2 * Resizable virtual memory filesystem for Linux.
4 * Copyright (C) 2000 Linus Torvalds.
6 * 2000-2001 Christoph Rohland
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
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>
19 * Copyright (c) 2004, 2008 Matt Mackall <mpm@selenic.com>
21 * This file is released under the GPL.
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>
32 #include <linux/random.h>
33 #include <linux/sched/signal.h>
34 #include <linux/export.h>
35 #include <linux/swap.h>
36 #include <linux/uio.h>
37 #include <linux/khugepaged.h>
38 #include <linux/hugetlb.h>
40 #include <asm/tlbflush.h> /* for arch/microblaze update_mmu_cache() */
42 static struct vfsmount *shm_mnt;
46 * This virtual memory filesystem is heavily based on the ramfs. It
47 * extends ramfs by the ability to use swap and honor resource limits
48 * which makes it a completely usable filesystem.
51 #include <linux/xattr.h>
52 #include <linux/exportfs.h>
53 #include <linux/posix_acl.h>
54 #include <linux/posix_acl_xattr.h>
55 #include <linux/mman.h>
56 #include <linux/string.h>
57 #include <linux/slab.h>
58 #include <linux/backing-dev.h>
59 #include <linux/shmem_fs.h>
60 #include <linux/writeback.h>
61 #include <linux/blkdev.h>
62 #include <linux/pagevec.h>
63 #include <linux/percpu_counter.h>
64 #include <linux/falloc.h>
65 #include <linux/splice.h>
66 #include <linux/security.h>
67 #include <linux/swapops.h>
68 #include <linux/mempolicy.h>
69 #include <linux/namei.h>
70 #include <linux/ctype.h>
71 #include <linux/migrate.h>
72 #include <linux/highmem.h>
73 #include <linux/seq_file.h>
74 #include <linux/magic.h>
75 #include <linux/syscalls.h>
76 #include <linux/fcntl.h>
77 #include <uapi/linux/memfd.h>
78 #include <linux/userfaultfd_k.h>
79 #include <linux/rmap.h>
80 #include <linux/uuid.h>
82 #include <linux/uaccess.h>
83 #include <asm/pgtable.h>
87 #define BLOCKS_PER_PAGE (PAGE_SIZE/512)
88 #define VM_ACCT(size) (PAGE_ALIGN(size) >> PAGE_SHIFT)
90 /* Pretend that each entry is of this size in directory's i_size */
91 #define BOGO_DIRENT_SIZE 20
93 /* Symlink up to this size is kmalloc'ed instead of using a swappable page */
94 #define SHORT_SYMLINK_LEN 128
97 * shmem_fallocate communicates with shmem_fault or shmem_writepage via
98 * inode->i_private (with i_mutex making sure that it has only one user at
99 * a time): we would prefer not to enlarge the shmem inode just for that.
101 struct shmem_falloc {
102 wait_queue_head_t *waitq; /* faults into hole wait for punch to end */
103 pgoff_t start; /* start of range currently being fallocated */
104 pgoff_t next; /* the next page offset to be fallocated */
105 pgoff_t nr_falloced; /* how many new pages have been fallocated */
106 pgoff_t nr_unswapped; /* how often writepage refused to swap out */
110 static unsigned long shmem_default_max_blocks(void)
112 return totalram_pages / 2;
115 static unsigned long shmem_default_max_inodes(void)
117 return min(totalram_pages - totalhigh_pages, totalram_pages / 2);
121 static bool shmem_should_replace_page(struct page *page, gfp_t gfp);
122 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
123 struct shmem_inode_info *info, pgoff_t index);
124 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
125 struct page **pagep, enum sgp_type sgp,
126 gfp_t gfp, struct vm_area_struct *vma,
127 struct vm_fault *vmf, vm_fault_t *fault_type);
129 int shmem_getpage(struct inode *inode, pgoff_t index,
130 struct page **pagep, enum sgp_type sgp)
132 return shmem_getpage_gfp(inode, index, pagep, sgp,
133 mapping_gfp_mask(inode->i_mapping), NULL, NULL, NULL);
136 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
138 return sb->s_fs_info;
142 * shmem_file_setup pre-accounts the whole fixed size of a VM object,
143 * for shared memory and for shared anonymous (/dev/zero) mappings
144 * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
145 * consistent with the pre-accounting of private mappings ...
147 static inline int shmem_acct_size(unsigned long flags, loff_t size)
149 return (flags & VM_NORESERVE) ?
150 0 : security_vm_enough_memory_mm(current->mm, VM_ACCT(size));
153 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
155 if (!(flags & VM_NORESERVE))
156 vm_unacct_memory(VM_ACCT(size));
159 static inline int shmem_reacct_size(unsigned long flags,
160 loff_t oldsize, loff_t newsize)
162 if (!(flags & VM_NORESERVE)) {
163 if (VM_ACCT(newsize) > VM_ACCT(oldsize))
164 return security_vm_enough_memory_mm(current->mm,
165 VM_ACCT(newsize) - VM_ACCT(oldsize));
166 else if (VM_ACCT(newsize) < VM_ACCT(oldsize))
167 vm_unacct_memory(VM_ACCT(oldsize) - VM_ACCT(newsize));
173 * ... whereas tmpfs objects are accounted incrementally as
174 * pages are allocated, in order to allow large sparse files.
175 * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
176 * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
178 static inline int shmem_acct_block(unsigned long flags, long pages)
180 if (!(flags & VM_NORESERVE))
183 return security_vm_enough_memory_mm(current->mm,
184 pages * VM_ACCT(PAGE_SIZE));
187 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
189 if (flags & VM_NORESERVE)
190 vm_unacct_memory(pages * VM_ACCT(PAGE_SIZE));
193 static inline bool shmem_inode_acct_block(struct inode *inode, long pages)
195 struct shmem_inode_info *info = SHMEM_I(inode);
196 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
198 if (shmem_acct_block(info->flags, pages))
201 if (sbinfo->max_blocks) {
202 if (percpu_counter_compare(&sbinfo->used_blocks,
203 sbinfo->max_blocks - pages) > 0)
205 percpu_counter_add(&sbinfo->used_blocks, pages);
211 shmem_unacct_blocks(info->flags, pages);
215 static inline void shmem_inode_unacct_blocks(struct inode *inode, long pages)
217 struct shmem_inode_info *info = SHMEM_I(inode);
218 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
220 if (sbinfo->max_blocks)
221 percpu_counter_sub(&sbinfo->used_blocks, pages);
222 shmem_unacct_blocks(info->flags, pages);
225 static const struct super_operations shmem_ops;
226 static const struct address_space_operations shmem_aops;
227 static const struct file_operations shmem_file_operations;
228 static const struct inode_operations shmem_inode_operations;
229 static const struct inode_operations shmem_dir_inode_operations;
230 static const struct inode_operations shmem_special_inode_operations;
231 static const struct vm_operations_struct shmem_vm_ops;
232 static struct file_system_type shmem_fs_type;
234 bool vma_is_shmem(struct vm_area_struct *vma)
236 return vma->vm_ops == &shmem_vm_ops;
239 static LIST_HEAD(shmem_swaplist);
240 static DEFINE_MUTEX(shmem_swaplist_mutex);
242 static int shmem_reserve_inode(struct super_block *sb)
244 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
245 if (sbinfo->max_inodes) {
246 spin_lock(&sbinfo->stat_lock);
247 if (!sbinfo->free_inodes) {
248 spin_unlock(&sbinfo->stat_lock);
251 sbinfo->free_inodes--;
252 spin_unlock(&sbinfo->stat_lock);
257 static void shmem_free_inode(struct super_block *sb)
259 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
260 if (sbinfo->max_inodes) {
261 spin_lock(&sbinfo->stat_lock);
262 sbinfo->free_inodes++;
263 spin_unlock(&sbinfo->stat_lock);
268 * shmem_recalc_inode - recalculate the block usage of an inode
269 * @inode: inode to recalc
271 * We have to calculate the free blocks since the mm can drop
272 * undirtied hole pages behind our back.
274 * But normally info->alloced == inode->i_mapping->nrpages + info->swapped
275 * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
277 * It has to be called with the spinlock held.
279 static void shmem_recalc_inode(struct inode *inode)
281 struct shmem_inode_info *info = SHMEM_I(inode);
284 freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
286 info->alloced -= freed;
287 inode->i_blocks -= freed * BLOCKS_PER_PAGE;
288 shmem_inode_unacct_blocks(inode, freed);
292 bool shmem_charge(struct inode *inode, long pages)
294 struct shmem_inode_info *info = SHMEM_I(inode);
297 if (!shmem_inode_acct_block(inode, pages))
300 /* nrpages adjustment first, then shmem_recalc_inode() when balanced */
301 inode->i_mapping->nrpages += pages;
303 spin_lock_irqsave(&info->lock, flags);
304 info->alloced += pages;
305 inode->i_blocks += pages * BLOCKS_PER_PAGE;
306 shmem_recalc_inode(inode);
307 spin_unlock_irqrestore(&info->lock, flags);
312 void shmem_uncharge(struct inode *inode, long pages)
314 struct shmem_inode_info *info = SHMEM_I(inode);
317 /* nrpages adjustment done by __delete_from_page_cache() or caller */
319 spin_lock_irqsave(&info->lock, flags);
320 info->alloced -= pages;
321 inode->i_blocks -= pages * BLOCKS_PER_PAGE;
322 shmem_recalc_inode(inode);
323 spin_unlock_irqrestore(&info->lock, flags);
325 shmem_inode_unacct_blocks(inode, pages);
329 * Replace item expected in radix tree by a new item, while holding tree lock.
331 static int shmem_radix_tree_replace(struct address_space *mapping,
332 pgoff_t index, void *expected, void *replacement)
334 struct radix_tree_node *node;
338 VM_BUG_ON(!expected);
339 VM_BUG_ON(!replacement);
340 item = __radix_tree_lookup(&mapping->i_pages, index, &node, &pslot);
343 if (item != expected)
345 __radix_tree_replace(&mapping->i_pages, node, pslot,
351 * Sometimes, before we decide whether to proceed or to fail, we must check
352 * that an entry was not already brought back from swap by a racing thread.
354 * Checking page is not enough: by the time a SwapCache page is locked, it
355 * might be reused, and again be SwapCache, using the same swap as before.
357 static bool shmem_confirm_swap(struct address_space *mapping,
358 pgoff_t index, swp_entry_t swap)
363 item = radix_tree_lookup(&mapping->i_pages, index);
365 return item == swp_to_radix_entry(swap);
369 * Definitions for "huge tmpfs": tmpfs mounted with the huge= option
372 * disables huge pages for the mount;
374 * enables huge pages for the mount;
375 * SHMEM_HUGE_WITHIN_SIZE:
376 * only allocate huge pages if the page will be fully within i_size,
377 * also respect fadvise()/madvise() hints;
379 * only allocate huge pages if requested with fadvise()/madvise();
382 #define SHMEM_HUGE_NEVER 0
383 #define SHMEM_HUGE_ALWAYS 1
384 #define SHMEM_HUGE_WITHIN_SIZE 2
385 #define SHMEM_HUGE_ADVISE 3
389 * Only can be set via /sys/kernel/mm/transparent_hugepage/shmem_enabled:
392 * disables huge on shm_mnt and all mounts, for emergency use;
394 * enables huge on shm_mnt and all mounts, w/o needing option, for testing;
397 #define SHMEM_HUGE_DENY (-1)
398 #define SHMEM_HUGE_FORCE (-2)
400 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
401 /* ifdef here to avoid bloating shmem.o when not necessary */
403 static int shmem_huge __read_mostly;
405 #if defined(CONFIG_SYSFS) || defined(CONFIG_TMPFS)
406 static int shmem_parse_huge(const char *str)
408 if (!strcmp(str, "never"))
409 return SHMEM_HUGE_NEVER;
410 if (!strcmp(str, "always"))
411 return SHMEM_HUGE_ALWAYS;
412 if (!strcmp(str, "within_size"))
413 return SHMEM_HUGE_WITHIN_SIZE;
414 if (!strcmp(str, "advise"))
415 return SHMEM_HUGE_ADVISE;
416 if (!strcmp(str, "deny"))
417 return SHMEM_HUGE_DENY;
418 if (!strcmp(str, "force"))
419 return SHMEM_HUGE_FORCE;
423 static const char *shmem_format_huge(int huge)
426 case SHMEM_HUGE_NEVER:
428 case SHMEM_HUGE_ALWAYS:
430 case SHMEM_HUGE_WITHIN_SIZE:
431 return "within_size";
432 case SHMEM_HUGE_ADVISE:
434 case SHMEM_HUGE_DENY:
436 case SHMEM_HUGE_FORCE:
445 static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
446 struct shrink_control *sc, unsigned long nr_to_split)
448 LIST_HEAD(list), *pos, *next;
449 LIST_HEAD(to_remove);
451 struct shmem_inode_info *info;
453 unsigned long batch = sc ? sc->nr_to_scan : 128;
454 int removed = 0, split = 0;
456 if (list_empty(&sbinfo->shrinklist))
459 spin_lock(&sbinfo->shrinklist_lock);
460 list_for_each_safe(pos, next, &sbinfo->shrinklist) {
461 info = list_entry(pos, struct shmem_inode_info, shrinklist);
464 inode = igrab(&info->vfs_inode);
466 /* inode is about to be evicted */
468 list_del_init(&info->shrinklist);
473 /* Check if there's anything to gain */
474 if (round_up(inode->i_size, PAGE_SIZE) ==
475 round_up(inode->i_size, HPAGE_PMD_SIZE)) {
476 list_move(&info->shrinklist, &to_remove);
481 list_move(&info->shrinklist, &list);
486 spin_unlock(&sbinfo->shrinklist_lock);
488 list_for_each_safe(pos, next, &to_remove) {
489 info = list_entry(pos, struct shmem_inode_info, shrinklist);
490 inode = &info->vfs_inode;
491 list_del_init(&info->shrinklist);
495 list_for_each_safe(pos, next, &list) {
498 info = list_entry(pos, struct shmem_inode_info, shrinklist);
499 inode = &info->vfs_inode;
501 if (nr_to_split && split >= nr_to_split)
504 page = find_get_page(inode->i_mapping,
505 (inode->i_size & HPAGE_PMD_MASK) >> PAGE_SHIFT);
509 /* No huge page at the end of the file: nothing to split */
510 if (!PageTransHuge(page)) {
516 * Leave the inode on the list if we failed to lock
517 * the page at this time.
