tmpfs: fix race between umount and swapoff
[linux-2.6.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) 2002-2005 Hugh Dickins.
10  * Copyright (C) 2002-2005 VERITAS Software Corporation.
11  * Copyright (C) 2004 Andi Kleen, SuSE Labs
12  *
13  * Extended attribute support for tmpfs:
14  * Copyright (c) 2004, Luke Kenneth Casson Leighton <lkcl@lkcl.net>
15  * Copyright (c) 2004 Red Hat, Inc., James Morris <jmorris@redhat.com>
16  *
17  * tiny-shmem:
18  * Copyright (c) 2004, 2008 Matt Mackall <mpm@selenic.com>
19  *
20  * This file is released under the GPL.
21  */
22
23 #include <linux/fs.h>
24 #include <linux/init.h>
25 #include <linux/vfs.h>
26 #include <linux/mount.h>
27 #include <linux/pagemap.h>
28 #include <linux/file.h>
29 #include <linux/mm.h>
30 #include <linux/module.h>
31 #include <linux/percpu_counter.h>
32 #include <linux/swap.h>
33
34 static struct vfsmount *shm_mnt;
35
36 #ifdef CONFIG_SHMEM
37 /*
38  * This virtual memory filesystem is heavily based on the ramfs. It
39  * extends ramfs by the ability to use swap and honor resource limits
40  * which makes it a completely usable filesystem.
41  */
42
43 #include <linux/xattr.h>
44 #include <linux/exportfs.h>
45 #include <linux/posix_acl.h>
46 #include <linux/generic_acl.h>
47 #include <linux/mman.h>
48 #include <linux/string.h>
49 #include <linux/slab.h>
50 #include <linux/backing-dev.h>
51 #include <linux/shmem_fs.h>
52 #include <linux/writeback.h>
53 #include <linux/blkdev.h>
54 #include <linux/security.h>
55 #include <linux/swapops.h>
56 #include <linux/mempolicy.h>
57 #include <linux/namei.h>
58 #include <linux/ctype.h>
59 #include <linux/migrate.h>
60 #include <linux/highmem.h>
61 #include <linux/seq_file.h>
62 #include <linux/magic.h>
63
64 #include <asm/uaccess.h>
65 #include <asm/div64.h>
66 #include <asm/pgtable.h>
67
68 /*
69  * The maximum size of a shmem/tmpfs file is limited by the maximum size of
70  * its triple-indirect swap vector - see illustration at shmem_swp_entry().
71  *
72  * With 4kB page size, maximum file size is just over 2TB on a 32-bit kernel,
73  * but one eighth of that on a 64-bit kernel.  With 8kB page size, maximum
74  * file size is just over 4TB on a 64-bit kernel, but 16TB on a 32-bit kernel,
75  * MAX_LFS_FILESIZE being then more restrictive than swap vector layout.
76  *
77  * We use / and * instead of shifts in the definitions below, so that the swap
78  * vector can be tested with small even values (e.g. 20) for ENTRIES_PER_PAGE.
