UPSTREAM mm: memcg: enable memcg OOM killer only for user faults
[linux-3.10.git] / mm / shmem.c
1 /*
2  * Resizable virtual memory filesystem for Linux.
3  *
4  * Copyright (C) 2000 Linus Torvalds.
5  *               2000 Transmeta Corp.
6  *               2000-2001 Christoph Rohland
7  *               2000-2001 SAP AG
8  *               2002 Red Hat Inc.
9  * Copyright (C) 2002-2011 Hugh Dickins.
10  * Copyright (C) 2011 Google Inc.
11  * Copyright (C) 2002-2005 VERITAS Software Corporation.
12  * Copyright (C) 2004 Andi Kleen, SuSE Labs
13  *
14  * Extended attribute support for tmpfs:
15  * Copyright (c) 2004, Luke Kenneth Casson Leighton <lkcl@lkcl.net>
16  * Copyright (c) 2004 Red Hat, Inc., James Morris <jmorris@redhat.com>
17  *
18  * tiny-shmem:
19  * Copyright (c) 2004, 2008 Matt Mackall <mpm@selenic.com>
20  *
21  * This file is released under the GPL.
22  */
23
24 #include <linux/fs.h>
25 #include <linux/init.h>
26 #include <linux/vfs.h>
27 #include <linux/mount.h>
28 #include <linux/ramfs.h>
29 #include <linux/pagemap.h>
30 #include <linux/file.h>
31 #include <linux/mm.h>
32 #include <linux/export.h>
33 #include <linux/swap.h>
34 #include <linux/aio.h>
35
36 static struct vfsmount *shm_mnt;
37
38 #ifdef CONFIG_SHMEM
39 /*
40  * This virtual memory filesystem is heavily based on the ramfs. It
41  * extends ramfs by the ability to use swap and honor resource limits
42  * which makes it a completely usable filesystem.
43  */
44
45 #include <linux/xattr.h>
46 #include <linux/exportfs.h>
47 #include <linux/posix_acl.h>
48 #include <linux/generic_acl.h>
49 #include <linux/mman.h>
50 #include <linux/string.h>
51 #include <linux/slab.h>
52 #include <linux/backing-dev.h>
53 #include <linux/shmem_fs.h>
54 #include <linux/writeback.h>
55 #include <linux/blkdev.h>
56 #include <linux/pagevec.h>
57 #include <linux/percpu_counter.h>
58 #include <linux/falloc.h>
59 #include <linux/splice.h>
60 #include <linux/security.h>
61 #include <linux/swapops.h>
62 #include <linux/mempolicy.h>
63 #include <linux/namei.h>
64 #include <linux/ctype.h>
65 #include <linux/migrate.h>
66 #include <linux/highmem.h>
67 #include <linux/seq_file.h>
68 #include <linux/magic.h>
69
70 #include <asm/uaccess.h>
71 #include <asm/pgtable.h>
72
73 #define BLOCKS_PER_PAGE  (PAGE_CACHE_SIZE/512)
74 #define VM_ACCT(size)    (PAGE_CACHE_ALIGN(size) >> PAGE_SHIFT)
75
76 /* Pretend that each entry is of this size in directory's i_size */
77 #define BOGO_DIRENT_SIZE 20
78
79 /* Symlink up to this size is kmalloc'ed instead of using a swappable page */
80 #define SHORT_SYMLINK_LEN 128
81
82 /*
83  * shmem_fallocate and shmem_writepage communicate via inode->i_private
84  * (with i_mutex making sure that it has only one user at a time):
85  * we would prefer not to enlarge the shmem inode just for that.
86  */
87 struct shmem_falloc {
88         pgoff_t start;          /* start of range currently being fallocated */
89         pgoff_t next;           /* the next page offset to be fallocated */
90         pgoff_t nr_falloced;    /* how many new pages have been fallocated */
91         pgoff_t nr_unswapped;   /* how often writepage refused to swap out */
92 };
93
94 /* Flag allocation requirements to shmem_getpage */
95 enum sgp_type {
96         SGP_READ,       /* don't exceed i_size, don't allocate page */
97         SGP_CACHE,      /* don't exceed i_size, may allocate page */
98         SGP_DIRTY,      /* like SGP_CACHE, but set new page dirty */
99         SGP_WRITE,      /* may exceed i_size, may allocate !Uptodate page */
100         SGP_FALLOC,     /* like SGP_WRITE, but make existing page Uptodate */
101 };
102
103 #ifdef CONFIG_TMPFS
104 static unsigned long shmem_default_max_blocks(void)
105 {
106         return totalram_pages / 2;
107 }
108
109 static unsigned long shmem_default_max_inodes(void)
110 {
111         return min(totalram_pages - totalhigh_pages, totalram_pages / 2);
112 }
113 #endif
114
115 static bool shmem_should_replace_page(struct page *page, gfp_t gfp);
116 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
117                                 struct shmem_inode_info *info, pgoff_t index);
118 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
119         struct page **pagep, enum sgp_type sgp, gfp_t gfp, int *fault_type);
120
121 static inline int shmem_getpage(struct inode *inode, pgoff_t index,
122         struct page **pagep, enum sgp_type sgp, int *fault_type)
123 {
124         return shmem_getpage_gfp(inode, index, pagep, sgp,
125                         mapping_gfp_mask(inode->i_mapping), fault_type);
126 }
127
128 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
129 {
130         return sb->s_fs_info;
131 }
132
133 /*
134  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
135  * for shared memory and for shared anonymous (/dev/zero) mappings
136  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
137  * consistent with the pre-accounting of private mappings ...
138  */
139 static inline int shmem_acct_size(unsigned long flags, loff_t size)
140 {
141         return (flags & VM_NORESERVE) ?
142                 0 : security_vm_enough_memory_mm(current->mm, VM_ACCT(size));
143 }
144
145 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
146 {
147         if (!(flags & VM_NORESERVE))
148                 vm_unacct_memory(VM_ACCT(size));
149 }
150
151 /*
152  * ... whereas tmpfs objects are accounted incrementally as
153  * pages are allocated, in order to allow huge sparse files.
154  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
155  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
156  */
157 static inline int shmem_acct_block(unsigned long flags)
158 {
159         return (flags & VM_NORESERVE) ?
160                 security_vm_enough_memory_mm(current->mm, VM_ACCT(PAGE_CACHE_SIZE)) : 0;
161 }
162
163 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
164 {
165         if (flags & VM_NORESERVE)
166                 vm_unacct_memory(pages * VM_ACCT(PAGE_CACHE_SIZE));
167 }
168
169 static const struct super_operations shmem_ops;
170 static const struct address_space_operations shmem_aops;
171 static const struct file_operations shmem_file_operations;
172 static const struct inode_operations shmem_inode_operations;
173 static const struct inode_operations shmem_dir_inode_operations;
174 static const struct inode_operations shmem_special_inode_operations;
175 static const struct vm_operations_struct shmem_vm_ops;
176
177 static struct backing_dev_info shmem_backing_dev_info  __read_mostly = {
178         .ra_pages       = 0,    /* No readahead */
179         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_SWAP_BACKED,
180 };
181
182 static LIST_HEAD(shmem_swaplist);
183 static DEFINE_MUTEX(shmem_swaplist_mutex);
184
185 static int shmem_reserve_inode(struct super_block *sb)
186 {
187         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
188         if (sbinfo->max_inodes) {
189                 spin_lock(&sbinfo->stat_lock);
190                 if (!sbinfo->free_inodes) {
191                         spin_unlock(&sbinfo->stat_lock);
192                         return -ENOSPC;
193                 }
194                 sbinfo->free_inodes--;
195                 spin_unlock(&sbinfo->stat_lock);
196         }
197         return 0;
198 }
199
200 static void shmem_free_inode(struct super_block *sb)
201 {
202         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
203         if (sbinfo->max_inodes) {
204                 spin_lock(&sbinfo->stat_lock);
205                 sbinfo->free_inodes++;
206                 spin_unlock(&sbinfo->stat_lock);
207         }
208 }
209
210 /**
211  * shmem_recalc_inode - recalculate the block usage of an inode
212  * @inode: inode to recalc
213  *
214  * We have to calculate the free blocks since the mm can drop
215  * undirtied hole pages behind our back.
216  *
217  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
218  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
219  *
220  * It has to be called with the spinlock held.
221  */
222 static void shmem_recalc_inode(struct inode *inode)
223 {
224         struct shmem_inode_info *info = SHMEM_I(inode);
225         long freed;
226
227         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
228         if (freed > 0) {
229                 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
230                 if (sbinfo->max_blocks)
231                         percpu_counter_add(&sbinfo->used_blocks, -freed);
232                 info->alloced -= freed;
233                 inode->i_blocks -= freed * BLOCKS_PER_PAGE;
234                 shmem_unacct_blocks(info->flags, freed);
235         }
236 }
237
238 /*
239  * Replace item expected in radix tree by a new item, while holding tree lock.
240  */
241 static int shmem_radix_tree_replace(struct address_space *mapping,
242                         pgoff_t index, void *expected, void *replacement)
243 {
244         void **pslot;
245         void *item = NULL;
246
247         VM_BUG_ON(!expected);
248         pslot = radix_tree_lookup_slot(&mapping->page_tree, index);
249         if (pslot)
250                 item = radix_tree_deref_slot_protected(pslot,
251                                                         &mapping->tree_lock);
252         if (item != expected)
253                 return -ENOENT;
254         if (replacement)
255                 radix_tree_replace_slot(pslot, replacement);
256         else
257                 radix_tree_delete(&mapping->page_tree, index);
258         return 0;
259 }
260
261 /*
262  * Sometimes, before we decide whether to proceed or to fail, we must check
263  * that an entry was not already brought back from swap by a racing thread.
264  *
265  * Checking page is not enough: by the time a SwapCache page is locked, it
266  * might be reused, and again be SwapCache, using the same swap as before.
267  */
268 static bool shmem_confirm_swap(struct address_space *mapping,
269                                pgoff_t index, swp_entry_t swap)
270 {
271         void *item;
272
273         rcu_read_lock();
274         item = radix_tree_lookup(&mapping->page_tree, index);
275         rcu_read_unlock();
276         return item == swp_to_radix_entry(swap);
277 }
278
279 /*
280  * Like add_to_page_cache_locked, but error if expected item has gone.
281  */
282 static int shmem_add_to_page_cache(struct page *page,
283                                    struct address_space *mapping,
284                                    pgoff_t index, gfp_t gfp, void *expected)
285 {
286         int error;
287
288         VM_BUG_ON(!PageLocked(page));
289         VM_BUG_ON(!PageSwapBacked(page));
290
291         page_cache_get(page);
292         page->mapping = mapping;
293         page->index = index;
294
295         spin_lock_irq(&mapping->tree_lock);
296         if (!expected)
297                 error = radix_tree_insert(&mapping->page_tree, index, page);
298         else
299                 error = shmem_radix_tree_replace(mapping, index, expected,
300                                                                  page);
301         if (!error) {
302                 mapping->nrpages++;
303                 __inc_zone_page_state(page, NR_FILE_PAGES);
304                 __inc_zone_page_state(page, NR_SHMEM);
305                 spin_unlock_irq(&mapping->tree_lock);
306         } else {
307                 page->mapping = NULL;
308                 spin_unlock_irq(&mapping->tree_lock);
309                 page_cache_release(page);
310         }
311         return error;
312 }
313
314 /*
315  * Like delete_from_page_cache, but substitutes swap for page.
316  */
317 static void shmem_delete_from_page_cache(struct page *page, void *radswap)
318 {
319         struct address_space *mapping = page->mapping;
320         int error;
321
322         spin_lock_irq(&mapping->tree_lock);
323         error = shmem_radix_tree_replace(mapping, page->index, page, radswap);
324         page->mapping = NULL;
325         mapping->nrpages--;
326         __dec_zone_page_state(page, NR_FILE_PAGES);
327         __dec_zone_page_state(page, NR_SHMEM);
328         spin_unlock_irq(&mapping->tree_lock);
329         page_cache_release(page);
330         BUG_ON(error);
331 }
332
333 /*
334  * Like find_get_pages, but collecting swap entries as well as pages.
335  */
336 static unsigned shmem_find_get_pages_and_swap(struct address_space *mapping,
337                                         pgoff_t start, unsigned int nr_pages,
338                                         struct page **pages, pgoff_t *indices)
339 {
340         void **slot;
341         unsigned int ret = 0;
342         struct radix_tree_iter iter;
343
344         if (!nr_pages)
345                 return 0;
346
347         rcu_read_lock();
348 restart:
349         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
350                 struct page *page;
351 repeat:
352                 page = radix_tree_deref_slot(slot);
353                 if (unlikely(!page))
354                         continue;
355                 if (radix_tree_exception(page)) {
356                         if (radix_tree_deref_retry(page))
357                                 goto restart;
358                         /*
359                          * Otherwise, we must be storing a swap entry
360                          * here as an exceptional entry: so return it
361                          * without attempting to raise page count.