519 * Waiting for the lock may lead to deadlock in the
522 if (!trylock_page(page)) {
527 ret = split_huge_page(page);
531 /* If split failed leave the inode on the list */
537 list_del_init(&info->shrinklist);
543 spin_lock(&sbinfo->shrinklist_lock);
544 list_splice_tail(&list, &sbinfo->shrinklist);
545 sbinfo->shrinklist_len -= removed;
546 spin_unlock(&sbinfo->shrinklist_lock);
551 static long shmem_unused_huge_scan(struct super_block *sb,
552 struct shrink_control *sc)
554 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
556 if (!READ_ONCE(sbinfo->shrinklist_len))
559 return shmem_unused_huge_shrink(sbinfo, sc, 0);
562 static long shmem_unused_huge_count(struct super_block *sb,
563 struct shrink_control *sc)
565 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
566 return READ_ONCE(sbinfo->shrinklist_len);
568 #else /* !CONFIG_TRANSPARENT_HUGE_PAGECACHE */
570 #define shmem_huge SHMEM_HUGE_DENY
572 static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
573 struct shrink_control *sc, unsigned long nr_to_split)
577 #endif /* CONFIG_TRANSPARENT_HUGE_PAGECACHE */
579 static inline bool is_huge_enabled(struct shmem_sb_info *sbinfo)
581 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE) &&
582 (shmem_huge == SHMEM_HUGE_FORCE || sbinfo->huge) &&
583 shmem_huge != SHMEM_HUGE_DENY)
589 * Like add_to_page_cache_locked, but error if expected item has gone.
591 static int shmem_add_to_page_cache(struct page *page,
592 struct address_space *mapping,
593 pgoff_t index, void *expected)
595 int error, nr = hpage_nr_pages(page);
597 VM_BUG_ON_PAGE(PageTail(page), page);
598 VM_BUG_ON_PAGE(index != round_down(index, nr), page);
599 VM_BUG_ON_PAGE(!PageLocked(page), page);
600 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
601 VM_BUG_ON(expected && PageTransHuge(page));
603 page_ref_add(page, nr);
604 page->mapping = mapping;
607 xa_lock_irq(&mapping->i_pages);
608 if (PageTransHuge(page)) {
609 void __rcu **results;
614 if (radix_tree_gang_lookup_slot(&mapping->i_pages,
615 &results, &idx, index, 1) &&
616 idx < index + HPAGE_PMD_NR) {
621 for (i = 0; i < HPAGE_PMD_NR; i++) {
622 error = radix_tree_insert(&mapping->i_pages,
623 index + i, page + i);
626 count_vm_event(THP_FILE_ALLOC);
628 } else if (!expected) {
629 error = radix_tree_insert(&mapping->i_pages, index, page);
631 error = shmem_radix_tree_replace(mapping, index, expected,
636 mapping->nrpages += nr;
637 if (PageTransHuge(page))
638 __inc_node_page_state(page, NR_SHMEM_THPS);
639 __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, nr);
640 __mod_node_page_state(page_pgdat(page), NR_SHMEM, nr);
641 xa_unlock_irq(&mapping->i_pages);
643 page->mapping = NULL;
644 xa_unlock_irq(&mapping->i_pages);
645 page_ref_sub(page, nr);
651 * Like delete_from_page_cache, but substitutes swap for page.
653 static void shmem_delete_from_page_cache(struct page *page, void *radswap)
655 struct address_space *mapping = page->mapping;
658 VM_BUG_ON_PAGE(PageCompound(page), page);
660 xa_lock_irq(&mapping->i_pages);
661 error = shmem_radix_tree_replace(mapping, page->index, page, radswap);
662 page->mapping = NULL;
664 __dec_node_page_state(page, NR_FILE_PAGES);
665 __dec_node_page_state(page, NR_SHMEM);
666 xa_unlock_irq(&mapping->i_pages);
672 * Remove swap entry from radix tree, free the swap and its page cache.
674 static int shmem_free_swap(struct address_space *mapping,
675 pgoff_t index, void *radswap)
679 xa_lock_irq(&mapping->i_pages);
680 old = radix_tree_delete_item(&mapping->i_pages, index, radswap);
681 xa_unlock_irq(&mapping->i_pages);
684 free_swap_and_cache(radix_to_swp_entry(radswap));
689 * Determine (in bytes) how many of the shmem object's pages mapped by the
690 * given offsets are swapped out.
692 * This is safe to call without i_mutex or the i_pages lock thanks to RCU,
693 * as long as the inode doesn't go away and racy results are not a problem.
695 unsigned long shmem_partial_swap_usage(struct address_space *mapping,
696 pgoff_t start, pgoff_t end)
698 struct radix_tree_iter iter;
701 unsigned long swapped = 0;
705 radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start) {
706 if (iter.index >= end)
709 page = radix_tree_deref_slot(slot);
711 if (radix_tree_deref_retry(page)) {
712 slot = radix_tree_iter_retry(&iter);
716 if (radix_tree_exceptional_entry(page))
719 if (need_resched()) {
720 slot = radix_tree_iter_resume(slot, &iter);
727 return swapped << PAGE_SHIFT;
731 * Determine (in bytes) how many of the shmem object's pages mapped by the
732 * given vma is swapped out.
734 * This is safe to call without i_mutex or the i_pages lock thanks to RCU,
735 * as long as the inode doesn't go away and racy results are not a problem.
737 unsigned long shmem_swap_usage(struct vm_area_struct *vma)
739 struct inode *inode = file_inode(vma->vm_file);
740 struct shmem_inode_info *info = SHMEM_I(inode);
741 struct address_space *mapping = inode->i_mapping;
742 unsigned long swapped;
744 /* Be careful as we don't hold info->lock */
745 swapped = READ_ONCE(info->swapped);
748 * The easier cases are when the shmem object has nothing in swap, or
749 * the vma maps it whole. Then we can simply use the stats that we
755 if (!vma->vm_pgoff && vma->vm_end - vma->vm_start >= inode->i_size)
756 return swapped << PAGE_SHIFT;
758 /* Here comes the more involved part */
759 return shmem_partial_swap_usage(mapping,
760 linear_page_index(vma, vma->vm_start),
761 linear_page_index(vma, vma->vm_end));
765 * SysV IPC SHM_UNLOCK restore Unevictable pages to their evictable lists.
767 void shmem_unlock_mapping(struct address_space *mapping)
770 pgoff_t indices[PAGEVEC_SIZE];
775 * Minor point, but we might as well stop if someone else SHM_LOCKs it.
777 while (!mapping_unevictable(mapping)) {
779 * Avoid pagevec_lookup(): find_get_pages() returns 0 as if it
780 * has finished, if it hits a row of PAGEVEC_SIZE swap entries.
782 pvec.nr = find_get_entries(mapping, index,
783 PAGEVEC_SIZE, pvec.pages, indices);
786 index = indices[pvec.nr - 1] + 1;
787 pagevec_remove_exceptionals(&pvec);
788 check_move_unevictable_pages(pvec.pages, pvec.nr);
789 pagevec_release(&pvec);
795 * Remove range of pages and swap entries from radix tree, and free them.
796 * If !unfalloc, truncate or punch hole; if unfalloc, undo failed fallocate.
798 static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
801 struct address_space *mapping = inode->i_mapping;
802 struct shmem_inode_info *info = SHMEM_I(inode);
803 pgoff_t start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
804 pgoff_t end = (lend + 1) >> PAGE_SHIFT;
805 unsigned int partial_start = lstart & (PAGE_SIZE - 1);
806 unsigned int partial_end = (lend + 1) & (PAGE_SIZE - 1);
808 pgoff_t indices[PAGEVEC_SIZE];
809 long nr_swaps_freed = 0;
814 end = -1; /* unsigned, so actually very big */
818 while (index < end) {
819 pvec.nr = find_get_entries(mapping, index,
820 min(end - index, (pgoff_t)PAGEVEC_SIZE),
821 pvec.pages, indices);
824 for (i = 0; i < pagevec_count(&pvec); i++) {
825 struct page *page = pvec.pages[i];
831 if (radix_tree_exceptional_entry(page)) {
834 nr_swaps_freed += !shmem_free_swap(mapping,
839 VM_BUG_ON_PAGE(page_to_pgoff(page) != index, page);
841 if (!trylock_page(page))
844 if (PageTransTail(page)) {
845 /* Middle of THP: zero out the page */
846 clear_highpage(page);
849 } else if (PageTransHuge(page)) {
850 if (index == round_down(end, HPAGE_PMD_NR)) {
852 * Range ends in the middle of THP:
855 clear_highpage(page);
859 index += HPAGE_PMD_NR - 1;
860 i += HPAGE_PMD_NR - 1;
863 if (!unfalloc || !PageUptodate(page)) {
864 VM_BUG_ON_PAGE(PageTail(page), page);
865 if (page_mapping(page) == mapping) {
866 VM_BUG_ON_PAGE(PageWriteback(page), page);
867 truncate_inode_page(mapping, page);
872 pagevec_remove_exceptionals(&pvec);
873 pagevec_release(&pvec);
879 struct page *page = NULL;
880 shmem_getpage(inode, start - 1, &page, SGP_READ);
882 unsigned int top = PAGE_SIZE;
887 zero_user_segment(page, partial_start, top);
888 set_page_dirty(page);
894 struct page *page = NULL;
895 shmem_getpage(inode, end, &page, SGP_READ);
897 zero_user_segment(page, 0, partial_end);
898 set_page_dirty(page);
907 while (index < end) {
910 pvec.nr = find_get_entries(mapping, index,
911 min(end - index, (pgoff_t)PAGEVEC_SIZE),
912 pvec.pages, indices);
914 /* If all gone or hole-punch or unfalloc, we're done */
915 if (index == start || end != -1)
917 /* But if truncating, restart to make sure all gone */
921 for (i = 0; i < pagevec_count(&pvec); i++) {
922 struct page *page = pvec.pages[i];
928 if (radix_tree_exceptional_entry(page)) {
931 if (shmem_free_swap(mapping, index, page)) {
932 /* Swap was replaced by page: retry */
942 if (PageTransTail(page)) {
943 /* Middle of THP: zero out the page */
944 clear_highpage(page);
947 * Partial thp truncate due 'start' in middle
948 * of THP: don't need to look on these pages
949 * again on !pvec.nr restart.
951 if (index != round_down(end, HPAGE_PMD_NR))
954 } else if (PageTransHuge(page)) {
955 if (index == round_down(end, HPAGE_PMD_NR)) {
957 * Range ends in the middle of THP:
960 clear_highpage(page);
964 index += HPAGE_PMD_NR - 1;
965 i += HPAGE_PMD_NR - 1;
968 if (!unfalloc || !PageUptodate(page)) {
969 VM_BUG_ON_PAGE(PageTail(page), page);
970 if (page_mapping(page) == mapping) {
971 VM_BUG_ON_PAGE(PageWriteback(page), page);
972 truncate_inode_page(mapping, page);
974 /* Page was replaced by swap: retry */
982 pagevec_remove_exceptionals(&pvec);
983 pagevec_release(&pvec);
987 spin_lock_irq(&info->lock);
988 info->swapped -= nr_swaps_freed;
989 shmem_recalc_inode(inode);
990 spin_unlock_irq(&info->lock);
993 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
995 shmem_undo_range(inode, lstart, lend, false);
996 inode->i_ctime = inode->i_mtime = current_time(inode);
998 EXPORT_SYMBOL_GPL(shmem_truncate_range);
1000 static int shmem_getattr(const struct path *path, struct kstat *stat,
1001 u32 request_mask, unsigned int query_flags)
1003 struct inode *inode = path->dentry->d_inode;
1004 struct shmem_inode_info *info = SHMEM_I(inode);
1005 struct shmem_sb_info *sb_info = SHMEM_SB(inode->i_sb);
1007 if (info->alloced - info->swapped != inode->i_mapping->nrpages) {
1008 spin_lock_irq(&info->lock);
1009 shmem_recalc_inode(inode);
1010 spin_unlock_irq(&info->lock);
1012 generic_fillattr(inode, stat);
1014 if (is_huge_enabled(sb_info))
1015 stat->blksize = HPAGE_PMD_SIZE;
1020 static int shmem_setattr(struct dentry *dentry, struct iattr *attr)
1022 struct inode *inode = d_inode(dentry);
1023 struct shmem_inode_info *info = SHMEM_I(inode);
1024 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1027 error = setattr_prepare(dentry, attr);
1031 if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
1032 loff_t oldsize = inode->i_size;
1033 loff_t newsize = attr->ia_size;
1035 /* protected by i_mutex */
1036 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
1037 (newsize > oldsize && (info->seals & F_SEAL_GROW)))
1040 if (newsize != oldsize) {
1041 error = shmem_reacct_size(SHMEM_I(inode)->flags,
1045 i_size_write(inode, newsize);
1046 inode->i_ctime = inode->i_mtime = current_time(inode);
1048 if (newsize <= oldsize) {
1049 loff_t holebegin = round_up(newsize, PAGE_SIZE);
1050 if (oldsize > holebegin)
1051 unmap_mapping_range(inode->i_mapping,
1054 shmem_truncate_range(inode,
1055 newsize, (loff_t)-1);
1056 /* unmap again to remove racily COWed private pages */
1057 if (oldsize > holebegin)
1058 unmap_mapping_range(inode->i_mapping,
1062 * Part of the huge page can be beyond i_size: subject
1063 * to shrink under memory pressure.
