pagevec_init(&pvec);
2544 pvec.nr = 1; /* start small: we may be there already */
2545 while (!done) {
2546 pvec.nr = find_get_entries(mapping, index,
2547 pvec.nr, pvec.pages, indices);
2548 if (!pvec.nr) {
2549 if (whence == SEEK_DATA)
2550 index = end;
2551 break;
2552 }
2553 for (i = 0; i < pvec.nr; i++, index++) {
2554 if (index < indices[i]) {
2555 if (whence == SEEK_HOLE) {
2556 done = true;
2557 break;
2558 }
2559 index = indices[i];
2560 }
2561 page = pvec.pages[i];
2562 if (page && !radix_tree_exceptional_entry(page)) {
2563 if (!PageUptodate(page))
2564 page = NULL;
2565 }
2566 if (index >= end ||
2567 (page && whence == SEEK_DATA) ||
2568 (!page && whence == SEEK_HOLE)) {
2569 done = true;
2570 break;
2571 }
2572 }
2573 pagevec_remove_exceptionals(&pvec);
2574 pagevec_release(&pvec);
2575 pvec.nr = PAGEVEC_SIZE;
2576 cond_resched();
2577 }
2578 return index;
2579}
2580
2581static loff_t shmem_file_llseek(struct file *file, loff_t offset, int whence)
2582{
2583 struct address_space *mapping = file->f_mapping;
2584 struct inode *inode = mapping->host;
2585 pgoff_t start, end;
2586 loff_t new_offset;
2587
2588 if (whence != SEEK_DATA && whence != SEEK_HOLE)
2589 return generic_file_llseek_size(file, offset, whence,
2590 MAX_LFS_FILESIZE, i_size_read(inode));
2591 inode_lock(inode);
2592 /* We're holding i_mutex so we can access i_size directly */
2593
2594 if (offset < 0)
2595 offset = -EINVAL;
2596 else if (offset >= inode->i_size)
2597 offset = -ENXIO;
2598 else {
2599 start = offset >> PAGE_SHIFT;
2600 end = (inode->i_size + PAGE_SIZE - 1) >> PAGE_SHIFT;
2601 new_offset = shmem_seek_hole_data(mapping, start, end, whence);
2602 new_offset <<= PAGE_SHIFT;
2603 if (new_offset > offset) {
2604 if (new_offset < inode->i_size)
2605 offset = new_offset;
2606 else if (whence == SEEK_DATA)
2607 offset = -ENXIO;
2608 else
2609 offset = inode->i_size;
2610 }
2611 }
2612
2613 if (offset >= 0)
2614 offset = vfs_setpos(file, offset, MAX_LFS_FILESIZE);
2615 inode_unlock(inode);
2616 return offset;
2617}
2618
2619/*
2620 * We need a tag: a new tag would expand every radix_tree_node by 8 bytes,
2621 * so reuse a tag which we firmly believe is never set or cleared on shmem.
2622 */
2623#define SHMEM_TAG_PINNED PAGECACHE_TAG_TOWRITE
2624#define LAST_SCAN 4 /* about 150ms max */
2625
2626static void shmem_tag_pins(struct address_space *mapping)
2627{
2628 struct radix_tree_iter iter;
2629 void **slot;
2630 pgoff_t start;
2631 struct page *page;
2632
2633 lru_add_drain();
2634 start = 0;
2635 rcu_read_lock();
2636
2637 radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start) {
2638 page = radix_tree_deref_slot(slot);
2639 if (!page || radix_tree_exception(page)) {
2640 if (radix_tree_deref_retry(page)) {
2641 slot = radix_tree_iter_retry(&iter);
2642 continue;
2643 }
2644 } else if (page_count(page) - page_mapcount(page) > 1) {
2645 xa_lock_irq(&mapping->i_pages);
2646 radix_tree_tag_set(&mapping->i_pages, iter.index,
2647 SHMEM_TAG_PINNED);
2648 xa_unlock_irq(&mapping->i_pages);
2649 }
2650
2651 if (need_resched()) {
2652 slot = radix_tree_iter_resume(slot, &iter);
2653 cond_resched_rcu();
2654 }
2655 }
2656 rcu_read_unlock();
2657}
2658
2659/*
2660 * Setting SEAL_WRITE requires us to verify there's no pending writer. However,
2661 * via get_user_pages(), drivers might have some pending I/O without any active
2662 * user-space mappings (eg., direct-IO, AIO). Therefore, we look at all pages
2663 * and see whether it has an elevated ref-count. If so, we tag them and wait for
2664 * them to be dropped.
2665 * The caller must guarantee that no new user will acquire writable references
2666 * to those pages to avoid races.
2667 */
2668static int shmem_wait_for_pins(struct address_space *mapping)
2669{
2670 struct radix_tree_iter iter;
2671 void **slot;
2672 pgoff_t start;
2673 struct page *page;
2674 int error, scan;
2675
2676 shmem_tag_pins(mapping);
2677
2678 error = 0;
2679 for (scan = 0; scan <= LAST_SCAN; scan++) {
2680 if (!radix_tree_tagged(&mapping->i_pages, SHMEM_TAG_PINNED))
2681 break;
2682
2683 if (!scan)
2684 lru_add_drain_all();
2685 else if (schedule_timeout_killable((HZ << scan) / 200))
2686 scan = LAST_SCAN;
2687
2688 start = 0;
2689 rcu_read_lock();
2690 radix_tree_for_each_tagged(slot, &mapping->i_pages, &iter,
2691 start, SHMEM_TAG_PINNED) {
2692
2693 page = radix_tree_deref_slot(slot);
2694 if (radix_tree_exception(page)) {
2695 if (radix_tree_deref_retry(page)) {
2696 slot = radix_tree_iter_retry(&iter);
2697 continue;
2698 }
2699
2700 page = NULL;
2701 }
2702
2703 if (page &&
2704 page_count(page) - page_mapcount(page) != 1) {
2705 if (scan < LAST_SCAN)
2706 goto continue_resched;
2707
2708 /*
2709 * On the last scan, we clean up all those tags
2710 * we inserted; but make a note that we still
2711 * found pages pinned.
2712 */
2713 error = -EBUSY;
2714 }
2715
2716 xa_lock_irq(&mapping->i_pages);
2717 radix_tree_tag_clear(&mapping->i_pages,
2718 iter.index, SHMEM_TAG_PINNED);
2719 xa_unlock_irq(&mapping->i_pages);
2720continue_resched:
2721 if (need_resched()) {
2722 slot = radix_tree_iter_resume(slot, &iter);
2723 cond_resched_rcu();
2724 }
2725 }
2726 rcu_read_unlock();
2727 }
2728
2729 return error;
2730}
2731
2732static unsigned int *memfd_file_seals_ptr(struct file *file)
2733{
2734 if (file->f_op == &shmem_file_operations)
2735 return &SHMEM_I(file_inode(file))->seals;
2736
2737#ifdef CONFIG_HUGETLBFS
2738 if (file->f_op == &hugetlbfs_file_operations)
2739 return &HUGETLBFS_I(file_inode(file))->seals;
2740#endif
2741
2742 return NULL;
2743}
2744
2745#define F_ALL_SEALS (F_SEAL_SEAL | \
2746 F_SEAL_SHRINK | \
2747 F_SEAL_GROW | \
2748 F_SEAL_WRITE)
2749
2750static int memfd_add_seals(struct file *file, unsigned int seals)
2751{
2752 struct inode *inode = file_inode(file);
2753 unsigned int *file_seals;
2754 int error;
2755
2756 /*
2757 * SEALING
2758 * Sealing allows multiple parties to share a shmem-file but restrict
2759 * access to a specific subset of file operations. Seals can only be
2760 * added, but never removed. This way, mutually untrusted parties can
2761 * share common memory regions with a well-defined policy. A malicious
2762 * peer can thus never perform unwanted operations on a shared object.
2763 *
2764 * Seals are only supported on special shmem-files and always affect
2765 * the whole underlying inode. Once a seal is set, it may prevent some
2766 * kinds of access to the file. Currently, the following seals are
2767 * defined:
2768 * SEAL_SEAL: Prevent further seals from being set on this file
2769 * SEAL_SHRINK: Prevent the file from shrinking
2770 * SEAL_GROW: Prevent the file from growing
2771 * SEAL_WRITE: Prevent write access to the file
2772 *
2773 * As we don't require any trust relationship between two parties, we
2774 * must prevent seals from being removed. Therefore, sealing a file
2775 * only adds a given set of seals to the file, it never touches
2776 * existing seals. Furthermore, the "setting seals"-operation can be
2777 * sealed itself, which basically prevents any further seal from being
2778 * added.
2779 *
2780 * Semantics of sealing are only defined on volatile files. Only
2781 * anonymous shmem files support sealing. More importantly, seals are
2782 * never written to disk. Therefore, there's no plan to support it on
2783 * other file types.
2784 */
2785
2786 if (!(file->f_mode & FMODE_WRITE))
2787 return -EPERM;
2788 if (seals & ~(unsigned int)F_ALL_SEALS)
2789 return -EINVAL;
2790
2791 inode_lock(inode);
2792
2793 file_seals = memfd_file_seals_ptr(file);
2794 if (!file_seals) {
2795 error = -EINVAL;
2796 goto unlock;
2797 }
2798
2799 if (*file_seals & F_SEAL_SEAL) {
2800 error = -EPERM;
2801 goto unlock;
2802 }
2803
2804 if ((seals & F_SEAL_WRITE) && !(*file_seals & F_SEAL_WRITE)) {
2805 error = mapping_deny_writable(file->f_mapping);
2806 if (error)
2807 goto unlock;
2808
2809 error = shmem_wait_for_pins(file->f_mapping);
2810 if (error) {
2811 mapping_allow_writable(file->f_mapping);
2812 goto unlock;
2813 }
2814 }
2815
2816 *file_seals |= seals;
2817 error = 0;
2818
2819unlock:
2820 inode_unlock(inode);
2821 return error;
2822}
2823
2824static int memfd_get_seals(struct file *file)
2825{
2826 unsigned int *seals = memfd_file_seals_ptr(file);
2827
2828 return seals ? *seals : -EINVAL;
2829}
2830
2831long memfd_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
2832{
2833 long error;
2834
2835 switch (cmd) {
2836 case F_ADD_SEALS:
2837 /* disallow upper 32bit */
2838 if (arg > UINT_MAX)
2839 return -EINVAL;
2840
2841 error = memfd_add_seals(file, arg);
2842 break;
2843 case F_GET_SEALS:
2844 error = memfd_get_seals(file);
2845 break;
2846 default:
2847 error = -EINVAL;
2848 break;
2849 }
2850
2851 return error;
2852}
2853
2854static long shmem_fallocate(struct file *file, int mode, loff_t offset,
2855 loff_t len)
2856{
2857 struct inode *inode = file_inode(file);
2858 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
2859 struct shmem_inode_info *info = SHMEM_I(inode);
2860 struct shmem_falloc shmem_falloc;
2861 pgoff_t start, index, end;
2862 int error;
2863
2864 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2865 return -EOPNOTSUPP;
2866
2867 inode_lock(inode);
2868
2869 if (mode & FALLOC_FL_PUNCH_HOLE) {
2870 struct address_space *mapping = file->f_mapping;
2871 loff_t unmap_start = round_up(offset, PAGE_SIZE);
2872 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
2873 DECLARE_WAIT_QUEUE_HEAD_ONSTACK(shmem_falloc_waitq);
2874
2875 /* protected by i_mutex */
2876 if (info->seals & F_SEAL_WRITE) {
2877 error = -EPERM;
2878 goto out;
2879 }
2880
2881 shmem_falloc.waitq = &shmem_falloc_waitq;
2882 shmem_falloc.start = unmap_start >> PAGE_SHIFT;
2883 shmem_falloc.next = (unmap_end + 1) >> PAGE_SHIFT;
2884 spin_lock(&inode->i_lock);
2885 inode->i_private = &shmem_falloc;
2886 spin_unlock(&inode->i_lock);
2887
2888 if ((u64)unmap_end > (u64)unmap_start)
2889 unmap_mapping_range(mapping, unmap_start,
2890 1 + unmap_end - unmap_start, 0);
2891 shmem_truncate_range(inode, offset, offset + len - 1);
2892 /* No need to unmap again: hole-punching leaves COWed pages */
2893
2894 spin_lock(&inode->i_lock);
2895 inode->i_private = NULL;
2896 wake_up_all(&shmem_falloc_waitq);
2897 WARN_ON_ONCE(!list_empty(&shmem_falloc_waitq.head));
2898 spin_unlock(&inode->i_lock);
2899 error = 0;
2900 goto out;
2901 }
2902
2903 /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
2904 error = inode_newsize_ok(inode, offset + len);
2905 if (error)
2906 goto out;
2907
2908 if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
2909 error = -EPERM;
2910 goto out;
2911 }
2912
2913 start = offset >> PAGE_SHIFT;
2914 end = (offset + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
2915 /* Try to avoid a swapstorm if len is impossible to satisfy */
2916 if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
2917 error = -ENOSPC;
2918 goto out;
2919 }
2920
2921 shmem_falloc.waitq = NULL;
2922 shmem_falloc.start = start;
2923 shmem_falloc.next = start;
2924 shmem_falloc.nr_falloced = 0;
2925 shmem_falloc.nr_unswapped = 0;
2926 spin_lock(&inode->i_lock);
2927 inode->i_private = &shmem_falloc;
2928 spin_unlock(&inode->i_lock);
2929
2930 for (index = start; index < end; index++) {
2931 struct page *page;
2932
2933 /*
2934 * Good, the fallocate(2) manpage permits EINTR: we may have
2935 * been interrupted because we are using up too much memory.
2936 */
2937 if (signal_pending(current))
2938 error = -EINTR;
2939 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
2940 error = -ENOMEM;
2941 else
2942 error = shmem_getpage(inode, index, &page, SGP_FALLOC);
2943 if (error) {
2944 /* Remove the !PageUptodate pages we added */
2945 if (index > start) {
2946 shmem_undo_range(inode,
2947 (loff_t)start << PAGE_SHIFT,
2948 ((loff_t)index << PAGE_SHIFT) - 1, true);
2949 }
2950 goto undone;
2951 }
2952
2953 /*
2954 * Inform shmem_writepage() how far we have reached.
2955 * No need for lock or barrier: we have the page lock.
2956 */
2957 shmem_falloc.next++;
2958 if (!PageUptodate(page))
2959 shmem_falloc.nr_falloced++;
2960
2961 /*
2962 * If !PageUptodate, leave it that way so that freeable pages
2963 * can be recognized if we need to rollback on error later.
2964 * But set_page_dirty so that memory pressure will swap rather
2965 * than free the pages we are allocating (and SGP_CACHE pages
2966 * might still be clean: we now need to mark those dirty too).
2967 */
2968 set_page_dirty(page);
2969 unlock_page(page);
2970 put_page(page);
2971 cond_resched();
2972 }
2973
2974 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
2975 i_size_write(inode, offset + len);
2976 inode->i_ctime = current_time(inode);
2977undone:
2978 spin_lock(&inode->i_lock);
2979 inode->i_private = NULL;
2980 spin_unlock(&inode->i_lock);
2981out:
2982 inode_unlock(inode);
2983 return error;
2984}
2985
2986static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
2987{
2988 struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
2989
2990 buf->f_type = TMPFS_MAGIC;
2991 buf->f_bsize = PAGE_SIZE;
2992 buf->f_namelen = NAME_MAX;
2993 if (sbinfo->max_blocks) {
2994 buf->f_blocks = sbinfo->max_blocks;
2995 buf->f_bavail =
2996 buf->f_bfree = sbinfo->max_blocks -
2997 percpu_counter_sum(&sbinfo->used_blocks);
2998 }
2999 if (sbinfo->max_inodes) {
3000 buf->f_files = sbinfo->max_inodes;
3001 buf->f_ffree = sbinfo->free_inodes;
3002 }
3003 /* else leave those fields 0 like simple_statfs */
3004 return 0;
3005}
3006
3007/*
3008 * File creation. Allocate an inode, and we're done..
3009 */
3010static int
3011shmem_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
3012{
3013 struct inode *inode;
3014 int error = -ENOSPC;
3015
3016 inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
3017 if (inode) {
3018 error = simple_acl_create(dir, inode);
3019 if (error)
3020 goto out_iput;
3021 error = security_inode_init_security(inode, dir,
3022 &dentry->d_name,
3023 shmem_initxattrs, NULL);
3024 if (error && error != -EOPNOTSUPP)
3025 goto out_iput;
3026
3027 error = 0;
3028 dir->i_size += BOGO_DIRENT_SIZE;
3029 dir->i_ctime = dir->i_mtime = current_time(dir);
3030 d_instantiate(dentry, inode);
3031 dget(dentry); /* Extra count - pin the dentry in core */
3032 }
3033 return error;
3034out_iput:
3035 iput(inode);
3036 return error;
3037}
3038
3039static int
3040shmem_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
3041{
3042 struct inode *inode;
3043 int error = -ENOSPC;
3044
3045 inode = shmem_get_inode(dir->i_sb, dir, mode, 0, VM_NORESERVE);
3046 if (inode) {
3047 error = security_inode_init_security(inode, dir,
3048 NULL,
3049 shmem_initxattrs, NULL);
3050 if (error && error != -EOPNOTSUPP)
3051 goto out_iput;
3052 error = simple_acl_create(dir, inode);
3053 if (error)
3054 goto out_iput;
3055 d_tmpfile(dentry, inode);
3056 }
3057 return error;
3058out_iput:
3059 iput(inode);
3060 return error;
3061}
3062
3063static int shmem_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
3064{
3065 int error;
3066
3067 if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
3068 return error;
3069 inc_nlink(dir);
3070 return 0;
3071}
3072
3073static int shmem_create(struct inode *dir, struct dentry *dentry, umode_t mode,
3074 bool excl)
3075{
3076 return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
3077}
3078
3079/*
3080 * Link a file..
3081 */
3082static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
3083{
3084 struct inode *inode = d_inode(old_dentry);
3085 int ret;
3086
3087 /*
3088 * No ordinary (disk based) filesystem counts links as inodes;
3089 * but each new link needs a new dentry, pinning lowmem, and
3090 * tmpfs dentries cannot be pruned until they are unlinked.
3091 */
3092 ret = shmem_reserve_inode(inode->i_sb);
3093 if (ret)
3094 goto out;
3095
3096 dir->i_size += BOGO_DIRENT_SIZE;
3097 inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
3098 inc_nlink(inode);
3099 ihold(inode); /* New dentry reference */
3100 dget(dentry); /* Extra pinning count for the created dentry */
3101 d_instantiate(dentry, inode);
3102out:
3103 return ret;
3104}
3105
3106static int shmem_unlink(struct inode *dir, struct dentry *dentry)
3107{
3108 struct inode *inode = d_inode(dentry);
3109
3110 if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
3111 shmem_free_inode(inode->i_sb);
3112
3113 dir->i_size -= BOGO_DIRENT_SIZE;
3114 inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
3115 drop_nlink(inode);
3116 dput(dentry); /* Undo the count from "create" - this does all the work */
3117 return 0;
3118}
3119
3120static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
3121{
3122 if (!simple_empty(dentry))
3123 return -ENOTEMPTY;
3124
3125 drop_nlink(d_inode(dentry));
3126 drop_nlink(dir);
3127 return shmem_unlink(dir, dentry);
3128}
3129
3130static int shmem_exchange(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
3131{
3132 bool old_is_dir = d_is_dir(old_dentry);
3133 bool new_is_dir = d_is_dir(new_dentry);
3134
3135 if (old_dir != new_dir && old_is_dir != new_is_dir) {
3136 if (old_is_dir) {
3137 drop_nlink(old_dir);
3138 inc_nlink(new_dir);
3139 } else {
3140 drop_nlink(new_dir);
3141 inc_nlink(old_dir);
3142 }
3143 }
3144 old_dir->i_ctime = old_dir->i_mtime =
3145 new_dir->i_ctime = new_dir->i_mtime =
3146 d_inode(old_dentry)->i_ctime =
3147 d_inode(new_dentry)->i_ctime = current_time(old_dir);
3148
3149 return 0;
3150}
3151
3152static int shmem_whiteout(struct inode *old_dir, struct dentry *old_dentry)
3153{
3154 struct dentry *whiteout;
3155 int error;
3156
3157 whiteout = d_alloc(old_dentry->d_parent, &old_dentry->d_name);
3158 if (!whiteout)
3159 return -ENOMEM;
3160
3161 error = shmem_mknod(old_dir, whiteout,
3162 S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
3163 dput(whiteout);
3164 if (error)
3165 return error;
3166
3167 /*
3168 * Cheat and hash the whiteout while the old dentry is still in
3169 * place, instead of playing games with FS_RENAME_DOES_D_MOVE.
3170 *
3171 * d_lookup() will consistently find one of them at this point,
3172 * not sure which one, but that isn't even important.
3173 */
3174 d_rehash(whiteout);
3175 return 0;
3176}
3177
3178/*
3179 * The VFS layer already does all the dentry stuff for rename,
3180 * we just have to decrement the usage count for the target if
3181 * it exists so that the VFS layer correctly free's it when it
3182 * gets overwritten.
3183 */
3184static int shmem_rename2(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry, unsigned int flags)
3185{
3186 struct inode *inode = d_inode(old_dentry);
3187 int they_are_dirs = S_ISDIR(inode->i_mode);
3188
3189 if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
3190 return -EINVAL;
3191
3192 if (flags & RENAME_EXCHANGE)
3193 return shmem_exchange(old_dir, old_dentry, new_dir, new_dentry);
3194
3195 if (!simple_empty(new_dentry))
3196 return -ENOTEMPTY;
3197
3198 if (flags & RENAME_WHITEOUT) {
3199 int error;
3200
3201 error = shmem_whiteout(old_dir, old_dentry);
3202 if (error)
3203 return error;
3204 }
3205
3206 if (d_really_is_positive(new_dentry)) {
3207 (void) shmem_unlink(new_dir, new_dentry);
3208 if (they_are_dirs) {
3209 drop_nlink(d_inode(new_dentry));
3210 drop_nlink(old_dir);
3211 }
3212 } else if (they_are_dirs) {
3213 drop_nlink(old_dir);
3214 inc_nlink(new_dir);
3215 }
3216
3217 old_dir->i_size -= BOGO_DIRENT_SIZE;
3218 new_dir->i_size += BOGO_DIRENT_SIZE;
3219 old_dir->i_ctime = old_dir->i_mtime =
3220 new_dir->i_ctime = new_dir->i_mtime =
3221 inode->i_ctime = current_time(old_dir);
3222 return 0;
3223}
3224
3225static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
3226{
3227 int error;
3228 int len;
3229 struct inode *inode;
3230 struct page *page;
3231
3232 len = strlen(symname) + 1;
3233 if (len > PAGE_SIZE)
3234 return -ENAMETOOLONG;
3235
3236 inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0, VM_NORESERVE);
3237 if (!inode)
3238 return -ENOSPC;
3239
3240 error = security_inode_init_security(inode, dir, &dentry->d_name,
3241 shmem_initxattrs, NULL);
3242 if (error) {
3243 if (error != -EOPNOTSUPP) {
3244 iput(inode);
3245 return error;
3246 }
3247 error = 0;
3248 }
3249
3250 inode->i_size = len-1;
3251 if (len <= SHORT_SYMLINK_LEN) {
3252 inode->i_link = kmemdup(symname, len, GFP_KERNEL);
3253 if (!inode->i_link) {
3254 iput(inode);
3255 return -ENOMEM;
3256 }
3257 inode->i_op = &shmem_short_symlink_operations;
3258 } else {
3259 inode_nohighmem(inode);
3260 error = shmem_getpage(inode, 0, &page, SGP_WRITE);
3261 if (error) {
3262 iput(inode);
3263 return error;
3264 }
3265 inode->i_mapping->a_ops = &shmem_aops;
3266 inode->i_op = &shmem_symlink_inode_operations;
3267 memcpy(page_address(page), symname, len);
3268 SetPageUptodate(page);
3269 set_page_dirty(page);
3270 unlock_page(page);
3271 put_page(page);
3272 }
3273 dir->i_size += BOGO_DIRENT_SIZE;
3274 dir->i_ctime = dir->i_mtime = current_time(dir);
3275 d_instantiate(dentry, inode);
3276 dget(dentry);
3277 return 0;
3278}
3279
3280static void shmem_put_link(void *arg)
3281{
3282 mark_page_accessed(arg);
3283 put_page(arg);
3284}
3285
3286static const char *shmem_get_link(struct dentry *dentry,
3287 struct inode *inode,
3288 struct delayed_call *done)
3289{
3290 struct page *page = NULL;
3291 int error;
3292 if (!dentry) {
3293 page = find_get_page(inode->i_mapping, 0);
3294 if (!page)
3295 return ERR_PTR(-ECHILD);
3296 if (!PageUptodate(page)) {
3297 put_page(page);
3298 return ERR_PTR(-ECHILD);
3299 }
3300 } else {
3301 error = shmem_getpage(inode, 0, &page, SGP_READ);
3302 if (error)
3303 return ERR_PTR(error);
3304 unlock_page(page);
3305 }
3306 set_delayed_call(done, shmem_put_link, page);
3307 return page_address(page);
3308}
3309
3310#ifdef CONFIG_TMPFS_XATTR
3311/*
3312 * Superblocks without xattr inode operations may get some security.* xattr
3313 * support from the LSM "for free". As soon as we have any other xattrs
3314 * like ACLs, we also need to implement the security.* handlers at
3315 * filesystem level, though.