362                          */
363                         goto export;
364                 }
365                 if (!page_cache_get_speculative(page))
366                         goto repeat;
367
368                 /* Has the page moved? */
369                 if (unlikely(page != *slot)) {
370                         page_cache_release(page);
371                         goto repeat;
372                 }
373 export:
374                 indices[ret] = iter.index;
375                 pages[ret] = page;
376                 if (++ret == nr_pages)
377                         break;
378         }
379         rcu_read_unlock();
380         return ret;
381 }
382
383 /*
384  * Remove swap entry from radix tree, free the swap and its page cache.
385  */
386 static int shmem_free_swap(struct address_space *mapping,
387                            pgoff_t index, void *radswap)
388 {
389         int error;
390
391         spin_lock_irq(&mapping->tree_lock);
392         error = shmem_radix_tree_replace(mapping, index, radswap, NULL);
393         spin_unlock_irq(&mapping->tree_lock);
394         if (!error)
395                 free_swap_and_cache(radix_to_swp_entry(radswap));
396         return error;
397 }
398
399 /*
400  * Pagevec may contain swap entries, so shuffle up pages before releasing.
401  */
402 static void shmem_deswap_pagevec(struct pagevec *pvec)
403 {
404         int i, j;
405
406         for (i = 0, j = 0; i < pagevec_count(pvec); i++) {
407                 struct page *page = pvec->pages[i];
408                 if (!radix_tree_exceptional_entry(page))
409                         pvec->pages[j++] = page;
410         }
411         pvec->nr = j;
412 }
413
414 /*
415  * SysV IPC SHM_UNLOCK restore Unevictable pages to their evictable lists.
416  */
417 void shmem_unlock_mapping(struct address_space *mapping)
418 {
419         struct pagevec pvec;
420         pgoff_t indices[PAGEVEC_SIZE];
421         pgoff_t index = 0;
422
423         pagevec_init(&pvec, 0);
424         /*
425          * Minor point, but we might as well stop if someone else SHM_LOCKs it.
426          */
427         while (!mapping_unevictable(mapping)) {
428                 /*
429                  * Avoid pagevec_lookup(): find_get_pages() returns 0 as if it
430                  * has finished, if it hits a row of PAGEVEC_SIZE swap entries.
431                  */
432                 pvec.nr = shmem_find_get_pages_and_swap(mapping, index,
433                                         PAGEVEC_SIZE, pvec.pages, indices);
434                 if (!pvec.nr)
435                         break;
436                 index = indices[pvec.nr - 1] + 1;
437                 shmem_deswap_pagevec(&pvec);
438                 check_move_unevictable_pages(pvec.pages, pvec.nr);
439                 pagevec_release(&pvec);
440                 cond_resched();
441         }
442 }
443
444 /*
445  * Remove range of pages and swap entries from radix tree, and free them.
446  * If !unfalloc, truncate or punch hole; if unfalloc, undo failed fallocate.
447  */
448 static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
449                                                                  bool unfalloc)
450 {
451         struct address_space *mapping = inode->i_mapping;
452         struct shmem_inode_info *info = SHMEM_I(inode);
453         pgoff_t start = (lstart + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
454         pgoff_t end = (lend + 1) >> PAGE_CACHE_SHIFT;
455         unsigned int partial_start = lstart & (PAGE_CACHE_SIZE - 1);
456         unsigned int partial_end = (lend + 1) & (PAGE_CACHE_SIZE - 1);
457         struct pagevec pvec;
458         pgoff_t indices[PAGEVEC_SIZE];
459         long nr_swaps_freed = 0;
460         pgoff_t index;
461         int i;
462
463         if (lend == -1)
464                 end = -1;       /* unsigned, so actually very big */
465
466         pagevec_init(&pvec, 0);
467         index = start;
468         while (index < end) {
469                 pvec.nr = shmem_find_get_pages_and_swap(mapping, index,
470                                 min(end - index, (pgoff_t)PAGEVEC_SIZE),
471                                                         pvec.pages, indices);
472                 if (!pvec.nr)
473                         break;
474                 mem_cgroup_uncharge_start();
475                 for (i = 0; i < pagevec_count(&pvec); i++) {
476                         struct page *page = pvec.pages[i];
477
478                         index = indices[i];
479                         if (index >= end)
480                                 break;
481
482                         if (radix_tree_exceptional_entry(page)) {
483                                 if (unfalloc)
484                                         continue;
485                                 nr_swaps_freed += !shmem_free_swap(mapping,
486                                                                 index, page);
487                                 continue;
488                         }
489
490                         if (!trylock_page(page))
491                                 continue;
492                         if (!unfalloc || !PageUptodate(page)) {
493                                 if (page->mapping == mapping) {
494                                         VM_BUG_ON(PageWriteback(page));
495                                         truncate_inode_page(mapping, page);
496                                 }
497                         }
498                         unlock_page(page);
499                 }
500                 shmem_deswap_pagevec(&pvec);
501                 pagevec_release(&pvec);
502                 mem_cgroup_uncharge_end();
503                 cond_resched();
504                 index++;
505         }
506
507         if (partial_start) {
508                 struct page *page = NULL;
509                 shmem_getpage(inode, start - 1, &page, SGP_READ, NULL);
510                 if (page) {
511                         unsigned int top = PAGE_CACHE_SIZE;
512                         if (start > end) {
513                                 top = partial_end;
514                                 partial_end = 0;
515                         }
516                         zero_user_segment(page, partial_start, top);
517                         set_page_dirty(page);
518                         unlock_page(page);
519                         page_cache_release(page);
520                 }
521         }
522         if (partial_end) {
523                 struct page *page = NULL;
524                 shmem_getpage(inode, end, &page, SGP_READ, NULL);
525                 if (page) {
526                         zero_user_segment(page, 0, partial_end);
527                         set_page_dirty(page);
528                         unlock_page(page);
529                         page_cache_release(page);
530                 }
531         }
532         if (start >= end)
533                 return;
534
535         index = start;
536         for ( ; ; ) {
537                 cond_resched();
538                 pvec.nr = shmem_find_get_pages_and_swap(mapping, index,
539                                 min(end - index, (pgoff_t)PAGEVEC_SIZE),
540                                                         pvec.pages, indices);
541                 if (!pvec.nr) {
542                         if (index == start || unfalloc)
543                                 break;
544                         index = start;
545                         continue;
546                 }
547                 if ((index == start || unfalloc) && indices[0] >= end) {
548                         shmem_deswap_pagevec(&pvec);
549                         pagevec_release(&pvec);
550                         break;
551                 }
552                 mem_cgroup_uncharge_start();
553                 for (i = 0; i < pagevec_count(&pvec); i++) {
554                         struct page *page = pvec.pages[i];
555
556                         index = indices[i];
557                         if (index >= end)
558                                 break;
559
560                         if (radix_tree_exceptional_entry(page)) {
561                                 if (unfalloc)
562                                         continue;
563                                 nr_swaps_freed += !shmem_free_swap(mapping,
564                                                                 index, page);
565                                 continue;
566                         }
567
568                         lock_page(page);
569                         if (!unfalloc || !PageUptodate(page)) {
570                                 if (page->mapping == mapping) {
571                                         VM_BUG_ON(PageWriteback(page));
572                                         truncate_inode_page(mapping, page);
573                                 }
574                         }
575                         unlock_page(page);
576                 }
577                 shmem_deswap_pagevec(&pvec);
578                 pagevec_release(&pvec);
579                 mem_cgroup_uncharge_end();
580                 index++;
581         }
582
583         spin_lock(&info->lock);
584         info->swapped -= nr_swaps_freed;
585         shmem_recalc_inode(inode);
586         spin_unlock(&info->lock);
587 }
588
589 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
590 {
591         shmem_undo_range(inode, lstart, lend, false);
592         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
593 }
594 EXPORT_SYMBOL_GPL(shmem_truncate_range);
595
596 static int shmem_setattr(struct dentry *dentry, struct iattr *attr)
597 {
598         struct inode *inode = dentry->d_inode;
599         int error;
600
601         error = inode_change_ok(inode, attr);
602         if (error)
603                 return error;
604
605         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
606                 loff_t oldsize = inode->i_size;
607                 loff_t newsize = attr->ia_size;
608
609                 if (newsize != oldsize) {
610                         i_size_write(inode, newsize);
611                         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
612                 }
613                 if (newsize < oldsize) {
614                         loff_t holebegin = round_up(newsize, PAGE_SIZE);
615                         unmap_mapping_range(inode->i_mapping, holebegin, 0, 1);
616                         shmem_truncate_range(inode, newsize, (loff_t)-1);
617                         /* unmap again to remove racily COWed private pages */
618                         unmap_mapping_range(inode->i_mapping, holebegin, 0, 1);
619                 }
620         }
621
622         setattr_copy(inode, attr);
623 #ifdef CONFIG_TMPFS_POSIX_ACL
624         if (attr->ia_valid & ATTR_MODE)
625                 error = generic_acl_chmod(inode);
626 #endif
627         return error;
628 }
629
630 static void shmem_evict_inode(struct inode *inode)
631 {
632         struct shmem_inode_info *info = SHMEM_I(inode);
633
634         if (inode->i_mapping->a_ops == &shmem_aops) {
635                 shmem_unacct_size(info->flags, inode->i_size);
636                 inode->i_size = 0;
637                 shmem_truncate_range(inode, 0, (loff_t)-1);
638                 if (!list_empty(&info->swaplist)) {
639                         mutex_lock(&shmem_swaplist_mutex);
640                         list_del_init(&info->swaplist);
641                         mutex_unlock(&shmem_swaplist_mutex);
642                 }
643         } else
644                 kfree(info->symlink);
645
646         simple_xattrs_free(&info->xattrs);
647         WARN_ON(inode->i_blocks);
648         shmem_free_inode(inode->i_sb);
649         clear_inode(inode);
650 }
651
652 /*
653  * If swap found in inode, free it and move page from swapcache to filecache.
654  */
655 static int shmem_unuse_inode(struct shmem_inode_info *info,
656                              swp_entry_t swap, struct page **pagep)
657 {
658         struct address_space *mapping = info->vfs_inode.i_mapping;
659         void *radswap;
660         pgoff_t index;
661         gfp_t gfp;
662         int error = 0;
663
664         radswap = swp_to_radix_entry(swap);
665         index = radix_tree_locate_item(&mapping->page_tree, radswap);
666         if (index == -1)
667                 return 0;
668
669         /*
670          * Move _head_ to start search for next from here.
671          * But be careful: shmem_evict_inode checks list_empty without taking
672          * mutex, and there's an instant in list_move_tail when info->swaplist
673          * would appear empty, if it were the only one on shmem_swaplist.
674          */
675         if (shmem_swaplist.next != &info->swaplist)
676                 list_move_tail(&shmem_swaplist, &info->swaplist);
677
678         gfp = mapping_gfp_mask(mapping);
679         if (shmem_should_replace_page(*pagep, gfp)) {
680                 mutex_unlock(&shmem_swaplist_mutex);
681                 error = shmem_replace_page(pagep, gfp, info, index);
682                 mutex_lock(&shmem_swaplist_mutex);
683                 /*
684                  * We needed to drop mutex to make that restrictive page
685                  * allocation, but the inode might have been freed while we
686                  * dropped it: although a racing shmem_evict_inode() cannot
687                  * complete without emptying the radix_tree, our page lock
688                  * on this swapcache page is not enough to prevent that -
689                  * free_swap_and_cache() of our swap entry will only
690                  * trylock_page(), removing swap from radix_tree whatever.
691                  *
692                  * We must not proceed to shmem_add_to_page_cache() if the
693                  * inode has been freed, but of course we cannot rely on
694                  * inode or mapping or info to check that.  However, we can
695                  * safely check if our swap entry is still in use (and here
696                  * it can't have got reused for another page): if it's still
697                  * in use, then the inode cannot have been freed yet, and we
698                  * can safely proceed (if it's no longer in use, that tells
699                  * nothing about the inode, but we don't need to unuse swap).
700                  */
701                 if (!page_swapcount(*pagep))
702                         error = -ENOENT;
703         }
704
705         /*
706          * We rely on shmem_swaplist_mutex, not only to protect the swaplist,
707          * but also to hold up shmem_evict_inode(): so inode cannot be freed
708          * beneath us (pagelock doesn't help until the page is in pagecache).
709          */
710         if (!error)
711                 error = shmem_add_to_page_cache(*pagep, mapping, index,
712                                                 GFP_NOWAIT, radswap);
713         if (error != -ENOMEM) {
714                 /*
715                  * Truncation and eviction use free_swap_and_cache(), which
716                  * only does trylock page: if we raced, best clean up here.
717                  */
718                 delete_from_swap_cache(*pagep);
719                 set_page_dirty(*pagep);
720                 if (!error) {
721                         spin_lock(&info->lock);
722                         info->swapped--;
723                         spin_unlock(&info->lock);
724                         swap_free(swap);
725                 }
726                 error = 1;      /* not an error, but entry was found */
727         }
728         return error;
729 }
730
731 /*
732  * Search through swapped inodes to find and replace swap by page.