1065 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
1066 spin_lock(&sbinfo->shrinklist_lock);
1068 * _careful to defend against unlocked access to
1069 * ->shrink_list in shmem_unused_huge_shrink()
1071 if (list_empty_careful(&info->shrinklist)) {
1072 list_add_tail(&info->shrinklist,
1073 &sbinfo->shrinklist);
1074 sbinfo->shrinklist_len++;
1076 spin_unlock(&sbinfo->shrinklist_lock);
1081 setattr_copy(inode, attr);
1082 if (attr->ia_valid & ATTR_MODE)
1083 error = posix_acl_chmod(inode, inode->i_mode);
1087 static void shmem_evict_inode(struct inode *inode)
1089 struct shmem_inode_info *info = SHMEM_I(inode);
1090 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1092 if (inode->i_mapping->a_ops == &shmem_aops) {
1093 shmem_unacct_size(info->flags, inode->i_size);
1095 shmem_truncate_range(inode, 0, (loff_t)-1);
1096 if (!list_empty(&info->shrinklist)) {
1097 spin_lock(&sbinfo->shrinklist_lock);
1098 if (!list_empty(&info->shrinklist)) {
1099 list_del_init(&info->shrinklist);
1100 sbinfo->shrinklist_len--;
1102 spin_unlock(&sbinfo->shrinklist_lock);
1104 if (!list_empty(&info->swaplist)) {
1105 mutex_lock(&shmem_swaplist_mutex);
1106 list_del_init(&info->swaplist);
1107 mutex_unlock(&shmem_swaplist_mutex);
1111 simple_xattrs_free(&info->xattrs);
1112 WARN_ON(inode->i_blocks);
1113 shmem_free_inode(inode->i_sb);
1117 static unsigned long find_swap_entry(struct radix_tree_root *root, void *item)
1119 struct radix_tree_iter iter;
1121 unsigned long found = -1;
1122 unsigned int checked = 0;
1125 radix_tree_for_each_slot(slot, root, &iter, 0) {
1126 void *entry = radix_tree_deref_slot(slot);
1128 if (radix_tree_deref_retry(entry)) {
1129 slot = radix_tree_iter_retry(&iter);
1132 if (entry == item) {
1137 if ((checked % 4096) != 0)
1139 slot = radix_tree_iter_resume(slot, &iter);
1148 * If swap found in inode, free it and move page from swapcache to filecache.
1150 static int shmem_unuse_inode(struct shmem_inode_info *info,
1151 swp_entry_t swap, struct page **pagep)
1153 struct address_space *mapping = info->vfs_inode.i_mapping;
1159 radswap = swp_to_radix_entry(swap);
1160 index = find_swap_entry(&mapping->i_pages, radswap);
1162 return -EAGAIN; /* tell shmem_unuse we found nothing */
1165 * Move _head_ to start search for next from here.
1166 * But be careful: shmem_evict_inode checks list_empty without taking
1167 * mutex, and there's an instant in list_move_tail when info->swaplist
1168 * would appear empty, if it were the only one on shmem_swaplist.
1170 if (shmem_swaplist.next != &info->swaplist)
1171 list_move_tail(&shmem_swaplist, &info->swaplist);
1173 gfp = mapping_gfp_mask(mapping);
1174 if (shmem_should_replace_page(*pagep, gfp)) {
1175 mutex_unlock(&shmem_swaplist_mutex);
1176 error = shmem_replace_page(pagep, gfp, info, index);
1177 mutex_lock(&shmem_swaplist_mutex);
1179 * We needed to drop mutex to make that restrictive page
1180 * allocation, but the inode might have been freed while we
1181 * dropped it: although a racing shmem_evict_inode() cannot
1182 * complete without emptying the radix_tree, our page lock
1183 * on this swapcache page is not enough to prevent that -
1184 * free_swap_and_cache() of our swap entry will only
1185 * trylock_page(), removing swap from radix_tree whatever.
1187 * We must not proceed to shmem_add_to_page_cache() if the
1188 * inode has been freed, but of course we cannot rely on
1189 * inode or mapping or info to check that. However, we can
1190 * safely check if our swap entry is still in use (and here
1191 * it can't have got reused for another page): if it's still
1192 * in use, then the inode cannot have been freed yet, and we
1193 * can safely proceed (if it's no longer in use, that tells
1194 * nothing about the inode, but we don't need to unuse swap).
1196 if (!page_swapcount(*pagep))
1201 * We rely on shmem_swaplist_mutex, not only to protect the swaplist,
1202 * but also to hold up shmem_evict_inode(): so inode cannot be freed
1203 * beneath us (pagelock doesn't help until the page is in pagecache).
1206 error = shmem_add_to_page_cache(*pagep, mapping, index,
1208 if (error != -ENOMEM) {
1210 * Truncation and eviction use free_swap_and_cache(), which
1211 * only does trylock page: if we raced, best clean up here.
1213 delete_from_swap_cache(*pagep);
1214 set_page_dirty(*pagep);
1216 spin_lock_irq(&info->lock);
1218 spin_unlock_irq(&info->lock);
1226 * Search through swapped inodes to find and replace swap by page.
1228 int shmem_unuse(swp_entry_t swap, struct page *page)
1230 struct list_head *this, *next;
1231 struct shmem_inode_info *info;
1232 struct mem_cgroup *memcg;
1236 * There's a faint possibility that swap page was replaced before
1237 * caller locked it: caller will come back later with the right page.
1239 if (unlikely(!PageSwapCache(page) || page_private(page) != swap.val))
1243 * Charge page using GFP_KERNEL while we can wait, before taking
1244 * the shmem_swaplist_mutex which might hold up shmem_writepage().
1245 * Charged back to the user (not to caller) when swap account is used.
1247 error = mem_cgroup_try_charge_delay(page, current->mm, GFP_KERNEL,
1251 /* No radix_tree_preload: swap entry keeps a place for page in tree */
1254 mutex_lock(&shmem_swaplist_mutex);
1255 list_for_each_safe(this, next, &shmem_swaplist) {
1256 info = list_entry(this, struct shmem_inode_info, swaplist);
1258 error = shmem_unuse_inode(info, swap, &page);
1260 list_del_init(&info->swaplist);
1262 if (error != -EAGAIN)
1264 /* found nothing in this: move on to search the next */
1266 mutex_unlock(&shmem_swaplist_mutex);
1269 if (error != -ENOMEM)
1271 mem_cgroup_cancel_charge(page, memcg, false);
1273 mem_cgroup_commit_charge(page, memcg, true, false);
1281 * Move the page from the page cache to the swap cache.
1283 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1285 struct shmem_inode_info *info;
1286 struct address_space *mapping;
1287 struct inode *inode;
1291 VM_BUG_ON_PAGE(PageCompound(page), page);
1292 BUG_ON(!PageLocked(page));
1293 mapping = page->mapping;
1294 index = page->index;
1295 inode = mapping->host;
1296 info = SHMEM_I(inode);
1297 if (info->flags & VM_LOCKED)
1299 if (!total_swap_pages)
1303 * Our capabilities prevent regular writeback or sync from ever calling
1304 * shmem_writepage; but a stacking filesystem might use ->writepage of
1305 * its underlying filesystem, in which case tmpfs should write out to
1306 * swap only in response to memory pressure, and not for the writeback
1309 if (!wbc->for_reclaim) {
1310 WARN_ON_ONCE(1); /* Still happens? Tell us about it! */
1315 * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
1316 * value into swapfile.c, the only way we can correctly account for a
1317 * fallocated page arriving here is now to initialize it and write it.
1319 * That's okay for a page already fallocated earlier, but if we have
1320 * not yet completed the fallocation, then (a) we want to keep track
1321 * of this page in case we have to undo it, and (b) it may not be a
1322 * good idea to continue anyway, once we're pushing into swap. So
1323 * reactivate the page, and let shmem_fallocate() quit when too many.
1325 if (!PageUptodate(page)) {
1326 if (inode->i_private) {
1327 struct shmem_falloc *shmem_falloc;
1328 spin_lock(&inode->i_lock);
1329 shmem_falloc = inode->i_private;
1331 !shmem_falloc->waitq &&
1332 index >= shmem_falloc->start &&
1333 index < shmem_falloc->next)
1334 shmem_falloc->nr_unswapped++;
1336 shmem_falloc = NULL;
1337 spin_unlock(&inode->i_lock);
1341 clear_highpage(page);
1342 flush_dcache_page(page);
1343 SetPageUptodate(page);
1346 swap = get_swap_page(page);
1351 * Add inode to shmem_unuse()'s list of swapped-out inodes,
1352 * if it's not already there. Do it now before the page is
1353 * moved to swap cache, when its pagelock no longer protects
1354 * the inode from eviction. But don't unlock the mutex until
1355 * we've incremented swapped, because shmem_unuse_inode() will
1356 * prune a !swapped inode from the swaplist under this mutex.
1358 mutex_lock(&shmem_swaplist_mutex);
1359 if (list_empty(&info->swaplist))
1360 list_add_tail(&info->swaplist, &shmem_swaplist);
1362 if (add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
1363 spin_lock_irq(&info->lock);
1364 shmem_recalc_inode(inode);
1366 spin_unlock_irq(&info->lock);
1368 swap_shmem_alloc(swap);
1369 shmem_delete_from_page_cache(page, swp_to_radix_entry(swap));
1371 mutex_unlock(&shmem_swaplist_mutex);
1372 BUG_ON(page_mapped(page));
1373 swap_writepage(page, wbc);
1377 mutex_unlock(&shmem_swaplist_mutex);
1378 put_swap_page(page, swap);
1380 set_page_dirty(page);
1381 if (wbc->for_reclaim)
1382 return AOP_WRITEPAGE_ACTIVATE; /* Return with page locked */
1387 #if defined(CONFIG_NUMA) && defined(CONFIG_TMPFS)
1388 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1392 if (!mpol || mpol->mode == MPOL_DEFAULT)
1393 return; /* show nothing */
1395 mpol_to_str(buffer, sizeof(buffer), mpol);
1397 seq_printf(seq, ",mpol=%s", buffer);
1400 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1402 struct mempolicy *mpol = NULL;
1404 spin_lock(&sbinfo->stat_lock); /* prevent replace/use races */
1405 mpol = sbinfo->mpol;
1407 spin_unlock(&sbinfo->stat_lock);
1411 #else /* !CONFIG_NUMA || !CONFIG_TMPFS */
1412 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1415 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1419 #endif /* CONFIG_NUMA && CONFIG_TMPFS */
1421 #define vm_policy vm_private_data
1424 static void shmem_pseudo_vma_init(struct vm_area_struct *vma,
1425 struct shmem_inode_info *info, pgoff_t index)
1427 /* Create a pseudo vma that just contains the policy */
1428 vma_init(vma, NULL);
1429 /* Bias interleave by inode number to distribute better across nodes */
1430 vma->vm_pgoff = index + info->vfs_inode.i_ino;
1431 vma->vm_policy = mpol_shared_policy_lookup(&info->policy, index);
1434 static void shmem_pseudo_vma_destroy(struct vm_area_struct *vma)
1436 /* Drop reference taken by mpol_shared_policy_lookup() */
1437 mpol_cond_put(vma->vm_policy);
1440 static struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
1441 struct shmem_inode_info *info, pgoff_t index)
1443 struct vm_area_struct pvma;
1445 struct vm_fault vmf;
1447 shmem_pseudo_vma_init(&pvma, info, index);
1450 page = swap_cluster_readahead(swap, gfp, &vmf);
1451 shmem_pseudo_vma_destroy(&pvma);
1456 static struct page *shmem_alloc_hugepage(gfp_t gfp,
1457 struct shmem_inode_info *info, pgoff_t index)
1459 struct vm_area_struct pvma;
1460 struct inode *inode = &info->vfs_inode;
1461 struct address_space *mapping = inode->i_mapping;
1462 pgoff_t idx, hindex;
1463 void __rcu **results;
1466 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
1469 hindex = round_down(index, HPAGE_PMD_NR);
1471 if (radix_tree_gang_lookup_slot(&mapping->i_pages, &results, &idx,
1472 hindex, 1) && idx < hindex + HPAGE_PMD_NR) {
1478 shmem_pseudo_vma_init(&pvma, info, hindex);
1479 page = alloc_pages_vma(gfp | __GFP_COMP | __GFP_NORETRY | __GFP_NOWARN,
1480 HPAGE_PMD_ORDER, &pvma, 0, numa_node_id(), true);
1481 shmem_pseudo_vma_destroy(&pvma);
1483 prep_transhuge_page(page);
1487 static struct page *shmem_alloc_page(gfp_t gfp,
1488 struct shmem_inode_info *info, pgoff_t index)
1490 struct vm_area_struct pvma;
1493 shmem_pseudo_vma_init(&pvma, info, index);
1494 page = alloc_page_vma(gfp, &pvma, 0);
1495 shmem_pseudo_vma_destroy(&pvma);
1500 static struct page *shmem_alloc_and_acct_page(gfp_t gfp,
1501 struct inode *inode,
1502 pgoff_t index, bool huge)
1504 struct shmem_inode_info *info = SHMEM_I(inode);
1509 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
1511 nr = huge ? HPAGE_PMD_NR : 1;
1513 if (!shmem_inode_acct_block(inode, nr))
1517 page = shmem_alloc_hugepage(gfp, info, index);
1519 page = shmem_alloc_page(gfp, info, index);
1521 __SetPageLocked(page);
1522 __SetPageSwapBacked(page);
1527 shmem_inode_unacct_blocks(inode, nr);
1529 return ERR_PTR(err);
1533 * When a page is moved from swapcache to shmem filecache (either by the
1534 * usual swapin of shmem_getpage_gfp(), or by the less common swapoff of
1535 * shmem_unuse_inode()), it may have been read in earlier from swap, in
1536 * ignorance of the mapping it belongs to. If that mapping has special
1537 * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
1538 * we may need to copy to a suitable page before moving to filecache.