79  */
80 #define ENTRIES_PER_PAGE (PAGE_CACHE_SIZE/sizeof(unsigned long))
81 #define ENTRIES_PER_PAGEPAGE ((unsigned long long)ENTRIES_PER_PAGE*ENTRIES_PER_PAGE)
82
83 #define SHMSWP_MAX_INDEX (SHMEM_NR_DIRECT + (ENTRIES_PER_PAGEPAGE/2) * (ENTRIES_PER_PAGE+1))
84 #define SHMSWP_MAX_BYTES (SHMSWP_MAX_INDEX << PAGE_CACHE_SHIFT)
85
86 #define SHMEM_MAX_BYTES  min_t(unsigned long long, SHMSWP_MAX_BYTES, MAX_LFS_FILESIZE)
87 #define SHMEM_MAX_INDEX  ((unsigned long)((SHMEM_MAX_BYTES+1) >> PAGE_CACHE_SHIFT))
88
89 #define BLOCKS_PER_PAGE  (PAGE_CACHE_SIZE/512)
90 #define VM_ACCT(size)    (PAGE_CACHE_ALIGN(size) >> PAGE_SHIFT)
91
92 /* info->flags needs VM_flags to handle pagein/truncate races efficiently */
93 #define SHMEM_PAGEIN     VM_READ
94 #define SHMEM_TRUNCATE   VM_WRITE
95
96 /* Definition to limit shmem_truncate's steps between cond_rescheds */
97 #define LATENCY_LIMIT    64
98
99 /* Pretend that each entry is of this size in directory's i_size */
100 #define BOGO_DIRENT_SIZE 20
101
102 /* Flag allocation requirements to shmem_getpage and shmem_swp_alloc */
103 enum sgp_type {
104         SGP_READ,       /* don't exceed i_size, don't allocate page */
105         SGP_CACHE,      /* don't exceed i_size, may allocate page */
106         SGP_DIRTY,      /* like SGP_CACHE, but set new page dirty */
107         SGP_WRITE,      /* may exceed i_size, may allocate page */
108 };
109
110 #ifdef CONFIG_TMPFS
111 static unsigned long shmem_default_max_blocks(void)
112 {
113         return totalram_pages / 2;
114 }
115
116 static unsigned long shmem_default_max_inodes(void)
117 {
118         return min(totalram_pages - totalhigh_pages, totalram_pages / 2);
119 }
120 #endif
121
122 static int shmem_getpage(struct inode *inode, unsigned long idx,
123                          struct page **pagep, enum sgp_type sgp, int *type);
124
125 static inline struct page *shmem_dir_alloc(gfp_t gfp_mask)
126 {
127         /*
128          * The above definition of ENTRIES_PER_PAGE, and the use of
129          * BLOCKS_PER_PAGE on indirect pages, assume PAGE_CACHE_SIZE:
130          * might be reconsidered if it ever diverges from PAGE_SIZE.
131          *
132          * Mobility flags are masked out as swap vectors cannot move
133          */
134         return alloc_pages((gfp_mask & ~GFP_MOVABLE_MASK) | __GFP_ZERO,
135                                 PAGE_CACHE_SHIFT-PAGE_SHIFT);
136 }
137
138 static inline void shmem_dir_free(struct page *page)
139 {
140         __free_pages(page, PAGE_CACHE_SHIFT-PAGE_SHIFT);
141 }
142
143 static struct page **shmem_dir_map(struct page *page)
144 {
145         return (struct page **)kmap_atomic(page, KM_USER0);
146 }
147
148 static inline void shmem_dir_unmap(struct page **dir)
149 {
150         kunmap_atomic(dir, KM_USER0);
151 }
152
153 static swp_entry_t *shmem_swp_map(struct page *page)
154 {
155         return (swp_entry_t *)kmap_atomic(page, KM_USER1);
156 }
157
158 static inline void shmem_swp_balance_unmap(void)
159 {
160         /*
161          * When passing a pointer to an i_direct entry, to code which
162          * also handles indirect entries and so will shmem_swp_unmap,
163          * we must arrange for the preempt count to remain in balance.
164          * What kmap_atomic of a lowmem page does depends on config
165          * and architecture, so pretend to kmap_atomic some lowmem page.
166          */
167         (void) kmap_atomic(ZERO_PAGE(0), KM_USER1);
168 }
169
170 static inline void shmem_swp_unmap(swp_entry_t *entry)
171 {
172         kunmap_atomic(entry, KM_USER1);
173 }
174
175 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
176 {
177         return sb->s_fs_info;
178 }
179
180 /*
181  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
182  * for shared memory and for shared anonymous (/dev/zero) mappings
183  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
184  * consistent with the pre-accounting of private mappings ...