733  */
734 int shmem_unuse(swp_entry_t swap, struct page *page)
735 {
736         struct list_head *this, *next;
737         struct shmem_inode_info *info;
738         int found = 0;
739         int error = 0;
740
741         /*
742          * There's a faint possibility that swap page was replaced before
743          * caller locked it: caller will come back later with the right page.
744          */
745         if (unlikely(!PageSwapCache(page) || page_private(page) != swap.val))
746                 goto out;
747
748         /*
749          * Charge page using GFP_KERNEL while we can wait, before taking
750          * the shmem_swaplist_mutex which might hold up shmem_writepage().
751          * Charged back to the user (not to caller) when swap account is used.
752          */
753         error = mem_cgroup_cache_charge(page, current->mm, GFP_KERNEL);
754         if (error)
755                 goto out;
756         /* No radix_tree_preload: swap entry keeps a place for page in tree */
757
758         mutex_lock(&shmem_swaplist_mutex);
759         list_for_each_safe(this, next, &shmem_swaplist) {
760                 info = list_entry(this, struct shmem_inode_info, swaplist);
761                 if (info->swapped)
762                         found = shmem_unuse_inode(info, swap, &page);
763                 else
764                         list_del_init(&info->swaplist);
765                 cond_resched();
766                 if (found)
767                         break;
768         }
769         mutex_unlock(&shmem_swaplist_mutex);
770
771         if (found < 0)
772                 error = found;
773 out:
774         unlock_page(page);
775         page_cache_release(page);
776         return error;
777 }
778
779 /*
780  * Move the page from the page cache to the swap cache.
781  */
782 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
783 {
784         struct shmem_inode_info *info;
785         struct address_space *mapping;
786         struct inode *inode;
787         swp_entry_t swap;
788         pgoff_t index;
789
790         BUG_ON(!PageLocked(page));
791         mapping = page->mapping;
792         index = page->index;
793         inode = mapping->host;
794         info = SHMEM_I(inode);
795         if (info->flags & VM_LOCKED)
796                 goto redirty;
797         if (!total_swap_pages)
798                 goto redirty;
799
800         /*
801          * shmem_backing_dev_info's capabilities prevent regular writeback or
802          * sync from ever calling shmem_writepage; but a stacking filesystem
803          * might use ->writepage of its underlying filesystem, in which case
804          * tmpfs should write out to swap only in response to memory pressure,
805          * and not for the writeback threads or sync.
806          */
807         if (!wbc->for_reclaim) {
808                 WARN_ON_ONCE(1);        /* Still happens? Tell us about it! */
809                 goto redirty;
810         }
811
812         /*
813          * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
814          * value into swapfile.c, the only way we can correctly account for a
815          * fallocated page arriving here is now to initialize it and write it.
816          *
817          * That's okay for a page already fallocated earlier, but if we have
818          * not yet completed the fallocation, then (a) we want to keep track
819          * of this page in case we have to undo it, and (b) it may not be a
820          * good idea to continue anyway, once we're pushing into swap.  So
821          * reactivate the page, and let shmem_fallocate() quit when too many.
822          */
823         if (!PageUptodate(page)) {
824                 if (inode->i_private) {
825                         struct shmem_falloc *shmem_falloc;
826                         spin_lock(&inode->i_lock);
827                         shmem_falloc = inode->i_private;
828                         if (shmem_falloc &&
829                             index >= shmem_falloc->start &&
830                             index < shmem_falloc->next)
831                                 shmem_falloc->nr_unswapped++;
832                         else
833                                 shmem_falloc = NULL;
834                         spin_unlock(&inode->i_lock);
835                         if (shmem_falloc)
836                                 goto redirty;
837                 }
838                 clear_highpage(page);
839                 flush_dcache_page(page);
840                 SetPageUptodate(page);
841         }
842
843         swap = get_swap_page();
844         if (!swap.val)
845                 goto redirty;
846
847         /*
848          * Add inode to shmem_unuse()'s list of swapped-out inodes,
849          * if it's not already there.  Do it now before the page is
850          * moved to swap cache, when its pagelock no longer protects
851          * the inode from eviction.  But don't unlock the mutex until
852          * we've incremented swapped, because shmem_unuse_inode() will
853          * prune a !swapped inode from the swaplist under this mutex.
854          */
855         mutex_lock(&shmem_swaplist_mutex);
856         if (list_empty(&info->swaplist))
857                 list_add_tail(&info->swaplist, &shmem_swaplist);
858
859         if (add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
860                 swap_shmem_alloc(swap);
861                 shmem_delete_from_page_cache(page, swp_to_radix_entry(swap));
862
863                 spin_lock(&info->lock);
864                 info->swapped++;
865                 shmem_recalc_inode(inode);
866                 spin_unlock(&info->lock);
867
868                 mutex_unlock(&shmem_swaplist_mutex);
869                 BUG_ON(page_mapped(page));
870                 swap_writepage(page, wbc);
871                 return 0;
872         }
873
874         mutex_unlock(&shmem_swaplist_mutex);
875         swapcache_free(swap, NULL);
876 redirty:
877         set_page_dirty(page);
878         if (wbc->for_reclaim)
879                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
880         unlock_page(page);
881         return 0;
882 }
883
884 #ifdef CONFIG_NUMA
885 #ifdef CONFIG_TMPFS
886 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
887 {
888         char buffer[64];
889
890         if (!mpol || mpol->mode == MPOL_DEFAULT)
891                 return;         /* show nothing */
892
893         mpol_to_str(buffer, sizeof(buffer), mpol);
894
895         seq_printf(seq, ",mpol=%s", buffer);
896 }
897
898 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
899 {
900         struct mempolicy *mpol = NULL;
901         if (sbinfo->mpol) {
902                 spin_lock(&sbinfo->stat_lock);  /* prevent replace/use races */
903                 mpol = sbinfo->mpol;
904                 mpol_get(mpol);
905                 spin_unlock(&sbinfo->stat_lock);
906         }
907         return mpol;
908 }
909 #endif /* CONFIG_TMPFS */
910
911 static struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
912                         struct shmem_inode_info *info, pgoff_t index)
913 {
914         struct vm_area_struct pvma;
915         struct page *page;
916
917         /* Create a pseudo vma that just contains the policy */
918         pvma.vm_start = 0;
919         /* Bias interleave by inode number to distribute better across nodes */
920         pvma.vm_pgoff = index + info->vfs_inode.i_ino;
921         pvma.vm_ops = NULL;
922         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, index);
923
924         page = swapin_readahead(swap, gfp, &pvma, 0);
925
926         /* Drop reference taken by mpol_shared_policy_lookup() */
927         mpol_cond_put(pvma.vm_policy);
928
929         return page;
930 }
931
932 static struct page *shmem_alloc_page(gfp_t gfp,
933                         struct shmem_inode_info *info, pgoff_t index)
934 {
935         struct vm_area_struct pvma;
936         struct page *page;
937
938         /* Create a pseudo vma that just contains the policy */
939         pvma.vm_start = 0;
940         /* Bias interleave by inode number to distribute better across nodes */
941         pvma.vm_pgoff = index + info->vfs_inode.i_ino;
942         pvma.vm_ops = NULL;
943         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, index);
944
945         page = alloc_page_vma(gfp, &pvma, 0);
946
947         /* Drop reference taken by mpol_shared_policy_lookup() */
948         mpol_cond_put(pvma.vm_policy);
949
950         return page;
951 }
952 #else /* !CONFIG_NUMA */
953 #ifdef CONFIG_TMPFS
954 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
955 {
956 }
957 #endif /* CONFIG_TMPFS */
958
959 static inline struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
960                         struct shmem_inode_info *info, pgoff_t index)
961 {
962         return swapin_readahead(swap, gfp, NULL, 0);
963 }
964
965 static inline struct page *shmem_alloc_page(gfp_t gfp,
966                         struct shmem_inode_info *info, pgoff_t index)
967 {
968         return alloc_page(gfp);
969 }
970 #endif /* CONFIG_NUMA */
971
972 #if !defined(CONFIG_NUMA) || !defined(CONFIG_TMPFS)
973 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
974 {
975         return NULL;
976 }
977 #endif
978
979 /*
980  * When a page is moved from swapcache to shmem filecache (either by the
981  * usual swapin of shmem_getpage_gfp(), or by the less common swapoff of
982  * shmem_unuse_inode()), it may have been read in earlier from swap, in
983  * ignorance of the mapping it belongs to.  If that mapping has special
984  * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
985  * we may need to copy to a suitable page before moving to filecache.
986  *
987  * In a future release, this may well be extended to respect cpuset and
988  * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
989  * but for now it is a simple matter of zone.
990  */
991 static bool shmem_should_replace_page(struct page *page, gfp_t gfp)
992 {
993         return page_zonenum(page) > gfp_zone(gfp);
994 }
995
996 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
997                                 struct shmem_inode_info *info, pgoff_t index)
998 {
999         struct page *oldpage, *newpage;
1000         struct address_space *swap_mapping;
1001         pgoff_t swap_index;
1002         int error;
1003
1004         oldpage = *pagep;
1005         swap_index = page_private(oldpage);
1006         swap_mapping = page_mapping(oldpage);
1007
1008         /*
1009          * We have arrived here because our zones are constrained, so don't
1010          * limit chance of success by further cpuset and node constraints.
1011          */
1012         gfp &= ~GFP_CONSTRAINT_MASK;
1013         newpage = shmem_alloc_page(gfp, info, index);
1014         if (!newpage)
1015                 return -ENOMEM;
1016
1017         page_cache_get(newpage);
1018         copy_highpage(newpage, oldpage);
1019         flush_dcache_page(newpage);
1020
1021         __set_page_locked(newpage);
1022         SetPageUptodate(newpage);
1023         SetPageSwapBacked(newpage);
1024         set_page_private(newpage, swap_index);
1025         SetPageSwapCache(newpage);
1026
1027         /*
1028          * Our caller will very soon move newpage out of swapcache, but it's
1029          * a nice clean interface for us to replace oldpage by newpage there.
1030          */
1031         spin_lock_irq(&swap_mapping->tree_lock);
1032         error = shmem_radix_tree_replace(swap_mapping, swap_index, oldpage,
1033                                                                    newpage);
1034         if (!error) {
1035                 __inc_zone_page_state(newpage, NR_FILE_PAGES);
1036                 __dec_zone_page_state(oldpage, NR_FILE_PAGES);
1037         }
1038         spin_unlock_irq(&swap_mapping->tree_lock);
1039
1040         if (unlikely(error)) {
1041                 /*
1042                  * Is this possible?  I think not, now that our callers check
1043                  * both PageSwapCache and page_private after getting page lock;
1044                  * but be defensive.  Reverse old to newpage for clear and free.
1045                  */
1046                 oldpage = newpage;
1047         } else {
1048                 mem_cgroup_replace_page_cache(oldpage, newpage);
1049                 lru_cache_add_anon(newpage);
1050                 *pagep = newpage;
1051         }
1052
1053         ClearPageSwapCache(oldpage);
1054         set_page_private(oldpage, 0);
1055
1056         unlock_page(oldpage);
1057         page_cache_release(oldpage);
1058         page_cache_release(oldpage);
1059         return error;
1060 }
1061
1062 /*
1063  * shmem_getpage_gfp - find page in cache, or get from swap, or allocate
1064  *
1065  * If we allocate a new one we do not mark it dirty. That's up to the
1066  * vm. If we swap it in we mark it dirty since we also free the swap
1067  * entry since a page cannot live in both the swap and page cache
1068  */
1069 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
1070         struct page **pagep, enum sgp_type sgp, gfp_t gfp, int *fault_type)
1071 {
1072         struct address_space *mapping = inode->i_mapping;
1073         struct shmem_inode_info *info;
1074         struct shmem_sb_info *sbinfo;
1075         struct page *page;
1076         swp_entry_t swap;
1077         int error;
1078         int once = 0;
1079         int alloced = 0;
1080
1081         if (index > (MAX_LFS_FILESIZE >> PAGE_CACHE_SHIFT))
1082                 return -EFBIG;
1083 repeat:
1084         swap.val = 0;
1085         page = find_lock_page(mapping, index);
1086         if (radix_tree_exceptional_entry(page)) {
1087                 swap = radix_to_swp_entry(page);
1088                 page = NULL;
1089         }
1090
1091         if (sgp != SGP_WRITE && sgp != SGP_FALLOC &&
1092             ((loff_t)index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
1093                 error = -EINVAL;
1094                 goto failed;
1095         }
1096
1097         /* fallocated page? */
1098         if (page && !PageUptodate(page)) {
1099                 if (sgp != SGP_READ)
1100                         goto clear;
1101                 unlock_page(page);
1102                 page_cache_release(page);
1103                 page = NULL;
1104         }
1105         if (page || (sgp == SGP_READ && !swap.val)) {
1106                 *pagep = page;
1107                 return 0;
1108         }
1109
1110         /*
1111          * Fast cache lookup did not find it:
1112          * bring it back from swap or allocate.