1540 * In a future release, this may well be extended to respect cpuset and
1541 * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
1542 * but for now it is a simple matter of zone.
1544 static bool shmem_should_replace_page(struct page *page, gfp_t gfp)
1546 return page_zonenum(page) > gfp_zone(gfp);
1549 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
1550 struct shmem_inode_info *info, pgoff_t index)
1552 struct page *oldpage, *newpage;
1553 struct address_space *swap_mapping;
1559 entry.val = page_private(oldpage);
1560 swap_index = swp_offset(entry);
1561 swap_mapping = page_mapping(oldpage);
1564 * We have arrived here because our zones are constrained, so don't
1565 * limit chance of success by further cpuset and node constraints.
1567 gfp &= ~GFP_CONSTRAINT_MASK;
1568 newpage = shmem_alloc_page(gfp, info, index);
1573 copy_highpage(newpage, oldpage);
1574 flush_dcache_page(newpage);
1576 __SetPageLocked(newpage);
1577 __SetPageSwapBacked(newpage);
1578 SetPageUptodate(newpage);
1579 set_page_private(newpage, entry.val);
1580 SetPageSwapCache(newpage);
1583 * Our caller will very soon move newpage out of swapcache, but it's
1584 * a nice clean interface for us to replace oldpage by newpage there.
1586 xa_lock_irq(&swap_mapping->i_pages);
1587 error = shmem_radix_tree_replace(swap_mapping, swap_index, oldpage,
1590 __inc_node_page_state(newpage, NR_FILE_PAGES);
1591 __dec_node_page_state(oldpage, NR_FILE_PAGES);
1593 xa_unlock_irq(&swap_mapping->i_pages);
1595 if (unlikely(error)) {
1597 * Is this possible? I think not, now that our callers check
1598 * both PageSwapCache and page_private after getting page lock;
1599 * but be defensive. Reverse old to newpage for clear and free.
1603 mem_cgroup_migrate(oldpage, newpage);
1604 lru_cache_add_anon(newpage);
1608 ClearPageSwapCache(oldpage);
1609 set_page_private(oldpage, 0);
1611 unlock_page(oldpage);
1618 * shmem_getpage_gfp - find page in cache, or get from swap, or allocate
1620 * If we allocate a new one we do not mark it dirty. That's up to the
1621 * vm. If we swap it in we mark it dirty since we also free the swap
1622 * entry since a page cannot live in both the swap and page cache.
1624 * fault_mm and fault_type are only supplied by shmem_fault:
1625 * otherwise they are NULL.
1627 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
1628 struct page **pagep, enum sgp_type sgp, gfp_t gfp,
1629 struct vm_area_struct *vma, struct vm_fault *vmf,
1630 vm_fault_t *fault_type)
1632 struct address_space *mapping = inode->i_mapping;
1633 struct shmem_inode_info *info = SHMEM_I(inode);
1634 struct shmem_sb_info *sbinfo;
1635 struct mm_struct *charge_mm;
1636 struct mem_cgroup *memcg;
1639 enum sgp_type sgp_huge = sgp;
1640 pgoff_t hindex = index;
1645 if (index > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
1647 if (sgp == SGP_NOHUGE || sgp == SGP_HUGE)
1651 page = find_lock_entry(mapping, index);
1652 if (radix_tree_exceptional_entry(page)) {
1653 swap = radix_to_swp_entry(page);
1657 if (sgp <= SGP_CACHE &&
1658 ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1663 if (page && sgp == SGP_WRITE)
1664 mark_page_accessed(page);
1666 /* fallocated page? */
1667 if (page && !PageUptodate(page)) {
1668 if (sgp != SGP_READ)
1674 if (page || (sgp == SGP_READ && !swap.val)) {
1680 * Fast cache lookup did not find it:
1681 * bring it back from swap or allocate.
1683 sbinfo = SHMEM_SB(inode->i_sb);
1684 charge_mm = vma ? vma->vm_mm : current->mm;
1687 /* Look it up and read it in.. */
1688 page = lookup_swap_cache(swap, NULL, 0);
1690 /* Or update major stats only when swapin succeeds?? */
1692 *fault_type |= VM_FAULT_MAJOR;
1693 count_vm_event(PGMAJFAULT);
1694 count_memcg_event_mm(charge_mm, PGMAJFAULT);
1696 /* Here we actually start the io */
1697 page = shmem_swapin(swap, gfp, info, index);
1704 /* We have to do this with page locked to prevent races */
1706 if (!PageSwapCache(page) || page_private(page) != swap.val ||
1707 !shmem_confirm_swap(mapping, index, swap)) {
1708 error = -EEXIST; /* try again */
1711 if (!PageUptodate(page)) {
1715 wait_on_page_writeback(page);
1717 if (shmem_should_replace_page(page, gfp)) {
1718 error = shmem_replace_page(&page, gfp, info, index);
1723 error = mem_cgroup_try_charge_delay(page, charge_mm, gfp, &memcg,
1726 error = shmem_add_to_page_cache(page, mapping, index,
1727 swp_to_radix_entry(swap));
1729 * We already confirmed swap under page lock, and make
1730 * no memory allocation here, so usually no possibility
1731 * of error; but free_swap_and_cache() only trylocks a
1732 * page, so it is just possible that the entry has been
1733 * truncated or holepunched since swap was confirmed.
1734 * shmem_undo_range() will have done some of the
1735 * unaccounting, now delete_from_swap_cache() will do
1737 * Reset swap.val? No, leave it so "failed" goes back to
1738 * "repeat": reading a hole and writing should succeed.
1741 mem_cgroup_cancel_charge(page, memcg, false);
1742 delete_from_swap_cache(page);
1748 mem_cgroup_commit_charge(page, memcg, true, false);
1750 spin_lock_irq(&info->lock);
1752 shmem_recalc_inode(inode);
1753 spin_unlock_irq(&info->lock);
1755 if (sgp == SGP_WRITE)
1756 mark_page_accessed(page);
1758 delete_from_swap_cache(page);
1759 set_page_dirty(page);
1763 if (vma && userfaultfd_missing(vma)) {
1764 *fault_type = handle_userfault(vmf, VM_UFFD_MISSING);
1768 /* shmem_symlink() */
1769 if (mapping->a_ops != &shmem_aops)
1771 if (shmem_huge == SHMEM_HUGE_DENY || sgp_huge == SGP_NOHUGE)
1773 if (shmem_huge == SHMEM_HUGE_FORCE)
1775 switch (sbinfo->huge) {
1778 case SHMEM_HUGE_NEVER:
1780 case SHMEM_HUGE_WITHIN_SIZE:
1781 off = round_up(index, HPAGE_PMD_NR);
1782 i_size = round_up(i_size_read(inode), PAGE_SIZE);
1783 if (i_size >= HPAGE_PMD_SIZE &&
1784 i_size >> PAGE_SHIFT >= off)
1787 case SHMEM_HUGE_ADVISE:
1788 if (sgp_huge == SGP_HUGE)
1790 /* TODO: implement fadvise() hints */
1795 page = shmem_alloc_and_acct_page(gfp, inode, index, true);
1797 alloc_nohuge: page = shmem_alloc_and_acct_page(gfp, inode,
1802 error = PTR_ERR(page);
1804 if (error != -ENOSPC)
1807 * Try to reclaim some spece by splitting a huge page
1808 * beyond i_size on the filesystem.
1812 ret = shmem_unused_huge_shrink(sbinfo, NULL, 1);
1813 if (ret == SHRINK_STOP)
1821 if (PageTransHuge(page))
1822 hindex = round_down(index, HPAGE_PMD_NR);
1826 if (sgp == SGP_WRITE)
1827 __SetPageReferenced(page);
1829 error = mem_cgroup_try_charge_delay(page, charge_mm, gfp, &memcg,
1830 PageTransHuge(page));
1833 error = radix_tree_maybe_preload_order(gfp & GFP_RECLAIM_MASK,
1834 compound_order(page));
1836 error = shmem_add_to_page_cache(page, mapping, hindex,
1838 radix_tree_preload_end();
1841 mem_cgroup_cancel_charge(page, memcg,
1842 PageTransHuge(page));
1845 mem_cgroup_commit_charge(page, memcg, false,
1846 PageTransHuge(page));
1847 lru_cache_add_anon(page);
1849 spin_lock_irq(&info->lock);
1850 info->alloced += 1 << compound_order(page);
1851 inode->i_blocks += BLOCKS_PER_PAGE << compound_order(page);
1852 shmem_recalc_inode(inode);
1853 spin_unlock_irq(&info->lock);
1856 if (PageTransHuge(page) &&
1857 DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE) <
1858 hindex + HPAGE_PMD_NR - 1) {
1860 * Part of the huge page is beyond i_size: subject
1861 * to shrink under memory pressure.
1863 spin_lock(&sbinfo->shrinklist_lock);
1865 * _careful to defend against unlocked access to
1866 * ->shrink_list in shmem_unused_huge_shrink()
1868 if (list_empty_careful(&info->shrinklist)) {
1869 list_add_tail(&info->shrinklist,
1870 &sbinfo->shrinklist);
1871 sbinfo->shrinklist_len++;
1873 spin_unlock(&sbinfo->shrinklist_lock);
1877 * Let SGP_FALLOC use the SGP_WRITE optimization on a new page.
1879 if (sgp == SGP_FALLOC)
1883 * Let SGP_WRITE caller clear ends if write does not fill page;
1884 * but SGP_FALLOC on a page fallocated earlier must initialize
1885 * it now, lest undo on failure cancel our earlier guarantee.
1887 if (sgp != SGP_WRITE && !PageUptodate(page)) {
1888 struct page *head = compound_head(page);
1891 for (i = 0; i < (1 << compound_order(head)); i++) {
1892 clear_highpage(head + i);
1893 flush_dcache_page(head + i);
1895 SetPageUptodate(head);
1899 /* Perhaps the file has been truncated since we checked */
1900 if (sgp <= SGP_CACHE &&
1901 ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1903 ClearPageDirty(page);
1904 delete_from_page_cache(page);
1905 spin_lock_irq(&info->lock);
1906 shmem_recalc_inode(inode);
1907 spin_unlock_irq(&info->lock);
1912 *pagep = page + index - hindex;
1919 shmem_inode_unacct_blocks(inode, 1 << compound_order(page));
1921 if (PageTransHuge(page)) {
1927 if (swap.val && !shmem_confirm_swap(mapping, index, swap))
1934 if (error == -ENOSPC && !once++) {
1935 spin_lock_irq(&info->lock);
1936 shmem_recalc_inode(inode);
1937 spin_unlock_irq(&info->lock);
1940 if (error == -EEXIST) /* from above or from radix_tree_insert */
1946 * This is like autoremove_wake_function, but it removes the wait queue
1947 * entry unconditionally - even if something else had already woken the
1950 static int synchronous_wake_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
1952 int ret = default_wake_function(wait, mode, sync, key);
1953 list_del_init(&wait->entry);
1957 static vm_fault_t shmem_fault(struct vm_fault *vmf)
1959 struct vm_area_struct *vma = vmf->vma;
1960 struct inode *inode = file_inode(vma->vm_file);
1961 gfp_t gfp = mapping_gfp_mask(inode->i_mapping);
1964 vm_fault_t ret = VM_FAULT_LOCKED;
1967 * Trinity finds that probing a hole which tmpfs is punching can
1968 * prevent the hole-punch from ever completing: which in turn
1969 * locks writers out with its hold on i_mutex. So refrain from
1970 * faulting pages into the hole while it's being punched. Although
1971 * shmem_undo_range() does remove the additions, it may be unable to
1972 * keep up, as each new page needs its own unmap_mapping_range() call,
1973 * and the i_mmap tree grows ever slower to scan if new vmas are added.
1975 * It does not matter if we sometimes reach this check just before the
1976 * hole-punch begins, so that one fault then races with the punch:
1977 * we just need to make racing faults a rare case.
1979 * The implementation below would be much simpler if we just used a
1980 * standard mutex or completion: but we cannot take i_mutex in fault,
1981 * and bloating every shmem inode for this unlikely case would be sad.
1983 if (unlikely(inode->i_private)) {
1984 struct shmem_falloc *shmem_falloc;
1986 spin_lock(&inode->i_lock);
1987 shmem_falloc = inode->i_private;
1989 shmem_falloc->waitq &&
1990 vmf->pgoff >= shmem_falloc->start &&
1991 vmf->pgoff < shmem_falloc->next) {
1992 wait_queue_head_t *shmem_falloc_waitq;
1993 DEFINE_WAIT_FUNC(shmem_fault_wait, synchronous_wake_function);
1995 ret = VM_FAULT_NOPAGE;
1996 if ((vmf->flags & FAULT_FLAG_ALLOW_RETRY) &&
1997 !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
1998 /* It's polite to up mmap_sem if we can */
1999 up_read(&vma->vm_mm->mmap_sem);
2000 ret = VM_FAULT_RETRY;
2003 shmem_falloc_waitq = shmem_falloc->waitq;
2004 prepare_to_wait(shmem_falloc_waitq, &shmem_fault_wait,
2005 TASK_UNINTERRUPTIBLE);
2006 spin_unlock(&inode->i_lock);
2010 * shmem_falloc_waitq points into the shmem_fallocate()
2011 * stack of the hole-punching task: shmem_falloc_waitq
2012 * is usually invalid by the time we reach here, but
2013 * finish_wait() does not dereference it in that case;
2014 * though i_lock needed lest racing with wake_up_all().