3316 */
3317
3318/*
3319 * Callback for security_inode_init_security() for acquiring xattrs.
3320 */
3321static int shmem_initxattrs(struct inode *inode,
3322 const struct xattr *xattr_array,
3323 void *fs_info)
3324{
3325 struct shmem_inode_info *info = SHMEM_I(inode);
3326 const struct xattr *xattr;
3327 struct simple_xattr *new_xattr;
3328 size_t len;
3329
3330 for (xattr = xattr_array; xattr->name != NULL; xattr++) {
3331 new_xattr = simple_xattr_alloc(xattr->value, xattr->value_len);
3332 if (!new_xattr)
3333 return -ENOMEM;
3334
3335 len = strlen(xattr->name) + 1;
3336 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
3337 GFP_KERNEL);
3338 if (!new_xattr->name) {
3339 kfree(new_xattr);
3340 return -ENOMEM;
3341 }
3342
3343 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
3344 XATTR_SECURITY_PREFIX_LEN);
3345 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
3346 xattr->name, len);
3347
3348 simple_xattr_list_add(&info->xattrs, new_xattr);
3349 }
3350
3351 return 0;
3352}
3353
3354static int shmem_xattr_handler_get(const struct xattr_handler *handler,
3355 struct dentry *unused, struct inode *inode,
3356 const char *name, void *buffer, size_t size)
3357{
3358 struct shmem_inode_info *info = SHMEM_I(inode);
3359
3360 name = xattr_full_name(handler, name);
3361 return simple_xattr_get(&info->xattrs, name, buffer, size);
3362}
3363
3364static int shmem_xattr_handler_set(const struct xattr_handler *handler,
3365 struct dentry *unused, struct inode *inode,
3366 const char *name, const void *value,
3367 size_t size, int flags)
3368{
3369 struct shmem_inode_info *info = SHMEM_I(inode);
3370
3371 name = xattr_full_name(handler, name);
3372 return simple_xattr_set(&info->xattrs, name, value, size, flags);
3373}
3374
3375static const struct xattr_handler shmem_security_xattr_handler = {
3376 .prefix = XATTR_SECURITY_PREFIX,
3377 .get = shmem_xattr_handler_get,
3378 .set = shmem_xattr_handler_set,
3379};
3380
3381static const struct xattr_handler shmem_trusted_xattr_handler = {
3382 .prefix = XATTR_TRUSTED_PREFIX,
3383 .get = shmem_xattr_handler_get,
3384 .set = shmem_xattr_handler_set,
3385};
3386
3387static const struct xattr_handler *shmem_xattr_handlers[] = {
3388#ifdef CONFIG_TMPFS_POSIX_ACL
3389 &posix_acl_access_xattr_handler,
3390 &posix_acl_default_xattr_handler,
3391#endif
3392 &shmem_security_xattr_handler,
3393 &shmem_trusted_xattr_handler,
3394 NULL
3395};
3396
3397static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
3398{
3399 struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
3400 return simple_xattr_list(d_inode(dentry), &info->xattrs, buffer, size);
3401}
3402#endif /* CONFIG_TMPFS_XATTR */
3403
3404static const struct inode_operations shmem_short_symlink_operations = {
3405 .get_link = simple_get_link,
3406#ifdef CONFIG_TMPFS_XATTR
3407 .listxattr = shmem_listxattr,
3408#endif
3409};
3410
3411static const struct inode_operations shmem_symlink_inode_operations = {
3412 .get_link = shmem_get_link,
3413#ifdef CONFIG_TMPFS_XATTR
3414 .listxattr = shmem_listxattr,
3415#endif
3416};
3417
3418static struct dentry *shmem_get_parent(struct dentry *child)
3419{
3420 return ERR_PTR(-ESTALE);
3421}
3422
3423static int shmem_match(struct inode *ino, void *vfh)
3424{
3425 __u32 *fh = vfh;
3426 __u64 inum = fh[2];
3427 inum = (inum << 32) | fh[1];
3428 return ino->i_ino == inum && fh[0] == ino->i_generation;
3429}
3430
3431static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
3432 struct fid *fid, int fh_len, int fh_type)
3433{
3434 struct inode *inode;
3435 struct dentry *dentry = NULL;
3436 u64 inum;
3437
3438 if (fh_len < 3)
3439 return NULL;
3440
3441 inum = fid->raw[2];
3442 inum = (inum << 32) | fid->raw[1];
3443
3444 inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
3445 shmem_match, fid->raw);
3446 if (inode) {
3447 dentry = d_find_alias(inode);
3448 iput(inode);
3449 }
3450
3451 return dentry;
3452}
3453
3454static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
3455 struct inode *parent)
3456{
3457 if (*len < 3) {
3458 *len = 3;
3459 return FILEID_INVALID;
3460 }
3461
3462 if (inode_unhashed(inode)) {
3463 /* Unfortunately insert_inode_hash is not idempotent,
3464 * so as we hash inodes here rather than at creation
3465 * time, we need a lock to ensure we only try
3466 * to do it once
3467 */
3468 static DEFINE_SPINLOCK(lock);
3469 spin_lock(&lock);
3470 if (inode_unhashed(inode))
3471 __insert_inode_hash(inode,
3472 inode->i_ino + inode->i_generation);
3473 spin_unlock(&lock);
3474 }
3475
3476 fh[0] = inode->i_generation;
3477 fh[1] = inode->i_ino;
3478 fh[2] = ((__u64)inode->i_ino) >> 32;
3479
3480 *len = 3;
3481 return 1;
3482}
3483
3484static const struct export_operations shmem_export_ops = {
3485 .get_parent = shmem_get_parent,
3486 .encode_fh = shmem_encode_fh,
3487 .fh_to_dentry = shmem_fh_to_dentry,
3488};
3489
3490static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
3491 bool remount)
3492{
3493 char *this_char, *value, *rest;
3494 struct mempolicy *mpol = NULL;
3495 uid_t uid;
3496 gid_t gid;
3497
3498 while (options != NULL) {
3499 this_char = options;
3500 for (;;) {
3501 /*
3502 * NUL-terminate this option: unfortunately,
3503 * mount options form a comma-separated list,
3504 * but mpol's nodelist may also contain commas.
3505 */
3506 options = strchr(options, ',');
3507 if (options == NULL)
3508 break;
3509 options++;
3510 if (!isdigit(*options)) {
3511 options[-1] = '\0';
3512 break;
3513 }
3514 }
3515 if (!*this_char)
3516 continue;
3517 if ((value = strchr(this_char,'=')) != NULL) {
3518 *value++ = 0;
3519 } else {
3520 pr_err("tmpfs: No value for mount option '%s'\n",
3521 this_char);
3522 goto error;
3523 }
3524
3525 if (!strcmp(this_char,"size")) {
3526 unsigned long long size;
3527 size = memparse(value,&rest);
3528 if (*rest == '%') {
3529 size <<= PAGE_SHIFT;
3530 size *= totalram_pages;
3531 do_div(size, 100);
3532 rest++;
3533 }
3534 if (*rest)
3535 goto bad_val;
3536 sbinfo->max_blocks =
3537 DIV_ROUND_UP(size, PAGE_SIZE);
3538 } else if (!strcmp(this_char,"nr_blocks")) {
3539 sbinfo->max_blocks = memparse(value, &rest);
3540 if (*rest)
3541 goto bad_val;
3542 } else if (!strcmp(this_char,"nr_inodes")) {
3543 sbinfo->max_inodes = memparse(value, &rest);
3544 if (*rest)
3545 goto bad_val;
3546 } else if (!strcmp(this_char,"mode")) {
3547 if (remount)
3548 continue;
3549 sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
3550 if (*rest)
3551 goto bad_val;
3552 } else if (!strcmp(this_char,"uid")) {
3553 if (remount)
3554 continue;
3555 uid = simple_strtoul(value, &rest, 0);
3556 if (*rest)
3557 goto bad_val;
3558 sbinfo->uid = make_kuid(current_user_ns(), uid);
3559 if (!uid_valid(sbinfo->uid))
3560 goto bad_val;
3561 } else if (!strcmp(this_char,"gid")) {
3562 if (remount)
3563 continue;
3564 gid = simple_strtoul(value, &rest, 0);
3565 if (*rest)
3566 goto bad_val;
3567 sbinfo->gid = make_kgid(current_user_ns(), gid);
3568 if (!gid_valid(sbinfo->gid))
3569 goto bad_val;
3570#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3571 } else if (!strcmp(this_char, "huge")) {
3572 int huge;
3573 huge = shmem_parse_huge(value);
3574 if (huge < 0)
3575 goto bad_val;
3576 if (!has_transparent_hugepage() &&
3577 huge != SHMEM_HUGE_NEVER)
3578 goto bad_val;
3579 sbinfo->huge = huge;
3580#endif
3581#ifdef CONFIG_NUMA
3582 } else if (!strcmp(this_char,"mpol")) {
3583 mpol_put(mpol);
3584 mpol = NULL;
3585 if (mpol_parse_str(value, &mpol))
3586 goto bad_val;
3587#endif
3588 } else {
3589 pr_err("tmpfs: Bad mount option %s\n", this_char);
3590 goto error;
3591 }
3592 }
3593 sbinfo->mpol = mpol;
3594 return 0;
3595
3596bad_val:
3597 pr_err("tmpfs: Bad value '%s' for mount option '%s'\n",
3598 value, this_char);
3599error:
3600 mpol_put(mpol);
3601 return 1;
3602
3603}
3604
3605static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
3606{
3607 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
3608 struct shmem_sb_info config = *sbinfo;
3609 unsigned long inodes;
3610 int error = -EINVAL;
3611
3612 config.mpol = NULL;
3613 if (shmem_parse_options(data, &config, true))
3614 return error;
3615
3616 spin_lock(&sbinfo->stat_lock);
3617 inodes = sbinfo->max_inodes - sbinfo->free_inodes;
3618 if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
3619 goto out;
3620 if (config.max_inodes < inodes)
3621 goto out;
3622 /*
3623 * Those tests disallow limited->unlimited while any are in use;
3624 * but we must separately disallow unlimited->limited, because
3625 * in that case we have no record of how much is already in use.
3626 */
3627 if (config.max_blocks && !sbinfo->max_blocks)
3628 goto out;
3629 if (config.max_inodes && !sbinfo->max_inodes)
3630 goto out;
3631
3632 error = 0;
3633 sbinfo->huge = config.huge;
3634 sbinfo->max_blocks = config.max_blocks;
3635 sbinfo->max_inodes = config.max_inodes;
3636 sbinfo->free_inodes = config.max_inodes - inodes;
3637
3638 /*
3639 * Preserve previous mempolicy unless mpol remount option was specified.
3640 */
3641 if (config.mpol) {
3642 mpol_put(sbinfo->mpol);
3643 sbinfo->mpol = config.mpol; /* transfers initial ref */
3644 }
3645out:
3646 spin_unlock(&sbinfo->stat_lock);
3647 return error;
3648}
3649
3650static int shmem_show_options(struct seq_file *seq, struct dentry *root)
3651{
3652 struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
3653
3654 if (sbinfo->max_blocks != shmem_default_max_blocks())
3655 seq_printf(seq, ",size=%luk",
3656 sbinfo->max_blocks << (PAGE_SHIFT - 10));
3657 if (sbinfo->max_inodes != shmem_default_max_inodes())
3658 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
3659 if (sbinfo->mode != (S_IRWXUGO | S_ISVTX))
3660 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
3661 if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
3662 seq_printf(seq, ",uid=%u",
3663 from_kuid_munged(&init_user_ns, sbinfo->uid));
3664 if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
3665 seq_printf(seq, ",gid=%u",
3666 from_kgid_munged(&init_user_ns, sbinfo->gid));
3667#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3668 /* Rightly or wrongly, show huge mount option unmasked by shmem_huge */
3669 if (sbinfo->huge)
3670 seq_printf(seq, ",huge=%s", shmem_format_huge(sbinfo->huge));
3671#endif
3672 shmem_show_mpol(seq, sbinfo->mpol);
3673 return 0;
3674}
3675
3676#define MFD_NAME_PREFIX "memfd:"
3677#define MFD_NAME_PREFIX_LEN (sizeof(MFD_NAME_PREFIX) - 1)
3678#define MFD_NAME_MAX_LEN (NAME_MAX - MFD_NAME_PREFIX_LEN)
3679
3680#define MFD_ALL_FLAGS (MFD_CLOEXEC | MFD_ALLOW_SEALING | MFD_HUGETLB)
3681
3682SYSCALL_DEFINE2(memfd_create,
3683 const char __user *, uname,
3684 unsigned int, flags)
3685{
3686 unsigned int *file_seals;
3687 struct file *file;
3688 int fd, error;
3689 char *name;
3690 long len;
3691
3692 if (!(flags & MFD_HUGETLB)) {
3693 if (flags & ~(unsigned int)MFD_ALL_FLAGS)
3694 return -EINVAL;
3695 } else {
3696 /* Allow huge page size encoding in flags. */
3697 if (flags & ~(unsigned int)(MFD_ALL_FLAGS |
3698 (MFD_HUGE_MASK << MFD_HUGE_SHIFT)))
3699 return -EINVAL;
3700 }
3701
3702 /* length includes terminating zero */
3703 len = strnlen_user(uname, MFD_NAME_MAX_LEN + 1);
3704 if (len <= 0)
3705 return -EFAULT;
3706 if (len > MFD_NAME_MAX_LEN + 1)
3707 return -EINVAL;
3708
3709 name = kmalloc(len + MFD_NAME_PREFIX_LEN, GFP_KERNEL);
3710 if (!name)
3711 return -ENOMEM;
3712
3713 strcpy(name, MFD_NAME_PREFIX);
3714 if (copy_from_user(&name[MFD_NAME_PREFIX_LEN], uname, len)) {
3715 error = -EFAULT;
3716 goto err_name;
3717 }
3718
3719 /* terminating-zero may have changed after strnlen_user() returned */
3720 if (name[len + MFD_NAME_PREFIX_LEN - 1]) {
3721 error = -EFAULT;
3722 goto err_name;
3723 }
3724
3725 fd = get_unused_fd_flags((flags & MFD_CLOEXEC) ? O_CLOEXEC : 0);
3726 if (fd < 0) {
3727 error = fd;
3728 goto err_name;
3729 }
3730
3731 if (flags & MFD_HUGETLB) {
3732 struct user_struct *user = NULL;
3733
3734 file = hugetlb_file_setup(name, 0, VM_NORESERVE, &user,
3735 HUGETLB_ANONHUGE_INODE,
3736 (flags >> MFD_HUGE_SHIFT) &
3737 MFD_HUGE_MASK);
3738 } else
3739 file = shmem_file_setup(name, 0, VM_NORESERVE);
3740 if (IS_ERR(file)) {
3741 error = PTR_ERR(file);
3742 goto err_fd;
3743 }
3744 file->f_mode |= FMODE_LSEEK | FMODE_PREAD | FMODE_PWRITE;
3745 file->f_flags |= O_RDWR | O_LARGEFILE;
3746
3747 if (flags & MFD_ALLOW_SEALING) {
3748 file_seals = memfd_file_seals_ptr(file);
3749 *file_seals &= ~F_SEAL_SEAL;
3750 }
3751
3752 fd_install(fd, file);
3753 kfree(name);
3754 return fd;
3755
3756err_fd:
3757 put_unused_fd(fd);
3758err_name:
3759 kfree(name);
3760 return error;
3761}
3762
3763#endif /* CONFIG_TMPFS */
3764
3765static void shmem_put_super(struct super_block *sb)
3766{
3767 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
3768
3769 percpu_counter_destroy(&sbinfo->used_blocks);
3770 mpol_put(sbinfo->mpol);
3771 kfree(sbinfo);
3772 sb->s_fs_info = NULL;
3773}
3774
3775int shmem_fill_super(struct super_block *sb, void *data, int silent)
3776{
3777 struct inode *inode;
3778 struct shmem_sb_info *sbinfo;
3779 int err = -ENOMEM;
3780
3781 /* Round up to L1_CACHE_BYTES to resist false sharing */
3782 sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
3783 L1_CACHE_BYTES), GFP_KERNEL);
3784 if (!sbinfo)
3785 return -ENOMEM;
3786
3787 sbinfo->mode = S_IRWXUGO | S_ISVTX;
3788 sbinfo->uid = current_fsuid();
3789 sbinfo->gid = current_fsgid();
3790 sb->s_fs_info = sbinfo;
3791
3792#ifdef CONFIG_TMPFS
3793 /*
3794 * Per default we only allow half of the physical ram per
3795 * tmpfs instance, limiting inodes to one per page of lowmem;
3796 * but the internal instance is left unlimited.
3797 */
3798 if (!(sb->s_flags & SB_KERNMOUNT)) {
3799 sbinfo->max_blocks = shmem_default_max_blocks();
3800 sbinfo->max_inodes = shmem_default_max_inodes();
3801 if (shmem_parse_options(data, sbinfo, false)) {
3802 err = -EINVAL;
3803 goto failed;
3804 }
3805 } else {
3806 sb->s_flags |= SB_NOUSER;
3807 }
3808 sb->s_export_op = &shmem_export_ops;
3809 sb->s_flags |= SB_NOSEC;
3810#else
3811 sb->s_flags |= SB_NOUSER;
3812#endif
3813
3814 spin_lock_init(&sbinfo->stat_lock);
3815 if (percpu_counter_init(&sbinfo->used_blocks, 0, GFP_KERNEL))
3816 goto failed;
3817 sbinfo->free_inodes = sbinfo->max_inodes;
3818 spin_lock_init(&sbinfo->shrinklist_lock);
3819 INIT_LIST_HEAD(&sbinfo->shrinklist);
3820
3821 sb->s_maxbytes = MAX_LFS_FILESIZE;
3822 sb->s_blocksize = PAGE_SIZE;
3823 sb->s_blocksize_bits = PAGE_SHIFT;
3824 sb->s_magic = TMPFS_MAGIC;
3825 sb->s_op = &shmem_ops;
3826 sb->s_time_gran = 1;
3827#ifdef CONFIG_TMPFS_XATTR
3828 sb->s_xattr = shmem_xattr_handlers;
3829#endif
3830#ifdef CONFIG_TMPFS_POSIX_ACL
3831 sb->s_flags |= SB_POSIXACL;
3832#endif
3833 uuid_gen(&sb->s_uuid);
3834
3835 inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
3836 if (!inode)
3837 goto failed;
3838 inode->i_uid = sbinfo->uid;
3839 inode->i_gid = sbinfo->gid;
3840 sb->s_root = d_make_root(inode);
3841 if (!sb->s_root)
3842 goto failed;
3843 return 0;
3844
3845failed:
3846 shmem_put_super(sb);
3847 return err;
3848}
3849
3850static struct kmem_cache *shmem_inode_cachep;
3851
3852static struct inode *shmem_alloc_inode(struct super_block *sb)
3853{
3854 struct shmem_inode_info *info;
3855 info = kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
3856 if (!info)
3857 return NULL;
3858 return &info->vfs_inode;
3859}
3860
3861static void shmem_destroy_callback(struct rcu_head *head)
3862{
3863 struct inode *inode = container_of(head, struct inode, i_rcu);
3864 if (S_ISLNK(inode->i_mode))
3865 kfree(inode->i_link);
3866 kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
3867}
3868
3869static void shmem_destroy_inode(struct inode *inode)
3870{
3871 if (S_ISREG(inode->i_mode))
3872 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
3873 call_rcu(&inode->i_rcu, shmem_destroy_callback);
3874}
3875
3876static void shmem_init_inode(void *foo)
3877{
3878 struct shmem_inode_info *info = foo;
3879 inode_init_once(&info->vfs_inode);
3880}
3881
3882static void shmem_init_inodecache(void)
3883{
3884 shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
3885 sizeof(struct shmem_inode_info),
3886 0, SLAB_PANIC|SLAB_ACCOUNT, shmem_init_inode);
3887}
3888
3889static void shmem_destroy_inodecache(void)
3890{
3891 kmem_cache_destroy(shmem_inode_cachep);
3892}
3893
3894static const struct address_space_operations shmem_aops = {
3895 .writepage = shmem_writepage,
3896 .set_page_dirty = __set_page_dirty_no_writeback,
3897#ifdef CONFIG_TMPFS
3898 .write_begin = shmem_write_begin,
3899 .write_end = shmem_write_end,
3900#endif
3901#ifdef CONFIG_MIGRATION
3902 .migratepage = migrate_page,
3903#endif
3904 .error_remove_page = generic_error_remove_page,
3905};
3906
3907static const struct file_operations shmem_file_operations = {
3908 .mmap = shmem_mmap,
3909 .get_unmapped_area = shmem_get_unmapped_area,
3910#ifdef CONFIG_TMPFS
3911 .llseek = shmem_file_llseek,
3912 .read_iter = shmem_file_read_iter,
3913 .write_iter = generic_file_write_iter,
3914 .fsync = noop_fsync,
3915 .splice_read = generic_file_splice_read,
3916 .splice_write = iter_file_splice_write,
3917 .fallocate = shmem_fallocate,
3918#endif
3919};
3920
3921static const struct inode_operations shmem_inode_operations = {
3922 .getattr = shmem_getattr,
3923 .setattr = shmem_setattr,
3924#ifdef CONFIG_TMPFS_XATTR
3925 .listxattr = shmem_listxattr,
3926 .set_acl = simple_set_acl,
3927#endif
3928};
3929
3930static const struct inode_operations shmem_dir_inode_operations = {
3931#ifdef CONFIG_TMPFS
3932 .create = shmem_create,
3933 .lookup = simple_lookup,
3934 .link = shmem_link,
3935 .unlink = shmem_unlink,
3936 .symlink = shmem_symlink,
3937 .mkdir = shmem_mkdir,
3938 .rmdir = shmem_rmdir,
3939 .mknod = shmem_mknod,
3940 .rename = shmem_rename2,
3941 .tmpfile = shmem_tmpfile,
3942#endif
3943#ifdef CONFIG_TMPFS_XATTR
3944 .listxattr = shmem_listxattr,
3945#endif
3946#ifdef CONFIG_TMPFS_POSIX_ACL
3947 .setattr = shmem_setattr,
3948 .set_acl = simple_set_acl,
3949#endif
3950};
3951
3952static const struct inode_operations shmem_special_inode_operations = {
3953#ifdef CONFIG_TMPFS_XATTR
3954 .listxattr = shmem_listxattr,
3955#endif
3956#ifdef CONFIG_TMPFS_POSIX_ACL
3957 .setattr = shmem_setattr,
3958 .set_acl = simple_set_acl,
3959#endif
3960};
3961
3962static const struct super_operations shmem_ops = {
3963 .alloc_inode = shmem_alloc_inode,
3964 .destroy_inode = shmem_destroy_inode,
3965#ifdef CONFIG_TMPFS
3966 .statfs = shmem_statfs,
3967 .remount_fs = shmem_remount_fs,
3968 .show_options = shmem_show_options,
3969#endif
3970 .evict_inode = shmem_evict_inode,
3971 .drop_inode = generic_delete_inode,
3972 .put_super = shmem_put_super,
3973#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3974 .nr_cached_objects = shmem_unused_huge_count,
3975 .free_cached_objects = shmem_unused_huge_scan,
3976#endif
3977};
3978
3979static const struct vm_operations_struct shmem_vm_ops = {
3980 .fault = shmem_fault,
3981 .map_pages = filemap_map_pages,
3982#ifdef CONFIG_NUMA
3983 .set_policy = shmem_set_policy,
3984 .get_policy = shmem_get_policy,
3985#endif
3986};
3987
3988static struct dentry *shmem_mount(struct file_system_type *fs_type,
3989 int flags, const char *dev_name, void *data)
3990{
3991 return mount_nodev(fs_type, flags, data, shmem_fill_super);
3992}
3993
3994static struct file_system_type shmem_fs_type = {
3995 .owner = THIS_MODULE,
3996 .name = "tmpfs",
3997 .mount = shmem_mount,
3998 .kill_sb = kill_litter_super,
3999 .fs_flags = FS_USERNS_MOUNT,
4000};
4001
4002int __init shmem_init(void)
4003{
4004 int error;
4005
4006 /* If rootfs called this, don't re-init */
4007 if (shmem_inode_cachep)
4008 return 0;
4009
4010 shmem_init_inodecache();
4011
4012 error = register_filesystem(&shmem_fs_type);
4013 if (error) {
4014 pr_err("Could not register tmpfs\n");
4015 goto out2;
4016 }
4017
4018 shm_mnt = kern_mount(&shmem_fs_type);
4019 if (IS_ERR(shm_mnt)) {
4020 error = PTR_ERR(shm_mnt);
4021 pr_err("Could not kern_mount tmpfs\n");
4022 goto out1;
4023 }
4024
4025#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
4026 if (has_transparent_hugepage() && shmem_huge > SHMEM_HUGE_DENY)
4027 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
4028 else
4029 shmem_huge = 0; /* just in case it was patched */
4030#endif
4031 return 0;
4032
4033out1:
4034 unregister_filesystem(&shmem_fs_type);
4035out2:
4036 shmem_destroy_inodecache();
4037 shm_mnt = ERR_PTR(error);
4038 return error;
4039}
4040
4041#if defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE) && defined(CONFIG_SYSFS)
4042static ssize_t shmem_enabled_show(struct kobject *kobj,
4043 struct kobj_attribute *attr, char *buf)
4044{
4045 int values[] = {
4046 SHMEM_HUGE_ALWAYS,
4047 SHMEM_HUGE_WITHIN_SIZE,
4048 SHMEM_HUGE_ADVISE,
4049 SHMEM_HUGE_NEVER,
4050 SHMEM_HUGE_DENY,
4051 SHMEM_HUGE_FORCE,
4052 };
4053 int i, count;
4054
4055 for (i = 0, count = 0; i < ARRAY_SIZE(values); i++) {
4056 const char *fmt = shmem_huge == values[i] ? "[%s] " : "%s ";
4057
4058 count += sprintf(buf + count, fmt,
4059 shmem_format_huge(values[i]));
4060 }
4061 buf[count - 1] = '\n';
4062 return count;
4063}
4064
4065static ssize_t shmem_enabled_store(struct kobject *kobj,
4066 struct kobj_attribute *attr, const char *buf, size_t count)
4067{
4068 char tmp[16];
4069 int huge;
4070
4071 if (count + 1 > sizeof(tmp))
4072 return -EINVAL;
4073 memcpy(tmp, buf, count);
4074 tmp[count] = '\0';
4075 if (count && tmp[count - 1] == '\n')
4076 tmp[count - 1] = '\0';
4077
4078 huge = shmem_parse_huge(tmp);
4079 if (huge == -EINVAL)
4080 return -EINVAL;
4081 if (!has_transparent_hugepage() &&
4082 huge != SHMEM_HUGE_NEVER && huge != SHMEM_HUGE_DENY)
4083 return -EINVAL;
4084
4085 shmem_huge = huge;
4086 if (shmem_huge > SHMEM_HUGE_DENY)
4087 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
4088 return count;
4089}
4090
4091struct kobj_attribute shmem_enabled_attr =
4092 __ATTR(shmem_enabled, 0644, shmem_enabled_show, shmem_enabled_store);
4093#endif /* CONFIG_TRANSPARENT_HUGE_PAGECACHE && CONFIG_SYSFS */
4094
4095#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
4096bool shmem_huge_enabled(struct vm_area_struct *vma)
4097{
4098 struct inode *inode = file_inode(vma->vm_file);
4099 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
4100 loff_t i_size;
4101 pgoff_t off;
4102
4103 if (shmem_huge == SHMEM_HUGE_FORCE)
4104 return true;
4105 if (shmem_huge == SHMEM_HUGE_DENY)
4106 return false;
4107 switch (sbinfo->huge) {
4108 case SHMEM_HUGE_NEVER:
4109 return false;
4110 case SHMEM_HUGE_ALWAYS:
4111 return true;
4112 case SHMEM_HUGE_WITHIN_SIZE:
4113 off = round_up(vma->vm_pgoff, HPAGE_PMD_NR);
4114 i_size = round_up(i_size_read(inode), PAGE_SIZE);
4115 if (i_size >= HPAGE_PMD_SIZE &&
4116 i_size >> PAGE_SHIFT >= off)
4117 return true;
4118 /* fall through */
4119 case SHMEM_HUGE_ADVISE:
4120 /* TODO: implement fadvise() hints */
4121 return (vma->vm_flags & VM_HUGEPAGE);
4122 default:
4123 VM_BUG_ON(1);
4124 return false;
4125 }
4126}
4127#endif /* CONFIG_TRANSPARENT_HUGE_PAGECACHE */
4128
4129#else /* !CONFIG_SHMEM */
4130
4131/*
4132 * tiny-shmem: simple shmemfs and tmpfs using ramfs code
4133 *
4134 * This is intended for small system where the benefits of the full
4135 * shmem code (swap-backed and resource-limited) are outweighed by
4136 * their complexity. On systems without swap this code should be
4137 * effectively equivalent, but much lighter weight.