1113          */
1114         info = SHMEM_I(inode);
1115         sbinfo = SHMEM_SB(inode->i_sb);
1116
1117         if (swap.val) {
1118                 /* Look it up and read it in.. */
1119                 page = lookup_swap_cache(swap);
1120                 if (!page) {
1121                         /* here we actually do the io */
1122                         if (fault_type)
1123                                 *fault_type |= VM_FAULT_MAJOR;
1124                         page = shmem_swapin(swap, gfp, info, index);
1125                         if (!page) {
1126                                 error = -ENOMEM;
1127                                 goto failed;
1128                         }
1129                 }
1130
1131                 /* We have to do this with page locked to prevent races */
1132                 lock_page(page);
1133                 if (!PageSwapCache(page) || page_private(page) != swap.val ||
1134                     !shmem_confirm_swap(mapping, index, swap)) {
1135                         error = -EEXIST;        /* try again */
1136                         goto unlock;
1137                 }
1138                 if (!PageUptodate(page)) {
1139                         error = -EIO;
1140                         goto failed;
1141                 }
1142                 wait_on_page_writeback(page);
1143
1144                 if (shmem_should_replace_page(page, gfp)) {
1145                         error = shmem_replace_page(&page, gfp, info, index);
1146                         if (error)
1147                                 goto failed;
1148                 }
1149
1150                 error = mem_cgroup_cache_charge(page, current->mm,
1151                                                 gfp & GFP_RECLAIM_MASK);
1152                 if (!error) {
1153                         error = shmem_add_to_page_cache(page, mapping, index,
1154                                                 gfp, swp_to_radix_entry(swap));
1155                         /*
1156                          * We already confirmed swap under page lock, and make
1157                          * no memory allocation here, so usually no possibility
1158                          * of error; but free_swap_and_cache() only trylocks a
1159                          * page, so it is just possible that the entry has been
1160                          * truncated or holepunched since swap was confirmed.
1161                          * shmem_undo_range() will have done some of the
1162                          * unaccounting, now delete_from_swap_cache() will do
1163                          * the rest (including mem_cgroup_uncharge_swapcache).
1164                          * Reset swap.val? No, leave it so "failed" goes back to
1165                          * "repeat": reading a hole and writing should succeed.
1166                          */
1167                         if (error)
1168                                 delete_from_swap_cache(page);
1169                 }
1170                 if (error)
1171                         goto failed;
1172
1173                 spin_lock(&info->lock);
1174                 info->swapped--;
1175                 shmem_recalc_inode(inode);
1176                 spin_unlock(&info->lock);
1177
1178                 delete_from_swap_cache(page);
1179                 set_page_dirty(page);
1180                 swap_free(swap);
1181
1182         } else {
1183                 if (shmem_acct_block(info->flags)) {
1184                         error = -ENOSPC;
1185                         goto failed;
1186                 }
1187                 if (sbinfo->max_blocks) {
1188                         if (percpu_counter_compare(&sbinfo->used_blocks,
1189                                                 sbinfo->max_blocks) >= 0) {
1190                                 error = -ENOSPC;
1191                                 goto unacct;
1192                         }
1193                         percpu_counter_inc(&sbinfo->used_blocks);
1194                 }
1195
1196                 page = shmem_alloc_page(gfp, info, index);
1197                 if (!page) {
1198                         error = -ENOMEM;
1199                         goto decused;
1200                 }
1201
1202                 SetPageSwapBacked(page);
1203                 __set_page_locked(page);
1204                 error = mem_cgroup_cache_charge(page, current->mm,
1205                                                 gfp & GFP_RECLAIM_MASK);
1206                 if (error)
1207                         goto decused;
1208                 error = radix_tree_preload(gfp & GFP_RECLAIM_MASK);
1209                 if (!error) {
1210                         error = shmem_add_to_page_cache(page, mapping, index,
1211                                                         gfp, NULL);
1212                         radix_tree_preload_end();
1213                 }
1214                 if (error) {
1215                         mem_cgroup_uncharge_cache_page(page);
1216                         goto decused;
1217                 }
1218                 lru_cache_add_anon(page);
1219
1220                 spin_lock(&info->lock);
1221                 info->alloced++;
1222                 inode->i_blocks += BLOCKS_PER_PAGE;
1223                 shmem_recalc_inode(inode);
1224                 spin_unlock(&info->lock);
1225                 alloced = true;
1226
1227                 /*
1228                  * Let SGP_FALLOC use the SGP_WRITE optimization on a new page.
1229                  */
1230                 if (sgp == SGP_FALLOC)
1231                         sgp = SGP_WRITE;
1232 clear:
1233                 /*
1234                  * Let SGP_WRITE caller clear ends if write does not fill page;
1235                  * but SGP_FALLOC on a page fallocated earlier must initialize
1236                  * it now, lest undo on failure cancel our earlier guarantee.
1237                  */
1238                 if (sgp != SGP_WRITE) {
1239                         clear_highpage(page);
1240                         flush_dcache_page(page);
1241                         SetPageUptodate(page);
1242                 }
1243                 if (sgp == SGP_DIRTY)
1244                         set_page_dirty(page);
1245         }
1246
1247         /* Perhaps the file has been truncated since we checked */
1248         if (sgp != SGP_WRITE && sgp != SGP_FALLOC &&
1249             ((loff_t)index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
1250                 error = -EINVAL;
1251                 if (alloced)
1252                         goto trunc;
1253                 else
1254                         goto failed;
1255         }
1256         *pagep = page;
1257         return 0;
1258
1259         /*
1260          * Error recovery.
1261          */
1262 trunc:
1263         info = SHMEM_I(inode);
1264         ClearPageDirty(page);
1265         delete_from_page_cache(page);
1266         spin_lock(&info->lock);
1267         info->alloced--;
1268         inode->i_blocks -= BLOCKS_PER_PAGE;
1269         spin_unlock(&info->lock);
1270 decused:
1271         sbinfo = SHMEM_SB(inode->i_sb);
1272         if (sbinfo->max_blocks)
1273                 percpu_counter_add(&sbinfo->used_blocks, -1);
1274 unacct:
1275         shmem_unacct_blocks(info->flags, 1);
1276 failed:
1277         if (swap.val && error != -EINVAL &&
1278             !shmem_confirm_swap(mapping, index, swap))
1279                 error = -EEXIST;
1280 unlock:
1281         if (page) {
1282                 unlock_page(page);
1283                 page_cache_release(page);
1284         }
1285         if (error == -ENOSPC && !once++) {
1286                 info = SHMEM_I(inode);
1287                 spin_lock(&info->lock);
1288                 shmem_recalc_inode(inode);
1289                 spin_unlock(&info->lock);
1290                 goto repeat;
1291         }
1292         if (error == -EEXIST)   /* from above or from radix_tree_insert */
1293                 goto repeat;
1294         return error;
1295 }
1296
1297 static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1298 {
1299         struct inode *inode = file_inode(vma->vm_file);
1300         int error;
1301         int ret = VM_FAULT_LOCKED;
1302
1303         error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_CACHE, &ret);
1304         if (error)
1305                 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1306
1307         if (ret & VM_FAULT_MAJOR) {
1308                 count_vm_event(PGMAJFAULT);
1309                 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1310         }
1311         return ret;
1312 }
1313
1314 #ifdef CONFIG_NUMA
1315 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
1316 {
1317         struct inode *inode = file_inode(vma->vm_file);
1318         return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
1319 }
1320
1321 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
1322                                           unsigned long addr)
1323 {
1324         struct inode *inode = file_inode(vma->vm_file);
1325         pgoff_t index;
1326
1327         index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1328         return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
1329 }
1330 #endif
1331
1332 int shmem_lock(struct file *file, int lock, struct user_struct *user)
1333 {
1334         struct inode *inode = file_inode(file);
1335         struct shmem_inode_info *info = SHMEM_I(inode);
1336         int retval = -ENOMEM;
1337
1338         spin_lock(&info->lock);
1339         if (lock && !(info->flags & VM_LOCKED)) {
1340                 if (!user_shm_lock(inode->i_size, user))
1341                         goto out_nomem;
1342                 info->flags |= VM_LOCKED;
1343                 mapping_set_unevictable(file->f_mapping);
1344         }
1345         if (!lock && (info->flags & VM_LOCKED) && user) {
1346                 user_shm_unlock(inode->i_size, user);
1347                 info->flags &= ~VM_LOCKED;
1348                 mapping_clear_unevictable(file->f_mapping);
1349         }
1350         retval = 0;
1351
1352 out_nomem:
1353         spin_unlock(&info->lock);
1354         return retval;
1355 }
1356
1357 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1358 {
1359         file_accessed(file);
1360         vma->vm_ops = &shmem_vm_ops;
1361         return 0;
1362 }
1363
1364 static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
1365                                      umode_t mode, dev_t dev, unsigned long flags)
1366 {
1367         struct inode *inode;
1368         struct shmem_inode_info *info;
1369         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1370
1371         if (shmem_reserve_inode(sb))
1372                 return NULL;
1373
1374         inode = new_inode(sb);
1375         if (inode) {
1376                 inode->i_ino = get_next_ino();
1377                 inode_init_owner(inode, dir, mode);
1378                 inode->i_blocks = 0;
1379                 inode->i_mapping->backing_dev_info = &shmem_backing_dev_info;
1380                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1381                 inode->i_generation = get_seconds();
1382                 info = SHMEM_I(inode);
1383                 memset(info, 0, (char *)inode - (char *)info);
1384                 spin_lock_init(&info->lock);
1385                 info->flags = flags & VM_NORESERVE;
1386                 INIT_LIST_HEAD(&info->swaplist);
1387                 simple_xattrs_init(&info->xattrs);
1388                 cache_no_acl(inode);
1389
1390                 switch (mode & S_IFMT) {
1391                 default:
1392                         inode->i_op = &shmem_special_inode_operations;
1393                         init_special_inode(inode, mode, dev);
1394                         break;
1395                 case S_IFREG:
1396                         inode->i_mapping->a_ops = &shmem_aops;
1397                         inode->i_op = &shmem_inode_operations;
1398                         inode->i_fop = &shmem_file_operations;
1399                         mpol_shared_policy_init(&info->policy,
1400                                                  shmem_get_sbmpol(sbinfo));
1401                         break;
1402                 case S_IFDIR:
1403                         inc_nlink(inode);
1404                         /* Some things misbehave if size == 0 on a directory */
1405                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
1406                         inode->i_op = &shmem_dir_inode_operations;
1407                         inode->i_fop = &simple_dir_operations;
1408                         break;
1409                 case S_IFLNK:
1410                         /*
1411                          * Must not load anything in the rbtree,
1412                          * mpol_free_shared_policy will not be called.
1413                          */
1414                         mpol_shared_policy_init(&info->policy, NULL);
1415                         break;
1416                 }
1417         } else
1418                 shmem_free_inode(sb);
1419         return inode;
1420 }
1421
1422 #ifdef CONFIG_TMPFS
1423 static const struct inode_operations shmem_symlink_inode_operations;
1424 static const struct inode_operations shmem_short_symlink_operations;
1425
1426 #ifdef CONFIG_TMPFS_XATTR
1427 static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
1428 #else
1429 #define shmem_initxattrs NULL
1430 #endif
1431
1432 static int
1433 shmem_write_begin(struct file *file, struct address_space *mapping,
1434                         loff_t pos, unsigned len, unsigned flags,
1435                         struct page **pagep, void **fsdata)
1436 {
1437         struct inode *inode = mapping->host;
1438         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1439         return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1440 }
1441
1442 static int
1443 shmem_write_end(struct file *file, struct address_space *mapping,
1444                         loff_t pos, unsigned len, unsigned copied,
1445                         struct page *page, void *fsdata)
1446 {
1447         struct inode *inode = mapping->host;
1448
1449         if (pos + copied > inode->i_size)
1450                 i_size_write(inode, pos + copied);
1451
1452         if (!PageUptodate(page)) {
1453                 if (copied < PAGE_CACHE_SIZE) {
1454                         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1455                         zero_user_segments(page, 0, from,
1456                                         from + copied, PAGE_CACHE_SIZE);
1457                 }
1458                 SetPageUptodate(page);
1459         }
1460         set_page_dirty(page);
1461         unlock_page(page);
1462         page_cache_release(page);
1463
1464         return copied;
1465 }
1466
1467 static void do_shmem_file_read(struct file *filp, loff_t *ppos, read_descriptor_t *desc, read_actor_t actor)
1468 {
1469         struct inode *inode = file_inode(filp);
1470         struct address_space *mapping = inode->i_mapping;
1471         pgoff_t index;
1472         unsigned long offset;
1473         enum sgp_type sgp = SGP_READ;
1474
1475         /*
1476          * Might this read be for a stacking filesystem?  Then when reading
1477          * holes of a sparse file, we actually need to allocate those pages,
1478          * and even mark them dirty, so it cannot exceed the max_blocks limit.