2016 spin_lock(&inode->i_lock);
2017 finish_wait(shmem_falloc_waitq, &shmem_fault_wait);
2018 spin_unlock(&inode->i_lock);
2021 spin_unlock(&inode->i_lock);
2026 if ((vma->vm_flags & VM_NOHUGEPAGE) ||
2027 test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
2029 else if (vma->vm_flags & VM_HUGEPAGE)
2032 err = shmem_getpage_gfp(inode, vmf->pgoff, &vmf->page, sgp,
2033 gfp, vma, vmf, &ret);
2035 return vmf_error(err);
2039 unsigned long shmem_get_unmapped_area(struct file *file,
2040 unsigned long uaddr, unsigned long len,
2041 unsigned long pgoff, unsigned long flags)
2043 unsigned long (*get_area)(struct file *,
2044 unsigned long, unsigned long, unsigned long, unsigned long);
2046 unsigned long offset;
2047 unsigned long inflated_len;
2048 unsigned long inflated_addr;
2049 unsigned long inflated_offset;
2051 if (len > TASK_SIZE)
2054 get_area = current->mm->get_unmapped_area;
2055 addr = get_area(file, uaddr, len, pgoff, flags);
2057 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
2059 if (IS_ERR_VALUE(addr))
2061 if (addr & ~PAGE_MASK)
2063 if (addr > TASK_SIZE - len)
2066 if (shmem_huge == SHMEM_HUGE_DENY)
2068 if (len < HPAGE_PMD_SIZE)
2070 if (flags & MAP_FIXED)
2073 * Our priority is to support MAP_SHARED mapped hugely;
2074 * and support MAP_PRIVATE mapped hugely too, until it is COWed.
2075 * But if caller specified an address hint and we allocated area there
2076 * successfully, respect that as before.
2081 if (shmem_huge != SHMEM_HUGE_FORCE) {
2082 struct super_block *sb;
2085 VM_BUG_ON(file->f_op != &shmem_file_operations);
2086 sb = file_inode(file)->i_sb;
2089 * Called directly from mm/mmap.c, or drivers/char/mem.c
2090 * for "/dev/zero", to create a shared anonymous object.
2092 if (IS_ERR(shm_mnt))
2094 sb = shm_mnt->mnt_sb;
2096 if (SHMEM_SB(sb)->huge == SHMEM_HUGE_NEVER)
2100 offset = (pgoff << PAGE_SHIFT) & (HPAGE_PMD_SIZE-1);
2101 if (offset && offset + len < 2 * HPAGE_PMD_SIZE)
2103 if ((addr & (HPAGE_PMD_SIZE-1)) == offset)
2106 inflated_len = len + HPAGE_PMD_SIZE - PAGE_SIZE;
2107 if (inflated_len > TASK_SIZE)
2109 if (inflated_len < len)
2112 inflated_addr = get_area(NULL, uaddr, inflated_len, 0, flags);
2113 if (IS_ERR_VALUE(inflated_addr))
2115 if (inflated_addr & ~PAGE_MASK)
2118 inflated_offset = inflated_addr & (HPAGE_PMD_SIZE-1);
2119 inflated_addr += offset - inflated_offset;
2120 if (inflated_offset > offset)
2121 inflated_addr += HPAGE_PMD_SIZE;
2123 if (inflated_addr > TASK_SIZE - len)
2125 return inflated_addr;
2129 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
2131 struct inode *inode = file_inode(vma->vm_file);
2132 return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
2135 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
2138 struct inode *inode = file_inode(vma->vm_file);
2141 index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2142 return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
2146 int shmem_lock(struct file *file, int lock, struct user_struct *user)
2148 struct inode *inode = file_inode(file);
2149 struct shmem_inode_info *info = SHMEM_I(inode);
2150 int retval = -ENOMEM;
2153 * What serializes the accesses to info->flags?
2154 * ipc_lock_object() when called from shmctl_do_lock(),
2155 * no serialization needed when called from shm_destroy().
2157 if (lock && !(info->flags & VM_LOCKED)) {
2158 if (!user_shm_lock(inode->i_size, user))
2160 info->flags |= VM_LOCKED;
2161 mapping_set_unevictable(file->f_mapping);
2163 if (!lock && (info->flags & VM_LOCKED) && user) {
2164 user_shm_unlock(inode->i_size, user);
2165 info->flags &= ~VM_LOCKED;
2166 mapping_clear_unevictable(file->f_mapping);
2174 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
2176 file_accessed(file);
2177 vma->vm_ops = &shmem_vm_ops;
2178 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE) &&
2179 ((vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK) <
2180 (vma->vm_end & HPAGE_PMD_MASK)) {
2181 khugepaged_enter(vma, vma->vm_flags);
2186 static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
2187 umode_t mode, dev_t dev, unsigned long flags)
2189 struct inode *inode;
2190 struct shmem_inode_info *info;
2191 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2193 if (shmem_reserve_inode(sb))
2196 inode = new_inode(sb);
2198 inode->i_ino = get_next_ino();
2199 inode_init_owner(inode, dir, mode);
2200 inode->i_blocks = 0;
2201 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
2202 inode->i_generation = prandom_u32();
2203 info = SHMEM_I(inode);
2204 memset(info, 0, (char *)inode - (char *)info);
2205 spin_lock_init(&info->lock);
2206 info->seals = F_SEAL_SEAL;
2207 info->flags = flags & VM_NORESERVE;
2208 INIT_LIST_HEAD(&info->shrinklist);
2209 INIT_LIST_HEAD(&info->swaplist);
2210 simple_xattrs_init(&info->xattrs);
2211 cache_no_acl(inode);
2213 switch (mode & S_IFMT) {
2215 inode->i_op = &shmem_special_inode_operations;
2216 init_special_inode(inode, mode, dev);
2219 inode->i_mapping->a_ops = &shmem_aops;
2220 inode->i_op = &shmem_inode_operations;
2221 inode->i_fop = &shmem_file_operations;
2222 mpol_shared_policy_init(&info->policy,
2223 shmem_get_sbmpol(sbinfo));
2227 /* Some things misbehave if size == 0 on a directory */
2228 inode->i_size = 2 * BOGO_DIRENT_SIZE;
2229 inode->i_op = &shmem_dir_inode_operations;
2230 inode->i_fop = &simple_dir_operations;
2234 * Must not load anything in the rbtree,
2235 * mpol_free_shared_policy will not be called.
2237 mpol_shared_policy_init(&info->policy, NULL);
2241 lockdep_annotate_inode_mutex_key(inode);
2243 shmem_free_inode(sb);
2247 bool shmem_mapping(struct address_space *mapping)
2249 return mapping->a_ops == &shmem_aops;
2252 static int shmem_mfill_atomic_pte(struct mm_struct *dst_mm,
2254 struct vm_area_struct *dst_vma,
2255 unsigned long dst_addr,
2256 unsigned long src_addr,
2258 struct page **pagep)
2260 struct inode *inode = file_inode(dst_vma->vm_file);
2261 struct shmem_inode_info *info = SHMEM_I(inode);
2262 struct address_space *mapping = inode->i_mapping;
2263 gfp_t gfp = mapping_gfp_mask(mapping);
2264 pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
2265 struct mem_cgroup *memcg;
2269 pte_t _dst_pte, *dst_pte;
2271 pgoff_t offset, max_off;
2274 if (!shmem_inode_acct_block(inode, 1)) {
2276 * We may have got a page, returned -ENOENT triggering a retry,
2277 * and now we find ourselves with -ENOMEM. Release the page, to
2278 * avoid a BUG_ON in our caller.
2280 if (unlikely(*pagep)) {
2288 page = shmem_alloc_page(gfp, info, pgoff);
2290 goto out_unacct_blocks;
2292 if (!zeropage) { /* mcopy_atomic */
2293 page_kaddr = kmap_atomic(page);
2294 ret = copy_from_user(page_kaddr,
2295 (const void __user *)src_addr,
2297 kunmap_atomic(page_kaddr);
2299 /* fallback to copy_from_user outside mmap_sem */
2300 if (unlikely(ret)) {
2302 shmem_inode_unacct_blocks(inode, 1);
2303 /* don't free the page */
2306 } else { /* mfill_zeropage_atomic */
2307 clear_highpage(page);
2314 VM_BUG_ON(PageLocked(page) || PageSwapBacked(page));
2315 __SetPageLocked(page);
2316 __SetPageSwapBacked(page);
2317 __SetPageUptodate(page);
2320 offset = linear_page_index(dst_vma, dst_addr);
2321 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2322 if (unlikely(offset >= max_off))
2325 ret = mem_cgroup_try_charge_delay(page, dst_mm, gfp, &memcg, false);
2329 ret = radix_tree_maybe_preload(gfp & GFP_RECLAIM_MASK);
2331 ret = shmem_add_to_page_cache(page, mapping, pgoff, NULL);
2332 radix_tree_preload_end();
2335 goto out_release_uncharge;
2337 mem_cgroup_commit_charge(page, memcg, false, false);
2339 _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
2340 if (dst_vma->vm_flags & VM_WRITE)
2341 _dst_pte = pte_mkwrite(pte_mkdirty(_dst_pte));
2344 * We don't set the pte dirty if the vma has no
2345 * VM_WRITE permission, so mark the page dirty or it
2346 * could be freed from under us. We could do it
2347 * unconditionally before unlock_page(), but doing it
2348 * only if VM_WRITE is not set is faster.
2350 set_page_dirty(page);
2353 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
2356 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
2357 if (unlikely(offset >= max_off))
2358 goto out_release_uncharge_unlock;
2361 if (!pte_none(*dst_pte))
2362 goto out_release_uncharge_unlock;
2364 lru_cache_add_anon(page);
2366 spin_lock_irq(&info->lock);
2368 inode->i_blocks += BLOCKS_PER_PAGE;
2369 shmem_recalc_inode(inode);
2370 spin_unlock_irq(&info->lock);
2372 inc_mm_counter(dst_mm, mm_counter_file(page));
2373 page_add_file_rmap(page, false);
2374 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
2376 /* No need to invalidate - it was non-present before */
2377 update_mmu_cache(dst_vma, dst_addr, dst_pte);
2378 pte_unmap_unlock(dst_pte, ptl);
2383 out_release_uncharge_unlock:
2384 pte_unmap_unlock(dst_pte, ptl);
2385 ClearPageDirty(page);
2386 delete_from_page_cache(page);
2387 out_release_uncharge:
2388 mem_cgroup_cancel_charge(page, memcg, false);
2393 shmem_inode_unacct_blocks(inode, 1);
2397 int shmem_mcopy_atomic_pte(struct mm_struct *dst_mm,
2399 struct vm_area_struct *dst_vma,
2400 unsigned long dst_addr,
2401 unsigned long src_addr,
2402 struct page **pagep)
2404 return shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
2405 dst_addr, src_addr, false, pagep);
2408 int shmem_mfill_zeropage_pte(struct mm_struct *dst_mm,
2410 struct vm_area_struct *dst_vma,
2411 unsigned long dst_addr)
2413 struct page *page = NULL;
2415 return shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
2416 dst_addr, 0, true, &page);
2420 static const struct inode_operations shmem_symlink_inode_operations;
2421 static const struct inode_operations shmem_short_symlink_operations;
2423 #ifdef CONFIG_TMPFS_XATTR
2424 static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
2426 #define shmem_initxattrs NULL
2430 shmem_write_begin(struct file *file, struct address_space *mapping,
2431 loff_t pos, unsigned len, unsigned flags,
2432 struct page **pagep, void **fsdata)
2434 struct inode *inode = mapping->host;
2435 struct shmem_inode_info *info = SHMEM_I(inode);
2436 pgoff_t index = pos >> PAGE_SHIFT;
2438 /* i_mutex is held by caller */
2439 if (unlikely(info->seals & (F_SEAL_WRITE | F_SEAL_GROW))) {
2440 if (info->seals & F_SEAL_WRITE)
2442 if ((info->seals & F_SEAL_GROW) && pos + len > inode->i_size)
2446 return shmem_getpage(inode, index, pagep, SGP_WRITE);
2450 shmem_write_end(struct file *file, struct address_space *mapping,
2451 loff_t pos, unsigned len, unsigned copied,
2452 struct page *page, void *fsdata)
2454 struct inode *inode = mapping->host;
2456 if (pos + copied > inode->i_size)
2457 i_size_write(inode, pos + copied);
2459 if (!PageUptodate(page)) {
2460 struct page *head = compound_head(page);
2461 if (PageTransCompound(page)) {
2464 for (i = 0; i < HPAGE_PMD_NR; i++) {
2465 if (head + i == page)
2467 clear_highpage(head + i);
2468 flush_dcache_page(head + i);
2471 if (copied < PAGE_SIZE) {
2472 unsigned from = pos & (PAGE_SIZE - 1);
2473 zero_user_segments(page, 0, from,
2474 from + copied, PAGE_SIZE);
2476 SetPageUptodate(head);
2478 set_page_dirty(page);
2485 static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
2487 struct file *file = iocb->ki_filp;
2488 struct inode *inode = file_inode(file);
2489 struct address_space *mapping = inode->i_mapping;
2491 unsigned long offset;
2492 enum sgp_type sgp = SGP_READ;
2495 loff_t *ppos = &iocb->ki_pos;
2498 * Might this read be for a stacking filesystem? Then when reading
2499 * holes of a sparse file, we actually need to allocate those pages,
2500 * and even mark them dirty, so it cannot exceed the max_blocks limit.