185  */
186 static inline int shmem_acct_size(unsigned long flags, loff_t size)
187 {
188         return (flags & VM_NORESERVE) ?
189                 0 : security_vm_enough_memory_kern(VM_ACCT(size));
190 }
191
192 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
193 {
194         if (!(flags & VM_NORESERVE))
195                 vm_unacct_memory(VM_ACCT(size));
196 }
197
198 /*
199  * ... whereas tmpfs objects are accounted incrementally as
200  * pages are allocated, in order to allow huge sparse files.
201  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
202  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
203  */
204 static inline int shmem_acct_block(unsigned long flags)
205 {
206         return (flags & VM_NORESERVE) ?
207                 security_vm_enough_memory_kern(VM_ACCT(PAGE_CACHE_SIZE)) : 0;
208 }
209
210 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
211 {
212         if (flags & VM_NORESERVE)
213                 vm_unacct_memory(pages * VM_ACCT(PAGE_CACHE_SIZE));
214 }
215
216 static const struct super_operations shmem_ops;
217 static const struct address_space_operations shmem_aops;
218 static const struct file_operations shmem_file_operations;
219 static const struct inode_operations shmem_inode_operations;
220 static const struct inode_operations shmem_dir_inode_operations;
221 static const struct inode_operations shmem_special_inode_operations;
222 static const struct vm_operations_struct shmem_vm_ops;
223
224 static struct backing_dev_info shmem_backing_dev_info  __read_mostly = {
225         .ra_pages       = 0,    /* No readahead */
226         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_SWAP_BACKED,
227 };
228
229 static LIST_HEAD(shmem_swaplist);
230 static DEFINE_MUTEX(shmem_swaplist_mutex);
231
232 static void shmem_free_blocks(struct inode *inode, long pages)
233 {
234         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
235         if (sbinfo->max_blocks) {
236                 percpu_counter_add(&sbinfo->used_blocks, -pages);
237                 spin_lock(&inode->i_lock);
238                 inode->i_blocks -= pages*BLOCKS_PER_PAGE;
239                 spin_unlock(&inode->i_lock);
240         }
241 }
242
243 static int shmem_reserve_inode(struct super_block *sb)
244 {
245         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
246         if (sbinfo->max_inodes) {
247                 spin_lock(&sbinfo->stat_lock);
248                 if (!sbinfo->free_inodes) {
249                         spin_unlock(&sbinfo->stat_lock);
250                         return -ENOSPC;
251                 }
252                 sbinfo->free_inodes--;
253                 spin_unlock(&sbinfo->stat_lock);
254         }
255         return 0;
256 }
257
258 static void shmem_free_inode(struct super_block *sb)
259 {
260         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
261         if (sbinfo->max_inodes) {
262                 spin_lock(&sbinfo->stat_lock);
263                 sbinfo->free_inodes++;
264                 spin_unlock(&sbinfo->stat_lock);
265         }
266 }
267
268 /**
269  * shmem_recalc_inode - recalculate the size of an inode
270  * @inode: inode to recalc
271  *
272  * We have to calculate the free blocks since the mm can drop
273  * undirtied hole pages behind our back.
274  *
275  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
276  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
277  *
278  * It has to be called with the spinlock held.
279  */
280 static void shmem_recalc_inode(struct inode *inode)
281 {
282         struct shmem_inode_info *info = SHMEM_I(inode);
283         long freed;
284
285         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
286         if (freed > 0) {
287                 info->alloced -= freed;
288                 shmem_unacct_blocks(info->flags, freed);
289                 shmem_free_blocks(inode, freed);
290         }
291 }
292
293 /**
294  * shmem_swp_entry - find the swap vector position in the info structure
295  * @info:  info structure for the inode
296  * @index: index of the page to find
297  * @page:  optional page to add to the structure. Has to be preset to
298  *         all zeros
299  *
300  * If there is no space allocated yet it will return NULL when
301  * page is NULL, else it will use the page for the needed block,
302  * setting it to NULL on return to indicate that it has been used.