4138 */
4139
4140static struct file_system_type shmem_fs_type = {
4141 .name = "tmpfs",
4142 .mount = ramfs_mount,
4143 .kill_sb = kill_litter_super,
4144 .fs_flags = FS_USERNS_MOUNT,
4145};
4146
4147int __init shmem_init(void)
4148{
4149 BUG_ON(register_filesystem(&shmem_fs_type) != 0);
4150
4151 shm_mnt = kern_mount(&shmem_fs_type);
4152 BUG_ON(IS_ERR(shm_mnt));
4153
4154 return 0;
4155}
4156
4157int shmem_unuse(swp_entry_t swap, struct page *page)
4158{
4159 return 0;
4160}
4161
4162int shmem_lock(struct file *file, int lock, struct user_struct *user)
4163{
4164 return 0;
4165}
4166
4167void shmem_unlock_mapping(struct address_space *mapping)
4168{
4169}
4170
4171#ifdef CONFIG_MMU
4172unsigned long shmem_get_unmapped_area(struct file *file,
4173 unsigned long addr, unsigned long len,
4174 unsigned long pgoff, unsigned long flags)
4175{
4176 return current->mm->get_unmapped_area(file, addr, len, pgoff, flags);
4177}
4178#endif
4179
4180void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
4181{
4182 truncate_inode_pages_range(inode->i_mapping, lstart, lend);
4183}
4184EXPORT_SYMBOL_GPL(shmem_truncate_range);
4185
4186#define shmem_vm_ops generic_file_vm_ops
4187#define shmem_file_operations ramfs_file_operations
4188#define shmem_get_inode(sb, dir, mode, dev, flags) ramfs_get_inode(sb, dir, mode, dev)
4189#define shmem_acct_size(flags, size) 0
4190#define shmem_unacct_size(flags, size) do {} while (0)
4191
4192#endif /* CONFIG_SHMEM */
4193
4194/* common code */
4195
4196static const struct dentry_operations anon_ops = {
4197 .d_dname = simple_dname
4198};
4199
4200static struct file *__shmem_file_setup(struct vfsmount *mnt, const char *name, loff_t size,
4201 unsigned long flags, unsigned int i_flags)
4202{
4203 struct file *res;
4204 struct inode *inode;
4205 struct path path;
4206 struct super_block *sb;
4207 struct qstr this;
4208
4209 if (IS_ERR(mnt))
4210 return ERR_CAST(mnt);
4211
4212 if (size < 0 || size > MAX_LFS_FILESIZE)
4213 return ERR_PTR(-EINVAL);
4214
4215 if (shmem_acct_size(flags, size))
4216 return ERR_PTR(-ENOMEM);
4217
4218 res = ERR_PTR(-ENOMEM);
4219 this.name = name;
4220 this.len = strlen(name);
4221 this.hash = 0; /* will go */
4222 sb = mnt->mnt_sb;
4223 path.mnt = mntget(mnt);
4224 path.dentry = d_alloc_pseudo(sb, &this);
4225 if (!path.dentry)
4226 goto put_memory;
4227 d_set_d_op(path.dentry, &anon_ops);
4228
4229 res = ERR_PTR(-ENOSPC);
4230 inode = shmem_get_inode(sb, NULL, S_IFREG | S_IRWXUGO, 0, flags);
4231 if (!inode)
4232 goto put_memory;
4233
4234 inode->i_flags |= i_flags;
4235 d_instantiate(path.dentry, inode);
4236 inode->i_size = size;
4237 clear_nlink(inode); /* It is unlinked */
4238 res = ERR_PTR(ramfs_nommu_expand_for_mapping(inode, size));
4239 if (IS_ERR(res))
4240 goto put_path;
4241
4242 res = alloc_file(&path, FMODE_WRITE | FMODE_READ,
4243 &shmem_file_operations);
4244 if (IS_ERR(res))
4245 goto put_path;
4246
4247 return res;
4248
4249put_memory:
4250 shmem_unacct_size(flags, size);
4251put_path:
4252 path_put(&path);
4253 return res;
4254}
4255
4256/**
4257 * shmem_kernel_file_setup - get an unlinked file living in tmpfs which must be
4258 * kernel internal. There will be NO LSM permission checks against the
4259 * underlying inode. So users of this interface must do LSM checks at a
4260 * higher layer. The users are the big_key and shm implementations. LSM
4261 * checks are provided at the key or shm level rather than the inode.
4262 * @name: name for dentry (to be seen in /proc/<pid>/maps
4263 * @size: size to be set for the file
4264 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4265 */
4266struct file *shmem_kernel_file_setup(const char *name, loff_t size, unsigned long flags)
4267{
4268 return __shmem_file_setup(shm_mnt, name, size, flags, S_PRIVATE);
4269}
4270
4271/**
4272 * shmem_file_setup - get an unlinked file living in tmpfs
4273 * @name: name for dentry (to be seen in /proc/<pid>/maps
4274 * @size: size to be set for the file
4275 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4276 */
4277struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
4278{
4279 return __shmem_file_setup(shm_mnt, name, size, flags, 0);
4280}
4281EXPORT_SYMBOL_GPL(shmem_file_setup);
4282
4283/**
4284 * shmem_file_setup_with_mnt - get an unlinked file living in tmpfs
4285 * @mnt: the tmpfs mount where the file will be created
4286 * @name: name for dentry (to be seen in /proc/<pid>/maps
4287 * @size: size to be set for the file
4288 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4289 */
4290struct file *shmem_file_setup_with_mnt(struct vfsmount *mnt, const char *name,
4291 loff_t size, unsigned long flags)
4292{
4293 return __shmem_file_setup(mnt, name, size, flags, 0);
4294}
4295EXPORT_SYMBOL_GPL(shmem_file_setup_with_mnt);
4296
4297/**
4298 * shmem_zero_setup - setup a shared anonymous mapping
4299 * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
4300 */
4301int shmem_zero_setup(struct vm_area_struct *vma)
4302{
4303 struct file *file;
4304 loff_t size = vma->vm_end - vma->vm_start;
4305
4306 /*
4307 * Cloning a new file under mmap_sem leads to a lock ordering conflict
4308 * between XFS directory reading and selinux: since this file is only
4309 * accessible to the user through its mapping, use S_PRIVATE flag to
4310 * bypass file security, in the same way as shmem_kernel_file_setup().
4311 */
4312 file = shmem_kernel_file_setup("dev/zero", size, vma->vm_flags);
4313 if (IS_ERR(file))
4314 return PTR_ERR(file);
4315
4316 if (vma->vm_file)
4317 fput(vma->vm_file);
4318 vma->vm_file = file;
4319 vma->vm_ops = &shmem_vm_ops;
4320
4321 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE) &&
4322 ((vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK) <
4323 (vma->vm_end & HPAGE_PMD_MASK)) {
4324 khugepaged_enter(vma, vma->vm_flags);
4325 }
4326
4327 return 0;
4328}
4329
4330/**
4331 * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
4332 * @mapping: the page's address_space
4333 * @index: the page index
4334 * @gfp: the page allocator flags to use if allocating
4335 *
4336 * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
4337 * with any new page allocations done using the specified allocation flags.
4338 * But read_cache_page_gfp() uses the ->readpage() method: which does not
4339 * suit tmpfs, since it may have pages in swapcache, and needs to find those
4340 * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
4341 *
4342 * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
4343 * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
4344 */
4345struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
4346 pgoff_t index, gfp_t gfp)
4347{
4348#ifdef CONFIG_SHMEM
4349 struct inode *inode = mapping->host;
4350 struct page *page;
4351 int error;
4352
4353 BUG_ON(mapping->a_ops != &shmem_aops);
4354 error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE,
4355 gfp, NULL, NULL, NULL);
4356 if (error)
4357 page = ERR_PTR(error);
4358 else
4359 unlock_page(page);
4360 return page;
4361#else
4362 /*
4363 * The tiny !SHMEM case uses ramfs without swap
4364 */
4365 return read_cache_page_gfp(mapping, index, gfp);
4366#endif
4367}
4368EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);
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#include <linux/khugepaged.h>
36
37static struct vfsmount *shm_mnt;
38
39#ifdef CONFIG_SHMEM
40/*
41 * This virtual memory filesystem is heavily based on the ramfs. It
42 * extends ramfs by the ability to use swap and honor resource limits
43 * which makes it a completely usable filesystem.
44 */
45
46#include <linux/xattr.h>
47#include <linux/exportfs.h>
48#include <linux/posix_acl.h>
49#include <linux/posix_acl_xattr.h>
50#include <linux/mman.h>
51#include <linux/string.h>
52#include <linux/slab.h>
53#include <linux/backing-dev.h>
54#include <linux/shmem_fs.h>
55#include <linux/writeback.h>
56#include <linux/blkdev.h>
57#include <linux/pagevec.h>
58#include <linux/percpu_counter.h>
59#include <linux/falloc.h>
60#include <linux/splice.h>
61#include <linux/security.h>
62#include <linux/swapops.h>
63#include <linux/mempolicy.h>
64#include <linux/namei.h>
65#include <linux/ctype.h>
66#include <linux/migrate.h>
67#include <linux/highmem.h>
68#include <linux/seq_file.h>
69#include <linux/magic.h>
70#include <linux/syscalls.h>
71#include <linux/fcntl.h>
72#include <uapi/linux/memfd.h>
73
74#include <linux/uaccess.h>
75#include <asm/pgtable.h>
76
77#include "internal.h"
78
79#define BLOCKS_PER_PAGE (PAGE_SIZE/512)
80#define VM_ACCT(size) (PAGE_ALIGN(size) >> PAGE_SHIFT)
81
82/* Pretend that each entry is of this size in directory's i_size */
83#define BOGO_DIRENT_SIZE 20
84
85/* Symlink up to this size is kmalloc'ed instead of using a swappable page */
86#define SHORT_SYMLINK_LEN 128
87
88/*
89 * shmem_fallocate communicates with shmem_fault or shmem_writepage via
90 * inode->i_private (with i_mutex making sure that it has only one user at
91 * a time): we would prefer not to enlarge the shmem inode just for that.
92 */
93struct shmem_falloc {
94 wait_queue_head_t *waitq; /* faults into hole wait for punch to end */
95 pgoff_t start; /* start of range currently being fallocated */
96 pgoff_t next; /* the next page offset to be fallocated */
97 pgoff_t nr_falloced; /* how many new pages have been fallocated */
98 pgoff_t nr_unswapped; /* how often writepage refused to swap out */
99};
100
101#ifdef CONFIG_TMPFS
102static unsigned long shmem_default_max_blocks(void)
103{
104 return totalram_pages / 2;
105}
106
107static unsigned long shmem_default_max_inodes(void)
108{
109 return min(totalram_pages - totalhigh_pages, totalram_pages / 2);
110}
111#endif
112
113static bool shmem_should_replace_page(struct page *page, gfp_t gfp);
114static int shmem_replace_page(struct page **pagep, gfp_t gfp,
115 struct shmem_inode_info *info, pgoff_t index);
116static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
117 struct page **pagep, enum sgp_type sgp,
118 gfp_t gfp, struct mm_struct *fault_mm, int *fault_type);
119
120int shmem_getpage(struct inode *inode, pgoff_t index,
121 struct page **pagep, enum sgp_type sgp)
122{
123 return shmem_getpage_gfp(inode, index, pagep, sgp,
124 mapping_gfp_mask(inode->i_mapping), NULL, NULL);
125}
126
127static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
128{
129 return sb->s_fs_info;
130}
131
132/*
133 * shmem_file_setup pre-accounts the whole fixed size of a VM object,
134 * for shared memory and for shared anonymous (/dev/zero) mappings
135 * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
136 * consistent with the pre-accounting of private mappings ...
137 */
138static inline int shmem_acct_size(unsigned long flags, loff_t size)
139{
140 return (flags & VM_NORESERVE) ?
141 0 : security_vm_enough_memory_mm(current->mm, VM_ACCT(size));
142}
143
144static inline void shmem_unacct_size(unsigned long flags, loff_t size)
145{
146 if (!(flags & VM_NORESERVE))
147 vm_unacct_memory(VM_ACCT(size));
148}
149
150static inline int shmem_reacct_size(unsigned long flags,
151 loff_t oldsize, loff_t newsize)
152{
153 if (!(flags & VM_NORESERVE)) {
154 if (VM_ACCT(newsize) > VM_ACCT(oldsize))
155 return security_vm_enough_memory_mm(current->mm,
156 VM_ACCT(newsize) - VM_ACCT(oldsize));
157 else if (VM_ACCT(newsize) < VM_ACCT(oldsize))
158 vm_unacct_memory(VM_ACCT(oldsize) - VM_ACCT(newsize));
159 }
160 return 0;
161}
162
163/*
164 * ... whereas tmpfs objects are accounted incrementally as
165 * pages are allocated, in order to allow large sparse files.
166 * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
167 * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
168 */
169static inline int shmem_acct_block(unsigned long flags, long pages)
170{
171 if (!(flags & VM_NORESERVE))
172 return 0;
173
174 return security_vm_enough_memory_mm(current->mm,
175 pages * VM_ACCT(PAGE_SIZE));
176}
177
178static inline void shmem_unacct_blocks(unsigned long flags, long pages)
179{
180 if (flags & VM_NORESERVE)
181 vm_unacct_memory(pages * VM_ACCT(PAGE_SIZE));
182}
183
184static const struct super_operations shmem_ops;
185static const struct address_space_operations shmem_aops;
186static const struct file_operations shmem_file_operations;
187static const struct inode_operations shmem_inode_operations;
188static const struct inode_operations shmem_dir_inode_operations;
189static const struct inode_operations shmem_special_inode_operations;
190static const struct vm_operations_struct shmem_vm_ops;
191static struct file_system_type shmem_fs_type;
192
193static LIST_HEAD(shmem_swaplist);
194static DEFINE_MUTEX(shmem_swaplist_mutex);
195
196static int shmem_reserve_inode(struct super_block *sb)
197{
198 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
199 if (sbinfo->max_inodes) {
200 spin_lock(&sbinfo->stat_lock);
201 if (!sbinfo->free_inodes) {
202 spin_unlock(&sbinfo->stat_lock);
203 return -ENOSPC;
204 }
205 sbinfo->free_inodes--;
206 spin_unlock(&sbinfo->stat_lock);
207 }
208 return 0;
209}
210
211static void shmem_free_inode(struct super_block *sb)
212{
213 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
214 if (sbinfo->max_inodes) {
215 spin_lock(&sbinfo->stat_lock);
216 sbinfo->free_inodes++;
217 spin_unlock(&sbinfo->stat_lock);
218 }
219}
220
221/**
222 * shmem_recalc_inode - recalculate the block usage of an inode
223 * @inode: inode to recalc
224 *
225 * We have to calculate the free blocks since the mm can drop
226 * undirtied hole pages behind our back.
227 *
228 * But normally info->alloced == inode->i_mapping->nrpages + info->swapped
229 * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
230 *
231 * It has to be called with the spinlock held.
232 */
233static void shmem_recalc_inode(struct inode *inode)
234{
235 struct shmem_inode_info *info = SHMEM_I(inode);
236 long freed;
237
238 freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
239 if (freed > 0) {
240 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
241 if (sbinfo->max_blocks)
242 percpu_counter_add(&sbinfo->used_blocks, -freed);
243 info->alloced -= freed;
244 inode->i_blocks -= freed * BLOCKS_PER_PAGE;
245 shmem_unacct_blocks(info->flags, freed);
246 }
247}
248
249bool shmem_charge(struct inode *inode, long pages)
250{
251 struct shmem_inode_info *info = SHMEM_I(inode);
252 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
253 unsigned long flags;
254
255 if (shmem_acct_block(info->flags, pages))
256 return false;
257 spin_lock_irqsave(&info->lock, flags);
258 info->alloced += pages;
259 inode->i_blocks += pages * BLOCKS_PER_PAGE;
260 shmem_recalc_inode(inode);
261 spin_unlock_irqrestore(&info->lock, flags);
262 inode->i_mapping->nrpages += pages;
263
264 if (!sbinfo->max_blocks)
265 return true;
266 if (percpu_counter_compare(&sbinfo->used_blocks,
267 sbinfo->max_blocks - pages) > 0) {
268 inode->i_mapping->nrpages -= pages;
269 spin_lock_irqsave(&info->lock, flags);
270 info->alloced -= pages;
271 shmem_recalc_inode(inode);
272 spin_unlock_irqrestore(&info->lock, flags);
273 shmem_unacct_blocks(info->flags, pages);
274 return false;
275 }
276 percpu_counter_add(&sbinfo->used_blocks, pages);
277 return true;
278}
279
280void shmem_uncharge(struct inode *inode, long pages)
281{
282 struct shmem_inode_info *info = SHMEM_I(inode);
283 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
284 unsigned long flags;
285
286 spin_lock_irqsave(&info->lock, flags);
287 info->alloced -= pages;
288 inode->i_blocks -= pages * BLOCKS_PER_PAGE;
289 shmem_recalc_inode(inode);
290 spin_unlock_irqrestore(&info->lock, flags);
291
292 if (sbinfo->max_blocks)
293 percpu_counter_sub(&sbinfo->used_blocks, pages);
294 shmem_unacct_blocks(info->flags, pages);
295}
296
297/*
298 * Replace item expected in radix tree by a new item, while holding tree lock.
299 */
300static int shmem_radix_tree_replace(struct address_space *mapping,
301 pgoff_t index, void *expected, void *replacement)
302{
303 struct radix_tree_node *node;
304 void **pslot;
305 void *item;
306
307 VM_BUG_ON(!expected);
308 VM_BUG_ON(!replacement);
309 item = __radix_tree_lookup(&mapping->page_tree, index, &node, &pslot);
310 if (!item)
311 return -ENOENT;
312 if (item != expected)
313 return -ENOENT;
314 __radix_tree_replace(&mapping->page_tree, node, pslot,
315 replacement, NULL, NULL);
316 return 0;
317}
318
319/*
320 * Sometimes, before we decide whether to proceed or to fail, we must check
321 * that an entry was not already brought back from swap by a racing thread.
322 *
323 * Checking page is not enough: by the time a SwapCache page is locked, it
324 * might be reused, and again be SwapCache, using the same swap as before.
325 */
326static bool shmem_confirm_swap(struct address_space *mapping,
327 pgoff_t index, swp_entry_t swap)
328{
329 void *item;
330
331 rcu_read_lock();
332 item = radix_tree_lookup(&mapping->page_tree, index);
333 rcu_read_unlock();
334 return item == swp_to_radix_entry(swap);
335}
336
337/*
338 * Definitions for "huge tmpfs": tmpfs mounted with the huge= option
339 *
340 * SHMEM_HUGE_NEVER:
341 * disables huge pages for the mount;
342 * SHMEM_HUGE_ALWAYS:
343 * enables huge pages for the mount;
344 * SHMEM_HUGE_WITHIN_SIZE:
345 * only allocate huge pages if the page will be fully within i_size,
346 * also respect fadvise()/madvise() hints;
347 * SHMEM_HUGE_ADVISE:
348 * only allocate huge pages if requested with fadvise()/madvise();
349 */
350
351#define SHMEM_HUGE_NEVER 0
352#define SHMEM_HUGE_ALWAYS 1
353#define SHMEM_HUGE_WITHIN_SIZE 2
354#define SHMEM_HUGE_ADVISE 3
355
356/*
357 * Special values.