2502 if (!iter_is_iovec(to))
2505 index = *ppos >> PAGE_SHIFT;
2506 offset = *ppos & ~PAGE_MASK;
2509 struct page *page = NULL;
2511 unsigned long nr, ret;
2512 loff_t i_size = i_size_read(inode);
2514 end_index = i_size >> PAGE_SHIFT;
2515 if (index > end_index)
2517 if (index == end_index) {
2518 nr = i_size & ~PAGE_MASK;
2523 error = shmem_getpage(inode, index, &page, sgp);
2525 if (error == -EINVAL)
2530 if (sgp == SGP_CACHE)
2531 set_page_dirty(page);
2536 * We must evaluate after, since reads (unlike writes)
2537 * are called without i_mutex protection against truncate
2540 i_size = i_size_read(inode);
2541 end_index = i_size >> PAGE_SHIFT;
2542 if (index == end_index) {
2543 nr = i_size & ~PAGE_MASK;
2554 * If users can be writing to this page using arbitrary
2555 * virtual addresses, take care about potential aliasing
2556 * before reading the page on the kernel side.
2558 if (mapping_writably_mapped(mapping))
2559 flush_dcache_page(page);
2561 * Mark the page accessed if we read the beginning.
2564 mark_page_accessed(page);
2566 page = ZERO_PAGE(0);
2571 * Ok, we have the page, and it's up-to-date, so
2572 * now we can copy it to user space...
2574 ret = copy_page_to_iter(page, offset, nr, to);
2577 index += offset >> PAGE_SHIFT;
2578 offset &= ~PAGE_MASK;
2581 if (!iov_iter_count(to))
2590 *ppos = ((loff_t) index << PAGE_SHIFT) + offset;
2591 file_accessed(file);
2592 return retval ? retval : error;
2596 * llseek SEEK_DATA or SEEK_HOLE through the radix_tree.
2598 static pgoff_t shmem_seek_hole_data(struct address_space *mapping,
2599 pgoff_t index, pgoff_t end, int whence)
2602 struct pagevec pvec;
2603 pgoff_t indices[PAGEVEC_SIZE];
2607 pagevec_init(&pvec);
2608 pvec.nr = 1; /* start small: we may be there already */
2610 pvec.nr = find_get_entries(mapping, index,
2611 pvec.nr, pvec.pages, indices);
2613 if (whence == SEEK_DATA)
2617 for (i = 0; i < pvec.nr; i++, index++) {
2618 if (index < indices[i]) {
2619 if (whence == SEEK_HOLE) {
2625 page = pvec.pages[i];
2626 if (page && !radix_tree_exceptional_entry(page)) {
2627 if (!PageUptodate(page))
2631 (page && whence == SEEK_DATA) ||
2632 (!page && whence == SEEK_HOLE)) {
2637 pagevec_remove_exceptionals(&pvec);
2638 pagevec_release(&pvec);
2639 pvec.nr = PAGEVEC_SIZE;
2645 static loff_t shmem_file_llseek(struct file *file, loff_t offset, int whence)
2647 struct address_space *mapping = file->f_mapping;
2648 struct inode *inode = mapping->host;
2652 if (whence != SEEK_DATA && whence != SEEK_HOLE)
2653 return generic_file_llseek_size(file, offset, whence,
2654 MAX_LFS_FILESIZE, i_size_read(inode));
2656 /* We're holding i_mutex so we can access i_size directly */
2658 if (offset < 0 || offset >= inode->i_size)
2661 start = offset >> PAGE_SHIFT;
2662 end = (inode->i_size + PAGE_SIZE - 1) >> PAGE_SHIFT;
2663 new_offset = shmem_seek_hole_data(mapping, start, end, whence);
2664 new_offset <<= PAGE_SHIFT;
2665 if (new_offset > offset) {
2666 if (new_offset < inode->i_size)
2667 offset = new_offset;
2668 else if (whence == SEEK_DATA)
2671 offset = inode->i_size;
2676 offset = vfs_setpos(file, offset, MAX_LFS_FILESIZE);
2677 inode_unlock(inode);
2681 static long shmem_fallocate(struct file *file, int mode, loff_t offset,
2684 struct inode *inode = file_inode(file);
2685 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
2686 struct shmem_inode_info *info = SHMEM_I(inode);
2687 struct shmem_falloc shmem_falloc;
2688 pgoff_t start, index, end;
2691 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2696 if (mode & FALLOC_FL_PUNCH_HOLE) {
2697 struct address_space *mapping = file->f_mapping;
2698 loff_t unmap_start = round_up(offset, PAGE_SIZE);
2699 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
2700 DECLARE_WAIT_QUEUE_HEAD_ONSTACK(shmem_falloc_waitq);
2702 /* protected by i_mutex */
2703 if (info->seals & F_SEAL_WRITE) {
2708 shmem_falloc.waitq = &shmem_falloc_waitq;
2709 shmem_falloc.start = (u64)unmap_start >> PAGE_SHIFT;
2710 shmem_falloc.next = (unmap_end + 1) >> PAGE_SHIFT;
2711 spin_lock(&inode->i_lock);
2712 inode->i_private = &shmem_falloc;
2713 spin_unlock(&inode->i_lock);
2715 if ((u64)unmap_end > (u64)unmap_start)
2716 unmap_mapping_range(mapping, unmap_start,
2717 1 + unmap_end - unmap_start, 0);
2718 shmem_truncate_range(inode, offset, offset + len - 1);
2719 /* No need to unmap again: hole-punching leaves COWed pages */
2721 spin_lock(&inode->i_lock);
2722 inode->i_private = NULL;
2723 wake_up_all(&shmem_falloc_waitq);
2724 WARN_ON_ONCE(!list_empty(&shmem_falloc_waitq.head));
2725 spin_unlock(&inode->i_lock);
2730 /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
2731 error = inode_newsize_ok(inode, offset + len);
2735 if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
2740 start = offset >> PAGE_SHIFT;
2741 end = (offset + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
2742 /* Try to avoid a swapstorm if len is impossible to satisfy */
2743 if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
2748 shmem_falloc.waitq = NULL;
2749 shmem_falloc.start = start;
2750 shmem_falloc.next = start;
2751 shmem_falloc.nr_falloced = 0;
2752 shmem_falloc.nr_unswapped = 0;
2753 spin_lock(&inode->i_lock);
2754 inode->i_private = &shmem_falloc;
2755 spin_unlock(&inode->i_lock);
2757 for (index = start; index < end; index++) {
2761 * Good, the fallocate(2) manpage permits EINTR: we may have
2762 * been interrupted because we are using up too much memory.
2764 if (signal_pending(current))
2766 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
2769 error = shmem_getpage(inode, index, &page, SGP_FALLOC);
2771 /* Remove the !PageUptodate pages we added */
2772 if (index > start) {
2773 shmem_undo_range(inode,
2774 (loff_t)start << PAGE_SHIFT,
2775 ((loff_t)index << PAGE_SHIFT) - 1, true);
2781 * Inform shmem_writepage() how far we have reached.
2782 * No need for lock or barrier: we have the page lock.
2784 shmem_falloc.next++;
2785 if (!PageUptodate(page))
2786 shmem_falloc.nr_falloced++;
2789 * If !PageUptodate, leave it that way so that freeable pages
2790 * can be recognized if we need to rollback on error later.
2791 * But set_page_dirty so that memory pressure will swap rather
2792 * than free the pages we are allocating (and SGP_CACHE pages
2793 * might still be clean: we now need to mark those dirty too).
2795 set_page_dirty(page);
2801 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
2802 i_size_write(inode, offset + len);
2803 inode->i_ctime = current_time(inode);
2805 spin_lock(&inode->i_lock);
2806 inode->i_private = NULL;
2807 spin_unlock(&inode->i_lock);
2809 inode_unlock(inode);
2813 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
2815 struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
2817 buf->f_type = TMPFS_MAGIC;
2818 buf->f_bsize = PAGE_SIZE;
2819 buf->f_namelen = NAME_MAX;
2820 if (sbinfo->max_blocks) {
2821 buf->f_blocks = sbinfo->max_blocks;
2823 buf->f_bfree = sbinfo->max_blocks -
2824 percpu_counter_sum(&sbinfo->used_blocks);
2826 if (sbinfo->max_inodes) {
2827 buf->f_files = sbinfo->max_inodes;
2828 buf->f_ffree = sbinfo->free_inodes;
2830 /* else leave those fields 0 like simple_statfs */
2835 * File creation. Allocate an inode, and we're done..
2838 shmem_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
2840 struct inode *inode;
2841 int error = -ENOSPC;
2843 inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
2845 error = simple_acl_create(dir, inode);
2848 error = security_inode_init_security(inode, dir,
2850 shmem_initxattrs, NULL);
2851 if (error && error != -EOPNOTSUPP)
2855 dir->i_size += BOGO_DIRENT_SIZE;
2856 dir->i_ctime = dir->i_mtime = current_time(dir);
2857 d_instantiate(dentry, inode);
2858 dget(dentry); /* Extra count - pin the dentry in core */
2867 shmem_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
2869 struct inode *inode;
2870 int error = -ENOSPC;
2872 inode = shmem_get_inode(dir->i_sb, dir, mode, 0, VM_NORESERVE);
2874 error = security_inode_init_security(inode, dir,
2876 shmem_initxattrs, NULL);
2877 if (error && error != -EOPNOTSUPP)
2879 error = simple_acl_create(dir, inode);
2882 d_tmpfile(dentry, inode);
2890 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
2894 if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
2900 static int shmem_create(struct inode *dir, struct dentry *dentry, umode_t mode,
2903 return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
2909 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
2911 struct inode *inode = d_inode(old_dentry);
2915 * No ordinary (disk based) filesystem counts links as inodes;
2916 * but each new link needs a new dentry, pinning lowmem, and
2917 * tmpfs dentries cannot be pruned until they are unlinked.
2918 * But if an O_TMPFILE file is linked into the tmpfs, the
2919 * first link must skip that, to get the accounting right.
2921 if (inode->i_nlink) {
2922 ret = shmem_reserve_inode(inode->i_sb);
2927 dir->i_size += BOGO_DIRENT_SIZE;
2928 inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
2930 ihold(inode); /* New dentry reference */
2931 dget(dentry); /* Extra pinning count for the created dentry */
2932 d_instantiate(dentry, inode);
2937 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
2939 struct inode *inode = d_inode(dentry);
2941 if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
2942 shmem_free_inode(inode->i_sb);
2944 dir->i_size -= BOGO_DIRENT_SIZE;
2945 inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
2947 dput(dentry); /* Undo the count from "create" - this does all the work */
2951 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
2953 if (!simple_empty(dentry))
2956 drop_nlink(d_inode(dentry));
2958 return shmem_unlink(dir, dentry);
2961 static int shmem_exchange(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
2963 bool old_is_dir = d_is_dir(old_dentry);
2964 bool new_is_dir = d_is_dir(new_dentry);
2966 if (old_dir != new_dir && old_is_dir != new_is_dir) {
2968 drop_nlink(old_dir);
2971 drop_nlink(new_dir);
2975 old_dir->i_ctime = old_dir->i_mtime =
2976 new_dir->i_ctime = new_dir->i_mtime =
2977 d_inode(old_dentry)->i_ctime =
2978 d_inode(new_dentry)->i_ctime = current_time(old_dir);
2983 static int shmem_whiteout(struct inode *old_dir, struct dentry *old_dentry)
2985 struct dentry *whiteout;
2988 whiteout = d_alloc(old_dentry->d_parent, &old_dentry->d_name);
2992 error = shmem_mknod(old_dir, whiteout,
2993 S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
2999 * Cheat and hash the whiteout while the old dentry is still in
3000 * place, instead of playing games with FS_RENAME_DOES_D_MOVE.
3002 * d_lookup() will consistently find one of them at this point,
3003 * not sure which one, but that isn't even important.
3010 * The VFS layer already does all the dentry stuff for rename,
3011 * we just have to decrement the usage count for the target if
3012 * it exists so that the VFS layer correctly free's it when it
3015 static int shmem_rename2(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry, unsigned int flags)
3017 struct inode *inode = d_inode(old_dentry);
3018 int they_are_dirs = S_ISDIR(inode->i_mode);
3020 if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
3023 if (flags & RENAME_EXCHANGE)
3024 return shmem_exchange(old_dir, old_dentry, new_dir, new_dentry);
3026 if (!simple_empty(new_dentry))
3029 if (flags & RENAME_WHITEOUT) {
3032 error = shmem_whiteout(old_dir, old_dentry);
3037 if (d_really_is_positive(new_dentry)) {
3038 (void) shmem_unlink(new_dir, new_dentry);
3039 if (they_are_dirs) {
3040 drop_nlink(d_inode(new_dentry));
3041 drop_nlink(old_dir);
3043 } else if (they_are_dirs) {
3044 drop_nlink(old_dir);
3048 old_dir->i_size -= BOGO_DIRENT_SIZE;
3049 new_dir->i_size += BOGO_DIRENT_SIZE;
3050 old_dir->i_ctime = old_dir->i_mtime =
3051 new_dir->i_ctime = new_dir->i_mtime =
3052 inode->i_ctime = current_time(old_dir);
3056 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
3060 struct inode *inode;
3063 len = strlen(symname) + 1;
3064 if (len > PAGE_SIZE)
3065 return -ENAMETOOLONG;
3067 inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK | 0777, 0,
3072 error = security_inode_init_security(inode, dir, &dentry->d_name,
3073 shmem_initxattrs, NULL);
3075 if (error != -EOPNOTSUPP) {
3082 inode->i_size = len-1;
3083 if (len <= SHORT_SYMLINK_LEN) {
3084 inode->i_link = kmemdup(symname, len, GFP_KERNEL);
3085 if (!inode->i_link) {
3089 inode->i_op = &shmem_short_symlink_operations;
3091 inode_nohighmem(inode);
3092 error = shmem_getpage(inode, 0, &page, SGP_WRITE);
3097 inode->i_mapping->a_ops = &shmem_aops;
3098 inode->i_op = &shmem_symlink_inode_operations;
3099 memcpy(page_address(page), symname, len);
3100 SetPageUptodate(page);
3101 set_page_dirty(page);
3105 dir->i_size += BOGO_DIRENT_SIZE;
3106 dir->i_ctime = dir->i_mtime = current_time(dir);
3107 d_instantiate(dentry, inode);
3112 static void shmem_put_link(void *arg)
3114 mark_page_accessed(arg);
3118 static const char *shmem_get_link(struct dentry *dentry,
3119 struct inode *inode,
3120 struct delayed_call *done)
3122 struct page *page = NULL;
3125 page = find_get_page(inode->i_mapping, 0);
3127 return ERR_PTR(-ECHILD);
3128 if (!PageUptodate(page)) {
3130 return ERR_PTR(-ECHILD);
3133 error = shmem_getpage(inode, 0, &page, SGP_READ);
3135 return ERR_PTR(error);
3138 set_delayed_call(done, shmem_put_link, page);
3139 return page_address(page);
3142 #ifdef CONFIG_TMPFS_XATTR
3144 * Superblocks without xattr inode operations may get some security.* xattr
3145 * support from the LSM "for free". As soon as we have any other xattrs
3146 * like ACLs, we also need to implement the security.* handlers at
3147 * filesystem level, though.