303  *
304  * The swap vector is organized the following way:
305  *
306  * There are SHMEM_NR_DIRECT entries directly stored in the
307  * shmem_inode_info structure. So small files do not need an addional
308  * allocation.
309  *
310  * For pages with index > SHMEM_NR_DIRECT there is the pointer
311  * i_indirect which points to a page which holds in the first half
312  * doubly indirect blocks, in the second half triple indirect blocks:
313  *
314  * For an artificial ENTRIES_PER_PAGE = 4 this would lead to the
315  * following layout (for SHMEM_NR_DIRECT == 16):
316  *
317  * i_indirect -> dir --> 16-19
318  *            |      +-> 20-23
319  *            |
320  *            +-->dir2 --> 24-27
321  *            |        +-> 28-31
322  *            |        +-> 32-35
323  *            |        +-> 36-39
324  *            |
325  *            +-->dir3 --> 40-43
326  *                     +-> 44-47
327  *                     +-> 48-51
328  *                     +-> 52-55
329  */
330 static swp_entry_t *shmem_swp_entry(struct shmem_inode_info *info, unsigned long index, struct page **page)
331 {
332         unsigned long offset;
333         struct page **dir;
334         struct page *subdir;
335
336         if (index < SHMEM_NR_DIRECT) {
337                 shmem_swp_balance_unmap();
338                 return info->i_direct+index;
339         }
340         if (!info->i_indirect) {
341                 if (page) {
342                         info->i_indirect = *page;
343                         *page = NULL;
344                 }
345                 return NULL;                    /* need another page */
346         }
347
348         index -= SHMEM_NR_DIRECT;
349         offset = index % ENTRIES_PER_PAGE;
350         index /= ENTRIES_PER_PAGE;
351         dir = shmem_dir_map(info->i_indirect);
352
353         if (index >= ENTRIES_PER_PAGE/2) {
354                 index -= ENTRIES_PER_PAGE/2;
355                 dir += ENTRIES_PER_PAGE/2 + index/ENTRIES_PER_PAGE;
356                 index %= ENTRIES_PER_PAGE;
357                 subdir = *dir;
358                 if (!subdir) {
359                         if (page) {
360                                 *dir = *page;
361                                 *page = NULL;
362                         }
363                         shmem_dir_unmap(dir);
364                         return NULL;            /* need another page */
365                 }
366                 shmem_dir_unmap(dir);
367                 dir = shmem_dir_map(subdir);
368         }
369
370         dir += index;
371         subdir = *dir;
372         if (!subdir) {
373                 if (!page || !(subdir = *page)) {
374                         shmem_dir_unmap(dir);
375                         return NULL;            /* need a page */
376                 }
377                 *dir = subdir;
378                 *page = NULL;
379         }
380         shmem_dir_unmap(dir);
381         return shmem_swp_map(subdir) + offset;
382 }
383
384 static void shmem_swp_set(struct shmem_inode_info *info, swp_entry_t *entry, unsigned long value)
385 {
386         long incdec = value? 1: -1;
387
388         entry->val = value;
389         info->swapped += incdec;
390         if ((unsigned long)(entry - info->i_direct) >= SHMEM_NR_DIRECT) {
391                 struct page *page = kmap_atomic_to_page(entry);
392                 set_page_private(page, page_private(page) + incdec);
393         }
394 }
395
396 /**
397  * shmem_swp_alloc - get the position of the swap entry for the page.
398  * @info:       info structure for the inode
399  * @index:      index of the page to find
400  * @sgp:        check and recheck i_size? skip allocation?
401  *
402  * If the entry does not exist, allocate it.