358 * Only can be set via /sys/kernel/mm/transparent_hugepage/shmem_enabled:
359 *
360 * SHMEM_HUGE_DENY:
361 * disables huge on shm_mnt and all mounts, for emergency use;
362 * SHMEM_HUGE_FORCE:
363 * enables huge on shm_mnt and all mounts, w/o needing option, for testing;
364 *
365 */
366#define SHMEM_HUGE_DENY (-1)
367#define SHMEM_HUGE_FORCE (-2)
368
369#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
370/* ifdef here to avoid bloating shmem.o when not necessary */
371
372int shmem_huge __read_mostly;
373
374#if defined(CONFIG_SYSFS) || defined(CONFIG_TMPFS)
375static int shmem_parse_huge(const char *str)
376{
377 if (!strcmp(str, "never"))
378 return SHMEM_HUGE_NEVER;
379 if (!strcmp(str, "always"))
380 return SHMEM_HUGE_ALWAYS;
381 if (!strcmp(str, "within_size"))
382 return SHMEM_HUGE_WITHIN_SIZE;
383 if (!strcmp(str, "advise"))
384 return SHMEM_HUGE_ADVISE;
385 if (!strcmp(str, "deny"))
386 return SHMEM_HUGE_DENY;
387 if (!strcmp(str, "force"))
388 return SHMEM_HUGE_FORCE;
389 return -EINVAL;
390}
391
392static const char *shmem_format_huge(int huge)
393{
394 switch (huge) {
395 case SHMEM_HUGE_NEVER:
396 return "never";
397 case SHMEM_HUGE_ALWAYS:
398 return "always";
399 case SHMEM_HUGE_WITHIN_SIZE:
400 return "within_size";
401 case SHMEM_HUGE_ADVISE:
402 return "advise";
403 case SHMEM_HUGE_DENY:
404 return "deny";
405 case SHMEM_HUGE_FORCE:
406 return "force";
407 default:
408 VM_BUG_ON(1);
409 return "bad_val";
410 }
411}
412#endif
413
414static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
415 struct shrink_control *sc, unsigned long nr_to_split)
416{
417 LIST_HEAD(list), *pos, *next;
418 LIST_HEAD(to_remove);
419 struct inode *inode;
420 struct shmem_inode_info *info;
421 struct page *page;
422 unsigned long batch = sc ? sc->nr_to_scan : 128;
423 int removed = 0, split = 0;
424
425 if (list_empty(&sbinfo->shrinklist))
426 return SHRINK_STOP;
427
428 spin_lock(&sbinfo->shrinklist_lock);
429 list_for_each_safe(pos, next, &sbinfo->shrinklist) {
430 info = list_entry(pos, struct shmem_inode_info, shrinklist);
431
432 /* pin the inode */
433 inode = igrab(&info->vfs_inode);
434
435 /* inode is about to be evicted */
436 if (!inode) {
437 list_del_init(&info->shrinklist);
438 removed++;
439 goto next;
440 }
441
442 /* Check if there's anything to gain */
443 if (round_up(inode->i_size, PAGE_SIZE) ==
444 round_up(inode->i_size, HPAGE_PMD_SIZE)) {
445 list_move(&info->shrinklist, &to_remove);
446 removed++;
447 goto next;
448 }
449
450 list_move(&info->shrinklist, &list);
451next:
452 if (!--batch)
453 break;
454 }
455 spin_unlock(&sbinfo->shrinklist_lock);
456
457 list_for_each_safe(pos, next, &to_remove) {
458 info = list_entry(pos, struct shmem_inode_info, shrinklist);
459 inode = &info->vfs_inode;
460 list_del_init(&info->shrinklist);
461 iput(inode);
462 }
463
464 list_for_each_safe(pos, next, &list) {
465 int ret;
466
467 info = list_entry(pos, struct shmem_inode_info, shrinklist);
468 inode = &info->vfs_inode;
469
470 if (nr_to_split && split >= nr_to_split) {
471 iput(inode);
472 continue;
473 }
474
475 page = find_lock_page(inode->i_mapping,
476 (inode->i_size & HPAGE_PMD_MASK) >> PAGE_SHIFT);
477 if (!page)
478 goto drop;
479
480 if (!PageTransHuge(page)) {
481 unlock_page(page);
482 put_page(page);
483 goto drop;
484 }
485
486 ret = split_huge_page(page);
487 unlock_page(page);
488 put_page(page);
489
490 if (ret) {
491 /* split failed: leave it on the list */
492 iput(inode);
493 continue;
494 }
495
496 split++;
497drop:
498 list_del_init(&info->shrinklist);
499 removed++;
500 iput(inode);
501 }
502
503 spin_lock(&sbinfo->shrinklist_lock);
504 list_splice_tail(&list, &sbinfo->shrinklist);
505 sbinfo->shrinklist_len -= removed;
506 spin_unlock(&sbinfo->shrinklist_lock);
507
508 return split;
509}
510
511static long shmem_unused_huge_scan(struct super_block *sb,
512 struct shrink_control *sc)
513{
514 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
515
516 if (!READ_ONCE(sbinfo->shrinklist_len))
517 return SHRINK_STOP;
518
519 return shmem_unused_huge_shrink(sbinfo, sc, 0);
520}
521
522static long shmem_unused_huge_count(struct super_block *sb,
523 struct shrink_control *sc)
524{
525 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
526 return READ_ONCE(sbinfo->shrinklist_len);
527}
528#else /* !CONFIG_TRANSPARENT_HUGE_PAGECACHE */
529
530#define shmem_huge SHMEM_HUGE_DENY
531
532static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
533 struct shrink_control *sc, unsigned long nr_to_split)
534{
535 return 0;
536}
537#endif /* CONFIG_TRANSPARENT_HUGE_PAGECACHE */
538
539/*
540 * Like add_to_page_cache_locked, but error if expected item has gone.
541 */
542static int shmem_add_to_page_cache(struct page *page,
543 struct address_space *mapping,
544 pgoff_t index, void *expected)
545{
546 int error, nr = hpage_nr_pages(page);
547
548 VM_BUG_ON_PAGE(PageTail(page), page);
549 VM_BUG_ON_PAGE(index != round_down(index, nr), page);
550 VM_BUG_ON_PAGE(!PageLocked(page), page);
551 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
552 VM_BUG_ON(expected && PageTransHuge(page));
553
554 page_ref_add(page, nr);
555 page->mapping = mapping;
556 page->index = index;
557
558 spin_lock_irq(&mapping->tree_lock);
559 if (PageTransHuge(page)) {
560 void __rcu **results;
561 pgoff_t idx;
562 int i;
563
564 error = 0;
565 if (radix_tree_gang_lookup_slot(&mapping->page_tree,
566 &results, &idx, index, 1) &&
567 idx < index + HPAGE_PMD_NR) {
568 error = -EEXIST;
569 }
570
571 if (!error) {
572 for (i = 0; i < HPAGE_PMD_NR; i++) {
573 error = radix_tree_insert(&mapping->page_tree,
574 index + i, page + i);
575 VM_BUG_ON(error);
576 }
577 count_vm_event(THP_FILE_ALLOC);
578 }
579 } else if (!expected) {
580 error = radix_tree_insert(&mapping->page_tree, index, page);
581 } else {
582 error = shmem_radix_tree_replace(mapping, index, expected,
583 page);
584 }
585
586 if (!error) {
587 mapping->nrpages += nr;
588 if (PageTransHuge(page))
589 __inc_node_page_state(page, NR_SHMEM_THPS);
590 __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, nr);
591 __mod_node_page_state(page_pgdat(page), NR_SHMEM, nr);
592 spin_unlock_irq(&mapping->tree_lock);
593 } else {
594 page->mapping = NULL;
595 spin_unlock_irq(&mapping->tree_lock);
596 page_ref_sub(page, nr);
597 }
598 return error;
599}
600
601/*
602 * Like delete_from_page_cache, but substitutes swap for page.
603 */
604static void shmem_delete_from_page_cache(struct page *page, void *radswap)
605{
606 struct address_space *mapping = page->mapping;
607 int error;
608
609 VM_BUG_ON_PAGE(PageCompound(page), page);
610
611 spin_lock_irq(&mapping->tree_lock);
612 error = shmem_radix_tree_replace(mapping, page->index, page, radswap);
613 page->mapping = NULL;
614 mapping->nrpages--;
615 __dec_node_page_state(page, NR_FILE_PAGES);
616 __dec_node_page_state(page, NR_SHMEM);
617 spin_unlock_irq(&mapping->tree_lock);
618 put_page(page);
619 BUG_ON(error);
620}
621
622/*
623 * Remove swap entry from radix tree, free the swap and its page cache.
624 */
625static int shmem_free_swap(struct address_space *mapping,
626 pgoff_t index, void *radswap)
627{
628 void *old;
629
630 spin_lock_irq(&mapping->tree_lock);
631 old = radix_tree_delete_item(&mapping->page_tree, index, radswap);
632 spin_unlock_irq(&mapping->tree_lock);
633 if (old != radswap)
634 return -ENOENT;
635 free_swap_and_cache(radix_to_swp_entry(radswap));
636 return 0;
637}
638
639/*
640 * Determine (in bytes) how many of the shmem object's pages mapped by the
641 * given offsets are swapped out.
642 *
643 * This is safe to call without i_mutex or mapping->tree_lock thanks to RCU,
644 * as long as the inode doesn't go away and racy results are not a problem.
645 */
646unsigned long shmem_partial_swap_usage(struct address_space *mapping,
647 pgoff_t start, pgoff_t end)
648{
649 struct radix_tree_iter iter;
650 void **slot;
651 struct page *page;
652 unsigned long swapped = 0;
653
654 rcu_read_lock();
655
656 radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
657 if (iter.index >= end)
658 break;
659
660 page = radix_tree_deref_slot(slot);
661
662 if (radix_tree_deref_retry(page)) {
663 slot = radix_tree_iter_retry(&iter);
664 continue;
665 }
666
667 if (radix_tree_exceptional_entry(page))
668 swapped++;
669
670 if (need_resched()) {
671 slot = radix_tree_iter_resume(slot, &iter);
672 cond_resched_rcu();
673 }
674 }
675
676 rcu_read_unlock();
677
678 return swapped << PAGE_SHIFT;
679}
680
681/*
682 * Determine (in bytes) how many of the shmem object's pages mapped by the
683 * given vma is swapped out.
684 *
685 * This is safe to call without i_mutex or mapping->tree_lock thanks to RCU,
686 * as long as the inode doesn't go away and racy results are not a problem.
687 */
688unsigned long shmem_swap_usage(struct vm_area_struct *vma)
689{
690 struct inode *inode = file_inode(vma->vm_file);
691 struct shmem_inode_info *info = SHMEM_I(inode);
692 struct address_space *mapping = inode->i_mapping;
693 unsigned long swapped;
694
695 /* Be careful as we don't hold info->lock */
696 swapped = READ_ONCE(info->swapped);
697
698 /*
699 * The easier cases are when the shmem object has nothing in swap, or
700 * the vma maps it whole. Then we can simply use the stats that we
701 * already track.
702 */
703 if (!swapped)
704 return 0;
705
706 if (!vma->vm_pgoff && vma->vm_end - vma->vm_start >= inode->i_size)
707 return swapped << PAGE_SHIFT;
708
709 /* Here comes the more involved part */
710 return shmem_partial_swap_usage(mapping,
711 linear_page_index(vma, vma->vm_start),
712 linear_page_index(vma, vma->vm_end));
713}
714
715/*
716 * SysV IPC SHM_UNLOCK restore Unevictable pages to their evictable lists.
717 */
718void shmem_unlock_mapping(struct address_space *mapping)
719{
720 struct pagevec pvec;
721 pgoff_t indices[PAGEVEC_SIZE];
722 pgoff_t index = 0;
723
724 pagevec_init(&pvec, 0);
725 /*
726 * Minor point, but we might as well stop if someone else SHM_LOCKs it.
727 */
728 while (!mapping_unevictable(mapping)) {
729 /*
730 * Avoid pagevec_lookup(): find_get_pages() returns 0 as if it
731 * has finished, if it hits a row of PAGEVEC_SIZE swap entries.
732 */
733 pvec.nr = find_get_entries(mapping, index,
734 PAGEVEC_SIZE, pvec.pages, indices);
735 if (!pvec.nr)
736 break;
737 index = indices[pvec.nr - 1] + 1;
738 pagevec_remove_exceptionals(&pvec);
739 check_move_unevictable_pages(pvec.pages, pvec.nr);
740 pagevec_release(&pvec);
741 cond_resched();
742 }
743}
744
745/*
746 * Remove range of pages and swap entries from radix tree, and free them.
747 * If !unfalloc, truncate or punch hole; if unfalloc, undo failed fallocate.
748 */
749static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
750 bool unfalloc)
751{
752 struct address_space *mapping = inode->i_mapping;
753 struct shmem_inode_info *info = SHMEM_I(inode);
754 pgoff_t start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
755 pgoff_t end = (lend + 1) >> PAGE_SHIFT;
756 unsigned int partial_start = lstart & (PAGE_SIZE - 1);
757 unsigned int partial_end = (lend + 1) & (PAGE_SIZE - 1);
758 struct pagevec pvec;
759 pgoff_t indices[PAGEVEC_SIZE];
760 long nr_swaps_freed = 0;
761 pgoff_t index;
762 int i;
763
764 if (lend == -1)
765 end = -1; /* unsigned, so actually very big */
766
767 pagevec_init(&pvec, 0);
768 index = start;
769 while (index < end) {
770 pvec.nr = find_get_entries(mapping, index,
771 min(end - index, (pgoff_t)PAGEVEC_SIZE),
772 pvec.pages, indices);
773 if (!pvec.nr)
774 break;
775 for (i = 0; i < pagevec_count(&pvec); i++) {
776 struct page *page = pvec.pages[i];
777
778 index = indices[i];
779 if (index >= end)
780 break;
781
782 if (radix_tree_exceptional_entry(page)) {
783 if (unfalloc)
784 continue;
785 nr_swaps_freed += !shmem_free_swap(mapping,
786 index, page);
787 continue;
788 }
789
790 VM_BUG_ON_PAGE(page_to_pgoff(page) != index, page);
791
792 if (!trylock_page(page))
793 continue;
794
795 if (PageTransTail(page)) {
796 /* Middle of THP: zero out the page */
797 clear_highpage(page);
798 unlock_page(page);
799 continue;
800 } else if (PageTransHuge(page)) {
801 if (index == round_down(end, HPAGE_PMD_NR)) {
802 /*
803 * Range ends in the middle of THP:
804 * zero out the page
805 */
806 clear_highpage(page);
807 unlock_page(page);
808 continue;
809 }
810 index += HPAGE_PMD_NR - 1;
811 i += HPAGE_PMD_NR - 1;
812 }
813
814 if (!unfalloc || !PageUptodate(page)) {
815 VM_BUG_ON_PAGE(PageTail(page), page);
816 if (page_mapping(page) == mapping) {
817 VM_BUG_ON_PAGE(PageWriteback(page), page);
818 truncate_inode_page(mapping, page);
819 }
820 }
821 unlock_page(page);
822 }
823 pagevec_remove_exceptionals(&pvec);
824 pagevec_release(&pvec);
825 cond_resched();
826 index++;
827 }
828
829 if (partial_start) {
830 struct page *page = NULL;
831 shmem_getpage(inode, start - 1, &page, SGP_READ);
832 if (page) {
833 unsigned int top = PAGE_SIZE;
834 if (start > end) {
835 top = partial_end;
836 partial_end = 0;
837 }
838 zero_user_segment(page, partial_start, top);
839 set_page_dirty(page);
840 unlock_page(page);
841 put_page(page);
842 }
843 }
844 if (partial_end) {
845 struct page *page = NULL;
846 shmem_getpage(inode, end, &page, SGP_READ);
847 if (page) {
848 zero_user_segment(page, 0, partial_end);
849 set_page_dirty(page);
850 unlock_page(page);
851 put_page(page);
852 }
853 }
854 if (start >= end)
855 return;
856
857 index = start;
858 while (index < end) {
859 cond_resched();
860
861 pvec.nr = find_get_entries(mapping, index,
862 min(end - index, (pgoff_t)PAGEVEC_SIZE),
863 pvec.pages, indices);
864 if (!pvec.nr) {
865 /* If all gone or hole-punch or unfalloc, we're done */
866 if (index == start || end != -1)
867 break;
868 /* But if truncating, restart to make sure all gone */
869 index = start;
870 continue;
871 }
872 for (i = 0; i < pagevec_count(&pvec); i++) {
873 struct page *page = pvec.pages[i];
874
875 index = indices[i];
876 if (index >= end)
877 break;
878
879 if (radix_tree_exceptional_entry(page)) {
880 if (unfalloc)
881 continue;
882 if (shmem_free_swap(mapping, index, page)) {
883 /* Swap was replaced by page: retry */
884 index--;
885 break;
886 }
887 nr_swaps_freed++;
888 continue;
889 }
890
891 lock_page(page);
892
893 if (PageTransTail(page)) {
894 /* Middle of THP: zero out the page */
895 clear_highpage(page);
896 unlock_page(page);
897 /*
898 * Partial thp truncate due 'start' in middle
899 * of THP: don't need to look on these pages
900 * again on !pvec.nr restart.
901 */
902 if (index != round_down(end, HPAGE_PMD_NR))
903 start++;
904 continue;
905 } else if (PageTransHuge(page)) {
906 if (index == round_down(end, HPAGE_PMD_NR)) {
907 /*
908 * Range ends in the middle of THP:
909 * zero out the page
910 */
911 clear_highpage(page);
912 unlock_page(page);
913 continue;
914 }
915 index += HPAGE_PMD_NR - 1;
916 i += HPAGE_PMD_NR - 1;
917 }
918
919 if (!unfalloc || !PageUptodate(page)) {
920 VM_BUG_ON_PAGE(PageTail(page), page);
921 if (page_mapping(page) == mapping) {
922 VM_BUG_ON_PAGE(PageWriteback(page), page);
923 truncate_inode_page(mapping, page);
924 } else {
925 /* Page was replaced by swap: retry */
926 unlock_page(page);
927 index--;
928 break;
929 }
930 }
931 unlock_page(page);
932 }
933 pagevec_remove_exceptionals(&pvec);
934 pagevec_release(&pvec);
935 index++;
936 }
937
938 spin_lock_irq(&info->lock);
939 info->swapped -= nr_swaps_freed;
940 shmem_recalc_inode(inode);
941 spin_unlock_irq(&info->lock);
942}
943
944void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
945{
946 shmem_undo_range(inode, lstart, lend, false);
947 inode->i_ctime = inode->i_mtime = current_time(inode);
948}
949EXPORT_SYMBOL_GPL(shmem_truncate_range);
950
951static int shmem_getattr(struct vfsmount *mnt, struct dentry *dentry,
952 struct kstat *stat)
953{
954 struct inode *inode = dentry->d_inode;
955 struct shmem_inode_info *info = SHMEM_I(inode);
956
957 if (info->alloced - info->swapped != inode->i_mapping->nrpages) {
958 spin_lock_irq(&info->lock);
959 shmem_recalc_inode(inode);
960 spin_unlock_irq(&info->lock);
961 }
962 generic_fillattr(inode, stat);
963 return 0;
964}
965
966static int shmem_setattr(struct dentry *dentry, struct iattr *attr)
967{
968 struct inode *inode = d_inode(dentry);
969 struct shmem_inode_info *info = SHMEM_I(inode);
970 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
971 int error;
972
973 error = setattr_prepare(dentry, attr);
974 if (error)
975 return error;
976
977 if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
978 loff_t oldsize = inode->i_size;
979 loff_t newsize = attr->ia_size;
980
981 /* protected by i_mutex */
982 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
983 (newsize > oldsize && (info->seals & F_SEAL_GROW)))
984 return -EPERM;
985
986 if (newsize != oldsize) {
987 error = shmem_reacct_size(SHMEM_I(inode)->flags,
988 oldsize, newsize);
989 if (error)
990 return error;
991 i_size_write(inode, newsize);
992 inode->i_ctime = inode->i_mtime = current_time(inode);
993 }
994 if (newsize <= oldsize) {
995 loff_t holebegin = round_up(newsize, PAGE_SIZE);
996 if (oldsize > holebegin)
997 unmap_mapping_range(inode->i_mapping,
998 holebegin, 0, 1);
999 if (info->alloced)
1000 shmem_truncate_range(inode,
1001 newsize, (loff_t)-1);
1002 /* unmap again to remove racily COWed private pages */
1003 if (oldsize > holebegin)
1004 unmap_mapping_range(inode->i_mapping,
1005 holebegin, 0, 1);
1006
1007 /*
1008 * Part of the huge page can be beyond i_size: subject
1009 * to shrink under memory pressure.
1010 */
1011 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
1012 spin_lock(&sbinfo->shrinklist_lock);
1013 if (list_empty(&info->shrinklist)) {
1014 list_add_tail(&info->shrinklist,
1015 &sbinfo->shrinklist);
1016 sbinfo->shrinklist_len++;
1017 }
1018 spin_unlock(&sbinfo->shrinklist_lock);
1019 }
1020 }
1021 }
1022
1023 setattr_copy(inode, attr);
1024 if (attr->ia_valid & ATTR_MODE)
1025 error = posix_acl_chmod(inode, inode->i_mode);
1026 return error;
1027}
1028
1029static void shmem_evict_inode(struct inode *inode)
1030{
1031 struct shmem_inode_info *info = SHMEM_I(inode);
1032 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1033
1034 if (inode->i_mapping->a_ops == &shmem_aops) {
1035 shmem_unacct_size(info->flags, inode->i_size);
1036 inode->i_size = 0;
1037 shmem_truncate_range(inode, 0, (loff_t)-1);
1038 if (!list_empty(&info->shrinklist)) {
1039 spin_lock(&sbinfo->shrinklist_lock);
1040 if (!list_empty(&info->shrinklist)) {
1041 list_del_init(&info->shrinklist);
1042 sbinfo->shrinklist_len--;
1043 }
1044 spin_unlock(&sbinfo->shrinklist_lock);
1045 }
1046 if (!list_empty(&info->swaplist)) {
1047 mutex_lock(&shmem_swaplist_mutex);
1048 list_del_init(&info->swaplist);
1049 mutex_unlock(&shmem_swaplist_mutex);
1050 }
1051 }
1052
1053 simple_xattrs_free(&info->xattrs);
1054 WARN_ON(inode->i_blocks);
1055 shmem_free_inode(inode->i_sb);
1056 clear_inode(inode);
1057}
1058
1059static unsigned long find_swap_entry(struct radix_tree_root *root, void *item)
1060{
1061 struct radix_tree_iter iter;
1062 void **slot;
1063 unsigned long found = -1;
1064 unsigned int checked = 0;
1065
1066 rcu_read_lock();
1067 radix_tree_for_each_slot(slot, root, &iter, 0) {
1068 if (*slot == item) {
1069 found = iter.index;
1070 break;
1071 }
1072 checked++;
1073 if ((checked % 4096) != 0)
1074 continue;
1075 slot = radix_tree_iter_resume(slot, &iter);
1076 cond_resched_rcu();
1077 }
1078
1079 rcu_read_unlock();
1080 return found;
1081}
1082
1083/*
1084 * If swap found in inode, free it and move page from swapcache to filecache.
1085 */
1086static int shmem_unuse_inode(struct shmem_inode_info *info,
1087 swp_entry_t swap, struct page **pagep)
1088{
1089 struct address_space *mapping = info->vfs_inode.i_mapping;
1090 void *radswap;
1091 pgoff_t index;
1092 gfp_t gfp;
1093 int error = 0;
1094
1095 radswap = swp_to_radix_entry(swap);
1096 index = find_swap_entry(&mapping->page_tree, radswap);
1097 if (index == -1)
1098 return -EAGAIN; /* tell shmem_unuse we found nothing */
1099
1100 /*
1101 * Move _head_ to start search for next from here.
1102 * But be careful: shmem_evict_inode checks list_empty without taking
1103 * mutex, and there's an instant in list_move_tail when info->swaplist
1104 * would appear empty, if it were the only one on shmem_swaplist.
1105 */
1106 if (shmem_swaplist.next != &info->swaplist)
1107 list_move_tail(&shmem_swaplist, &info->swaplist);
1108
1109 gfp = mapping_gfp_mask(mapping);
1110 if (shmem_should_replace_page(*pagep, gfp)) {
1111 mutex_unlock(&shmem_swaplist_mutex);
1112 error = shmem_replace_page(pagep, gfp, info, index);
1113 mutex_lock(&shmem_swaplist_mutex);
1114 /*
1115 * We needed to drop mutex to make that restrictive page
1116 * allocation, but the inode might have been freed while we
1117 * dropped it: although a racing shmem_evict_inode() cannot
1118 * complete without emptying the radix_tree, our page lock
1119 * on this swapcache page is not enough to prevent that -
1120 * free_swap_and_cache() of our swap entry will only
1121 * trylock_page(), removing swap from radix_tree whatever.
1122 *
1123 * We must not proceed to shmem_add_to_page_cache() if the
1124 * inode has been freed, but of course we cannot rely on
1125 * inode or mapping or info to check that. However, we can
1126 * safely check if our swap entry is still in use (and here
1127 * it can't have got reused for another page): if it's still
1128 * in use, then the inode cannot have been freed yet, and we
1129 * can safely proceed (if it's no longer in use, that tells
1130 * nothing about the inode, but we don't need to unuse swap).
1131 */
1132 if (!page_swapcount(*pagep))
1133 error = -ENOENT;
1134 }
1135
1136 /*
1137 * We rely on shmem_swaplist_mutex, not only to protect the swaplist,
1138 * but also to hold up shmem_evict_inode(): so inode cannot be freed
1139 * beneath us (pagelock doesn't help until the page is in pagecache).
1140 */
1141 if (!error)
1142 error = shmem_add_to_page_cache(*pagep, mapping, index,
1143 radswap);
1144 if (error != -ENOMEM) {
1145 /*
1146 * Truncation and eviction use free_swap_and_cache(), which
1147 * only does trylock page: if we raced, best clean up here.
1148 */
1149 delete_from_swap_cache(*pagep);
1150 set_page_dirty(*pagep);
1151 if (!error) {
1152 spin_lock_irq(&info->lock);
1153 info->swapped--;
1154 spin_unlock_irq(&info->lock);
1155 swap_free(swap);
1156 }
1157 }
1158 return error;
1159}
1160
1161/*
1162 * Search through swapped inodes to find and replace swap by page.
1163 */
1164int shmem_unuse(swp_entry_t swap, struct page *page)
1165{
1166 struct list_head *this, *next;
1167 struct shmem_inode_info *info;
1168 struct mem_cgroup *memcg;
1169 int error = 0;
1170
1171 /*
1172 * There's a faint possibility that swap page was replaced before
1173 * caller locked it: caller will come back later with the right page.
1174 */
1175 if (unlikely(!PageSwapCache(page) || page_private(page) != swap.val))
1176 goto out;
1177
1178 /*
1179 * Charge page using GFP_KERNEL while we can wait, before taking
1180 * the shmem_swaplist_mutex which might hold up shmem_writepage().
1181 * Charged back to the user (not to caller) when swap account is used.
1182 */
1183 error = mem_cgroup_try_charge(page, current->mm, GFP_KERNEL, &memcg,
1184 false);
1185 if (error)
1186 goto out;
1187 /* No radix_tree_preload: swap entry keeps a place for page in tree */
1188 error = -EAGAIN;
1189
1190 mutex_lock(&shmem_swaplist_mutex);
1191 list_for_each_safe(this, next, &shmem_swaplist) {
1192 info = list_entry(this, struct shmem_inode_info, swaplist);
1193 if (info->swapped)
1194 error = shmem_unuse_inode(info, swap, &page);
1195 else
1196 list_del_init(&info->swaplist);
1197 cond_resched();
1198 if (error != -EAGAIN)
1199 break;
1200 /* found nothing in this: move on to search the next */
1201 }
1202 mutex_unlock(&shmem_swaplist_mutex);
1203
1204 if (error) {
1205 if (error != -ENOMEM)
1206 error = 0;
1207 mem_cgroup_cancel_charge(page, memcg, false);
1208 } else
1209 mem_cgroup_commit_charge(page, memcg, true, false);
1210out:
1211 unlock_page(page);
1212 put_page(page);
1213 return error;
1214}
1215
1216/*
1217 * Move the page from the page cache to the swap cache.