1479          */
1480         if (segment_eq(get_fs(), KERNEL_DS))
1481                 sgp = SGP_DIRTY;
1482
1483         index = *ppos >> PAGE_CACHE_SHIFT;
1484         offset = *ppos & ~PAGE_CACHE_MASK;
1485
1486         for (;;) {
1487                 struct page *page = NULL;
1488                 pgoff_t end_index;
1489                 unsigned long nr, ret;
1490                 loff_t i_size = i_size_read(inode);
1491
1492                 end_index = i_size >> PAGE_CACHE_SHIFT;
1493                 if (index > end_index)
1494                         break;
1495                 if (index == end_index) {
1496                         nr = i_size & ~PAGE_CACHE_MASK;
1497                         if (nr <= offset)
1498                                 break;
1499                 }
1500
1501                 desc->error = shmem_getpage(inode, index, &page, sgp, NULL);
1502                 if (desc->error) {
1503                         if (desc->error == -EINVAL)
1504                                 desc->error = 0;
1505                         break;
1506                 }
1507                 if (page)
1508                         unlock_page(page);
1509
1510                 /*
1511                  * We must evaluate after, since reads (unlike writes)
1512                  * are called without i_mutex protection against truncate
1513                  */
1514                 nr = PAGE_CACHE_SIZE;
1515                 i_size = i_size_read(inode);
1516                 end_index = i_size >> PAGE_CACHE_SHIFT;
1517                 if (index == end_index) {
1518                         nr = i_size & ~PAGE_CACHE_MASK;
1519                         if (nr <= offset) {
1520                                 if (page)
1521                                         page_cache_release(page);
1522                                 break;
1523                         }
1524                 }
1525                 nr -= offset;
1526
1527                 if (page) {
1528                         /*
1529                          * If users can be writing to this page using arbitrary
1530                          * virtual addresses, take care about potential aliasing
1531                          * before reading the page on the kernel side.
1532                          */
1533                         if (mapping_writably_mapped(mapping))
1534                                 flush_dcache_page(page);
1535                         /*
1536                          * Mark the page accessed if we read the beginning.
1537                          */
1538                         if (!offset)
1539                                 mark_page_accessed(page);
1540                 } else {
1541                         page = ZERO_PAGE(0);
1542                         page_cache_get(page);
1543                 }
1544
1545                 /*
1546                  * Ok, we have the page, and it's up-to-date, so
1547                  * now we can copy it to user space...
1548                  *
1549                  * The actor routine returns how many bytes were actually used..
1550                  * NOTE! This may not be the same as how much of a user buffer
1551                  * we filled up (we may be padding etc), so we can only update
1552                  * "pos" here (the actor routine has to update the user buffer
1553                  * pointers and the remaining count).
1554                  */
1555                 ret = actor(desc, page, offset, nr);
1556                 offset += ret;
1557                 index += offset >> PAGE_CACHE_SHIFT;
1558                 offset &= ~PAGE_CACHE_MASK;
1559
1560                 page_cache_release(page);
1561                 if (ret != nr || !desc->count)
1562                         break;
1563
1564                 cond_resched();
1565         }
1566
1567         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1568         file_accessed(filp);
1569 }
1570
1571 static ssize_t shmem_file_aio_read(struct kiocb *iocb,
1572                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
1573 {
1574         struct file *filp = iocb->ki_filp;
1575         ssize_t retval;
1576         unsigned long seg;
1577         size_t count;
1578         loff_t *ppos = &iocb->ki_pos;
1579
1580         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1581         if (retval)
1582                 return retval;
1583
1584         for (seg = 0; seg < nr_segs; seg++) {
1585                 read_descriptor_t desc;
1586
1587                 desc.written = 0;
1588                 desc.arg.buf = iov[seg].iov_base;
1589                 desc.count = iov[seg].iov_len;
1590                 if (desc.count == 0)
1591                         continue;
1592                 desc.error = 0;
1593                 do_shmem_file_read(filp, ppos, &desc, file_read_actor);
1594                 retval += desc.written;
1595                 if (desc.error) {
1596                         retval = retval ?: desc.error;
1597                         break;
1598                 }
1599                 if (desc.count > 0)
1600                         break;
1601         }
1602         return retval;
1603 }
1604
1605 static ssize_t shmem_file_splice_read(struct file *in, loff_t *ppos,
1606                                 struct pipe_inode_info *pipe, size_t len,
1607                                 unsigned int flags)
1608 {
1609         struct address_space *mapping = in->f_mapping;
1610         struct inode *inode = mapping->host;
1611         unsigned int loff, nr_pages, req_pages;
1612         struct page *pages[PIPE_DEF_BUFFERS];
1613         struct partial_page partial[PIPE_DEF_BUFFERS];
1614         struct page *page;
1615         pgoff_t index, end_index;
1616         loff_t isize, left;
1617         int error, page_nr;
1618         struct splice_pipe_desc spd = {
1619                 .pages = pages,
1620                 .partial = partial,
1621                 .nr_pages_max = PIPE_DEF_BUFFERS,
1622                 .flags = flags,
1623                 .ops = &page_cache_pipe_buf_ops,
1624                 .spd_release = spd_release_page,
1625         };
1626
1627         isize = i_size_read(inode);
1628         if (unlikely(*ppos >= isize))
1629                 return 0;
1630
1631         left = isize - *ppos;
1632         if (unlikely(left < len))
1633                 len = left;
1634
1635         if (splice_grow_spd(pipe, &spd))
1636                 return -ENOMEM;
1637
1638         index = *ppos >> PAGE_CACHE_SHIFT;
1639         loff = *ppos & ~PAGE_CACHE_MASK;
1640         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1641         nr_pages = min(req_pages, pipe->buffers);
1642
1643         spd.nr_pages = find_get_pages_contig(mapping, index,
1644                                                 nr_pages, spd.pages);
1645         index += spd.nr_pages;
1646         error = 0;
1647
1648         while (spd.nr_pages < nr_pages) {
1649                 error = shmem_getpage(inode, index, &page, SGP_CACHE, NULL);
1650                 if (error)
1651                         break;
1652                 unlock_page(page);
1653                 spd.pages[spd.nr_pages++] = page;
1654                 index++;
1655         }
1656
1657         index = *ppos >> PAGE_CACHE_SHIFT;
1658         nr_pages = spd.nr_pages;
1659         spd.nr_pages = 0;
1660
1661         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
1662                 unsigned int this_len;
1663
1664                 if (!len)
1665                         break;
1666
1667                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
1668                 page = spd.pages[page_nr];
1669
1670                 if (!PageUptodate(page) || page->mapping != mapping) {
1671                         error = shmem_getpage(inode, index, &page,
1672                                                         SGP_CACHE, NULL);
1673                         if (error)
1674                                 break;
1675                         unlock_page(page);
1676                         page_cache_release(spd.pages[page_nr]);
1677                         spd.pages[page_nr] = page;
1678                 }
1679
1680                 isize = i_size_read(inode);
1681                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1682                 if (unlikely(!isize || index > end_index))
1683                         break;
1684
1685                 if (end_index == index) {
1686                         unsigned int plen;
1687
1688                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1689                         if (plen <= loff)
1690                                 break;
1691
1692                         this_len = min(this_len, plen - loff);
1693                         len = this_len;
1694                 }
1695
1696                 spd.partial[page_nr].offset = loff;
1697                 spd.partial[page_nr].len = this_len;
1698                 len -= this_len;
1699                 loff = 0;
1700                 spd.nr_pages++;
1701                 index++;
1702         }
1703
1704         while (page_nr < nr_pages)
1705                 page_cache_release(spd.pages[page_nr++]);
1706
1707         if (spd.nr_pages)
1708                 error = splice_to_pipe(pipe, &spd);
1709
1710         splice_shrink_spd(&spd);
1711
1712         if (error > 0) {
1713                 *ppos += error;
1714                 file_accessed(in);
1715         }
1716         return error;
1717 }
1718
1719 /*
1720  * llseek SEEK_DATA or SEEK_HOLE through the radix_tree.
1721  */
1722 static pgoff_t shmem_seek_hole_data(struct address_space *mapping,
1723                                     pgoff_t index, pgoff_t end, int whence)
1724 {
1725         struct page *page;
1726         struct pagevec pvec;
1727         pgoff_t indices[PAGEVEC_SIZE];
1728         bool done = false;
1729         int i;
1730
1731         pagevec_init(&pvec, 0);
1732         pvec.nr = 1;            /* start small: we may be there already */
1733         while (!done) {
1734                 pvec.nr = shmem_find_get_pages_and_swap(mapping, index,
1735                                         pvec.nr, pvec.pages, indices);
1736                 if (!pvec.nr) {
1737                         if (whence == SEEK_DATA)
1738                                 index = end;
1739                         break;
1740                 }
1741                 for (i = 0; i < pvec.nr; i++, index++) {
1742                         if (index < indices[i]) {
1743                                 if (whence == SEEK_HOLE) {
1744                                         done = true;
1745                                         break;
1746                                 }
1747                                 index = indices[i];
1748                         }
1749                         page = pvec.pages[i];
1750                         if (page && !radix_tree_exceptional_entry(page)) {
1751                                 if (!PageUptodate(page))
1752                                         page = NULL;
1753                         }
1754                         if (index >= end ||
1755                             (page && whence == SEEK_DATA) ||
1756                             (!page && whence == SEEK_HOLE)) {
1757                                 done = true;
1758                                 break;
1759                         }
1760                 }
1761                 shmem_deswap_pagevec(&pvec);
1762                 pagevec_release(&pvec);
1763                 pvec.nr = PAGEVEC_SIZE;
1764                 cond_resched();
1765         }
1766         return index;
1767 }
1768
1769 static loff_t shmem_file_llseek(struct file *file, loff_t offset, int whence)
1770 {
1771         struct address_space *mapping = file->f_mapping;
1772         struct inode *inode = mapping->host;
1773         pgoff_t start, end;
1774         loff_t new_offset;
1775
1776         if (whence != SEEK_DATA && whence != SEEK_HOLE)
1777                 return generic_file_llseek_size(file, offset, whence,
1778                                         MAX_LFS_FILESIZE, i_size_read(inode));
1779         mutex_lock(&inode->i_mutex);
1780         /* We're holding i_mutex so we can access i_size directly */
1781
1782         if (offset < 0)
1783                 offset = -EINVAL;
1784         else if (offset >= inode->i_size)
1785                 offset = -ENXIO;
1786         else {
1787                 start = offset >> PAGE_CACHE_SHIFT;
1788                 end = (inode->i_size + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1789                 new_offset = shmem_seek_hole_data(mapping, start, end, whence);
1790                 new_offset <<= PAGE_CACHE_SHIFT;
1791                 if (new_offset > offset) {
1792                         if (new_offset < inode->i_size)
1793                                 offset = new_offset;
1794                         else if (whence == SEEK_DATA)
1795                                 offset = -ENXIO;
1796                         else
1797                                 offset = inode->i_size;
1798                 }
1799         }
1800
1801         if (offset >= 0 && offset != file->f_pos) {
1802                 file->f_pos = offset;
1803                 file->f_version = 0;
1804         }
1805         mutex_unlock(&inode->i_mutex);
1806         return offset;
1807 }
1808
1809 static long shmem_fallocate(struct file *file, int mode, loff_t offset,
1810                                                          loff_t len)
1811 {
1812         struct inode *inode = file_inode(file);
1813         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1814         struct shmem_falloc shmem_falloc;
1815         pgoff_t start, index, end;
1816         int error;
1817
1818         mutex_lock(&inode->i_mutex);
1819
1820         if (mode & FALLOC_FL_PUNCH_HOLE) {
1821                 struct address_space *mapping = file->f_mapping;
1822                 loff_t unmap_start = round_up(offset, PAGE_SIZE);
1823                 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
1824
1825                 if ((u64)unmap_end > (u64)unmap_start)
1826                         unmap_mapping_range(mapping, unmap_start,
1827                                             1 + unmap_end - unmap_start, 0);
1828                 shmem_truncate_range(inode, offset, offset + len - 1);
1829                 /* No need to unmap again: hole-punching leaves COWed pages */
1830                 error = 0;
1831                 goto out;
1832         }
1833
1834         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
1835         error = inode_newsize_ok(inode, offset + len);
1836         if (error)
1837                 goto out;
1838
1839         start = offset >> PAGE_CACHE_SHIFT;
1840         end = (offset + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1841         /* Try to avoid a swapstorm if len is impossible to satisfy */
1842         if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
1843                 error = -ENOSPC;
1844                 goto out;
1845         }
1846
1847         shmem_falloc.start = start;
1848         shmem_falloc.next  = start;
1849         shmem_falloc.nr_falloced = 0;
1850         shmem_falloc.nr_unswapped = 0;
1851         spin_lock(&inode->i_lock);
1852         inode->i_private = &shmem_falloc;
1853         spin_unlock(&inode->i_lock);
1854
1855         for (index = start; index < end; index++) {
1856                 struct page *page;
1857
1858                 /*
1859                  * Good, the fallocate(2) manpage permits EINTR: we may have
1860                  * been interrupted because we are using up too much memory.
1861                  */
1862                 if (signal_pending(current))
1863                         error = -EINTR;
1864                 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
1865                         error = -ENOMEM;
1866                 else
1867                         error = shmem_getpage(inode, index, &page, SGP_FALLOC,
1868                                                                         NULL);
1869                 if (error) {
1870                         /* Remove the !PageUptodate pages we added */
1871                         shmem_undo_range(inode,
1872                                 (loff_t)start << PAGE_CACHE_SHIFT,
1873                                 (loff_t)index << PAGE_CACHE_SHIFT, true);
1874                         goto undone;
1875                 }
1876
1877                 /*
1878                  * Inform shmem_writepage() how far we have reached.