403  */
404 static swp_entry_t *shmem_swp_alloc(struct shmem_inode_info *info, unsigned long index, enum sgp_type sgp)
405 {
406         struct inode *inode = &info->vfs_inode;
407         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
408         struct page *page = NULL;
409         swp_entry_t *entry;
410
411         if (sgp != SGP_WRITE &&
412             ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode))
413                 return ERR_PTR(-EINVAL);
414
415         while (!(entry = shmem_swp_entry(info, index, &page))) {
416                 if (sgp == SGP_READ)
417                         return shmem_swp_map(ZERO_PAGE(0));
418                 /*
419                  * Test used_blocks against 1 less max_blocks, since we have 1 data
420                  * page (and perhaps indirect index pages) yet to allocate:
421                  * a waste to allocate index if we cannot allocate data.
422                  */
423                 if (sbinfo->max_blocks) {
424                         if (percpu_counter_compare(&sbinfo->used_blocks,
425                                                 sbinfo->max_blocks - 1) >= 0)
426                                 return ERR_PTR(-ENOSPC);
427                         percpu_counter_inc(&sbinfo->used_blocks);
428                         spin_lock(&inode->i_lock);
429                         inode->i_blocks += BLOCKS_PER_PAGE;
430                         spin_unlock(&inode->i_lock);
431                 }
432
433                 spin_unlock(&info->lock);
434                 page = shmem_dir_alloc(mapping_gfp_mask(inode->i_mapping));
435                 spin_lock(&info->lock);
436
437                 if (!page) {
438                         shmem_free_blocks(inode, 1);
439                         return ERR_PTR(-ENOMEM);
440                 }
441                 if (sgp != SGP_WRITE &&
442                     ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
443                         entry = ERR_PTR(-EINVAL);
444                         break;
445                 }
446                 if (info->next_index <= index)
447                         info->next_index = index + 1;
448         }
449         if (page) {
450                 /* another task gave its page, or truncated the file */
451                 shmem_free_blocks(inode, 1);
452                 shmem_dir_free(page);
453         }
454         if (info->next_index <= index && !IS_ERR(entry))
455                 info->next_index = index + 1;
456         return entry;
457 }
458
459 /**
460  * shmem_free_swp - free some swap entries in a directory
461  * @dir:        pointer to the directory
462  * @edir:       pointer after last entry of the directory
463  * @punch_lock: pointer to spinlock when needed for the holepunch case
464  */
465 static int shmem_free_swp(swp_entry_t *dir, swp_entry_t *edir,
466                                                 spinlock_t *punch_lock)
467 {
468         spinlock_t *punch_unlock = NULL;
469         swp_entry_t *ptr;
470         int freed = 0;
471
472         for (ptr = dir; ptr < edir; ptr++) {
473                 if (ptr->val) {
474                         if (unlikely(punch_lock)) {
475                                 punch_unlock = punch_lock;
476                                 punch_lock = NULL;
477                                 spin_lock(punch_unlock);
478                                 if (!ptr->val)
479                                         continue;
480                         }
481                         free_swap_and_cache(*ptr);
482                         *ptr = (swp_entry_t){0};
483                         freed++;
484                 }
485         }
486         if (punch_unlock)
487                 spin_unlock(punch_unlock);
488         return freed;
489 }
490
491 static int shmem_map_and_free_swp(struct page *subdir, int offset,
492                 int limit, struct page ***dir, spinlock_t *punch_lock)
493 {
494         swp_entry_t *ptr;
495         int freed = 0;
496
497         ptr = shmem_swp_map(subdir);
498         for (; offset < limit; offset += LATENCY_LIMIT) {
499                 int size = limit - offset;
500                 if (size > LATENCY_LIMIT)
501                         size = LATENCY_LIMIT;
502                 freed += shmem_free_swp(ptr+offset, ptr+offset+size,
503                                                         punch_lock);
504                 if (need_resched()) {
505                         shmem_swp_unmap(ptr);
506                         if (*dir) {