1218 */
1219static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1220{
1221 struct shmem_inode_info *info;
1222 struct address_space *mapping;
1223 struct inode *inode;
1224 swp_entry_t swap;
1225 pgoff_t index;
1226
1227 VM_BUG_ON_PAGE(PageCompound(page), page);
1228 BUG_ON(!PageLocked(page));
1229 mapping = page->mapping;
1230 index = page->index;
1231 inode = mapping->host;
1232 info = SHMEM_I(inode);
1233 if (info->flags & VM_LOCKED)
1234 goto redirty;
1235 if (!total_swap_pages)
1236 goto redirty;
1237
1238 /*
1239 * Our capabilities prevent regular writeback or sync from ever calling
1240 * shmem_writepage; but a stacking filesystem might use ->writepage of
1241 * its underlying filesystem, in which case tmpfs should write out to
1242 * swap only in response to memory pressure, and not for the writeback
1243 * threads or sync.
1244 */
1245 if (!wbc->for_reclaim) {
1246 WARN_ON_ONCE(1); /* Still happens? Tell us about it! */
1247 goto redirty;
1248 }
1249
1250 /*
1251 * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
1252 * value into swapfile.c, the only way we can correctly account for a
1253 * fallocated page arriving here is now to initialize it and write it.
1254 *
1255 * That's okay for a page already fallocated earlier, but if we have
1256 * not yet completed the fallocation, then (a) we want to keep track
1257 * of this page in case we have to undo it, and (b) it may not be a
1258 * good idea to continue anyway, once we're pushing into swap. So
1259 * reactivate the page, and let shmem_fallocate() quit when too many.
1260 */
1261 if (!PageUptodate(page)) {
1262 if (inode->i_private) {
1263 struct shmem_falloc *shmem_falloc;
1264 spin_lock(&inode->i_lock);
1265 shmem_falloc = inode->i_private;
1266 if (shmem_falloc &&
1267 !shmem_falloc->waitq &&
1268 index >= shmem_falloc->start &&
1269 index < shmem_falloc->next)
1270 shmem_falloc->nr_unswapped++;
1271 else
1272 shmem_falloc = NULL;
1273 spin_unlock(&inode->i_lock);
1274 if (shmem_falloc)
1275 goto redirty;
1276 }
1277 clear_highpage(page);
1278 flush_dcache_page(page);
1279 SetPageUptodate(page);
1280 }
1281
1282 swap = get_swap_page();
1283 if (!swap.val)
1284 goto redirty;
1285
1286 if (mem_cgroup_try_charge_swap(page, swap))
1287 goto free_swap;
1288
1289 /*
1290 * Add inode to shmem_unuse()'s list of swapped-out inodes,
1291 * if it's not already there. Do it now before the page is
1292 * moved to swap cache, when its pagelock no longer protects
1293 * the inode from eviction. But don't unlock the mutex until
1294 * we've incremented swapped, because shmem_unuse_inode() will
1295 * prune a !swapped inode from the swaplist under this mutex.
1296 */
1297 mutex_lock(&shmem_swaplist_mutex);
1298 if (list_empty(&info->swaplist))
1299 list_add_tail(&info->swaplist, &shmem_swaplist);
1300
1301 if (add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
1302 spin_lock_irq(&info->lock);
1303 shmem_recalc_inode(inode);
1304 info->swapped++;
1305 spin_unlock_irq(&info->lock);
1306
1307 swap_shmem_alloc(swap);
1308 shmem_delete_from_page_cache(page, swp_to_radix_entry(swap));
1309
1310 mutex_unlock(&shmem_swaplist_mutex);
1311 BUG_ON(page_mapped(page));
1312 swap_writepage(page, wbc);
1313 return 0;
1314 }
1315
1316 mutex_unlock(&shmem_swaplist_mutex);
1317free_swap:
1318 swapcache_free(swap);
1319redirty:
1320 set_page_dirty(page);
1321 if (wbc->for_reclaim)
1322 return AOP_WRITEPAGE_ACTIVATE; /* Return with page locked */
1323 unlock_page(page);
1324 return 0;
1325}
1326
1327#if defined(CONFIG_NUMA) && defined(CONFIG_TMPFS)
1328static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1329{
1330 char buffer[64];
1331
1332 if (!mpol || mpol->mode == MPOL_DEFAULT)
1333 return; /* show nothing */
1334
1335 mpol_to_str(buffer, sizeof(buffer), mpol);
1336
1337 seq_printf(seq, ",mpol=%s", buffer);
1338}
1339
1340static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1341{
1342 struct mempolicy *mpol = NULL;
1343 if (sbinfo->mpol) {
1344 spin_lock(&sbinfo->stat_lock); /* prevent replace/use races */
1345 mpol = sbinfo->mpol;
1346 mpol_get(mpol);
1347 spin_unlock(&sbinfo->stat_lock);
1348 }
1349 return mpol;
1350}
1351#else /* !CONFIG_NUMA || !CONFIG_TMPFS */
1352static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1353{
1354}
1355static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1356{
1357 return NULL;
1358}
1359#endif /* CONFIG_NUMA && CONFIG_TMPFS */
1360#ifndef CONFIG_NUMA
1361#define vm_policy vm_private_data
1362#endif
1363
1364static void shmem_pseudo_vma_init(struct vm_area_struct *vma,
1365 struct shmem_inode_info *info, pgoff_t index)
1366{
1367 /* Create a pseudo vma that just contains the policy */
1368 vma->vm_start = 0;
1369 /* Bias interleave by inode number to distribute better across nodes */
1370 vma->vm_pgoff = index + info->vfs_inode.i_ino;
1371 vma->vm_ops = NULL;
1372 vma->vm_policy = mpol_shared_policy_lookup(&info->policy, index);
1373}
1374
1375static void shmem_pseudo_vma_destroy(struct vm_area_struct *vma)
1376{
1377 /* Drop reference taken by mpol_shared_policy_lookup() */
1378 mpol_cond_put(vma->vm_policy);
1379}
1380
1381static struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
1382 struct shmem_inode_info *info, pgoff_t index)
1383{
1384 struct vm_area_struct pvma;
1385 struct page *page;
1386
1387 shmem_pseudo_vma_init(&pvma, info, index);
1388 page = swapin_readahead(swap, gfp, &pvma, 0);
1389 shmem_pseudo_vma_destroy(&pvma);
1390
1391 return page;
1392}
1393
1394static struct page *shmem_alloc_hugepage(gfp_t gfp,
1395 struct shmem_inode_info *info, pgoff_t index)
1396{
1397 struct vm_area_struct pvma;
1398 struct inode *inode = &info->vfs_inode;
1399 struct address_space *mapping = inode->i_mapping;
1400 pgoff_t idx, hindex;
1401 void __rcu **results;
1402 struct page *page;
1403
1404 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
1405 return NULL;
1406
1407 hindex = round_down(index, HPAGE_PMD_NR);
1408 rcu_read_lock();
1409 if (radix_tree_gang_lookup_slot(&mapping->page_tree, &results, &idx,
1410 hindex, 1) && idx < hindex + HPAGE_PMD_NR) {
1411 rcu_read_unlock();
1412 return NULL;
1413 }
1414 rcu_read_unlock();
1415
1416 shmem_pseudo_vma_init(&pvma, info, hindex);
1417 page = alloc_pages_vma(gfp | __GFP_COMP | __GFP_NORETRY | __GFP_NOWARN,
1418 HPAGE_PMD_ORDER, &pvma, 0, numa_node_id(), true);
1419 shmem_pseudo_vma_destroy(&pvma);
1420 if (page)
1421 prep_transhuge_page(page);
1422 return page;
1423}
1424
1425static struct page *shmem_alloc_page(gfp_t gfp,
1426 struct shmem_inode_info *info, pgoff_t index)
1427{
1428 struct vm_area_struct pvma;
1429 struct page *page;
1430
1431 shmem_pseudo_vma_init(&pvma, info, index);
1432 page = alloc_page_vma(gfp, &pvma, 0);
1433 shmem_pseudo_vma_destroy(&pvma);
1434
1435 return page;
1436}
1437
1438static struct page *shmem_alloc_and_acct_page(gfp_t gfp,
1439 struct shmem_inode_info *info, struct shmem_sb_info *sbinfo,
1440 pgoff_t index, bool huge)
1441{
1442 struct page *page;
1443 int nr;
1444 int err = -ENOSPC;
1445
1446 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
1447 huge = false;
1448 nr = huge ? HPAGE_PMD_NR : 1;
1449
1450 if (shmem_acct_block(info->flags, nr))
1451 goto failed;
1452 if (sbinfo->max_blocks) {
1453 if (percpu_counter_compare(&sbinfo->used_blocks,
1454 sbinfo->max_blocks - nr) > 0)
1455 goto unacct;
1456 percpu_counter_add(&sbinfo->used_blocks, nr);
1457 }
1458
1459 if (huge)
1460 page = shmem_alloc_hugepage(gfp, info, index);
1461 else
1462 page = shmem_alloc_page(gfp, info, index);
1463 if (page) {
1464 __SetPageLocked(page);
1465 __SetPageSwapBacked(page);
1466 return page;
1467 }
1468
1469 err = -ENOMEM;
1470 if (sbinfo->max_blocks)
1471 percpu_counter_add(&sbinfo->used_blocks, -nr);
1472unacct:
1473 shmem_unacct_blocks(info->flags, nr);
1474failed:
1475 return ERR_PTR(err);
1476}
1477
1478/*
1479 * When a page is moved from swapcache to shmem filecache (either by the
1480 * usual swapin of shmem_getpage_gfp(), or by the less common swapoff of
1481 * shmem_unuse_inode()), it may have been read in earlier from swap, in
1482 * ignorance of the mapping it belongs to. If that mapping has special
1483 * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
1484 * we may need to copy to a suitable page before moving to filecache.
1485 *
1486 * In a future release, this may well be extended to respect cpuset and
1487 * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
1488 * but for now it is a simple matter of zone.
1489 */
1490static bool shmem_should_replace_page(struct page *page, gfp_t gfp)
1491{
1492 return page_zonenum(page) > gfp_zone(gfp);
1493}
1494
1495static int shmem_replace_page(struct page **pagep, gfp_t gfp,
1496 struct shmem_inode_info *info, pgoff_t index)
1497{
1498 struct page *oldpage, *newpage;
1499 struct address_space *swap_mapping;
1500 pgoff_t swap_index;
1501 int error;
1502
1503 oldpage = *pagep;
1504 swap_index = page_private(oldpage);
1505 swap_mapping = page_mapping(oldpage);
1506
1507 /*
1508 * We have arrived here because our zones are constrained, so don't
1509 * limit chance of success by further cpuset and node constraints.
1510 */
1511 gfp &= ~GFP_CONSTRAINT_MASK;
1512 newpage = shmem_alloc_page(gfp, info, index);
1513 if (!newpage)
1514 return -ENOMEM;
1515
1516 get_page(newpage);
1517 copy_highpage(newpage, oldpage);
1518 flush_dcache_page(newpage);
1519
1520 __SetPageLocked(newpage);
1521 __SetPageSwapBacked(newpage);
1522 SetPageUptodate(newpage);
1523 set_page_private(newpage, swap_index);
1524 SetPageSwapCache(newpage);
1525
1526 /*
1527 * Our caller will very soon move newpage out of swapcache, but it's
1528 * a nice clean interface for us to replace oldpage by newpage there.
1529 */
1530 spin_lock_irq(&swap_mapping->tree_lock);
1531 error = shmem_radix_tree_replace(swap_mapping, swap_index, oldpage,
1532 newpage);
1533 if (!error) {
1534 __inc_node_page_state(newpage, NR_FILE_PAGES);
1535 __dec_node_page_state(oldpage, NR_FILE_PAGES);
1536 }
1537 spin_unlock_irq(&swap_mapping->tree_lock);
1538
1539 if (unlikely(error)) {
1540 /*
1541 * Is this possible? I think not, now that our callers check
1542 * both PageSwapCache and page_private after getting page lock;
1543 * but be defensive. Reverse old to newpage for clear and free.
1544 */
1545 oldpage = newpage;
1546 } else {
1547 mem_cgroup_migrate(oldpage, newpage);
1548 lru_cache_add_anon(newpage);
1549 *pagep = newpage;
1550 }
1551
1552 ClearPageSwapCache(oldpage);
1553 set_page_private(oldpage, 0);
1554
1555 unlock_page(oldpage);
1556 put_page(oldpage);
1557 put_page(oldpage);
1558 return error;
1559}
1560
1561/*
1562 * shmem_getpage_gfp - find page in cache, or get from swap, or allocate
1563 *
1564 * If we allocate a new one we do not mark it dirty. That's up to the
1565 * vm. If we swap it in we mark it dirty since we also free the swap
1566 * entry since a page cannot live in both the swap and page cache.
1567 *
1568 * fault_mm and fault_type are only supplied by shmem_fault:
1569 * otherwise they are NULL.
1570 */
1571static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
1572 struct page **pagep, enum sgp_type sgp, gfp_t gfp,
1573 struct mm_struct *fault_mm, int *fault_type)
1574{
1575 struct address_space *mapping = inode->i_mapping;
1576 struct shmem_inode_info *info = SHMEM_I(inode);
1577 struct shmem_sb_info *sbinfo;
1578 struct mm_struct *charge_mm;
1579 struct mem_cgroup *memcg;
1580 struct page *page;
1581 swp_entry_t swap;
1582 enum sgp_type sgp_huge = sgp;
1583 pgoff_t hindex = index;
1584 int error;
1585 int once = 0;
1586 int alloced = 0;
1587
1588 if (index > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
1589 return -EFBIG;
1590 if (sgp == SGP_NOHUGE || sgp == SGP_HUGE)
1591 sgp = SGP_CACHE;
1592repeat:
1593 swap.val = 0;
1594 page = find_lock_entry(mapping, index);
1595 if (radix_tree_exceptional_entry(page)) {
1596 swap = radix_to_swp_entry(page);
1597 page = NULL;
1598 }
1599
1600 if (sgp <= SGP_CACHE &&
1601 ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1602 error = -EINVAL;
1603 goto unlock;
1604 }
1605
1606 if (page && sgp == SGP_WRITE)
1607 mark_page_accessed(page);
1608
1609 /* fallocated page? */
1610 if (page && !PageUptodate(page)) {
1611 if (sgp != SGP_READ)
1612 goto clear;
1613 unlock_page(page);
1614 put_page(page);
1615 page = NULL;
1616 }
1617 if (page || (sgp == SGP_READ && !swap.val)) {
1618 *pagep = page;
1619 return 0;
1620 }
1621
1622 /*
1623 * Fast cache lookup did not find it:
1624 * bring it back from swap or allocate.
1625 */
1626 sbinfo = SHMEM_SB(inode->i_sb);
1627 charge_mm = fault_mm ? : current->mm;
1628
1629 if (swap.val) {
1630 /* Look it up and read it in.. */
1631 page = lookup_swap_cache(swap);
1632 if (!page) {
1633 /* Or update major stats only when swapin succeeds?? */
1634 if (fault_type) {
1635 *fault_type |= VM_FAULT_MAJOR;
1636 count_vm_event(PGMAJFAULT);
1637 mem_cgroup_count_vm_event(fault_mm, PGMAJFAULT);
1638 }
1639 /* Here we actually start the io */
1640 page = shmem_swapin(swap, gfp, info, index);
1641 if (!page) {
1642 error = -ENOMEM;
1643 goto failed;
1644 }
1645 }
1646
1647 /* We have to do this with page locked to prevent races */
1648 lock_page(page);
1649 if (!PageSwapCache(page) || page_private(page) != swap.val ||
1650 !shmem_confirm_swap(mapping, index, swap)) {
1651 error = -EEXIST; /* try again */
1652 goto unlock;
1653 }
1654 if (!PageUptodate(page)) {
1655 error = -EIO;
1656 goto failed;
1657 }
1658 wait_on_page_writeback(page);
1659
1660 if (shmem_should_replace_page(page, gfp)) {
1661 error = shmem_replace_page(&page, gfp, info, index);
1662 if (error)
1663 goto failed;
1664 }
1665
1666 error = mem_cgroup_try_charge(page, charge_mm, gfp, &memcg,
1667 false);
1668 if (!error) {
1669 error = shmem_add_to_page_cache(page, mapping, index,
1670 swp_to_radix_entry(swap));
1671 /*
1672 * We already confirmed swap under page lock, and make
1673 * no memory allocation here, so usually no possibility
1674 * of error; but free_swap_and_cache() only trylocks a
1675 * page, so it is just possible that the entry has been
1676 * truncated or holepunched since swap was confirmed.
1677 * shmem_undo_range() will have done some of the
1678 * unaccounting, now delete_from_swap_cache() will do
1679 * the rest.
1680 * Reset swap.val? No, leave it so "failed" goes back to
1681 * "repeat": reading a hole and writing should succeed.
1682 */
1683 if (error) {
1684 mem_cgroup_cancel_charge(page, memcg, false);
1685 delete_from_swap_cache(page);
1686 }
1687 }
1688 if (error)
1689 goto failed;
1690
1691 mem_cgroup_commit_charge(page, memcg, true, false);
1692
1693 spin_lock_irq(&info->lock);
1694 info->swapped--;
1695 shmem_recalc_inode(inode);
1696 spin_unlock_irq(&info->lock);
1697
1698 if (sgp == SGP_WRITE)
1699 mark_page_accessed(page);
1700
1701 delete_from_swap_cache(page);
1702 set_page_dirty(page);
1703 swap_free(swap);
1704
1705 } else {
1706 /* shmem_symlink() */
1707 if (mapping->a_ops != &shmem_aops)
1708 goto alloc_nohuge;
1709 if (shmem_huge == SHMEM_HUGE_DENY || sgp_huge == SGP_NOHUGE)
1710 goto alloc_nohuge;
1711 if (shmem_huge == SHMEM_HUGE_FORCE)
1712 goto alloc_huge;
1713 switch (sbinfo->huge) {
1714 loff_t i_size;
1715 pgoff_t off;
1716 case SHMEM_HUGE_NEVER:
1717 goto alloc_nohuge;
1718 case SHMEM_HUGE_WITHIN_SIZE:
1719 off = round_up(index, HPAGE_PMD_NR);
1720 i_size = round_up(i_size_read(inode), PAGE_SIZE);
1721 if (i_size >= HPAGE_PMD_SIZE &&
1722 i_size >> PAGE_SHIFT >= off)
1723 goto alloc_huge;
1724 /* fallthrough */
1725 case SHMEM_HUGE_ADVISE:
1726 if (sgp_huge == SGP_HUGE)
1727 goto alloc_huge;
1728 /* TODO: implement fadvise() hints */
1729 goto alloc_nohuge;
1730 }
1731
1732alloc_huge:
1733 page = shmem_alloc_and_acct_page(gfp, info, sbinfo,
1734 index, true);
1735 if (IS_ERR(page)) {
1736alloc_nohuge: page = shmem_alloc_and_acct_page(gfp, info, sbinfo,
1737 index, false);
1738 }
1739 if (IS_ERR(page)) {
1740 int retry = 5;
1741 error = PTR_ERR(page);
1742 page = NULL;
1743 if (error != -ENOSPC)
1744 goto failed;
1745 /*
1746 * Try to reclaim some spece by splitting a huge page
1747 * beyond i_size on the filesystem.
1748 */
1749 while (retry--) {
1750 int ret;
1751 ret = shmem_unused_huge_shrink(sbinfo, NULL, 1);
1752 if (ret == SHRINK_STOP)
1753 break;
1754 if (ret)
1755 goto alloc_nohuge;
1756 }
1757 goto failed;
1758 }
1759
1760 if (PageTransHuge(page))
1761 hindex = round_down(index, HPAGE_PMD_NR);
1762 else
1763 hindex = index;
1764
1765 if (sgp == SGP_WRITE)
1766 __SetPageReferenced(page);
1767
1768 error = mem_cgroup_try_charge(page, charge_mm, gfp, &memcg,
1769 PageTransHuge(page));
1770 if (error)
1771 goto unacct;
1772 error = radix_tree_maybe_preload_order(gfp & GFP_RECLAIM_MASK,
1773 compound_order(page));
1774 if (!error) {
1775 error = shmem_add_to_page_cache(page, mapping, hindex,
1776 NULL);
1777 radix_tree_preload_end();
1778 }
1779 if (error) {
1780 mem_cgroup_cancel_charge(page, memcg,
1781 PageTransHuge(page));
1782 goto unacct;
1783 }
1784 mem_cgroup_commit_charge(page, memcg, false,
1785 PageTransHuge(page));
1786 lru_cache_add_anon(page);
1787
1788 spin_lock_irq(&info->lock);
1789 info->alloced += 1 << compound_order(page);
1790 inode->i_blocks += BLOCKS_PER_PAGE << compound_order(page);
1791 shmem_recalc_inode(inode);
1792 spin_unlock_irq(&info->lock);
1793 alloced = true;
1794
1795 if (PageTransHuge(page) &&
1796 DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE) <
1797 hindex + HPAGE_PMD_NR - 1) {
1798 /*
1799 * Part of the huge page is beyond i_size: subject
1800 * to shrink under memory pressure.
1801 */
1802 spin_lock(&sbinfo->shrinklist_lock);
1803 if (list_empty(&info->shrinklist)) {
1804 list_add_tail(&info->shrinklist,
1805 &sbinfo->shrinklist);
1806 sbinfo->shrinklist_len++;
1807 }
1808 spin_unlock(&sbinfo->shrinklist_lock);
1809 }
1810
1811 /*
1812 * Let SGP_FALLOC use the SGP_WRITE optimization on a new page.
1813 */
1814 if (sgp == SGP_FALLOC)
1815 sgp = SGP_WRITE;
1816clear:
1817 /*
1818 * Let SGP_WRITE caller clear ends if write does not fill page;
1819 * but SGP_FALLOC on a page fallocated earlier must initialize
1820 * it now, lest undo on failure cancel our earlier guarantee.
1821 */
1822 if (sgp != SGP_WRITE && !PageUptodate(page)) {
1823 struct page *head = compound_head(page);
1824 int i;
1825
1826 for (i = 0; i < (1 << compound_order(head)); i++) {
1827 clear_highpage(head + i);
1828 flush_dcache_page(head + i);
1829 }
1830 SetPageUptodate(head);
1831 }
1832 }
1833
1834 /* Perhaps the file has been truncated since we checked */
1835 if (sgp <= SGP_CACHE &&
1836 ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1837 if (alloced) {
1838 ClearPageDirty(page);
1839 delete_from_page_cache(page);
1840 spin_lock_irq(&info->lock);
1841 shmem_recalc_inode(inode);
1842 spin_unlock_irq(&info->lock);
1843 }
1844 error = -EINVAL;
1845 goto unlock;
1846 }
1847 *pagep = page + index - hindex;
1848 return 0;
1849
1850 /*
1851 * Error recovery.
1852 */
1853unacct:
1854 if (sbinfo->max_blocks)
1855 percpu_counter_sub(&sbinfo->used_blocks,
1856 1 << compound_order(page));
1857 shmem_unacct_blocks(info->flags, 1 << compound_order(page));
1858
1859 if (PageTransHuge(page)) {
1860 unlock_page(page);
1861 put_page(page);
1862 goto alloc_nohuge;
1863 }
1864failed:
1865 if (swap.val && !shmem_confirm_swap(mapping, index, swap))
1866 error = -EEXIST;
1867unlock:
1868 if (page) {
1869 unlock_page(page);
1870 put_page(page);
1871 }
1872 if (error == -ENOSPC && !once++) {
1873 spin_lock_irq(&info->lock);
1874 shmem_recalc_inode(inode);
1875 spin_unlock_irq(&info->lock);
1876 goto repeat;
1877 }
1878 if (error == -EEXIST) /* from above or from radix_tree_insert */
1879 goto repeat;
1880 return error;
1881}
1882
1883/*
1884 * This is like autoremove_wake_function, but it removes the wait queue
1885 * entry unconditionally - even if something else had already woken the
1886 * target.
1887 */
1888static int synchronous_wake_function(wait_queue_t *wait, unsigned mode, int sync, void *key)
1889{
1890 int ret = default_wake_function(wait, mode, sync, key);
1891 list_del_init(&wait->task_list);
1892 return ret;
1893}
1894
1895static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1896{
1897 struct inode *inode = file_inode(vma->vm_file);
1898 gfp_t gfp = mapping_gfp_mask(inode->i_mapping);
1899 enum sgp_type sgp;
1900 int error;
1901 int ret = VM_FAULT_LOCKED;
1902
1903 /*
1904 * Trinity finds that probing a hole which tmpfs is punching can
1905 * prevent the hole-punch from ever completing: which in turn
1906 * locks writers out with its hold on i_mutex. So refrain from
1907 * faulting pages into the hole while it's being punched. Although
1908 * shmem_undo_range() does remove the additions, it may be unable to
1909 * keep up, as each new page needs its own unmap_mapping_range() call,
1910 * and the i_mmap tree grows ever slower to scan if new vmas are added.
1911 *
1912 * It does not matter if we sometimes reach this check just before the
1913 * hole-punch begins, so that one fault then races with the punch:
1914 * we just need to make racing faults a rare case.