3151 * Callback for security_inode_init_security() for acquiring xattrs.
3153 static int shmem_initxattrs(struct inode *inode,
3154 const struct xattr *xattr_array,
3157 struct shmem_inode_info *info = SHMEM_I(inode);
3158 const struct xattr *xattr;
3159 struct simple_xattr *new_xattr;
3162 for (xattr = xattr_array; xattr->name != NULL; xattr++) {
3163 new_xattr = simple_xattr_alloc(xattr->value, xattr->value_len);
3167 len = strlen(xattr->name) + 1;
3168 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
3170 if (!new_xattr->name) {
3175 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
3176 XATTR_SECURITY_PREFIX_LEN);
3177 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
3180 simple_xattr_list_add(&info->xattrs, new_xattr);
3186 static int shmem_xattr_handler_get(const struct xattr_handler *handler,
3187 struct dentry *unused, struct inode *inode,
3188 const char *name, void *buffer, size_t size)
3190 struct shmem_inode_info *info = SHMEM_I(inode);
3192 name = xattr_full_name(handler, name);
3193 return simple_xattr_get(&info->xattrs, name, buffer, size);
3196 static int shmem_xattr_handler_set(const struct xattr_handler *handler,
3197 struct dentry *unused, struct inode *inode,
3198 const char *name, const void *value,
3199 size_t size, int flags)
3201 struct shmem_inode_info *info = SHMEM_I(inode);
3203 name = xattr_full_name(handler, name);
3204 return simple_xattr_set(&info->xattrs, name, value, size, flags);
3207 static const struct xattr_handler shmem_security_xattr_handler = {
3208 .prefix = XATTR_SECURITY_PREFIX,
3209 .get = shmem_xattr_handler_get,
3210 .set = shmem_xattr_handler_set,
3213 static const struct xattr_handler shmem_trusted_xattr_handler = {
3214 .prefix = XATTR_TRUSTED_PREFIX,
3215 .get = shmem_xattr_handler_get,
3216 .set = shmem_xattr_handler_set,
3219 static const struct xattr_handler *shmem_xattr_handlers[] = {
3220 #ifdef CONFIG_TMPFS_POSIX_ACL
3221 &posix_acl_access_xattr_handler,
3222 &posix_acl_default_xattr_handler,
3224 &shmem_security_xattr_handler,
3225 &shmem_trusted_xattr_handler,
3229 static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
3231 struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
3232 return simple_xattr_list(d_inode(dentry), &info->xattrs, buffer, size);
3234 #endif /* CONFIG_TMPFS_XATTR */
3236 static const struct inode_operations shmem_short_symlink_operations = {
3237 .get_link = simple_get_link,
3238 #ifdef CONFIG_TMPFS_XATTR
3239 .listxattr = shmem_listxattr,
3243 static const struct inode_operations shmem_symlink_inode_operations = {
3244 .get_link = shmem_get_link,
3245 #ifdef CONFIG_TMPFS_XATTR
3246 .listxattr = shmem_listxattr,
3250 static struct dentry *shmem_get_parent(struct dentry *child)
3252 return ERR_PTR(-ESTALE);
3255 static int shmem_match(struct inode *ino, void *vfh)
3259 inum = (inum << 32) | fh[1];
3260 return ino->i_ino == inum && fh[0] == ino->i_generation;
3263 /* Find any alias of inode, but prefer a hashed alias */
3264 static struct dentry *shmem_find_alias(struct inode *inode)
3266 struct dentry *alias = d_find_alias(inode);
3268 return alias ?: d_find_any_alias(inode);
3272 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
3273 struct fid *fid, int fh_len, int fh_type)
3275 struct inode *inode;
3276 struct dentry *dentry = NULL;
3283 inum = (inum << 32) | fid->raw[1];
3285 inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
3286 shmem_match, fid->raw);
3288 dentry = shmem_find_alias(inode);
3295 static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
3296 struct inode *parent)
3300 return FILEID_INVALID;
3303 if (inode_unhashed(inode)) {
3304 /* Unfortunately insert_inode_hash is not idempotent,
3305 * so as we hash inodes here rather than at creation
3306 * time, we need a lock to ensure we only try
3309 static DEFINE_SPINLOCK(lock);
3311 if (inode_unhashed(inode))
3312 __insert_inode_hash(inode,
3313 inode->i_ino + inode->i_generation);
3317 fh[0] = inode->i_generation;
3318 fh[1] = inode->i_ino;
3319 fh[2] = ((__u64)inode->i_ino) >> 32;
3325 static const struct export_operations shmem_export_ops = {
3326 .get_parent = shmem_get_parent,
3327 .encode_fh = shmem_encode_fh,
3328 .fh_to_dentry = shmem_fh_to_dentry,
3331 static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
3334 char *this_char, *value, *rest;
3335 struct mempolicy *mpol = NULL;
3339 while (options != NULL) {
3340 this_char = options;
3343 * NUL-terminate this option: unfortunately,
3344 * mount options form a comma-separated list,
3345 * but mpol's nodelist may also contain commas.
3347 options = strchr(options, ',');
3348 if (options == NULL)
3351 if (!isdigit(*options)) {
3358 if ((value = strchr(this_char,'=')) != NULL) {
3361 pr_err("tmpfs: No value for mount option '%s'\n",
3366 if (!strcmp(this_char,"size")) {
3367 unsigned long long size;
3368 size = memparse(value,&rest);
3370 size <<= PAGE_SHIFT;
3371 size *= totalram_pages;
3377 sbinfo->max_blocks =
3378 DIV_ROUND_UP(size, PAGE_SIZE);
3379 } else if (!strcmp(this_char,"nr_blocks")) {
3380 sbinfo->max_blocks = memparse(value, &rest);
3383 } else if (!strcmp(this_char,"nr_inodes")) {
3384 sbinfo->max_inodes = memparse(value, &rest);
3387 } else if (!strcmp(this_char,"mode")) {
3390 sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
3393 } else if (!strcmp(this_char,"uid")) {
3396 uid = simple_strtoul(value, &rest, 0);
3399 sbinfo->uid = make_kuid(current_user_ns(), uid);
3400 if (!uid_valid(sbinfo->uid))
3402 } else if (!strcmp(this_char,"gid")) {
3405 gid = simple_strtoul(value, &rest, 0);
3408 sbinfo->gid = make_kgid(current_user_ns(), gid);
3409 if (!gid_valid(sbinfo->gid))
3411 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3412 } else if (!strcmp(this_char, "huge")) {
3414 huge = shmem_parse_huge(value);
3417 if (!has_transparent_hugepage() &&
3418 huge != SHMEM_HUGE_NEVER)
3420 sbinfo->huge = huge;
3423 } else if (!strcmp(this_char,"mpol")) {
3426 if (mpol_parse_str(value, &mpol))
3430 pr_err("tmpfs: Bad mount option %s\n", this_char);
3434 sbinfo->mpol = mpol;
3438 pr_err("tmpfs: Bad value '%s' for mount option '%s'\n",
3446 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
3448 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
3449 struct shmem_sb_info config = *sbinfo;
3450 unsigned long inodes;
3451 int error = -EINVAL;
3454 if (shmem_parse_options(data, &config, true))
3457 spin_lock(&sbinfo->stat_lock);
3458 inodes = sbinfo->max_inodes - sbinfo->free_inodes;
3459 if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
3461 if (config.max_inodes < inodes)
3464 * Those tests disallow limited->unlimited while any are in use;
3465 * but we must separately disallow unlimited->limited, because
3466 * in that case we have no record of how much is already in use.
3468 if (config.max_blocks && !sbinfo->max_blocks)
3470 if (config.max_inodes && !sbinfo->max_inodes)
3474 sbinfo->huge = config.huge;
3475 sbinfo->max_blocks = config.max_blocks;
3476 sbinfo->max_inodes = config.max_inodes;
3477 sbinfo->free_inodes = config.max_inodes - inodes;
3480 * Preserve previous mempolicy unless mpol remount option was specified.
3483 mpol_put(sbinfo->mpol);
3484 sbinfo->mpol = config.mpol; /* transfers initial ref */
3487 spin_unlock(&sbinfo->stat_lock);
3491 static int shmem_show_options(struct seq_file *seq, struct dentry *root)
3493 struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
3495 if (sbinfo->max_blocks != shmem_default_max_blocks())
3496 seq_printf(seq, ",size=%luk",
3497 sbinfo->max_blocks << (PAGE_SHIFT - 10));
3498 if (sbinfo->max_inodes != shmem_default_max_inodes())
3499 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
3500 if (sbinfo->mode != (0777 | S_ISVTX))
3501 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
3502 if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
3503 seq_printf(seq, ",uid=%u",
3504 from_kuid_munged(&init_user_ns, sbinfo->uid));
3505 if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
3506 seq_printf(seq, ",gid=%u",
3507 from_kgid_munged(&init_user_ns, sbinfo->gid));
3508 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3509 /* Rightly or wrongly, show huge mount option unmasked by shmem_huge */
3511 seq_printf(seq, ",huge=%s", shmem_format_huge(sbinfo->huge));
3513 shmem_show_mpol(seq, sbinfo->mpol);
3517 #endif /* CONFIG_TMPFS */
3519 static void shmem_put_super(struct super_block *sb)
3521 struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
3523 percpu_counter_destroy(&sbinfo->used_blocks);
3524 mpol_put(sbinfo->mpol);
3526 sb->s_fs_info = NULL;
3529 int shmem_fill_super(struct super_block *sb, void *data, int silent)
3531 struct inode *inode;
3532 struct shmem_sb_info *sbinfo;
3535 /* Round up to L1_CACHE_BYTES to resist false sharing */
3536 sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
3537 L1_CACHE_BYTES), GFP_KERNEL);
3541 sbinfo->mode = 0777 | S_ISVTX;
3542 sbinfo->uid = current_fsuid();
3543 sbinfo->gid = current_fsgid();
3544 sb->s_fs_info = sbinfo;
3548 * Per default we only allow half of the physical ram per
3549 * tmpfs instance, limiting inodes to one per page of lowmem;
3550 * but the internal instance is left unlimited.