507                                 shmem_dir_unmap(*dir);
508                                 *dir = NULL;
509                         }
510                         cond_resched();
511                         ptr = shmem_swp_map(subdir);
512                 }
513         }
514         shmem_swp_unmap(ptr);
515         return freed;
516 }
517
518 static void shmem_free_pages(struct list_head *next)
519 {
520         struct page *page;
521         int freed = 0;
522
523         do {
524                 page = container_of(next, struct page, lru);
525                 next = next->next;
526                 shmem_dir_free(page);
527                 freed++;
528                 if (freed >= LATENCY_LIMIT) {
529                         cond_resched();
530                         freed = 0;
531                 }
532         } while (next);
533 }
534
535 static void shmem_truncate_range(struct inode *inode, loff_t start, loff_t end)
536 {
537         struct shmem_inode_info *info = SHMEM_I(inode);
538         unsigned long idx;
539         unsigned long size;
540         unsigned long limit;
541         unsigned long stage;
542         unsigned long diroff;
543         struct page **dir;
544         struct page *topdir;
545         struct page *middir;
546         struct page *subdir;
547         swp_entry_t *ptr;
548         LIST_HEAD(pages_to_free);
549         long nr_pages_to_free = 0;
550         long nr_swaps_freed = 0;
551         int offset;
552         int freed;
553         int punch_hole;
554         spinlock_t *needs_lock;
555         spinlock_t *punch_lock;
556         unsigned long upper_limit;
557
558         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
559         idx = (start + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
560         if (idx >= info->next_index)
561                 return;
562
563         spin_lock(&info->lock);
564         info->flags |= SHMEM_TRUNCATE;
565         if (likely(end == (loff_t) -1)) {
566                 limit = info->next_index;
567                 upper_limit = SHMEM_MAX_INDEX;
568                 info->next_index = idx;
569                 needs_lock = NULL;
570                 punch_hole = 0;
571         } else {
572                 if (end + 1 >= inode->i_size) { /* we may free a little more */
573                         limit = (inode->i_size + PAGE_CACHE_SIZE - 1) >>
574                                                         PAGE_CACHE_SHIFT;
575                         upper_limit = SHMEM_MAX_INDEX;
576                 } else {
577                         limit = (end + 1) >> PAGE_CACHE_SHIFT;
578                         upper_limit = limit;
579                 }
580                 needs_lock = &info->lock;
581                 punch_hole = 1;
582         }
583
584         topdir = info->i_indirect;
585         if (topdir && idx <= SHMEM_NR_DIRECT && !punch_hole) {
586                 info->i_indirect = NULL;
587                 nr_pages_to_free++;
588                 list_add(&topdir->lru, &pages_to_free);
589         }
590         spin_unlock(&info->lock);
591
592         if (info->swapped && idx < SHMEM_NR_DIRECT) {
593                 ptr = info->i_direct;
594                 size = limit;
595                 if (size > SHMEM_NR_DIRECT)
596                         size = SHMEM_NR_DIRECT;
597                 nr_swaps_freed = shmem_free_swp(ptr+idx, ptr+size, needs_lock);
598         }
599
600         /*
601          * If there are no indirect blocks or we are punching a hole
602          * below indirect blocks, nothing to be done.
603          */
604         if (!topdir || limit <= SHMEM_NR_DIRECT)
605                 goto done2;
606
607         /*
608          * The truncation case has already dropped info->lock, and we're safe
609          * because i_size and next_index have already been lowered, preventing
610          * access beyond.  But in the punch_hole case, we still need to take
611          * the lock when updating the swap directory, because there might be
612          * racing accesses by shmem_getpage(SGP_CACHE), shmem_unuse_inode or
613          * shmem_writepage.  However, whenever we find we can remove a whole
614          * directory page (not at the misaligned start or end of the range),
615          * we first NULLify its pointer in the level above, and then have no
616          * need to take the lock when updating its contents: needs_lock and
617          * punch_lock (either pointing to info->lock or NULL) manage this.