1915 *
1916 * The implementation below would be much simpler if we just used a
1917 * standard mutex or completion: but we cannot take i_mutex in fault,
1918 * and bloating every shmem inode for this unlikely case would be sad.
1919 */
1920 if (unlikely(inode->i_private)) {
1921 struct shmem_falloc *shmem_falloc;
1922
1923 spin_lock(&inode->i_lock);
1924 shmem_falloc = inode->i_private;
1925 if (shmem_falloc &&
1926 shmem_falloc->waitq &&
1927 vmf->pgoff >= shmem_falloc->start &&
1928 vmf->pgoff < shmem_falloc->next) {
1929 wait_queue_head_t *shmem_falloc_waitq;
1930 DEFINE_WAIT_FUNC(shmem_fault_wait, synchronous_wake_function);
1931
1932 ret = VM_FAULT_NOPAGE;
1933 if ((vmf->flags & FAULT_FLAG_ALLOW_RETRY) &&
1934 !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
1935 /* It's polite to up mmap_sem if we can */
1936 up_read(&vma->vm_mm->mmap_sem);
1937 ret = VM_FAULT_RETRY;
1938 }
1939
1940 shmem_falloc_waitq = shmem_falloc->waitq;
1941 prepare_to_wait(shmem_falloc_waitq, &shmem_fault_wait,
1942 TASK_UNINTERRUPTIBLE);
1943 spin_unlock(&inode->i_lock);
1944 schedule();
1945
1946 /*
1947 * shmem_falloc_waitq points into the shmem_fallocate()
1948 * stack of the hole-punching task: shmem_falloc_waitq
1949 * is usually invalid by the time we reach here, but
1950 * finish_wait() does not dereference it in that case;
1951 * though i_lock needed lest racing with wake_up_all().
1952 */
1953 spin_lock(&inode->i_lock);
1954 finish_wait(shmem_falloc_waitq, &shmem_fault_wait);
1955 spin_unlock(&inode->i_lock);
1956 return ret;
1957 }
1958 spin_unlock(&inode->i_lock);
1959 }
1960
1961 sgp = SGP_CACHE;
1962 if (vma->vm_flags & VM_HUGEPAGE)
1963 sgp = SGP_HUGE;
1964 else if (vma->vm_flags & VM_NOHUGEPAGE)
1965 sgp = SGP_NOHUGE;
1966
1967 error = shmem_getpage_gfp(inode, vmf->pgoff, &vmf->page, sgp,
1968 gfp, vma->vm_mm, &ret);
1969 if (error)
1970 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1971 return ret;
1972}
1973
1974unsigned long shmem_get_unmapped_area(struct file *file,
1975 unsigned long uaddr, unsigned long len,
1976 unsigned long pgoff, unsigned long flags)
1977{
1978 unsigned long (*get_area)(struct file *,
1979 unsigned long, unsigned long, unsigned long, unsigned long);
1980 unsigned long addr;
1981 unsigned long offset;
1982 unsigned long inflated_len;
1983 unsigned long inflated_addr;
1984 unsigned long inflated_offset;
1985
1986 if (len > TASK_SIZE)
1987 return -ENOMEM;
1988
1989 get_area = current->mm->get_unmapped_area;
1990 addr = get_area(file, uaddr, len, pgoff, flags);
1991
1992 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
1993 return addr;
1994 if (IS_ERR_VALUE(addr))
1995 return addr;
1996 if (addr & ~PAGE_MASK)
1997 return addr;
1998 if (addr > TASK_SIZE - len)
1999 return addr;
2000
2001 if (shmem_huge == SHMEM_HUGE_DENY)
2002 return addr;
2003 if (len < HPAGE_PMD_SIZE)
2004 return addr;
2005 if (flags & MAP_FIXED)
2006 return addr;
2007 /*
2008 * Our priority is to support MAP_SHARED mapped hugely;
2009 * and support MAP_PRIVATE mapped hugely too, until it is COWed.
2010 * But if caller specified an address hint, respect that as before.
2011 */
2012 if (uaddr)
2013 return addr;
2014
2015 if (shmem_huge != SHMEM_HUGE_FORCE) {
2016 struct super_block *sb;
2017
2018 if (file) {
2019 VM_BUG_ON(file->f_op != &shmem_file_operations);
2020 sb = file_inode(file)->i_sb;
2021 } else {
2022 /*
2023 * Called directly from mm/mmap.c, or drivers/char/mem.c
2024 * for "/dev/zero", to create a shared anonymous object.
2025 */
2026 if (IS_ERR(shm_mnt))
2027 return addr;
2028 sb = shm_mnt->mnt_sb;
2029 }
2030 if (SHMEM_SB(sb)->huge == SHMEM_HUGE_NEVER)
2031 return addr;
2032 }
2033
2034 offset = (pgoff << PAGE_SHIFT) & (HPAGE_PMD_SIZE-1);
2035 if (offset && offset + len < 2 * HPAGE_PMD_SIZE)
2036 return addr;
2037 if ((addr & (HPAGE_PMD_SIZE-1)) == offset)
2038 return addr;
2039
2040 inflated_len = len + HPAGE_PMD_SIZE - PAGE_SIZE;
2041 if (inflated_len > TASK_SIZE)
2042 return addr;
2043 if (inflated_len < len)
2044 return addr;
2045
2046 inflated_addr = get_area(NULL, 0, inflated_len, 0, flags);
2047 if (IS_ERR_VALUE(inflated_addr))
2048 return addr;
2049 if (inflated_addr & ~PAGE_MASK)
2050 return addr;
2051
2052 inflated_offset = inflated_addr & (HPAGE_PMD_SIZE-1);
2053 inflated_addr += offset - inflated_offset;
2054 if (inflated_offset > offset)
2055 inflated_addr += HPAGE_PMD_SIZE;
2056
2057 if (inflated_addr > TASK_SIZE - len)
2058 return addr;
2059 return inflated_addr;
2060}
2061
2062#ifdef CONFIG_NUMA
2063static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
2064{
2065 struct inode *inode = file_inode(vma->vm_file);
2066 return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
2067}
2068
2069static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
2070 unsigned long addr)
2071{
2072 struct inode *inode = file_inode(vma->vm_file);
2073 pgoff_t index;
2074
2075 index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2076 return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
2077}
2078#endif
2079
2080int shmem_lock(struct file *file, int lock, struct user_struct *user)
2081{
2082 struct inode *inode = file_inode(file);
2083 struct shmem_inode_info *info = SHMEM_I(inode);
2084 int retval = -ENOMEM;
2085
2086 spin_lock_irq(&info->lock);
2087 if (lock && !(info->flags & VM_LOCKED)) {
2088 if (!user_shm_lock(inode->i_size, user))
2089 goto out_nomem;
2090 info->flags |= VM_LOCKED;
2091 mapping_set_unevictable(file->f_mapping);
2092 }
2093 if (!lock && (info->flags & VM_LOCKED) && user) {
2094 user_shm_unlock(inode->i_size, user);
2095 info->flags &= ~VM_LOCKED;
2096 mapping_clear_unevictable(file->f_mapping);
2097 }
2098 retval = 0;
2099
2100out_nomem:
2101 spin_unlock_irq(&info->lock);
2102 return retval;
2103}
2104
2105static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
2106{
2107 file_accessed(file);
2108 vma->vm_ops = &shmem_vm_ops;
2109 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE) &&
2110 ((vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK) <
2111 (vma->vm_end & HPAGE_PMD_MASK)) {
2112 khugepaged_enter(vma, vma->vm_flags);
2113 }
2114 return 0;
2115}
2116
2117static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
2118 umode_t mode, dev_t dev, unsigned long flags)
2119{
2120 struct inode *inode;
2121 struct shmem_inode_info *info;
2122 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2123
2124 if (shmem_reserve_inode(sb))
2125 return NULL;
2126
2127 inode = new_inode(sb);
2128 if (inode) {
2129 inode->i_ino = get_next_ino();
2130 inode_init_owner(inode, dir, mode);
2131 inode->i_blocks = 0;
2132 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
2133 inode->i_generation = get_seconds();
2134 info = SHMEM_I(inode);
2135 memset(info, 0, (char *)inode - (char *)info);
2136 spin_lock_init(&info->lock);
2137 info->seals = F_SEAL_SEAL;
2138 info->flags = flags & VM_NORESERVE;
2139 INIT_LIST_HEAD(&info->shrinklist);
2140 INIT_LIST_HEAD(&info->swaplist);
2141 simple_xattrs_init(&info->xattrs);
2142 cache_no_acl(inode);
2143
2144 switch (mode & S_IFMT) {
2145 default:
2146 inode->i_op = &shmem_special_inode_operations;
2147 init_special_inode(inode, mode, dev);
2148 break;
2149 case S_IFREG:
2150 inode->i_mapping->a_ops = &shmem_aops;
2151 inode->i_op = &shmem_inode_operations;
2152 inode->i_fop = &shmem_file_operations;
2153 mpol_shared_policy_init(&info->policy,
2154 shmem_get_sbmpol(sbinfo));
2155 break;
2156 case S_IFDIR:
2157 inc_nlink(inode);
2158 /* Some things misbehave if size == 0 on a directory */
2159 inode->i_size = 2 * BOGO_DIRENT_SIZE;
2160 inode->i_op = &shmem_dir_inode_operations;
2161 inode->i_fop = &simple_dir_operations;
2162 break;
2163 case S_IFLNK:
2164 /*
2165 * Must not load anything in the rbtree,
2166 * mpol_free_shared_policy will not be called.
2167 */
2168 mpol_shared_policy_init(&info->policy, NULL);
2169 break;
2170 }
2171 } else
2172 shmem_free_inode(sb);
2173 return inode;
2174}
2175
2176bool shmem_mapping(struct address_space *mapping)
2177{
2178 if (!mapping->host)
2179 return false;
2180
2181 return mapping->host->i_sb->s_op == &shmem_ops;
2182}
2183
2184#ifdef CONFIG_TMPFS
2185static const struct inode_operations shmem_symlink_inode_operations;
2186static const struct inode_operations shmem_short_symlink_operations;
2187
2188#ifdef CONFIG_TMPFS_XATTR
2189static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
2190#else
2191#define shmem_initxattrs NULL
2192#endif
2193
2194static int
2195shmem_write_begin(struct file *file, struct address_space *mapping,
2196 loff_t pos, unsigned len, unsigned flags,
2197 struct page **pagep, void **fsdata)
2198{
2199 struct inode *inode = mapping->host;
2200 struct shmem_inode_info *info = SHMEM_I(inode);
2201 pgoff_t index = pos >> PAGE_SHIFT;
2202
2203 /* i_mutex is held by caller */
2204 if (unlikely(info->seals)) {
2205 if (info->seals & F_SEAL_WRITE)
2206 return -EPERM;
2207 if ((info->seals & F_SEAL_GROW) && pos + len > inode->i_size)
2208 return -EPERM;
2209 }
2210
2211 return shmem_getpage(inode, index, pagep, SGP_WRITE);
2212}
2213
2214static int
2215shmem_write_end(struct file *file, struct address_space *mapping,
2216 loff_t pos, unsigned len, unsigned copied,
2217 struct page *page, void *fsdata)
2218{
2219 struct inode *inode = mapping->host;
2220
2221 if (pos + copied > inode->i_size)
2222 i_size_write(inode, pos + copied);
2223
2224 if (!PageUptodate(page)) {
2225 struct page *head = compound_head(page);
2226 if (PageTransCompound(page)) {
2227 int i;
2228
2229 for (i = 0; i < HPAGE_PMD_NR; i++) {
2230 if (head + i == page)
2231 continue;
2232 clear_highpage(head + i);
2233 flush_dcache_page(head + i);
2234 }
2235 }
2236 if (copied < PAGE_SIZE) {
2237 unsigned from = pos & (PAGE_SIZE - 1);
2238 zero_user_segments(page, 0, from,
2239 from + copied, PAGE_SIZE);
2240 }
2241 SetPageUptodate(head);
2242 }
2243 set_page_dirty(page);
2244 unlock_page(page);
2245 put_page(page);
2246
2247 return copied;
2248}
2249
2250static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
2251{
2252 struct file *file = iocb->ki_filp;
2253 struct inode *inode = file_inode(file);
2254 struct address_space *mapping = inode->i_mapping;
2255 pgoff_t index;
2256 unsigned long offset;
2257 enum sgp_type sgp = SGP_READ;
2258 int error = 0;
2259 ssize_t retval = 0;
2260 loff_t *ppos = &iocb->ki_pos;
2261
2262 /*
2263 * Might this read be for a stacking filesystem? Then when reading
2264 * holes of a sparse file, we actually need to allocate those pages,
2265 * and even mark them dirty, so it cannot exceed the max_blocks limit.
2266 */
2267 if (!iter_is_iovec(to))
2268 sgp = SGP_CACHE;
2269
2270 index = *ppos >> PAGE_SHIFT;
2271 offset = *ppos & ~PAGE_MASK;
2272
2273 for (;;) {
2274 struct page *page = NULL;
2275 pgoff_t end_index;
2276 unsigned long nr, ret;
2277 loff_t i_size = i_size_read(inode);
2278
2279 end_index = i_size >> PAGE_SHIFT;
2280 if (index > end_index)
2281 break;
2282 if (index == end_index) {
2283 nr = i_size & ~PAGE_MASK;
2284 if (nr <= offset)
2285 break;
2286 }
2287
2288 error = shmem_getpage(inode, index, &page, sgp);
2289 if (error) {
2290 if (error == -EINVAL)
2291 error = 0;
2292 break;
2293 }
2294 if (page) {
2295 if (sgp == SGP_CACHE)
2296 set_page_dirty(page);
2297 unlock_page(page);
2298 }
2299
2300 /*
2301 * We must evaluate after, since reads (unlike writes)
2302 * are called without i_mutex protection against truncate
2303 */
2304 nr = PAGE_SIZE;
2305 i_size = i_size_read(inode);
2306 end_index = i_size >> PAGE_SHIFT;
2307 if (index == end_index) {
2308 nr = i_size & ~PAGE_MASK;
2309 if (nr <= offset) {
2310 if (page)
2311 put_page(page);
2312 break;
2313 }
2314 }
2315 nr -= offset;
2316
2317 if (page) {
2318 /*
2319 * If users can be writing to this page using arbitrary
2320 * virtual addresses, take care about potential aliasing
2321 * before reading the page on the kernel side.
2322 */
2323 if (mapping_writably_mapped(mapping))
2324 flush_dcache_page(page);
2325 /*
2326 * Mark the page accessed if we read the beginning.
2327 */
2328 if (!offset)
2329 mark_page_accessed(page);
2330 } else {
2331 page = ZERO_PAGE(0);
2332 get_page(page);
2333 }
2334
2335 /*
2336 * Ok, we have the page, and it's up-to-date, so
2337 * now we can copy it to user space...
2338 */
2339 ret = copy_page_to_iter(page, offset, nr, to);
2340 retval += ret;
2341 offset += ret;
2342 index += offset >> PAGE_SHIFT;
2343 offset &= ~PAGE_MASK;
2344
2345 put_page(page);
2346 if (!iov_iter_count(to))
2347 break;
2348 if (ret < nr) {
2349 error = -EFAULT;
2350 break;
2351 }
2352 cond_resched();
2353 }
2354
2355 *ppos = ((loff_t) index << PAGE_SHIFT) + offset;
2356 file_accessed(file);
2357 return retval ? retval : error;
2358}
2359
2360/*
2361 * llseek SEEK_DATA or SEEK_HOLE through the radix_tree.
2362 */
2363static pgoff_t shmem_seek_hole_data(struct address_space *mapping,
2364 pgoff_t index, pgoff_t end, int whence)
2365{
2366 struct page *page;
2367 struct pagevec pvec;
2368 pgoff_t indices[PAGEVEC_SIZE];
2369 bool done = false;
2370 int i;
2371
2372 pagevec_init(&pvec, 0);
2373 pvec.nr = 1; /* start small: we may be there already */
2374 while (!done) {
2375 pvec.nr = find_get_entries(mapping, index,
2376 pvec.nr, pvec.pages, indices);
2377 if (!pvec.nr) {
2378 if (whence == SEEK_DATA)
2379 index = end;
2380 break;
2381 }
2382 for (i = 0; i < pvec.nr; i++, index++) {
2383 if (index < indices[i]) {
2384 if (whence == SEEK_HOLE) {
2385 done = true;
2386 break;
2387 }
2388 index = indices[i];
2389 }
2390 page = pvec.pages[i];
2391 if (page && !radix_tree_exceptional_entry(page)) {
2392 if (!PageUptodate(page))
2393 page = NULL;
2394 }
2395 if (index >= end ||
2396 (page && whence == SEEK_DATA) ||
2397 (!page && whence == SEEK_HOLE)) {
2398 done = true;
2399 break;
2400 }
2401 }
2402 pagevec_remove_exceptionals(&pvec);
2403 pagevec_release(&pvec);
2404 pvec.nr = PAGEVEC_SIZE;
2405 cond_resched();
2406 }
2407 return index;
2408}
2409
2410static loff_t shmem_file_llseek(struct file *file, loff_t offset, int whence)
2411{
2412 struct address_space *mapping = file->f_mapping;
2413 struct inode *inode = mapping->host;
2414 pgoff_t start, end;
2415 loff_t new_offset;
2416
2417 if (whence != SEEK_DATA && whence != SEEK_HOLE)
2418 return generic_file_llseek_size(file, offset, whence,
2419 MAX_LFS_FILESIZE, i_size_read(inode));
2420 inode_lock(inode);
2421 /* We're holding i_mutex so we can access i_size directly */
2422
2423 if (offset < 0)
2424 offset = -EINVAL;
2425 else if (offset >= inode->i_size)
2426 offset = -ENXIO;
2427 else {
2428 start = offset >> PAGE_SHIFT;
2429 end = (inode->i_size + PAGE_SIZE - 1) >> PAGE_SHIFT;
2430 new_offset = shmem_seek_hole_data(mapping, start, end, whence);
2431 new_offset <<= PAGE_SHIFT;
2432 if (new_offset > offset) {
2433 if (new_offset < inode->i_size)
2434 offset = new_offset;
2435 else if (whence == SEEK_DATA)
2436 offset = -ENXIO;
2437 else
2438 offset = inode->i_size;
2439 }
2440 }
2441
2442 if (offset >= 0)
2443 offset = vfs_setpos(file, offset, MAX_LFS_FILESIZE);
2444 inode_unlock(inode);
2445 return offset;
2446}
2447
2448/*
2449 * We need a tag: a new tag would expand every radix_tree_node by 8 bytes,
2450 * so reuse a tag which we firmly believe is never set or cleared on shmem.
2451 */
2452#define SHMEM_TAG_PINNED PAGECACHE_TAG_TOWRITE
2453#define LAST_SCAN 4 /* about 150ms max */
2454
2455static void shmem_tag_pins(struct address_space *mapping)
2456{
2457 struct radix_tree_iter iter;
2458 void **slot;
2459 pgoff_t start;
2460 struct page *page;
2461
2462 lru_add_drain();
2463 start = 0;
2464 rcu_read_lock();
2465
2466 radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
2467 page = radix_tree_deref_slot(slot);
2468 if (!page || radix_tree_exception(page)) {
2469 if (radix_tree_deref_retry(page)) {
2470 slot = radix_tree_iter_retry(&iter);
2471 continue;
2472 }
2473 } else if (page_count(page) - page_mapcount(page) > 1) {
2474 spin_lock_irq(&mapping->tree_lock);
2475 radix_tree_tag_set(&mapping->page_tree, iter.index,
2476 SHMEM_TAG_PINNED);
2477 spin_unlock_irq(&mapping->tree_lock);
2478 }
2479
2480 if (need_resched()) {
2481 slot = radix_tree_iter_resume(slot, &iter);
2482 cond_resched_rcu();
2483 }
2484 }
2485 rcu_read_unlock();
2486}
2487
2488/*
2489 * Setting SEAL_WRITE requires us to verify there's no pending writer. However,
2490 * via get_user_pages(), drivers might have some pending I/O without any active
2491 * user-space mappings (eg., direct-IO, AIO). Therefore, we look at all pages
2492 * and see whether it has an elevated ref-count. If so, we tag them and wait for
2493 * them to be dropped.
2494 * The caller must guarantee that no new user will acquire writable references
2495 * to those pages to avoid races.
2496 */
2497static int shmem_wait_for_pins(struct address_space *mapping)
2498{
2499 struct radix_tree_iter iter;
2500 void **slot;
2501 pgoff_t start;
2502 struct page *page;
2503 int error, scan;
2504
2505 shmem_tag_pins(mapping);
2506
2507 error = 0;
2508 for (scan = 0; scan <= LAST_SCAN; scan++) {
2509 if (!radix_tree_tagged(&mapping->page_tree, SHMEM_TAG_PINNED))
2510 break;
2511
2512 if (!scan)
2513 lru_add_drain_all();
2514 else if (schedule_timeout_killable((HZ << scan) / 200))
2515 scan = LAST_SCAN;
2516
2517 start = 0;
2518 rcu_read_lock();
2519 radix_tree_for_each_tagged(slot, &mapping->page_tree, &iter,
2520 start, SHMEM_TAG_PINNED) {
2521
2522 page = radix_tree_deref_slot(slot);
2523 if (radix_tree_exception(page)) {
2524 if (radix_tree_deref_retry(page)) {
2525 slot = radix_tree_iter_retry(&iter);
2526 continue;
2527 }
2528
2529 page = NULL;
2530 }
2531
2532 if (page &&
2533 page_count(page) - page_mapcount(page) != 1) {
2534 if (scan < LAST_SCAN)
2535 goto continue_resched;
2536
2537 /*
2538 * On the last scan, we clean up all those tags
2539 * we inserted; but make a note that we still
2540 * found pages pinned.
2541 */
2542 error = -EBUSY;
2543 }
2544
2545 spin_lock_irq(&mapping->tree_lock);
2546 radix_tree_tag_clear(&mapping->page_tree,
2547 iter.index, SHMEM_TAG_PINNED);
2548 spin_unlock_irq(&mapping->tree_lock);
2549continue_resched:
2550 if (need_resched()) {
2551 slot = radix_tree_iter_resume(slot, &iter);
2552 cond_resched_rcu();
2553 }
2554 }
2555 rcu_read_unlock();
2556 }
2557
2558 return error;
2559}
2560
2561#define F_ALL_SEALS (F_SEAL_SEAL | \
2562 F_SEAL_SHRINK | \
2563 F_SEAL_GROW | \
2564 F_SEAL_WRITE)
2565
2566int shmem_add_seals(struct file *file, unsigned int seals)
2567{
2568 struct inode *inode = file_inode(file);
2569 struct shmem_inode_info *info = SHMEM_I(inode);
2570 int error;
2571
2572 /*
2573 * SEALING
2574 * Sealing allows multiple parties to share a shmem-file but restrict
2575 * access to a specific subset of file operations. Seals can only be
2576 * added, but never removed. This way, mutually untrusted parties can
2577 * share common memory regions with a well-defined policy. A malicious
2578 * peer can thus never perform unwanted operations on a shared object.
2579 *
2580 * Seals are only supported on special shmem-files and always affect
2581 * the whole underlying inode. Once a seal is set, it may prevent some
2582 * kinds of access to the file. Currently, the following seals are
2583 * defined:
2584 * SEAL_SEAL: Prevent further seals from being set on this file
2585 * SEAL_SHRINK: Prevent the file from shrinking
2586 * SEAL_GROW: Prevent the file from growing
2587 * SEAL_WRITE: Prevent write access to the file
2588 *
2589 * As we don't require any trust relationship between two parties, we
2590 * must prevent seals from being removed. Therefore, sealing a file
2591 * only adds a given set of seals to the file, it never touches
2592 * existing seals. Furthermore, the "setting seals"-operation can be
2593 * sealed itself, which basically prevents any further seal from being
2594 * added.
2595 *
2596 * Semantics of sealing are only defined on volatile files. Only
2597 * anonymous shmem files support sealing. More importantly, seals are
2598 * never written to disk. Therefore, there's no plan to support it on
2599 * other file types.