1879                  * No need for lock or barrier: we have the page lock.
1880                  */
1881                 shmem_falloc.next++;
1882                 if (!PageUptodate(page))
1883                         shmem_falloc.nr_falloced++;
1884
1885                 /*
1886                  * If !PageUptodate, leave it that way so that freeable pages
1887                  * can be recognized if we need to rollback on error later.
1888                  * But set_page_dirty so that memory pressure will swap rather
1889                  * than free the pages we are allocating (and SGP_CACHE pages
1890                  * might still be clean: we now need to mark those dirty too).
1891                  */
1892                 set_page_dirty(page);
1893                 unlock_page(page);
1894                 page_cache_release(page);
1895                 cond_resched();
1896         }
1897
1898         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
1899                 i_size_write(inode, offset + len);
1900         inode->i_ctime = CURRENT_TIME;
1901 undone:
1902         spin_lock(&inode->i_lock);
1903         inode->i_private = NULL;
1904         spin_unlock(&inode->i_lock);
1905 out:
1906         mutex_unlock(&inode->i_mutex);
1907         return error;
1908 }
1909
1910 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
1911 {
1912         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
1913
1914         buf->f_type = TMPFS_MAGIC;
1915         buf->f_bsize = PAGE_CACHE_SIZE;
1916         buf->f_namelen = NAME_MAX;
1917         if (sbinfo->max_blocks) {
1918                 buf->f_blocks = sbinfo->max_blocks;
1919                 buf->f_bavail =
1920                 buf->f_bfree  = sbinfo->max_blocks -
1921                                 percpu_counter_sum(&sbinfo->used_blocks);
1922         }
1923         if (sbinfo->max_inodes) {
1924                 buf->f_files = sbinfo->max_inodes;
1925                 buf->f_ffree = sbinfo->free_inodes;
1926         }
1927         /* else leave those fields 0 like simple_statfs */
1928         return 0;
1929 }
1930
1931 /*
1932  * File creation. Allocate an inode, and we're done..
1933  */
1934 static int
1935 shmem_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
1936 {
1937         struct inode *inode;
1938         int error = -ENOSPC;
1939
1940         inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
1941         if (inode) {
1942                 error = security_inode_init_security(inode, dir,
1943                                                      &dentry->d_name,
1944                                                      shmem_initxattrs, NULL);
1945                 if (error) {
1946                         if (error != -EOPNOTSUPP) {
1947                                 iput(inode);
1948                                 return error;
1949                         }
1950                 }
1951 #ifdef CONFIG_TMPFS_POSIX_ACL
1952                 error = generic_acl_init(inode, dir);
1953                 if (error) {
1954                         iput(inode);
1955                         return error;
1956                 }
1957 #else
1958                 error = 0;
1959 #endif
1960                 dir->i_size += BOGO_DIRENT_SIZE;
1961                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1962                 d_instantiate(dentry, inode);
1963                 dget(dentry); /* Extra count - pin the dentry in core */
1964         }
1965         return error;
1966 }
1967
1968 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
1969 {
1970         int error;
1971
1972         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
1973                 return error;
1974         inc_nlink(dir);
1975         return 0;
1976 }
1977
1978 static int shmem_create(struct inode *dir, struct dentry *dentry, umode_t mode,
1979                 bool excl)
1980 {
1981         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
1982 }
1983
1984 /*
1985  * Link a file..
1986  */
1987 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
1988 {
1989         struct inode *inode = old_dentry->d_inode;
1990         int ret;
1991
1992         /*
1993          * No ordinary (disk based) filesystem counts links as inodes;
1994          * but each new link needs a new dentry, pinning lowmem, and
1995          * tmpfs dentries cannot be pruned until they are unlinked.
1996          */
1997         ret = shmem_reserve_inode(inode->i_sb);
1998         if (ret)
1999                 goto out;
2000
2001         dir->i_size += BOGO_DIRENT_SIZE;
2002         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2003         inc_nlink(inode);
2004         ihold(inode);   /* New dentry reference */
2005         dget(dentry);           /* Extra pinning count for the created dentry */
2006         d_instantiate(dentry, inode);
2007 out:
2008         return ret;
2009 }
2010
2011 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
2012 {
2013         struct inode *inode = dentry->d_inode;
2014
2015         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
2016                 shmem_free_inode(inode->i_sb);
2017
2018         dir->i_size -= BOGO_DIRENT_SIZE;
2019         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2020         drop_nlink(inode);
2021         dput(dentry);   /* Undo the count from "create" - this does all the work */
2022         return 0;
2023 }
2024
2025 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
2026 {
2027         if (!simple_empty(dentry))
2028                 return -ENOTEMPTY;
2029
2030         drop_nlink(dentry->d_inode);
2031         drop_nlink(dir);
2032         return shmem_unlink(dir, dentry);
2033 }
2034
2035 /*
2036  * The VFS layer already does all the dentry stuff for rename,
2037  * we just have to decrement the usage count for the target if
2038  * it exists so that the VFS layer correctly free's it when it
2039  * gets overwritten.
2040  */
2041 static int shmem_rename(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
2042 {
2043         struct inode *inode = old_dentry->d_inode;
2044         int they_are_dirs = S_ISDIR(inode->i_mode);
2045
2046         if (!simple_empty(new_dentry))
2047                 return -ENOTEMPTY;
2048
2049         if (new_dentry->d_inode) {
2050                 (void) shmem_unlink(new_dir, new_dentry);
2051                 if (they_are_dirs)
2052                         drop_nlink(old_dir);
2053         } else if (they_are_dirs) {
2054                 drop_nlink(old_dir);
2055                 inc_nlink(new_dir);
2056         }
2057
2058         old_dir->i_size -= BOGO_DIRENT_SIZE;
2059         new_dir->i_size += BOGO_DIRENT_SIZE;
2060         old_dir->i_ctime = old_dir->i_mtime =
2061         new_dir->i_ctime = new_dir->i_mtime =
2062         inode->i_ctime = CURRENT_TIME;
2063         return 0;
2064 }
2065
2066 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
2067 {
2068         int error;
2069         int len;
2070         struct inode *inode;
2071         struct page *page;
2072         char *kaddr;
2073         struct shmem_inode_info *info;
2074
2075         len = strlen(symname) + 1;
2076         if (len > PAGE_CACHE_SIZE)
2077                 return -ENAMETOOLONG;
2078
2079         inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0, VM_NORESERVE);
2080         if (!inode)
2081                 return -ENOSPC;
2082
2083         error = security_inode_init_security(inode, dir, &dentry->d_name,
2084                                              shmem_initxattrs, NULL);
2085         if (error) {
2086                 if (error != -EOPNOTSUPP) {
2087                         iput(inode);
2088                         return error;
2089                 }
2090                 error = 0;
2091         }
2092
2093         info = SHMEM_I(inode);
2094         inode->i_size = len-1;
2095         if (len <= SHORT_SYMLINK_LEN) {
2096                 info->symlink = kmemdup(symname, len, GFP_KERNEL);
2097                 if (!info->symlink) {
2098                         iput(inode);
2099                         return -ENOMEM;
2100                 }
2101                 inode->i_op = &shmem_short_symlink_operations;
2102         } else {
2103                 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
2104                 if (error) {
2105                         iput(inode);
2106                         return error;
2107                 }
2108                 inode->i_mapping->a_ops = &shmem_aops;
2109                 inode->i_op = &shmem_symlink_inode_operations;
2110                 kaddr = kmap_atomic(page);
2111                 memcpy(kaddr, symname, len);
2112                 kunmap_atomic(kaddr);
2113                 SetPageUptodate(page);
2114                 set_page_dirty(page);
2115                 unlock_page(page);
2116                 page_cache_release(page);
2117         }
2118         dir->i_size += BOGO_DIRENT_SIZE;
2119         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2120         d_instantiate(dentry, inode);
2121         dget(dentry);
2122         return 0;
2123 }
2124
2125 static void *shmem_follow_short_symlink(struct dentry *dentry, struct nameidata *nd)
2126 {
2127         nd_set_link(nd, SHMEM_I(dentry->d_inode)->symlink);
2128         return NULL;
2129 }
2130
2131 static void *shmem_follow_link(struct dentry *dentry, struct nameidata *nd)
2132 {
2133         struct page *page = NULL;
2134         int error = shmem_getpage(dentry->d_inode, 0, &page, SGP_READ, NULL);
2135         nd_set_link(nd, error ? ERR_PTR(error) : kmap(page));
2136         if (page)
2137                 unlock_page(page);
2138         return page;
2139 }
2140
2141 static void shmem_put_link(struct dentry *dentry, struct nameidata *nd, void *cookie)
2142 {
2143         if (!IS_ERR(nd_get_link(nd))) {
2144                 struct page *page = cookie;
2145                 kunmap(page);
2146                 mark_page_accessed(page);
2147                 page_cache_release(page);
2148         }
2149 }
2150
2151 #ifdef CONFIG_TMPFS_XATTR
2152 /*
2153  * Superblocks without xattr inode operations may get some security.* xattr
2154  * support from the LSM "for free". As soon as we have any other xattrs
2155  * like ACLs, we also need to implement the security.* handlers at
2156  * filesystem level, though.
2157  */
2158
2159 /*
2160  * Callback for security_inode_init_security() for acquiring xattrs.
2161  */
2162 static int shmem_initxattrs(struct inode *inode,
2163                             const struct xattr *xattr_array,
2164                             void *fs_info)
2165 {
2166         struct shmem_inode_info *info = SHMEM_I(inode);
2167         const struct xattr *xattr;
2168         struct simple_xattr *new_xattr;
2169         size_t len;
2170
2171         for (xattr = xattr_array; xattr->name != NULL; xattr++) {
2172                 new_xattr = simple_xattr_alloc(xattr->value, xattr->value_len);
2173                 if (!new_xattr)
2174                         return -ENOMEM;
2175
2176                 len = strlen(xattr->name) + 1;
2177                 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
2178                                           GFP_KERNEL);
2179                 if (!new_xattr->name) {
2180                         kfree(new_xattr);
2181                         return -ENOMEM;
2182                 }
2183
2184                 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
2185                        XATTR_SECURITY_PREFIX_LEN);
2186                 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
2187                        xattr->name, len);
2188
2189                 simple_xattr_list_add(&info->xattrs, new_xattr);
2190         }
2191
2192         return 0;
2193 }
2194
2195 static const struct xattr_handler *shmem_xattr_handlers[] = {
2196 #ifdef CONFIG_TMPFS_POSIX_ACL
2197         &generic_acl_access_handler,
2198         &generic_acl_default_handler,
2199 #endif
2200         NULL
2201 };
2202
2203 static int shmem_xattr_validate(const char *name)
2204 {
2205         struct { const char *prefix; size_t len; } arr[] = {
2206                 { XATTR_SECURITY_PREFIX, XATTR_SECURITY_PREFIX_LEN },
2207                 { XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN }
2208         };
2209         int i;
2210
2211         for (i = 0; i < ARRAY_SIZE(arr); i++) {
2212                 size_t preflen = arr[i].len;
2213                 if (strncmp(name, arr[i].prefix, preflen) == 0) {
2214                         if (!name[preflen])
2215                                 return -EINVAL;
2216                         return 0;
2217                 }
2218         }
2219         return -EOPNOTSUPP;
2220 }
2221
2222 static ssize_t shmem_getxattr(struct dentry *dentry, const char *name,
2223                               void *buffer, size_t size)
2224 {
2225         struct shmem_inode_info *info = SHMEM_I(dentry->d_inode);
2226         int err;
2227
2228         /*
2229          * If this is a request for a synthetic attribute in the system.*
2230          * namespace use the generic infrastructure to resolve a handler
2231          * for it via sb->s_xattr.
2232          */
2233         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2234                 return generic_getxattr(dentry, name, buffer, size);
2235
2236         err = shmem_xattr_validate(name);
2237         if (err)
2238                 return err;
2239
2240         return simple_xattr_get(&info->xattrs, name, buffer, size);
2241 }
2242
2243 static int shmem_setxattr(struct dentry *dentry, const char *name,
2244                           const void *value, size_t size, int flags)
2245 {
2246         struct shmem_inode_info *info = SHMEM_I(dentry->d_inode);
2247         int err;
2248
2249         /*
2250          * If this is a request for a synthetic attribute in the system.*
2251          * namespace use the generic infrastructure to resolve a handler
2252          * for it via sb->s_xattr.
2253          */
2254         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2255                 return generic_setxattr(dentry, name, value, size, flags);
2256
2257         err = shmem_xattr_validate(name);
2258         if (err)
2259                 return err;
2260
2261         return simple_xattr_set(&info->xattrs, name, value, size, flags);
2262 }
2263
2264 static int shmem_removexattr(struct dentry *dentry, const char *name)
2265 {
2266         struct shmem_inode_info *info = SHMEM_I(dentry->d_inode);
2267         int err;
2268
2269         /*
2270          * If this is a request for a synthetic attribute in the system.*
2271          * namespace use the generic infrastructure to resolve a handler
2272          * for it via sb->s_xattr.