3552 if (!(sb->s_flags & SB_KERNMOUNT)) {
3553 sbinfo->max_blocks = shmem_default_max_blocks();
3554 sbinfo->max_inodes = shmem_default_max_inodes();
3555 if (shmem_parse_options(data, sbinfo, false)) {
3560 sb->s_flags |= SB_NOUSER;
3562 sb->s_export_op = &shmem_export_ops;
3563 sb->s_flags |= SB_NOSEC;
3565 sb->s_flags |= SB_NOUSER;
3568 spin_lock_init(&sbinfo->stat_lock);
3569 if (percpu_counter_init(&sbinfo->used_blocks, 0, GFP_KERNEL))
3571 sbinfo->free_inodes = sbinfo->max_inodes;
3572 spin_lock_init(&sbinfo->shrinklist_lock);
3573 INIT_LIST_HEAD(&sbinfo->shrinklist);
3575 sb->s_maxbytes = MAX_LFS_FILESIZE;
3576 sb->s_blocksize = PAGE_SIZE;
3577 sb->s_blocksize_bits = PAGE_SHIFT;
3578 sb->s_magic = TMPFS_MAGIC;
3579 sb->s_op = &shmem_ops;
3580 sb->s_time_gran = 1;
3581 #ifdef CONFIG_TMPFS_XATTR
3582 sb->s_xattr = shmem_xattr_handlers;
3584 #ifdef CONFIG_TMPFS_POSIX_ACL
3585 sb->s_flags |= SB_POSIXACL;
3587 uuid_gen(&sb->s_uuid);
3589 inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
3592 inode->i_uid = sbinfo->uid;
3593 inode->i_gid = sbinfo->gid;
3594 sb->s_root = d_make_root(inode);
3600 shmem_put_super(sb);
3604 static struct kmem_cache *shmem_inode_cachep;
3606 static struct inode *shmem_alloc_inode(struct super_block *sb)
3608 struct shmem_inode_info *info;
3609 info = kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
3612 return &info->vfs_inode;
3615 static void shmem_destroy_callback(struct rcu_head *head)
3617 struct inode *inode = container_of(head, struct inode, i_rcu);
3618 if (S_ISLNK(inode->i_mode))
3619 kfree(inode->i_link);
3620 kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
3623 static void shmem_destroy_inode(struct inode *inode)
3625 if (S_ISREG(inode->i_mode))
3626 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
3627 call_rcu(&inode->i_rcu, shmem_destroy_callback);
3630 static void shmem_init_inode(void *foo)
3632 struct shmem_inode_info *info = foo;
3633 inode_init_once(&info->vfs_inode);
3636 static void shmem_init_inodecache(void)
3638 shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
3639 sizeof(struct shmem_inode_info),
3640 0, SLAB_PANIC|SLAB_ACCOUNT, shmem_init_inode);
3643 static void shmem_destroy_inodecache(void)
3645 kmem_cache_destroy(shmem_inode_cachep);
3648 static const struct address_space_operations shmem_aops = {
3649 .writepage = shmem_writepage,
3650 .set_page_dirty = __set_page_dirty_no_writeback,
3652 .write_begin = shmem_write_begin,
3653 .write_end = shmem_write_end,
3655 #ifdef CONFIG_MIGRATION
3656 .migratepage = migrate_page,
3658 .error_remove_page = generic_error_remove_page,
3661 static const struct file_operations shmem_file_operations = {
3663 .get_unmapped_area = shmem_get_unmapped_area,
3665 .llseek = shmem_file_llseek,
3666 .read_iter = shmem_file_read_iter,
3667 .write_iter = generic_file_write_iter,
3668 .fsync = noop_fsync,
3669 .splice_read = generic_file_splice_read,
3670 .splice_write = iter_file_splice_write,
3671 .fallocate = shmem_fallocate,
3675 static const struct inode_operations shmem_inode_operations = {
3676 .getattr = shmem_getattr,
3677 .setattr = shmem_setattr,
3678 #ifdef CONFIG_TMPFS_XATTR
3679 .listxattr = shmem_listxattr,
3680 .set_acl = simple_set_acl,
3684 static const struct inode_operations shmem_dir_inode_operations = {
3686 .create = shmem_create,
3687 .lookup = simple_lookup,
3689 .unlink = shmem_unlink,
3690 .symlink = shmem_symlink,
3691 .mkdir = shmem_mkdir,
3692 .rmdir = shmem_rmdir,
3693 .mknod = shmem_mknod,
3694 .rename = shmem_rename2,
3695 .tmpfile = shmem_tmpfile,
3697 #ifdef CONFIG_TMPFS_XATTR
3698 .listxattr = shmem_listxattr,
3700 #ifdef CONFIG_TMPFS_POSIX_ACL
3701 .setattr = shmem_setattr,
3702 .set_acl = simple_set_acl,
3706 static const struct inode_operations shmem_special_inode_operations = {
3707 #ifdef CONFIG_TMPFS_XATTR
3708 .listxattr = shmem_listxattr,
3710 #ifdef CONFIG_TMPFS_POSIX_ACL
3711 .setattr = shmem_setattr,
3712 .set_acl = simple_set_acl,
3716 static const struct super_operations shmem_ops = {
3717 .alloc_inode = shmem_alloc_inode,
3718 .destroy_inode = shmem_destroy_inode,
3720 .statfs = shmem_statfs,
3721 .remount_fs = shmem_remount_fs,
3722 .show_options = shmem_show_options,
3724 .evict_inode = shmem_evict_inode,
3725 .drop_inode = generic_delete_inode,
3726 .put_super = shmem_put_super,
3727 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3728 .nr_cached_objects = shmem_unused_huge_count,
3729 .free_cached_objects = shmem_unused_huge_scan,
3733 static const struct vm_operations_struct shmem_vm_ops = {
3734 .fault = shmem_fault,
3735 .map_pages = filemap_map_pages,
3737 .set_policy = shmem_set_policy,
3738 .get_policy = shmem_get_policy,
3742 static struct dentry *shmem_mount(struct file_system_type *fs_type,
3743 int flags, const char *dev_name, void *data)
3745 return mount_nodev(fs_type, flags, data, shmem_fill_super);
3748 static struct file_system_type shmem_fs_type = {
3749 .owner = THIS_MODULE,
3751 .mount = shmem_mount,
3752 .kill_sb = kill_litter_super,
3753 .fs_flags = FS_USERNS_MOUNT,
3756 int __init shmem_init(void)
3760 /* If rootfs called this, don't re-init */
3761 if (shmem_inode_cachep)
3764 shmem_init_inodecache();
3766 error = register_filesystem(&shmem_fs_type);
3768 pr_err("Could not register tmpfs\n");
3772 shm_mnt = kern_mount(&shmem_fs_type);
3773 if (IS_ERR(shm_mnt)) {
3774 error = PTR_ERR(shm_mnt);
3775 pr_err("Could not kern_mount tmpfs\n");
3779 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3780 if (has_transparent_hugepage() && shmem_huge > SHMEM_HUGE_DENY)
3781 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
3783 shmem_huge = 0; /* just in case it was patched */
3788 unregister_filesystem(&shmem_fs_type);
3790 shmem_destroy_inodecache();
3791 shm_mnt = ERR_PTR(error);
3795 #if defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE) && defined(CONFIG_SYSFS)
3796 static ssize_t shmem_enabled_show(struct kobject *kobj,
3797 struct kobj_attribute *attr, char *buf)
3801 SHMEM_HUGE_WITHIN_SIZE,
3809 for (i = 0, count = 0; i < ARRAY_SIZE(values); i++) {
3810 const char *fmt = shmem_huge == values[i] ? "[%s] " : "%s ";
3812 count += sprintf(buf + count, fmt,
3813 shmem_format_huge(values[i]));
3815 buf[count - 1] = '\n';
3819 static ssize_t shmem_enabled_store(struct kobject *kobj,
3820 struct kobj_attribute *attr, const char *buf, size_t count)
3825 if (count + 1 > sizeof(tmp))
3827 memcpy(tmp, buf, count);
3829 if (count && tmp[count - 1] == '\n')
3830 tmp[count - 1] = '\0';
3832 huge = shmem_parse_huge(tmp);
3833 if (huge == -EINVAL)
3835 if (!has_transparent_hugepage() &&
3836 huge != SHMEM_HUGE_NEVER && huge != SHMEM_HUGE_DENY)
3840 if (shmem_huge > SHMEM_HUGE_DENY)
3841 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
3845 struct kobj_attribute shmem_enabled_attr =
3846 __ATTR(shmem_enabled, 0644, shmem_enabled_show, shmem_enabled_store);
3847 #endif /* CONFIG_TRANSPARENT_HUGE_PAGECACHE && CONFIG_SYSFS */
3849 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3850 bool shmem_huge_enabled(struct vm_area_struct *vma)
3852 struct inode *inode = file_inode(vma->vm_file);
3853 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
3857 if (shmem_huge == SHMEM_HUGE_FORCE)
3859 if (shmem_huge == SHMEM_HUGE_DENY)
3861 switch (sbinfo->huge) {
3862 case SHMEM_HUGE_NEVER:
3864 case SHMEM_HUGE_ALWAYS:
3866 case SHMEM_HUGE_WITHIN_SIZE:
3867 off = round_up(vma->vm_pgoff, HPAGE_PMD_NR);
3868 i_size = round_up(i_size_read(inode), PAGE_SIZE);
3869 if (i_size >= HPAGE_PMD_SIZE &&
3870 i_size >> PAGE_SHIFT >= off)
3873 case SHMEM_HUGE_ADVISE:
3874 /* TODO: implement fadvise() hints */
3875 return (vma->vm_flags & VM_HUGEPAGE);
3881 #endif /* CONFIG_TRANSPARENT_HUGE_PAGECACHE */
3883 #else /* !CONFIG_SHMEM */
3886 * tiny-shmem: simple shmemfs and tmpfs using ramfs code
3888 * This is intended for small system where the benefits of the full
3889 * shmem code (swap-backed and resource-limited) are outweighed by
3890 * their complexity. On systems without swap this code should be
3891 * effectively equivalent, but much lighter weight.
3894 static struct file_system_type shmem_fs_type = {
3896 .mount = ramfs_mount,
3897 .kill_sb = kill_litter_super,
3898 .fs_flags = FS_USERNS_MOUNT,
3901 int __init shmem_init(void)
3903 BUG_ON(register_filesystem(&shmem_fs_type) != 0);
3905 shm_mnt = kern_mount(&shmem_fs_type);
3906 BUG_ON(IS_ERR(shm_mnt));
3911 int shmem_unuse(swp_entry_t swap, struct page *page)
3916 int shmem_lock(struct file *file, int lock, struct user_struct *user)
3921 void shmem_unlock_mapping(struct address_space *mapping)
3926 unsigned long shmem_get_unmapped_area(struct file *file,
3927 unsigned long addr, unsigned long len,
3928 unsigned long pgoff, unsigned long flags)
3930 return current->mm->get_unmapped_area(file, addr, len, pgoff, flags);
3934 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
3936 truncate_inode_pages_range(inode->i_mapping, lstart, lend);
3938 EXPORT_SYMBOL_GPL(shmem_truncate_range);
3940 #define shmem_vm_ops generic_file_vm_ops
3941 #define shmem_file_operations ramfs_file_operations
3942 #define shmem_get_inode(sb, dir, mode, dev, flags) ramfs_get_inode(sb, dir, mode, dev)
3943 #define shmem_acct_size(flags, size) 0
3944 #define shmem_unacct_size(flags, size) do {} while (0)
3946 #endif /* CONFIG_SHMEM */
3950 static struct file *__shmem_file_setup(struct vfsmount *mnt, const char *name, loff_t size,
3951 unsigned long flags, unsigned int i_flags)
3953 struct inode *inode;
3957 return ERR_CAST(mnt);
3959 if (size < 0 || size > MAX_LFS_FILESIZE)
3960 return ERR_PTR(-EINVAL);
3962 if (shmem_acct_size(flags, size))
3963 return ERR_PTR(-ENOMEM);
3965 inode = shmem_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0,
3967 if (unlikely(!inode)) {
3968 shmem_unacct_size(flags, size);
3969 return ERR_PTR(-ENOSPC);
3971 inode->i_flags |= i_flags;
3972 inode->i_size = size;
3973 clear_nlink(inode); /* It is unlinked */
3974 res = ERR_PTR(ramfs_nommu_expand_for_mapping(inode, size));
3976 res = alloc_file_pseudo(inode, mnt, name, O_RDWR,
3977 &shmem_file_operations);
3984 * shmem_kernel_file_setup - get an unlinked file living in tmpfs which must be
3985 * kernel internal. There will be NO LSM permission checks against the
3986 * underlying inode. So users of this interface must do LSM checks at a
3987 * higher layer. The users are the big_key and shm implementations. LSM
3988 * checks are provided at the key or shm level rather than the inode.
3989 * @name: name for dentry (to be seen in /proc/<pid>/maps
3990 * @size: size to be set for the file
3991 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
3993 struct file *shmem_kernel_file_setup(const char *name, loff_t size, unsigned long flags)
3995 return __shmem_file_setup(shm_mnt, name, size, flags, S_PRIVATE);
3999 * shmem_file_setup - get an unlinked file living in tmpfs
4000 * @name: name for dentry (to be seen in /proc/<pid>/maps
4001 * @size: size to be set for the file
4002 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4004 struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
4006 return __shmem_file_setup(shm_mnt, name, size, flags, 0);
4008 EXPORT_SYMBOL_GPL(shmem_file_setup);
4011 * shmem_file_setup_with_mnt - get an unlinked file living in tmpfs
4012 * @mnt: the tmpfs mount where the file will be created
4013 * @name: name for dentry (to be seen in /proc/<pid>/maps
4014 * @size: size to be set for the file
4015 * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
4017 struct file *shmem_file_setup_with_mnt(struct vfsmount *mnt, const char *name,
4018 loff_t size, unsigned long flags)
4020 return __shmem_file_setup(mnt, name, size, flags, 0);
4022 EXPORT_SYMBOL_GPL(shmem_file_setup_with_mnt);
4025 * shmem_zero_setup - setup a shared anonymous mapping
4026 * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
4028 int shmem_zero_setup(struct vm_area_struct *vma)
4031 loff_t size = vma->vm_end - vma->vm_start;
4034 * Cloning a new file under mmap_sem leads to a lock ordering conflict
4035 * between XFS directory reading and selinux: since this file is only
4036 * accessible to the user through its mapping, use S_PRIVATE flag to
4037 * bypass file security, in the same way as shmem_kernel_file_setup().
4039 file = shmem_kernel_file_setup("dev/zero", size, vma->vm_flags);
4041 return PTR_ERR(file);
4045 vma->vm_file = file;
4046 vma->vm_ops = &shmem_vm_ops;
4048 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE) &&
4049 ((vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK) <
4050 (vma->vm_end & HPAGE_PMD_MASK)) {
4051 khugepaged_enter(vma, vma->vm_flags);
4058 * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
4059 * @mapping: the page's address_space
4060 * @index: the page index
4061 * @gfp: the page allocator flags to use if allocating
4063 * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
4064 * with any new page allocations done using the specified allocation flags.
4065 * But read_cache_page_gfp() uses the ->readpage() method: which does not
4066 * suit tmpfs, since it may have pages in swapcache, and needs to find those
4067 * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
4069 * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
4070 * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
4072 struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
4073 pgoff_t index, gfp_t gfp)
4076 struct inode *inode = mapping->host;
4080 BUG_ON(mapping->a_ops != &shmem_aops);
4081 error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE,
4082 gfp, NULL, NULL, NULL);
4084 page = ERR_PTR(error);
4090 * The tiny !SHMEM case uses ramfs without swap
4092 return read_cache_page_gfp(mapping, index, gfp);
4095 EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);