2600 */
2601
2602 if (file->f_op != &shmem_file_operations)
2603 return -EINVAL;
2604 if (!(file->f_mode & FMODE_WRITE))
2605 return -EPERM;
2606 if (seals & ~(unsigned int)F_ALL_SEALS)
2607 return -EINVAL;
2608
2609 inode_lock(inode);
2610
2611 if (info->seals & F_SEAL_SEAL) {
2612 error = -EPERM;
2613 goto unlock;
2614 }
2615
2616 if ((seals & F_SEAL_WRITE) && !(info->seals & F_SEAL_WRITE)) {
2617 error = mapping_deny_writable(file->f_mapping);
2618 if (error)
2619 goto unlock;
2620
2621 error = shmem_wait_for_pins(file->f_mapping);
2622 if (error) {
2623 mapping_allow_writable(file->f_mapping);
2624 goto unlock;
2625 }
2626 }
2627
2628 info->seals |= seals;
2629 error = 0;
2630
2631unlock:
2632 inode_unlock(inode);
2633 return error;
2634}
2635EXPORT_SYMBOL_GPL(shmem_add_seals);
2636
2637int shmem_get_seals(struct file *file)
2638{
2639 if (file->f_op != &shmem_file_operations)
2640 return -EINVAL;
2641
2642 return SHMEM_I(file_inode(file))->seals;
2643}
2644EXPORT_SYMBOL_GPL(shmem_get_seals);
2645
2646long shmem_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
2647{
2648 long error;
2649
2650 switch (cmd) {
2651 case F_ADD_SEALS:
2652 /* disallow upper 32bit */
2653 if (arg > UINT_MAX)
2654 return -EINVAL;
2655
2656 error = shmem_add_seals(file, arg);
2657 break;
2658 case F_GET_SEALS:
2659 error = shmem_get_seals(file);
2660 break;
2661 default:
2662 error = -EINVAL;
2663 break;
2664 }
2665
2666 return error;
2667}
2668
2669static long shmem_fallocate(struct file *file, int mode, loff_t offset,
2670 loff_t len)
2671{
2672 struct inode *inode = file_inode(file);
2673 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
2674 struct shmem_inode_info *info = SHMEM_I(inode);
2675 struct shmem_falloc shmem_falloc;
2676 pgoff_t start, index, end;
2677 int error;
2678
2679 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2680 return -EOPNOTSUPP;
2681
2682 inode_lock(inode);
2683
2684 if (mode & FALLOC_FL_PUNCH_HOLE) {
2685 struct address_space *mapping = file->f_mapping;
2686 loff_t unmap_start = round_up(offset, PAGE_SIZE);
2687 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
2688 DECLARE_WAIT_QUEUE_HEAD_ONSTACK(shmem_falloc_waitq);
2689
2690 /* protected by i_mutex */
2691 if (info->seals & F_SEAL_WRITE) {
2692 error = -EPERM;
2693 goto out;
2694 }
2695
2696 shmem_falloc.waitq = &shmem_falloc_waitq;
2697 shmem_falloc.start = unmap_start >> PAGE_SHIFT;
2698 shmem_falloc.next = (unmap_end + 1) >> PAGE_SHIFT;
2699 spin_lock(&inode->i_lock);
2700 inode->i_private = &shmem_falloc;
2701 spin_unlock(&inode->i_lock);
2702
2703 if ((u64)unmap_end > (u64)unmap_start)
2704 unmap_mapping_range(mapping, unmap_start,
2705 1 + unmap_end - unmap_start, 0);
2706 shmem_truncate_range(inode, offset, offset + len - 1);
2707 /* No need to unmap again: hole-punching leaves COWed pages */
2708
2709 spin_lock(&inode->i_lock);
2710 inode->i_private = NULL;
2711 wake_up_all(&shmem_falloc_waitq);
2712 WARN_ON_ONCE(!list_empty(&shmem_falloc_waitq.task_list));
2713 spin_unlock(&inode->i_lock);
2714 error = 0;
2715 goto out;
2716 }
2717
2718 /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
2719 error = inode_newsize_ok(inode, offset + len);
2720 if (error)
2721 goto out;
2722
2723 if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
2724 error = -EPERM;
2725 goto out;
2726 }
2727
2728 start = offset >> PAGE_SHIFT;
2729 end = (offset + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
2730 /* Try to avoid a swapstorm if len is impossible to satisfy */
2731 if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
2732 error = -ENOSPC;
2733 goto out;
2734 }
2735
2736 shmem_falloc.waitq = NULL;
2737 shmem_falloc.start = start;
2738 shmem_falloc.next = start;
2739 shmem_falloc.nr_falloced = 0;
2740 shmem_falloc.nr_unswapped = 0;
2741 spin_lock(&inode->i_lock);
2742 inode->i_private = &shmem_falloc;
2743 spin_unlock(&inode->i_lock);
2744
2745 for (index = start; index < end; index++) {
2746 struct page *page;
2747
2748 /*
2749 * Good, the fallocate(2) manpage permits EINTR: we may have
2750 * been interrupted because we are using up too much memory.
2751 */
2752 if (signal_pending(current))
2753 error = -EINTR;
2754 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
2755 error = -ENOMEM;
2756 else
2757 error = shmem_getpage(inode, index, &page, SGP_FALLOC);
2758 if (error) {
2759 /* Remove the !PageUptodate pages we added */
2760 if (index > start) {
2761 shmem_undo_range(inode,
2762 (loff_t)start << PAGE_SHIFT,
2763 ((loff_t)index << PAGE_SHIFT) - 1, true);
2764 }
2765 goto undone;
2766 }
2767
2768 /*
2769 * Inform shmem_writepage() how far we have reached.
2770 * No need for lock or barrier: we have the page lock.
2771 */
2772 shmem_falloc.next++;
2773 if (!PageUptodate(page))
2774 shmem_falloc.nr_falloced++;
2775
2776 /*
2777 * If !PageUptodate, leave it that way so that freeable pages
2778 * can be recognized if we need to rollback on error later.
2779 * But set_page_dirty so that memory pressure will swap rather
2780 * than free the pages we are allocating (and SGP_CACHE pages
2781 * might still be clean: we now need to mark those dirty too).
2782 */
2783 set_page_dirty(page);
2784 unlock_page(page);
2785 put_page(page);
2786 cond_resched();
2787 }
2788
2789 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
2790 i_size_write(inode, offset + len);
2791 inode->i_ctime = current_time(inode);
2792undone:
2793 spin_lock(&inode->i_lock);
2794 inode->i_private = NULL;
2795 spin_unlock(&inode->i_lock);
2796out:
2797 inode_unlock(inode);
2798 return error;
2799}
2800
2801static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
2802{
2803 struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
2804
2805 buf->f_type = TMPFS_MAGIC;
2806 buf->f_bsize = PAGE_SIZE;
2807 buf->f_namelen = NAME_MAX;
2808 if (sbinfo->max_blocks) {
2809 buf->f_blocks = sbinfo->max_blocks;
2810 buf->f_bavail =
2811 buf->f_bfree = sbinfo->max_blocks -
2812 percpu_counter_sum(&sbinfo->used_blocks);
2813 }
2814 if (sbinfo->max_inodes) {
2815 buf->f_files = sbinfo->max_inodes;
2816 buf->f_ffree = sbinfo->free_inodes;
2817 }
2818 /* else leave those fields 0 like simple_statfs */
2819 return 0;
2820}
2821
2822/*
2823 * File creation. Allocate an inode, and we're done..
2824 */
2825static int
2826shmem_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
2827{
2828 struct inode *inode;
2829 int error = -ENOSPC;
2830
2831 inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
2832 if (inode) {
2833 error = simple_acl_create(dir, inode);
2834 if (error)
2835 goto out_iput;
2836 error = security_inode_init_security(inode, dir,
2837 &dentry->d_name,
2838 shmem_initxattrs, NULL);
2839 if (error && error != -EOPNOTSUPP)
2840 goto out_iput;
2841
2842 error = 0;
2843 dir->i_size += BOGO_DIRENT_SIZE;
2844 dir->i_ctime = dir->i_mtime = current_time(dir);
2845 d_instantiate(dentry, inode);
2846 dget(dentry); /* Extra count - pin the dentry in core */
2847 }
2848 return error;
2849out_iput:
2850 iput(inode);
2851 return error;
2852}
2853
2854static int
2855shmem_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
2856{
2857 struct inode *inode;
2858 int error = -ENOSPC;
2859
2860 inode = shmem_get_inode(dir->i_sb, dir, mode, 0, VM_NORESERVE);
2861 if (inode) {
2862 error = security_inode_init_security(inode, dir,
2863 NULL,
2864 shmem_initxattrs, NULL);
2865 if (error && error != -EOPNOTSUPP)
2866 goto out_iput;
2867 error = simple_acl_create(dir, inode);
2868 if (error)
2869 goto out_iput;
2870 d_tmpfile(dentry, inode);
2871 }
2872 return error;
2873out_iput:
2874 iput(inode);
2875 return error;
2876}
2877
2878static int shmem_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
2879{
2880 int error;
2881
2882 if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
2883 return error;
2884 inc_nlink(dir);
2885 return 0;
2886}
2887
2888static int shmem_create(struct inode *dir, struct dentry *dentry, umode_t mode,
2889 bool excl)
2890{
2891 return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
2892}
2893
2894/*
2895 * Link a file..
2896 */
2897static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
2898{
2899 struct inode *inode = d_inode(old_dentry);
2900 int ret;
2901
2902 /*
2903 * No ordinary (disk based) filesystem counts links as inodes;
2904 * but each new link needs a new dentry, pinning lowmem, and
2905 * tmpfs dentries cannot be pruned until they are unlinked.
2906 */
2907 ret = shmem_reserve_inode(inode->i_sb);
2908 if (ret)
2909 goto out;
2910
2911 dir->i_size += BOGO_DIRENT_SIZE;
2912 inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
2913 inc_nlink(inode);
2914 ihold(inode); /* New dentry reference */
2915 dget(dentry); /* Extra pinning count for the created dentry */
2916 d_instantiate(dentry, inode);
2917out:
2918 return ret;
2919}
2920
2921static int shmem_unlink(struct inode *dir, struct dentry *dentry)
2922{
2923 struct inode *inode = d_inode(dentry);
2924
2925 if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
2926 shmem_free_inode(inode->i_sb);
2927
2928 dir->i_size -= BOGO_DIRENT_SIZE;
2929 inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
2930 drop_nlink(inode);
2931 dput(dentry); /* Undo the count from "create" - this does all the work */
2932 return 0;
2933}
2934
2935static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
2936{
2937 if (!simple_empty(dentry))
2938 return -ENOTEMPTY;
2939
2940 drop_nlink(d_inode(dentry));
2941 drop_nlink(dir);
2942 return shmem_unlink(dir, dentry);
2943}
2944
2945static int shmem_exchange(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
2946{
2947 bool old_is_dir = d_is_dir(old_dentry);
2948 bool new_is_dir = d_is_dir(new_dentry);
2949
2950 if (old_dir != new_dir && old_is_dir != new_is_dir) {
2951 if (old_is_dir) {
2952 drop_nlink(old_dir);
2953 inc_nlink(new_dir);
2954 } else {
2955 drop_nlink(new_dir);
2956 inc_nlink(old_dir);
2957 }
2958 }
2959 old_dir->i_ctime = old_dir->i_mtime =
2960 new_dir->i_ctime = new_dir->i_mtime =
2961 d_inode(old_dentry)->i_ctime =
2962 d_inode(new_dentry)->i_ctime = current_time(old_dir);
2963
2964 return 0;
2965}
2966
2967static int shmem_whiteout(struct inode *old_dir, struct dentry *old_dentry)
2968{
2969 struct dentry *whiteout;
2970 int error;
2971
2972 whiteout = d_alloc(old_dentry->d_parent, &old_dentry->d_name);
2973 if (!whiteout)
2974 return -ENOMEM;
2975
2976 error = shmem_mknod(old_dir, whiteout,
2977 S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
2978 dput(whiteout);
2979 if (error)
2980 return error;
2981
2982 /*
2983 * Cheat and hash the whiteout while the old dentry is still in
2984 * place, instead of playing games with FS_RENAME_DOES_D_MOVE.
2985 *
2986 * d_lookup() will consistently find one of them at this point,
2987 * not sure which one, but that isn't even important.
2988 */
2989 d_rehash(whiteout);
2990 return 0;
2991}
2992
2993/*
2994 * The VFS layer already does all the dentry stuff for rename,
2995 * we just have to decrement the usage count for the target if
2996 * it exists so that the VFS layer correctly free's it when it
2997 * gets overwritten.
2998 */
2999static int shmem_rename2(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry, unsigned int flags)
3000{
3001 struct inode *inode = d_inode(old_dentry);
3002 int they_are_dirs = S_ISDIR(inode->i_mode);
3003
3004 if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
3005 return -EINVAL;
3006
3007 if (flags & RENAME_EXCHANGE)
3008 return shmem_exchange(old_dir, old_dentry, new_dir, new_dentry);
3009
3010 if (!simple_empty(new_dentry))
3011 return -ENOTEMPTY;
3012
3013 if (flags & RENAME_WHITEOUT) {
3014 int error;
3015
3016 error = shmem_whiteout(old_dir, old_dentry);
3017 if (error)
3018 return error;
3019 }
3020
3021 if (d_really_is_positive(new_dentry)) {
3022 (void) shmem_unlink(new_dir, new_dentry);
3023 if (they_are_dirs) {
3024 drop_nlink(d_inode(new_dentry));
3025 drop_nlink(old_dir);
3026 }
3027 } else if (they_are_dirs) {
3028 drop_nlink(old_dir);
3029 inc_nlink(new_dir);
3030 }
3031
3032 old_dir->i_size -= BOGO_DIRENT_SIZE;
3033 new_dir->i_size += BOGO_DIRENT_SIZE;
3034 old_dir->i_ctime = old_dir->i_mtime =
3035 new_dir->i_ctime = new_dir->i_mtime =
3036 inode->i_ctime = current_time(old_dir);
3037 return 0;
3038}
3039
3040static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
3041{
3042 int error;
3043 int len;
3044 struct inode *inode;
3045 struct page *page;
3046 struct shmem_inode_info *info;
3047
3048 len = strlen(symname) + 1;
3049 if (len > PAGE_SIZE)
3050 return -ENAMETOOLONG;
3051
3052 inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0, VM_NORESERVE);
3053 if (!inode)
3054 return -ENOSPC;
3055
3056 error = security_inode_init_security(inode, dir, &dentry->d_name,
3057 shmem_initxattrs, NULL);
3058 if (error) {
3059 if (error != -EOPNOTSUPP) {
3060 iput(inode);
3061 return error;
3062 }
3063 error = 0;
3064 }
3065
3066 info = SHMEM_I(inode);
3067 inode->i_size = len-1;
3068 if (len <= SHORT_SYMLINK_LEN) {
3069 inode->i_link = kmemdup(symname, len, GFP_KERNEL);
3070 if (!inode->i_link) {
3071 iput(inode);
3072 return -ENOMEM;
3073 }
3074 inode->i_op = &shmem_short_symlink_operations;
3075 } else {
3076 inode_nohighmem(inode);
3077 error = shmem_getpage(inode, 0, &page, SGP_WRITE);
3078 if (error) {
3079 iput(inode);
3080 return error;
3081 }
3082 inode->i_mapping->a_ops = &shmem_aops;
3083 inode->i_op = &shmem_symlink_inode_operations;
3084 memcpy(page_address(page), symname, len);
3085 SetPageUptodate(page);
3086 set_page_dirty(page);
3087 unlock_page(page);
3088 put_page(page);
3089 }
3090 dir->i_size += BOGO_DIRENT_SIZE;
3091 dir->i_ctime = dir->i_mtime = current_time(dir);
3092 d_instantiate(dentry, inode);
3093 dget(dentry);
3094 return 0;
3095}
3096
3097static void shmem_put_link(void *arg)
3098{
3099 mark_page_accessed(arg);
3100 put_page(arg);
3101}
3102
3103static const char *shmem_get_link(struct dentry *dentry,
3104 struct inode *inode,
3105 struct delayed_call *done)
3106{
3107 struct page *page = NULL;
3108 int error;
3109 if (!dentry) {
3110 page = find_get_page(inode->i_mapping, 0);
3111 if (!page)
3112 return ERR_PTR(-ECHILD);
3113 if (!PageUptodate(page)) {
3114 put_page(page);
3115 return ERR_PTR(-ECHILD);
3116 }
3117 } else {
3118 error = shmem_getpage(inode, 0, &page, SGP_READ);
3119 if (error)
3120 return ERR_PTR(error);
3121 unlock_page(page);
3122 }
3123 set_delayed_call(done, shmem_put_link, page);
3124 return page_address(page);
3125}
3126
3127#ifdef CONFIG_TMPFS_XATTR
3128/*
3129 * Superblocks without xattr inode operations may get some security.* xattr
3130 * support from the LSM "for free". As soon as we have any other xattrs
3131 * like ACLs, we also need to implement the security.* handlers at
3132 * filesystem level, though.
3133 */
3134
3135/*
3136 * Callback for security_inode_init_security() for acquiring xattrs.
3137 */
3138static int shmem_initxattrs(struct inode *inode,
3139 const struct xattr *xattr_array,
3140 void *fs_info)
3141{
3142 struct shmem_inode_info *info = SHMEM_I(inode);
3143 const struct xattr *xattr;
3144 struct simple_xattr *new_xattr;
3145 size_t len;
3146
3147 for (xattr = xattr_array; xattr->name != NULL; xattr++) {
3148 new_xattr = simple_xattr_alloc(xattr->value, xattr->value_len);
3149 if (!new_xattr)
3150 return -ENOMEM;
3151
3152 len = strlen(xattr->name) + 1;
3153 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
3154 GFP_KERNEL);
3155 if (!new_xattr->name) {
3156 kfree(new_xattr);
3157 return -ENOMEM;
3158 }
3159
3160 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
3161 XATTR_SECURITY_PREFIX_LEN);
3162 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
3163 xattr->name, len);
3164
3165 simple_xattr_list_add(&info->xattrs, new_xattr);
3166 }
3167
3168 return 0;
3169}
3170
3171static int shmem_xattr_handler_get(const struct xattr_handler *handler,
3172 struct dentry *unused, struct inode *inode,
3173 const char *name, void *buffer, size_t size)
3174{
3175 struct shmem_inode_info *info = SHMEM_I(inode);
3176
3177 name = xattr_full_name(handler, name);
3178 return simple_xattr_get(&info->xattrs, name, buffer, size);
3179}
3180
3181static int shmem_xattr_handler_set(const struct xattr_handler *handler,
3182 struct dentry *unused, struct inode *inode,
3183 const char *name, const void *value,
3184 size_t size, int flags)
3185{
3186 struct shmem_inode_info *info = SHMEM_I(inode);
3187
3188 name = xattr_full_name(handler, name);
3189 return simple_xattr_set(&info->xattrs, name, value, size, flags);
3190}
3191
3192static const struct xattr_handler shmem_security_xattr_handler = {
3193 .prefix = XATTR_SECURITY_PREFIX,
3194 .get = shmem_xattr_handler_get,
3195 .set = shmem_xattr_handler_set,
3196};
3197
3198static const struct xattr_handler shmem_trusted_xattr_handler = {
3199 .prefix = XATTR_TRUSTED_PREFIX,
3200 .get = shmem_xattr_handler_get,
3201 .set = shmem_xattr_handler_set,
3202};
3203
3204static const struct xattr_handler *shmem_xattr_handlers[] = {
3205#ifdef CONFIG_TMPFS_POSIX_ACL
3206 &posix_acl_access_xattr_handler,
3207 &posix_acl_default_xattr_handler,
3208#endif
3209 &shmem_security_xattr_handler,
3210 &shmem_trusted_xattr_handler,
3211 NULL
3212};
3213
3214static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
3215{
3216 struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
3217 return simple_xattr_list(d_inode(dentry), &info->xattrs, buffer, size);
3218}
3219#endif /* CONFIG_TMPFS_XATTR */
3220
3221static const struct inode_operations shmem_short_symlink_operations = {
3222 .get_link = simple_get_link,
3223#ifdef CONFIG_TMPFS_XATTR
3224 .listxattr = shmem_listxattr,
3225#endif
3226};
3227
3228static const struct inode_operations shmem_symlink_inode_operations = {
3229 .get_link = shmem_get_link,
3230#ifdef CONFIG_TMPFS_XATTR
3231 .listxattr = shmem_listxattr,
3232#endif
3233};
3234
3235static struct dentry *shmem_get_parent(struct dentry *child)
3236{
3237 return ERR_PTR(-ESTALE);
3238}
3239
3240static int shmem_match(struct inode *ino, void *vfh)
3241{
3242 __u32 *fh = vfh;
3243 __u64 inum = fh[2];
3244 inum = (inum << 32) | fh[1];
3245 return ino->i_ino == inum && fh[0] == ino->i_generation;
3246}
3247
3248static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
3249 struct fid *fid, int fh_len, int fh_type)
3250{
3251 struct inode *inode;
3252 struct dentry *dentry = NULL;
3253 u64 inum;
3254
3255 if (fh_len < 3)
3256 return NULL;
3257
3258 inum = fid->raw[2];
3259 inum = (inum << 32) | fid->raw[1];
3260
3261 inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
3262 shmem_match, fid->raw);
3263 if (inode) {
3264 dentry = d_find_alias(inode);
3265 iput(inode);
3266 }
3267
3268 return dentry;
3269}
3270
3271static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
3272 struct inode *parent)
3273{
3274 if (*len < 3) {
3275 *len = 3;
3276 return FILEID_INVALID;
3277 }
3278
3279 if (inode_unhashed(inode)) {
3280 /* Unfortunately insert_inode_hash is not idempotent,
3281 * so as we hash inodes here rather than at creation
3282 * time, we need a lock to ensure we only try
3283 * to do it once
3284 */
3285 static DEFINE_SPINLOCK(lock);
3286 spin_lock(&lock);
3287 if (inode_unhashed(inode))
3288 __insert_inode_hash(inode,
3289 inode->i_ino + inode->i_generation);
3290 spin_unlock(&lock);
3291 }
3292
3293 fh[0] = inode->i_generation;
3294 fh[1] = inode->i_ino;
3295 fh[2] = ((__u64)inode->i_ino) >> 32;
3296
3297 *len = 3;
3298 return 1;
3299}
3300
3301static const struct export_operations shmem_export_ops = {
3302 .get_parent = shmem_get_parent,
3303 .encode_fh = shmem_encode_fh,
3304 .fh_to_dentry = shmem_fh_to_dentry,
3305};
3306
3307static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
3308 bool remount)
3309{
3310 char *this_char, *value, *rest;
3311 struct mempolicy *mpol = NULL;
3312 uid_t uid;
3313 gid_t gid;
3314
3315 while (options != NULL) {
3316 this_char = options;
3317 for (;;) {
3318 /*
3319 * NUL-terminate this option: unfortunately,
3320 * mount options form a comma-separated list,
3321 * but mpol's nodelist may also contain commas.
3322 */
3323 options = strchr(options, ',');
3324 if (options == NULL)
3325 break;
3326 options++;
3327 if (!isdigit(*options)) {
3328 options[-1] = '\0';
3329 break;
3330 }
3331 }
3332 if (!*this_char)
3333 continue;
3334 if ((value = strchr(this_char,'=')) != NULL) {
3335 *value++ = 0;
3336 } else {
3337 pr_err("tmpfs: No value for mount option '%s'\n",
3338 this_char);
3339 goto error;
3340 }
3341
3342 if (!strcmp(this_char,"size")) {
3343 unsigned long long size;
3344 size = memparse(value,&rest);
3345 if (*rest == '%') {
3346 size <<= PAGE_SHIFT;
3347 size *= totalram_pages;
3348 do_div(size, 100);
3349 rest++;
3350 }
3351 if (*rest)
3352 goto bad_val;
3353 sbinfo->max_blocks =
3354 DIV_ROUND_UP(size, PAGE_SIZE);
3355 } else if (!strcmp(this_char,"nr_blocks")) {
3356 sbinfo->max_blocks = memparse(value, &rest);
3357 if (*rest)
3358 goto bad_val;
3359 } else if (!strcmp(this_char,"nr_inodes")) {
3360 sbinfo->max_inodes = memparse(value, &rest);
3361 if (*rest)
3362 goto bad_val;
3363 } else if (!strcmp(this_char,"mode")) {
3364 if (remount)
3365 continue;
3366 sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
3367 if (*rest)
3368 goto bad_val;
3369 } else if (!strcmp(this_char,"uid")) {
3370 if (remount)
3371 continue;
3372 uid = simple_strtoul(value, &rest, 0);
3373 if (*rest)
3374 goto bad_val;
3375 sbinfo->uid = make_kuid(current_user_ns(), uid);
3376 if (!uid_valid(sbinfo->uid))
3377 goto bad_val;
3378 } else if (!strcmp(this_char,"gid")) {
3379 if (remount)
3380 continue;
3381 gid = simple_strtoul(value, &rest, 0);
3382 if (*rest)
3383 goto bad_val;
3384 sbinfo->gid = make_kgid(current_user_ns(), gid);
3385 if (!gid_valid(sbinfo->gid))
3386 goto bad_val;
3387#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3388 } else if (!strcmp(this_char, "huge")) {
3389 int huge;
3390 huge = shmem_parse_huge(value);
3391 if (huge < 0)
3392 goto bad_val;
3393 if (!has_transparent_hugepage() &&
3394 huge != SHMEM_HUGE_NEVER)
3395 goto bad_val;
3396 sbinfo->huge = huge;
3397#endif
3398#ifdef CONFIG_NUMA
3399 } else if (!strcmp(this_char,"mpol")) {
3400 mpol_put(mpol);
3401 mpol = NULL;
3402 if (mpol_parse_str(value, &mpol))
3403 goto bad_val;
3404#endif
3405 } else {
3406 pr_err("tmpfs: Bad mount option %s\n", this_char);
3407 goto error;
3408 }
3409 }
3410 sbinfo->mpol = mpol;
3411 return 0;
3412
3413bad_val:
3414 pr_err("tmpfs: Bad value '%s' for mount option '%s'\n",
3415 value, this_char);
3416error:
3417 mpol_put(mpol);
3418 return 1;
3419
3420}
3421
3422static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
3423{
3424 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
3425 struct shmem_sb_info config = *sbinfo;
3426 unsigned long inodes;
3427 int error = -EINVAL;
3428
3429 config.mpol = NULL;
3430 if (shmem_parse_options(data, &config, true))
3431 return error;
3432
3433 spin_lock(&sbinfo->stat_lock);
3434 inodes = sbinfo->max_inodes - sbinfo->free_inodes;
3435 if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
3436 goto out;
3437 if (config.max_inodes < inodes)
3438 goto out;
3439 /*
3440 * Those tests disallow limited->unlimited while any are in use;
3441 * but we must separately disallow unlimited->limited, because
3442 * in that case we have no record of how much is already in use.