2273          */
2274         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2275                 return generic_removexattr(dentry, name);
2276
2277         err = shmem_xattr_validate(name);
2278         if (err)
2279                 return err;
2280
2281         return simple_xattr_remove(&info->xattrs, name);
2282 }
2283
2284 static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
2285 {
2286         struct shmem_inode_info *info = SHMEM_I(dentry->d_inode);
2287         return simple_xattr_list(&info->xattrs, buffer, size);
2288 }
2289 #endif /* CONFIG_TMPFS_XATTR */
2290
2291 static const struct inode_operations shmem_short_symlink_operations = {
2292         .readlink       = generic_readlink,
2293         .follow_link    = shmem_follow_short_symlink,
2294 #ifdef CONFIG_TMPFS_XATTR
2295         .setxattr       = shmem_setxattr,
2296         .getxattr       = shmem_getxattr,
2297         .listxattr      = shmem_listxattr,
2298         .removexattr    = shmem_removexattr,
2299 #endif
2300 };
2301
2302 static const struct inode_operations shmem_symlink_inode_operations = {
2303         .readlink       = generic_readlink,
2304         .follow_link    = shmem_follow_link,
2305         .put_link       = shmem_put_link,
2306 #ifdef CONFIG_TMPFS_XATTR
2307         .setxattr       = shmem_setxattr,
2308         .getxattr       = shmem_getxattr,
2309         .listxattr      = shmem_listxattr,
2310         .removexattr    = shmem_removexattr,
2311 #endif
2312 };
2313
2314 static struct dentry *shmem_get_parent(struct dentry *child)
2315 {
2316         return ERR_PTR(-ESTALE);
2317 }
2318
2319 static int shmem_match(struct inode *ino, void *vfh)
2320 {
2321         __u32 *fh = vfh;
2322         __u64 inum = fh[2];
2323         inum = (inum << 32) | fh[1];
2324         return ino->i_ino == inum && fh[0] == ino->i_generation;
2325 }
2326
2327 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
2328                 struct fid *fid, int fh_len, int fh_type)
2329 {
2330         struct inode *inode;
2331         struct dentry *dentry = NULL;
2332         u64 inum;
2333
2334         if (fh_len < 3)
2335                 return NULL;
2336
2337         inum = fid->raw[2];
2338         inum = (inum << 32) | fid->raw[1];
2339
2340         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
2341                         shmem_match, fid->raw);
2342         if (inode) {
2343                 dentry = d_find_alias(inode);
2344                 iput(inode);
2345         }
2346
2347         return dentry;
2348 }
2349
2350 static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
2351                                 struct inode *parent)
2352 {
2353         if (*len < 3) {
2354                 *len = 3;
2355                 return FILEID_INVALID;
2356         }
2357
2358         if (inode_unhashed(inode)) {
2359                 /* Unfortunately insert_inode_hash is not idempotent,
2360                  * so as we hash inodes here rather than at creation
2361                  * time, we need a lock to ensure we only try
2362                  * to do it once
2363                  */
2364                 static DEFINE_SPINLOCK(lock);
2365                 spin_lock(&lock);
2366                 if (inode_unhashed(inode))
2367                         __insert_inode_hash(inode,
2368                                             inode->i_ino + inode->i_generation);
2369                 spin_unlock(&lock);
2370         }
2371
2372         fh[0] = inode->i_generation;
2373         fh[1] = inode->i_ino;
2374         fh[2] = ((__u64)inode->i_ino) >> 32;
2375
2376         *len = 3;
2377         return 1;
2378 }
2379
2380 static const struct export_operations shmem_export_ops = {
2381         .get_parent     = shmem_get_parent,
2382         .encode_fh      = shmem_encode_fh,
2383         .fh_to_dentry   = shmem_fh_to_dentry,
2384 };
2385
2386 static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
2387                                bool remount)
2388 {
2389         char *this_char, *value, *rest;
2390         struct mempolicy *mpol = NULL;
2391         uid_t uid;
2392         gid_t gid;
2393
2394         while (options != NULL) {
2395                 this_char = options;
2396                 for (;;) {
2397                         /*
2398                          * NUL-terminate this option: unfortunately,
2399                          * mount options form a comma-separated list,
2400                          * but mpol's nodelist may also contain commas.
2401                          */
2402                         options = strchr(options, ',');
2403                         if (options == NULL)
2404                                 break;
2405                         options++;
2406                         if (!isdigit(*options)) {
2407                                 options[-1] = '\0';
2408                                 break;
2409                         }
2410                 }
2411                 if (!*this_char)
2412                         continue;
2413                 if ((value = strchr(this_char,'=')) != NULL) {
2414                         *value++ = 0;
2415                 } else {
2416                         printk(KERN_ERR
2417                             "tmpfs: No value for mount option '%s'\n",
2418                             this_char);
2419                         goto error;
2420                 }
2421
2422                 if (!strcmp(this_char,"size")) {
2423                         unsigned long long size;
2424                         size = memparse(value,&rest);
2425                         if (*rest == '%') {
2426                                 size <<= PAGE_SHIFT;
2427                                 size *= totalram_pages;
2428                                 do_div(size, 100);
2429                                 rest++;
2430                         }
2431                         if (*rest)
2432                                 goto bad_val;
2433                         sbinfo->max_blocks =
2434                                 DIV_ROUND_UP(size, PAGE_CACHE_SIZE);
2435                 } else if (!strcmp(this_char,"nr_blocks")) {
2436                         sbinfo->max_blocks = memparse(value, &rest);
2437                         if (*rest)
2438                                 goto bad_val;
2439                 } else if (!strcmp(this_char,"nr_inodes")) {
2440                         sbinfo->max_inodes = memparse(value, &rest);
2441                         if (*rest)
2442                                 goto bad_val;
2443                 } else if (!strcmp(this_char,"mode")) {
2444                         if (remount)
2445                                 continue;
2446                         sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
2447                         if (*rest)
2448                                 goto bad_val;
2449                 } else if (!strcmp(this_char,"uid")) {
2450                         if (remount)
2451                                 continue;
2452                         uid = simple_strtoul(value, &rest, 0);
2453                         if (*rest)
2454                                 goto bad_val;
2455                         sbinfo->uid = make_kuid(current_user_ns(), uid);
2456                         if (!uid_valid(sbinfo->uid))
2457                                 goto bad_val;
2458                 } else if (!strcmp(this_char,"gid")) {
2459                         if (remount)
2460                                 continue;
2461                         gid = simple_strtoul(value, &rest, 0);
2462                         if (*rest)
2463                                 goto bad_val;
2464                         sbinfo->gid = make_kgid(current_user_ns(), gid);
2465                         if (!gid_valid(sbinfo->gid))
2466                                 goto bad_val;
2467                 } else if (!strcmp(this_char,"mpol")) {
2468                         mpol_put(mpol);
2469                         mpol = NULL;
2470                         if (mpol_parse_str(value, &mpol))
2471                                 goto bad_val;
2472                 } else {
2473                         printk(KERN_ERR "tmpfs: Bad mount option %s\n",
2474                                this_char);
2475                         goto error;
2476                 }
2477         }
2478         sbinfo->mpol = mpol;
2479         return 0;
2480
2481 bad_val:
2482         printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
2483                value, this_char);
2484 error:
2485         mpol_put(mpol);
2486         return 1;
2487
2488 }
2489
2490 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2491 {
2492         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2493         struct shmem_sb_info config = *sbinfo;
2494         unsigned long inodes;
2495         int error = -EINVAL;
2496
2497         config.mpol = NULL;
2498         if (shmem_parse_options(data, &config, true))
2499                 return error;
2500
2501         spin_lock(&sbinfo->stat_lock);
2502         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2503         if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
2504                 goto out;
2505         if (config.max_inodes < inodes)
2506                 goto out;
2507         /*
2508          * Those tests disallow limited->unlimited while any are in use;
2509          * but we must separately disallow unlimited->limited, because
2510          * in that case we have no record of how much is already in use.
2511          */
2512         if (config.max_blocks && !sbinfo->max_blocks)
2513                 goto out;
2514         if (config.max_inodes && !sbinfo->max_inodes)
2515                 goto out;
2516
2517         error = 0;
2518         sbinfo->max_blocks  = config.max_blocks;
2519         sbinfo->max_inodes  = config.max_inodes;
2520         sbinfo->free_inodes = config.max_inodes - inodes;
2521
2522         /*
2523          * Preserve previous mempolicy unless mpol remount option was specified.
2524          */
2525         if (config.mpol) {
2526                 mpol_put(sbinfo->mpol);
2527                 sbinfo->mpol = config.mpol;     /* transfers initial ref */
2528         }
2529 out:
2530         spin_unlock(&sbinfo->stat_lock);
2531         return error;
2532 }
2533
2534 static int shmem_show_options(struct seq_file *seq, struct dentry *root)
2535 {
2536         struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
2537
2538         if (sbinfo->max_blocks != shmem_default_max_blocks())
2539                 seq_printf(seq, ",size=%luk",
2540                         sbinfo->max_blocks << (PAGE_CACHE_SHIFT - 10));
2541         if (sbinfo->max_inodes != shmem_default_max_inodes())
2542                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
2543         if (sbinfo->mode != (S_IRWXUGO | S_ISVTX))
2544                 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
2545         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
2546                 seq_printf(seq, ",uid=%u",
2547                                 from_kuid_munged(&init_user_ns, sbinfo->uid));
2548         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
2549                 seq_printf(seq, ",gid=%u",
2550                                 from_kgid_munged(&init_user_ns, sbinfo->gid));
2551         shmem_show_mpol(seq, sbinfo->mpol);
2552         return 0;
2553 }
2554 #endif /* CONFIG_TMPFS */
2555
2556 static void shmem_put_super(struct super_block *sb)
2557 {
2558         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2559
2560         percpu_counter_destroy(&sbinfo->used_blocks);
2561         mpol_put(sbinfo->mpol);
2562         kfree(sbinfo);
2563         sb->s_fs_info = NULL;
2564 }
2565
2566 int shmem_fill_super(struct super_block *sb, void *data, int silent)
2567 {
2568         struct inode *inode;
2569         struct shmem_sb_info *sbinfo;
2570         int err = -ENOMEM;
2571
2572         /* Round up to L1_CACHE_BYTES to resist false sharing */
2573         sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
2574                                 L1_CACHE_BYTES), GFP_KERNEL);
2575         if (!sbinfo)
2576                 return -ENOMEM;
2577
2578         sbinfo->mode = S_IRWXUGO | S_ISVTX;
2579         sbinfo->uid = current_fsuid();
2580         sbinfo->gid = current_fsgid();
2581         sb->s_fs_info = sbinfo;
2582
2583 #ifdef CONFIG_TMPFS
2584         /*
2585          * Per default we only allow half of the physical ram per
2586          * tmpfs instance, limiting inodes to one per page of lowmem;
2587          * but the internal instance is left unlimited.