3443 */
3444 if (config.max_blocks && !sbinfo->max_blocks)
3445 goto out;
3446 if (config.max_inodes && !sbinfo->max_inodes)
3447 goto out;
3448
3449 error = 0;
3450 sbinfo->huge = config.huge;
3451 sbinfo->max_blocks = config.max_blocks;
3452 sbinfo->max_inodes = config.max_inodes;
3453 sbinfo->free_inodes = config.max_inodes - inodes;
3454
3455 /*
3456 * Preserve previous mempolicy unless mpol remount option was specified.
3457 */
3458 if (config.mpol) {
3459 mpol_put(sbinfo->mpol);
3460 sbinfo->mpol = config.mpol; /* transfers initial ref */
3461 }
3462out:
3463 spin_unlock(&sbinfo->stat_lock);
3464 return error;
3465}
3466
3467static int shmem_show_options(struct seq_file *seq, struct dentry *root)
3468{
3469 struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
3470
3471 if (sbinfo->max_blocks != shmem_default_max_blocks())
3472 seq_printf(seq, ",size=%luk",
3473 sbinfo->max_blocks << (PAGE_SHIFT - 10));
3474 if (sbinfo->max_inodes != shmem_default_max_inodes())
3475 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
3476 if (sbinfo->mode != (S_IRWXUGO | S_ISVTX))
3477 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
3478 if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
3479 seq_printf(seq, ",uid=%u",
3480 from_kuid_munged(&init_user_ns, sbinfo->uid));
3481 if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
3482 seq_printf(seq, ",gid=%u",
3483 from_kgid_munged(&init_user_ns, sbinfo->gid));
3484#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3485 /* Rightly or wrongly, show huge mount option unmasked by shmem_huge */
3486 if (sbinfo->huge)
3487 seq_printf(seq, ",huge=%s", shmem_format_huge(sbinfo->huge));
3488#endif
3489 shmem_show_mpol(seq, sbinfo->mpol);
3490 return 0;
3491}
3492
3493#define MFD_NAME_PREFIX "memfd:"
3494#define MFD_NAME_PREFIX_LEN (sizeof(MFD_NAME_PREFIX) - 1)
3495#define MFD_NAME_MAX_LEN (NAME_MAX - MFD_NAME_PREFIX_LEN)
3496
3497#define MFD_ALL_FLAGS (MFD_CLOEXEC | MFD_ALLOW_SEALING)
3498
3499SYSCALL_DEFINE2(memfd_create,
3500 const char __user *, uname,
3501 unsigned int, flags)
3502{
3503 struct shmem_inode_info *info;
3504 struct file *file;
3505 int fd, error;
3506 char *name;
3507 long len;
3508
3509 if (flags & ~(unsigned int)MFD_ALL_FLAGS)
3510 return -EINVAL;
3511
3512 /* length includes terminating zero */
3513 len = strnlen_user(uname, MFD_NAME_MAX_LEN + 1);
3514 if (len <= 0)
3515 return -EFAULT;
3516 if (len > MFD_NAME_MAX_LEN + 1)
3517 return -EINVAL;
3518
3519 name = kmalloc(len + MFD_NAME_PREFIX_LEN, GFP_TEMPORARY);
3520 if (!name)
3521 return -ENOMEM;
3522
3523 strcpy(name, MFD_NAME_PREFIX);
3524 if (copy_from_user(&name[MFD_NAME_PREFIX_LEN], uname, len)) {
3525 error = -EFAULT;
3526 goto err_name;
3527 }
3528
3529 /* terminating-zero may have changed after strnlen_user() returned */
3530 if (name[len + MFD_NAME_PREFIX_LEN - 1]) {
3531 error = -EFAULT;
3532 goto err_name;
3533 }
3534
3535 fd = get_unused_fd_flags((flags & MFD_CLOEXEC) ? O_CLOEXEC : 0);
3536 if (fd < 0) {
3537 error = fd;
3538 goto err_name;
3539 }
3540
3541 file = shmem_file_setup(name, 0, VM_NORESERVE);
3542 if (IS_ERR(file)) {
3543 error = PTR_ERR(file);
3544 goto err_fd;
3545 }
3546 info = SHMEM_I(file_inode(file));
3547 file->f_mode |= FMODE_LSEEK | FMODE_PREAD | FMODE_PWRITE;
3548 file->f_flags |= O_RDWR | O_LARGEFILE;
3549 if (flags & MFD_ALLOW_SEALING)
3550 info->seals &= ~F_SEAL_SEAL;
3551
3552 fd_install(fd, file);
3553 kfree(name);
3554 return fd;
3555
3556err_fd:
3557 put_unused_fd(fd);
3558err_name:
3559 kfree(name);
3560 return error;
3561}
3562
3563#endif /* CONFIG_TMPFS */
3564
3565static void shmem_put_super(struct super_block *sb)
3566{
3567 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
3568
3569 percpu_counter_destroy(&sbinfo->used_blocks);
3570 mpol_put(sbinfo->mpol);
3571 kfree(sbinfo);
3572 sb->s_fs_info = NULL;
3573}
3574
3575int shmem_fill_super(struct super_block *sb, void *data, int silent)
3576{
3577 struct inode *inode;
3578 struct shmem_sb_info *sbinfo;
3579 int err = -ENOMEM;
3580
3581 /* Round up to L1_CACHE_BYTES to resist false sharing */
3582 sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
3583 L1_CACHE_BYTES), GFP_KERNEL);
3584 if (!sbinfo)
3585 return -ENOMEM;
3586
3587 sbinfo->mode = S_IRWXUGO | S_ISVTX;
3588 sbinfo->uid = current_fsuid();
3589 sbinfo->gid = current_fsgid();
3590 sb->s_fs_info = sbinfo;
3591
3592#ifdef CONFIG_TMPFS
3593 /*
3594 * Per default we only allow half of the physical ram per
3595 * tmpfs instance, limiting inodes to one per page of lowmem;
3596 * but the internal instance is left unlimited.
3597 */
3598 if (!(sb->s_flags & MS_KERNMOUNT)) {
3599 sbinfo->max_blocks = shmem_default_max_blocks();
3600 sbinfo->max_inodes = shmem_default_max_inodes();
3601 if (shmem_parse_options(data, sbinfo, false)) {
3602 err = -EINVAL;
3603 goto failed;
3604 }
3605 } else {
3606 sb->s_flags |= MS_NOUSER;
3607 }
3608 sb->s_export_op = &shmem_export_ops;
3609 sb->s_flags |= MS_NOSEC;
3610#else
3611 sb->s_flags |= MS_NOUSER;
3612#endif
3613
3614 spin_lock_init(&sbinfo->stat_lock);
3615 if (percpu_counter_init(&sbinfo->used_blocks, 0, GFP_KERNEL))
3616 goto failed;
3617 sbinfo->free_inodes = sbinfo->max_inodes;
3618 spin_lock_init(&sbinfo->shrinklist_lock);
3619 INIT_LIST_HEAD(&sbinfo->shrinklist);
3620
3621 sb->s_maxbytes = MAX_LFS_FILESIZE;
3622 sb->s_blocksize = PAGE_SIZE;
3623 sb->s_blocksize_bits = PAGE_SHIFT;
3624 sb->s_magic = TMPFS_MAGIC;
3625 sb->s_op = &shmem_ops;
3626 sb->s_time_gran = 1;
3627#ifdef CONFIG_TMPFS_XATTR
3628 sb->s_xattr = shmem_xattr_handlers;
3629#endif
3630#ifdef CONFIG_TMPFS_POSIX_ACL
3631 sb->s_flags |= MS_POSIXACL;
3632#endif
3633
3634 inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
3635 if (!inode)
3636 goto failed;
3637 inode->i_uid = sbinfo->uid;
3638 inode->i_gid = sbinfo->gid;
3639 sb->s_root = d_make_root(inode);
3640 if (!sb->s_root)
3641 goto failed;
3642 return 0;
3643
3644failed:
3645 shmem_put_super(sb);
3646 return err;
3647}
3648
3649static struct kmem_cache *shmem_inode_cachep;
3650
3651static struct inode *shmem_alloc_inode(struct super_block *sb)
3652{
3653 struct shmem_inode_info *info;
3654 info = kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
3655 if (!info)
3656 return NULL;
3657 return &info->vfs_inode;
3658}
3659
3660static void shmem_destroy_callback(struct rcu_head *head)
3661{
3662 struct inode *inode = container_of(head, struct inode, i_rcu);
3663 if (S_ISLNK(inode->i_mode))
3664 kfree(inode->i_link);
3665 kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
3666}
3667
3668static void shmem_destroy_inode(struct inode *inode)
3669{
3670 if (S_ISREG(inode->i_mode))
3671 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
3672 call_rcu(&inode->i_rcu, shmem_destroy_callback);
3673}
3674
3675static void shmem_init_inode(void *foo)
3676{
3677 struct shmem_inode_info *info = foo;
3678 inode_init_once(&info->vfs_inode);
3679}
3680
3681static int shmem_init_inodecache(void)
3682{
3683 shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
3684 sizeof(struct shmem_inode_info),
3685 0, SLAB_PANIC|SLAB_ACCOUNT, shmem_init_inode);
3686 return 0;
3687}
3688
3689static void shmem_destroy_inodecache(void)
3690{
3691 kmem_cache_destroy(shmem_inode_cachep);
3692}
3693
3694static const struct address_space_operations shmem_aops = {
3695 .writepage = shmem_writepage,
3696 .set_page_dirty = __set_page_dirty_no_writeback,
3697#ifdef CONFIG_TMPFS
3698 .write_begin = shmem_write_begin,
3699 .write_end = shmem_write_end,
3700#endif
3701#ifdef CONFIG_MIGRATION
3702 .migratepage = migrate_page,
3703#endif
3704 .error_remove_page = generic_error_remove_page,
3705};
3706
3707static const struct file_operations shmem_file_operations = {
3708 .mmap = shmem_mmap,
3709 .get_unmapped_area = shmem_get_unmapped_area,
3710#ifdef CONFIG_TMPFS
3711 .llseek = shmem_file_llseek,
3712 .read_iter = shmem_file_read_iter,
3713 .write_iter = generic_file_write_iter,
3714 .fsync = noop_fsync,
3715 .splice_read = generic_file_splice_read,
3716 .splice_write = iter_file_splice_write,
3717 .fallocate = shmem_fallocate,
3718#endif
3719};
3720
3721static const struct inode_operations shmem_inode_operations = {
3722 .getattr = shmem_getattr,
3723 .setattr = shmem_setattr,
3724#ifdef CONFIG_TMPFS_XATTR
3725 .listxattr = shmem_listxattr,
3726 .set_acl = simple_set_acl,
3727#endif
3728};
3729
3730static const struct inode_operations shmem_dir_inode_operations = {
3731#ifdef CONFIG_TMPFS
3732 .create = shmem_create,
3733 .lookup = simple_lookup,
3734 .link = shmem_link,
3735 .unlink = shmem_unlink,
3736 .symlink = shmem_symlink,
3737 .mkdir = shmem_mkdir,
3738 .rmdir = shmem_rmdir,
3739 .mknod = shmem_mknod,
3740 .rename = shmem_rename2,
3741 .tmpfile = shmem_tmpfile,
3742#endif
3743#ifdef CONFIG_TMPFS_XATTR
3744 .listxattr = shmem_listxattr,
3745#endif
3746#ifdef CONFIG_TMPFS_POSIX_ACL
3747 .setattr = shmem_setattr,
3748 .set_acl = simple_set_acl,
3749#endif
3750};
3751
3752static const struct inode_operations shmem_special_inode_operations = {
3753#ifdef CONFIG_TMPFS_XATTR
3754 .listxattr = shmem_listxattr,
3755#endif
3756#ifdef CONFIG_TMPFS_POSIX_ACL
3757 .setattr = shmem_setattr,
3758 .set_acl = simple_set_acl,
3759#endif
3760};
3761
3762static const struct super_operations shmem_ops = {
3763 .alloc_inode = shmem_alloc_inode,
3764 .destroy_inode = shmem_destroy_inode,
3765#ifdef CONFIG_TMPFS
3766 .statfs = shmem_statfs,
3767 .remount_fs = shmem_remount_fs,
3768 .show_options = shmem_show_options,
3769#endif
3770 .evict_inode = shmem_evict_inode,
3771 .drop_inode = generic_delete_inode,
3772 .put_super = shmem_put_super,
3773#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3774 .nr_cached_objects = shmem_unused_huge_count,
3775 .free_cached_objects = shmem_unused_huge_scan,
3776#endif
3777};
3778
3779static const struct vm_operations_struct shmem_vm_ops = {
3780 .fault = shmem_fault,
3781 .map_pages = filemap_map_pages,
3782#ifdef CONFIG_NUMA
3783 .set_policy = shmem_set_policy,
3784 .get_policy = shmem_get_policy,
3785#endif
3786};
3787
3788static struct dentry *shmem_mount(struct file_system_type *fs_type,
3789 int flags, const char *dev_name, void *data)
3790{
3791 return mount_nodev(fs_type, flags, data, shmem_fill_super);
3792}
3793
3794static struct file_system_type shmem_fs_type = {
3795 .owner = THIS_MODULE,
3796 .name = "tmpfs",
3797 .mount = shmem_mount,
3798 .kill_sb = kill_litter_super,
3799 .fs_flags = FS_USERNS_MOUNT,
3800};
3801
3802int __init shmem_init(void)
3803{
3804 int error;
3805
3806 /* If rootfs called this, don't re-init */
3807 if (shmem_inode_cachep)
3808 return 0;
3809
3810 error = shmem_init_inodecache();
3811 if (error)
3812 goto out3;
3813
3814 error = register_filesystem(&shmem_fs_type);
3815 if (error) {
3816 pr_err("Could not register tmpfs\n");
3817 goto out2;
3818 }
3819
3820 shm_mnt = kern_mount(&shmem_fs_type);
3821 if (IS_ERR(shm_mnt)) {
3822 error = PTR_ERR(shm_mnt);
3823 pr_err("Could not kern_mount tmpfs\n");
3824 goto out1;
3825 }
3826
3827#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3828 if (has_transparent_hugepage() && shmem_huge < SHMEM_HUGE_DENY)
3829 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
3830 else
3831 shmem_huge = 0; /* just in case it was patched */
3832#endif
3833 return 0;
3834
3835out1:
3836 unregister_filesystem(&shmem_fs_type);
3837out2:
3838 shmem_destroy_inodecache();
3839out3:
3840 shm_mnt = ERR_PTR(error);
3841 return error;
3842}
3843
3844#if defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE) && defined(CONFIG_SYSFS)
3845static ssize_t shmem_enabled_show(struct kobject *kobj,
3846 struct kobj_attribute *attr, char *buf)
3847{
3848 int values[] = {
3849 SHMEM_HUGE_ALWAYS,
3850 SHMEM_HUGE_WITHIN_SIZE,
3851 SHMEM_HUGE_ADVISE,
3852 SHMEM_HUGE_NEVER,
3853 SHMEM_HUGE_DENY,
3854 SHMEM_HUGE_FORCE,
3855 };
3856 int i, count;
3857
3858 for (i = 0, count = 0; i < ARRAY_SIZE(values); i++) {
3859 const char *fmt = shmem_huge == values[i] ? "[%s] " : "%s ";
3860
3861 count += sprintf(buf + count, fmt,
3862 shmem_format_huge(values[i]));
3863 }
3864 buf[count - 1] = '\n';
3865 return count;
3866}
3867
3868static ssize_t shmem_enabled_store(struct kobject *kobj,
3869 struct kobj_attribute *attr, const char *buf, size_t count)
3870{
3871 char tmp[16];
3872 int huge;
3873
3874 if (count + 1 > sizeof(tmp))
3875 return -EINVAL;
3876 memcpy(tmp, buf, count);
3877 tmp[count] = '\0';
3878 if (count && tmp[count - 1] == '\n')
3879 tmp[count - 1] = '\0';
3880
3881 huge = shmem_parse_huge(tmp);
3882 if (huge == -EINVAL)
3883 return -EINVAL;
3884 if (!has_transparent_hugepage() &&
3885 huge != SHMEM_HUGE_NEVER && huge != SHMEM_HUGE_DENY)
3886 return -EINVAL;
3887
3888 shmem_huge = huge;
3889 if (shmem_huge < SHMEM_HUGE_DENY)
3890 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
3891 return count;
3892}
3893
3894struct kobj_attribute shmem_enabled_attr =
3895 __ATTR(shmem_enabled, 0644, shmem_enabled_show, shmem_enabled_store);
3896#endif /* CONFIG_TRANSPARENT_HUGE_PAGECACHE && CONFIG_SYSFS */
3897
3898#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3899bool shmem_huge_enabled(struct vm_area_struct *vma)
3900{
3901 struct inode *inode = file_inode(vma->vm_file);
3902 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
3903 loff_t i_size;
3904 pgoff_t off;
3905
3906 if (shmem_huge == SHMEM_HUGE_FORCE)
3907 return true;
3908 if (shmem_huge == SHMEM_HUGE_DENY)
3909 return false;
3910 switch (sbinfo->huge) {
3911 case SHMEM_HUGE_NEVER:
3912 return false;
3913 case SHMEM_HUGE_ALWAYS:
3914 return true;
3915 case SHMEM_HUGE_WITHIN_SIZE:
3916 off = round_up(vma->vm_pgoff, HPAGE_PMD_NR);
3917 i_size = round_up(i_size_read(inode), PAGE_SIZE);
3918 if (i_size >= HPAGE_PMD_SIZE &&
3919 i_size >> PAGE_SHIFT >= off)
3920 return true;
3921 case SHMEM_HUGE_ADVISE:
3922 /* TODO: implement fadvise() hints */
3923 return (vma->vm_flags & VM_HUGEPAGE);
3924 default:
3925 VM_BUG_ON(1);
3926 return false;
3927 }
3928}
3929#endif /* CONFIG_TRANSPARENT_HUGE_PAGECACHE */
3930
3931#else /* !CONFIG_SHMEM */
3932
3933/*
3934 * tiny-shmem: simple shmemfs and tmpfs using ramfs code
3935 *
3936 * This is intended for small system where the benefits of the full
3937 * shmem code (swap-backed and resource-limited) are outweighed by
3938 * their complexity. On systems without swap this code should be
3939 * effectively equivalent, but much lighter weight.
3940 */
3941
3942static struct file_system_type shmem_fs_type = {
3943 .name = "tmpfs",
3944 .mount = ramfs_mount,
3945 .kill_sb = kill_litter_super,
3946 .fs_flags = FS_USERNS_MOUNT,
3947};
3948
3949int __init shmem_init(void)
3950{
3951 BUG_ON(register_filesystem(&shmem_fs_type) != 0);
3952
3953 shm_mnt = kern_mount(&shmem_fs_type);
3954 BUG_ON(IS_ERR(shm_mnt));
3955
3956 return 0;
3957}
3958
3959int shmem_unuse(swp_entry_t swap, struct page *page)
3960{
3961 return 0;
3962}
3963
3964int shmem_lock(struct file *file, int lock, struct user_struct *user)
3965{
3966 return 0;
3967}
3968
3969void shmem_unlock_mapping(struct address_space *mapping)
3970{
3971}
3972
3973#ifdef CONFIG_MMU
3974unsigned long shmem_get_unmapped_area(struct file *file,
3975 unsigned long addr, unsigned long len,
3976 unsigned long pgoff, unsigned long flags)
3977{
3978 return current->mm->get_unmapped_area(file, addr, len, pgoff, flags);
3979}
3980#endif
3981
3982void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
3983{
3984 truncate_inode_pages_range(inode->i_mapping, lstart, lend);
3985}
3986EXPORT_SYMBOL_GPL(shmem_truncate_range);
3987
3988#define shmem_vm_ops generic_file_vm_ops
3989#define shmem_file_operations ramfs_file_operations
3990#define shmem_get_inode(sb, dir, mode, dev, flags) ramfs_get_inode(sb, dir, mode, dev)
3991#define shmem_acct_size(flags, size) 0
3992#define shmem_unacct_size(flags, size) do {} while (0)
3993
3994#endif /* CONFIG_SHMEM */
3995
3996/* common code */
3997
3998static const struct dentry_operations anon_ops = {
3999 .d_dname = simple_dname
4000};
4001
4002static struct file *__shmem_file_setup(const char *name, loff_t size,
4003 unsigned long flags, unsigned int i_flags)
4004{
4005 struct file *res;
4006 struct inode *inode;
4007 struct path path;
4008 struct super_block *sb;
4009 struct qstr this;
4010
4011 if (IS_ERR(shm_mnt))
4012 return ERR_CAST(shm_mnt);
4013
4014 if (size < 0 || size > MAX_LFS_FILESIZE)
4015 return ERR_PTR(-EINVAL);
4016
4017 if (shmem_acct_size(flags, size))
4018 return ERR_PTR(-ENOMEM);
4019
4020 res = ERR_PTR(-ENOMEM);
4021 this.name = name;
4022 this.len = strlen(name);
4023 this.hash = 0; /* will go */
4024 sb = shm_mnt->mnt_sb;
4025 path.mnt = mntget(shm_mnt);
4026 path.dentry = d_alloc_pseudo(sb, &this);
4027 if (!path.dentry)
4028 goto put_memory;
4029 d_set_d_op(path.dentry, &anon_ops);
4030
4031 res = ERR_PTR(-ENOSPC);
4032 inode = shmem_get_inode(sb, NULL, S_IFREG | S_IRWXUGO, 0, flags);
4033 if (!inode)
4034 goto put_memory;
4035
4036 inode->i_flags |= i_flags;
4037 d_instantiate(path.dentry, inode);
4038 inode->i_size = size;
4039 clear_nlink(inode); /* It is unlinked */
4040 res = ERR_PTR(ramfs_nommu_expand_for_mapping(inode, size));
4041 if (IS_ERR(res))
4042 goto put_path;
4043
4044 res = alloc_file(&path, FMODE_WRITE | FMODE_READ,
4045 &shmem_file_operations);
4046 if (IS_ERR(res))
4047 goto put_path;
4048
4049 return res;
4050
4051put_memory:
4052 shmem_unacct_size(flags, size);
4053put_path:
4054 path_put(&path);
4055 return res;
4056}
4057
4058/**
4059 * shmem_kernel_file_setup - get an unlinked file living in tmpfs which must be
4060 * kernel internal. There will be NO LSM permission checks against the
4061 * underlying inode. So users of this interface must do LSM checks at a
4062 * higher layer. The users are the big_key and shm implementations. LSM
4063 * checks are provided at the key or shm level rather than the inode.
4064 * @name: name for dentry (to be seen in /proc/<pid>/maps
4065 * @size: size to be set for the file
4066 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4067 */
4068struct file *shmem_kernel_file_setup(const char *name, loff_t size, unsigned long flags)
4069{
4070 return __shmem_file_setup(name, size, flags, S_PRIVATE);
4071}
4072
4073/**
4074 * shmem_file_setup - get an unlinked file living in tmpfs
4075 * @name: name for dentry (to be seen in /proc/<pid>/maps
4076 * @size: size to be set for the file
4077 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4078 */
4079struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
4080{
4081 return __shmem_file_setup(name, size, flags, 0);
4082}
4083EXPORT_SYMBOL_GPL(shmem_file_setup);
4084
4085/**
4086 * shmem_zero_setup - setup a shared anonymous mapping
4087 * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
4088 */
4089int shmem_zero_setup(struct vm_area_struct *vma)
4090{
4091 struct file *file;
4092 loff_t size = vma->vm_end - vma->vm_start;
4093
4094 /*
4095 * Cloning a new file under mmap_sem leads to a lock ordering conflict
4096 * between XFS directory reading and selinux: since this file is only
4097 * accessible to the user through its mapping, use S_PRIVATE flag to
4098 * bypass file security, in the same way as shmem_kernel_file_setup().
4099 */
4100 file = __shmem_file_setup("dev/zero", size, vma->vm_flags, S_PRIVATE);
4101 if (IS_ERR(file))
4102 return PTR_ERR(file);
4103
4104 if (vma->vm_file)
4105 fput(vma->vm_file);
4106 vma->vm_file = file;
4107 vma->vm_ops = &shmem_vm_ops;
4108
4109 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE) &&
4110 ((vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK) <
4111 (vma->vm_end & HPAGE_PMD_MASK)) {
4112 khugepaged_enter(vma, vma->vm_flags);
4113 }
4114
4115 return 0;
4116}
4117
4118/**
4119 * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
4120 * @mapping: the page's address_space
4121 * @index: the page index
4122 * @gfp: the page allocator flags to use if allocating
4123 *
4124 * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
4125 * with any new page allocations done using the specified allocation flags.
4126 * But read_cache_page_gfp() uses the ->readpage() method: which does not
4127 * suit tmpfs, since it may have pages in swapcache, and needs to find those
4128 * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
4129 *
4130 * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
4131 * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
4132 */
4133struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
4134 pgoff_t index, gfp_t gfp)
4135{
4136#ifdef CONFIG_SHMEM
4137 struct inode *inode = mapping->host;
4138 struct page *page;
4139 int error;
4140
4141 BUG_ON(mapping->a_ops != &shmem_aops);
4142 error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE,
4143 gfp, NULL, NULL);
4144 if (error)
4145 page = ERR_PTR(error);
4146 else
4147 unlock_page(page);
4148 return page;
4149#else
4150 /*
4151 * The tiny !SHMEM case uses ramfs without swap
4152 */
4153 return read_cache_page_gfp(mapping, index, gfp);
4154#endif
4155}
4156EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);