2588          */
2589         if (!(sb->s_flags & MS_NOUSER)) {
2590                 sbinfo->max_blocks = shmem_default_max_blocks();
2591                 sbinfo->max_inodes = shmem_default_max_inodes();
2592                 if (shmem_parse_options(data, sbinfo, false)) {
2593                         err = -EINVAL;
2594                         goto failed;
2595                 }
2596         }
2597         sb->s_export_op = &shmem_export_ops;
2598         sb->s_flags |= MS_NOSEC;
2599 #else
2600         sb->s_flags |= MS_NOUSER;
2601 #endif
2602
2603         spin_lock_init(&sbinfo->stat_lock);
2604         if (percpu_counter_init(&sbinfo->used_blocks, 0))
2605                 goto failed;
2606         sbinfo->free_inodes = sbinfo->max_inodes;
2607
2608         sb->s_maxbytes = MAX_LFS_FILESIZE;
2609         sb->s_blocksize = PAGE_CACHE_SIZE;
2610         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
2611         sb->s_magic = TMPFS_MAGIC;
2612         sb->s_op = &shmem_ops;
2613         sb->s_time_gran = 1;
2614 #ifdef CONFIG_TMPFS_XATTR
2615         sb->s_xattr = shmem_xattr_handlers;
2616 #endif
2617 #ifdef CONFIG_TMPFS_POSIX_ACL
2618         sb->s_flags |= MS_POSIXACL;
2619 #endif
2620
2621         inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
2622         if (!inode)
2623                 goto failed;
2624         inode->i_uid = sbinfo->uid;
2625         inode->i_gid = sbinfo->gid;
2626         sb->s_root = d_make_root(inode);
2627         if (!sb->s_root)
2628                 goto failed;
2629         return 0;
2630
2631 failed:
2632         shmem_put_super(sb);
2633         return err;
2634 }
2635
2636 static struct kmem_cache *shmem_inode_cachep;
2637
2638 static struct inode *shmem_alloc_inode(struct super_block *sb)
2639 {
2640         struct shmem_inode_info *info;
2641         info = kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
2642         if (!info)
2643                 return NULL;
2644         return &info->vfs_inode;
2645 }
2646
2647 static void shmem_destroy_callback(struct rcu_head *head)
2648 {
2649         struct inode *inode = container_of(head, struct inode, i_rcu);
2650         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
2651 }
2652
2653 static void shmem_destroy_inode(struct inode *inode)
2654 {
2655         if (S_ISREG(inode->i_mode))
2656                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
2657         call_rcu(&inode->i_rcu, shmem_destroy_callback);
2658 }
2659
2660 static void shmem_init_inode(void *foo)
2661 {
2662         struct shmem_inode_info *info = foo;
2663         inode_init_once(&info->vfs_inode);
2664 }
2665
2666 static int shmem_init_inodecache(void)
2667 {
2668         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
2669                                 sizeof(struct shmem_inode_info),
2670                                 0, SLAB_PANIC, shmem_init_inode);
2671         return 0;
2672 }
2673
2674 static void shmem_destroy_inodecache(void)
2675 {
2676         kmem_cache_destroy(shmem_inode_cachep);
2677 }
2678
2679 static const struct address_space_operations shmem_aops = {
2680         .writepage      = shmem_writepage,
2681         .set_page_dirty = __set_page_dirty_no_writeback,
2682 #ifdef CONFIG_TMPFS
2683         .write_begin    = shmem_write_begin,
2684         .write_end      = shmem_write_end,
2685 #endif
2686         .migratepage    = migrate_page,
2687         .error_remove_page = generic_error_remove_page,
2688 };
2689
2690 static const struct file_operations shmem_file_operations = {
2691         .mmap           = shmem_mmap,
2692 #ifdef CONFIG_TMPFS
2693         .llseek         = shmem_file_llseek,
2694         .read           = do_sync_read,
2695         .write          = do_sync_write,
2696         .aio_read       = shmem_file_aio_read,
2697         .aio_write      = generic_file_aio_write,
2698         .fsync          = noop_fsync,
2699         .splice_read    = shmem_file_splice_read,
2700         .splice_write   = generic_file_splice_write,
2701         .fallocate      = shmem_fallocate,
2702 #endif
2703 };
2704
2705 static const struct inode_operations shmem_inode_operations = {
2706         .setattr        = shmem_setattr,
2707 #ifdef CONFIG_TMPFS_XATTR
2708         .setxattr       = shmem_setxattr,
2709         .getxattr       = shmem_getxattr,
2710         .listxattr      = shmem_listxattr,
2711         .removexattr    = shmem_removexattr,
2712 #endif
2713 };
2714
2715 static const struct inode_operations shmem_dir_inode_operations = {
2716 #ifdef CONFIG_TMPFS
2717         .create         = shmem_create,
2718         .lookup         = simple_lookup,
2719         .link           = shmem_link,
2720         .unlink         = shmem_unlink,
2721         .symlink        = shmem_symlink,
2722         .mkdir          = shmem_mkdir,
2723         .rmdir          = shmem_rmdir,
2724         .mknod          = shmem_mknod,
2725         .rename         = shmem_rename,
2726 #endif
2727 #ifdef CONFIG_TMPFS_XATTR
2728         .setxattr       = shmem_setxattr,
2729         .getxattr       = shmem_getxattr,
2730         .listxattr      = shmem_listxattr,
2731         .removexattr    = shmem_removexattr,
2732 #endif
2733 #ifdef CONFIG_TMPFS_POSIX_ACL
2734         .setattr        = shmem_setattr,
2735 #endif
2736 };
2737
2738 static const struct inode_operations shmem_special_inode_operations = {
2739 #ifdef CONFIG_TMPFS_XATTR
2740         .setxattr       = shmem_setxattr,
2741         .getxattr       = shmem_getxattr,
2742         .listxattr      = shmem_listxattr,
2743         .removexattr    = shmem_removexattr,
2744 #endif
2745 #ifdef CONFIG_TMPFS_POSIX_ACL
2746         .setattr        = shmem_setattr,
2747 #endif
2748 };
2749
2750 static const struct super_operations shmem_ops = {
2751         .alloc_inode    = shmem_alloc_inode,
2752         .destroy_inode  = shmem_destroy_inode,
2753 #ifdef CONFIG_TMPFS
2754         .statfs         = shmem_statfs,
2755         .remount_fs     = shmem_remount_fs,
2756         .show_options   = shmem_show_options,
2757 #endif
2758         .evict_inode    = shmem_evict_inode,
2759         .drop_inode     = generic_delete_inode,
2760         .put_super      = shmem_put_super,
2761 };
2762
2763 static const struct vm_operations_struct shmem_vm_ops = {
2764         .fault          = shmem_fault,
2765 #ifdef CONFIG_NUMA
2766         .set_policy     = shmem_set_policy,
2767         .get_policy     = shmem_get_policy,
2768 #endif
2769         .remap_pages    = generic_file_remap_pages,
2770 };
2771
2772 static struct dentry *shmem_mount(struct file_system_type *fs_type,
2773         int flags, const char *dev_name, void *data)
2774 {
2775         return mount_nodev(fs_type, flags, data, shmem_fill_super);
2776 }
2777
2778 static struct file_system_type shmem_fs_type = {
2779         .owner          = THIS_MODULE,
2780         .name           = "tmpfs",
2781         .mount          = shmem_mount,
2782         .kill_sb        = kill_litter_super,
2783         .fs_flags       = FS_USERNS_MOUNT,
2784 };
2785
2786 int __init shmem_init(void)
2787 {
2788         int error;
2789
2790         error = bdi_init(&shmem_backing_dev_info);
2791         if (error)
2792                 goto out4;
2793
2794         error = shmem_init_inodecache();
2795         if (error)
2796                 goto out3;
2797
2798         error = register_filesystem(&shmem_fs_type);
2799         if (error) {
2800                 printk(KERN_ERR "Could not register tmpfs\n");
2801                 goto out2;
2802         }
2803
2804         shm_mnt = vfs_kern_mount(&shmem_fs_type, MS_NOUSER,
2805                                  shmem_fs_type.name, NULL);
2806         if (IS_ERR(shm_mnt)) {
2807                 error = PTR_ERR(shm_mnt);
2808                 printk(KERN_ERR "Could not kern_mount tmpfs\n");
2809                 goto out1;
2810         }
2811         return 0;
2812
2813 out1:
2814         unregister_filesystem(&shmem_fs_type);
2815 out2:
2816         shmem_destroy_inodecache();
2817 out3:
2818         bdi_destroy(&shmem_backing_dev_info);
2819 out4:
2820         shm_mnt = ERR_PTR(error);
2821         return error;
2822 }
2823
2824 #else /* !CONFIG_SHMEM */
2825
2826 /*
2827  * tiny-shmem: simple shmemfs and tmpfs using ramfs code
2828  *
2829  * This is intended for small system where the benefits of the full
2830  * shmem code (swap-backed and resource-limited) are outweighed by
2831  * their complexity. On systems without swap this code should be
2832  * effectively equivalent, but much lighter weight.
2833  */
2834
2835 static struct file_system_type shmem_fs_type = {
2836         .name           = "tmpfs",
2837         .mount          = ramfs_mount,
2838         .kill_sb        = kill_litter_super,
2839         .fs_flags       = FS_USERNS_MOUNT,
2840 };
2841
2842 int __init shmem_init(void)
2843 {
2844         BUG_ON(register_filesystem(&shmem_fs_type) != 0);
2845
2846         shm_mnt = kern_mount(&shmem_fs_type);
2847         BUG_ON(IS_ERR(shm_mnt));
2848
2849         return 0;
2850 }
2851
2852 int shmem_unuse(swp_entry_t swap, struct page *page)
2853 {
2854         return 0;
2855 }
2856
2857 int shmem_lock(struct file *file, int lock, struct user_struct *user)
2858 {
2859         return 0;
2860 }
2861
2862 void shmem_unlock_mapping(struct address_space *mapping)
2863 {
2864 }
2865
2866 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
2867 {
2868         truncate_inode_pages_range(inode->i_mapping, lstart, lend);
2869 }
2870 EXPORT_SYMBOL_GPL(shmem_truncate_range);
2871
2872 #define shmem_vm_ops                            generic_file_vm_ops
2873 #define shmem_file_operations                   ramfs_file_operations
2874 #define shmem_get_inode(sb, dir, mode, dev, flags)      ramfs_get_inode(sb, dir, mode, dev)
2875 #define shmem_acct_size(flags, size)            0
2876 #define shmem_unacct_size(flags, size)          do {} while (0)
2877
2878 #endif /* CONFIG_SHMEM */
2879
2880 /* common code */
2881
2882 static char *shmem_dname(struct dentry *dentry, char *buffer, int buflen)
2883 {
2884         return dynamic_dname(dentry, buffer, buflen, "/%s (deleted)",
2885                                 dentry->d_name.name);
2886 }
2887
2888 static struct dentry_operations anon_ops = {
2889         .d_dname = shmem_dname
2890 };
2891
2892 /**
2893  * shmem_file_setup - get an unlinked file living in tmpfs
2894  * @name: name for dentry (to be seen in /proc/<pid>/maps
2895  * @size: size to be set for the file
2896  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
2897  */
2898 struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
2899 {
2900         struct file *res;
2901         struct inode *inode;
2902         struct path path;
2903         struct super_block *sb;
2904         struct qstr this;
2905
2906         if (IS_ERR(shm_mnt))
2907                 return ERR_CAST(shm_mnt);
2908
2909         if (size < 0 || size > MAX_LFS_FILESIZE)
2910                 return ERR_PTR(-EINVAL);
2911
2912         if (shmem_acct_size(flags, size))
2913                 return ERR_PTR(-ENOMEM);
2914
2915         res = ERR_PTR(-ENOMEM);
2916         this.name = name;
2917         this.len = strlen(name);
2918         this.hash = 0; /* will go */
2919         sb = shm_mnt->mnt_sb;
2920         path.dentry = d_alloc_pseudo(sb, &this);
2921         if (!path.dentry)
2922                 goto put_memory;
2923         d_set_d_op(path.dentry, &anon_ops);
2924         path.mnt = mntget(shm_mnt);
2925
2926         res = ERR_PTR(-ENOSPC);
2927         inode = shmem_get_inode(sb, NULL, S_IFREG | S_IRWXUGO, 0, flags);
2928         if (!inode)
2929                 goto put_dentry;
2930
2931         d_instantiate(path.dentry, inode);
2932         inode->i_size = size;
2933         clear_nlink(inode);     /* It is unlinked */
2934         res = ERR_PTR(ramfs_nommu_expand_for_mapping(inode, size));
2935         if (IS_ERR(res))
2936                 goto put_dentry;
2937
2938         res = alloc_file(&path, FMODE_WRITE | FMODE_READ,
2939                   &shmem_file_operations);
2940         if (IS_ERR(res))
2941                 goto put_dentry;
2942
2943         return res;
2944
2945 put_dentry:
2946         path_put(&path);
2947 put_memory:
2948         shmem_unacct_size(flags, size);
2949         return res;
2950 }
2951 EXPORT_SYMBOL_GPL(shmem_file_setup);
2952
2953 void shmem_set_file(struct vm_area_struct *vma, struct file *file)
2954 {
2955         if (vma->vm_file)
2956                 fput(vma->vm_file);
2957         vma->vm_file = file;
2958         vma->vm_ops = &shmem_vm_ops;
2959 }
2960
2961 /**
2962  * shmem_zero_setup - setup a shared anonymous mapping
2963  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
2964  */
2965 int shmem_zero_setup(struct vm_area_struct *vma)
2966 {
2967         struct file *file;
2968         loff_t size = vma->vm_end - vma->vm_start;
2969
2970         file = shmem_file_setup("dev/zero", size, vma->vm_flags);
2971         if (IS_ERR(file))
2972                 return PTR_ERR(file);
2973
2974         shmem_set_file(vma, file);
2975         return 0;
2976 }
2977
2978 /**
2979  * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
2980  * @mapping:    the page's address_space
2981  * @index:      the page index
2982  * @gfp:        the page allocator flags to use if allocating
2983  *
2984  * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
2985  * with any new page allocations done using the specified allocation flags.
2986  * But read_cache_page_gfp() uses the ->readpage() method: which does not
2987  * suit tmpfs, since it may have pages in swapcache, and needs to find those
2988  * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
2989  *
2990  * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
2991  * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
2992  */
2993 struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
2994                                          pgoff_t index, gfp_t gfp)
2995 {
2996 #ifdef CONFIG_SHMEM
2997         struct inode *inode = mapping->host;
2998         struct page *page;
2999         int error;
3000
3001         BUG_ON(mapping->a_ops != &shmem_aops);
3002         error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE, gfp, NULL);
3003         if (error)
3004                 page = ERR_PTR(error);
3005         else
3006                 unlock_page(page);
3007         return page;
3008 #else
3009         /*
3010          * The tiny !SHMEM case uses ramfs without swap
3011          */
3012         return read_cache_page_gfp(mapping, index, gfp);
3013 #endif
3014 }
3015 EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);