618          */
619
620         upper_limit -= SHMEM_NR_DIRECT;
621         limit -= SHMEM_NR_DIRECT;
622         idx = (idx > SHMEM_NR_DIRECT)? (idx - SHMEM_NR_DIRECT): 0;
623         offset = idx % ENTRIES_PER_PAGE;
624         idx -= offset;
625
626         dir = shmem_dir_map(topdir);
627         stage = ENTRIES_PER_PAGEPAGE/2;
628         if (idx < ENTRIES_PER_PAGEPAGE/2) {
629                 middir = topdir;
630                 diroff = idx/ENTRIES_PER_PAGE;
631         } else {
632                 dir += ENTRIES_PER_PAGE/2;
633                 dir += (idx - ENTRIES_PER_PAGEPAGE/2)/ENTRIES_PER_PAGEPAGE;
634                 while (stage <= idx)
635                         stage += ENTRIES_PER_PAGEPAGE;
636                 middir = *dir;
637                 if (*dir) {
638                         diroff = ((idx - ENTRIES_PER_PAGEPAGE/2) %
639                                 ENTRIES_PER_PAGEPAGE) / ENTRIES_PER_PAGE;
640                         if (!diroff && !offset && upper_limit >= stage) {
641                                 if (needs_lock) {
642                                         spin_lock(needs_lock);
643                                         *dir = NULL;
644                                         spin_unlock(needs_lock);
645                                         needs_lock = NULL;
646                                 } else
647                                         *dir = NULL;
648                                 nr_pages_to_free++;
649                                 list_add(&middir->lru, &pages_to_free);
650                         }
651                         shmem_dir_unmap(dir);
652                         dir = shmem_dir_map(middir);
653                 } else {
654                         diroff = 0;
655                         offset = 0;
656                         idx = stage;
657                 }
658         }
659
660         for (; idx < limit; idx += ENTRIES_PER_PAGE, diroff++) {
661                 if (unlikely(idx == stage)) {
662                         shmem_dir_unmap(dir);
663                         dir = shmem_dir_map(topdir) +
664                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
665                         while (!*dir) {
666                                 dir++;
667                                 idx += ENTRIES_PER_PAGEPAGE;
668                                 if (idx >= limit)
669                                         goto done1;
670                         }
671                         stage = idx + ENTRIES_PER_PAGEPAGE;
672                         middir = *dir;
673                         if (punch_hole)
674                                 needs_lock = &info->lock;
675                         if (upper_limit >= stage) {
676                                 if (needs_lock) {
677                                         spin_lock(needs_lock);
678                                         *dir = NULL;
679                                         spin_unlock(needs_lock);
680                                         needs_lock = NULL;
681                                 } else
682                                         *dir = NULL;
683                                 nr_pages_to_free++;
684                                 list_add(&middir->lru, &pages_to_free);
685                         }
686                         shmem_dir_unmap(dir);
687                         cond_resched();
688                         dir = shmem_dir_map(middir);
689                         diroff = 0;
690                 }
691                 punch_lock = needs_lock;
692                 subdir = dir[diroff];
693                 if (subdir && !offset && upper_limit-idx >= ENTRIES_PER_PAGE) {
694                         if (needs_lock) {
695                                 spin_lock(needs_lock);
696                                 dir[diroff] = NULL;
697                                 spin_unlock(needs_lock);
698                                 punch_lock = NULL;
699                         } else
700                                 dir[diroff] = NULL;
701                         nr_pages_to_free++;
702                         list_add(&subdir->lru, &pages_to_free);
703                 }
704                 if (subdir && page_private(subdir) /* has swap entries */) {
705                         size = limit - idx;
706                         if (size > ENTRIES_PER_PAGE)
707                                 size = ENTRIES_PER_PAGE;
708                         freed = shmem_map_and_free_swp(subdir,
709                                         offset, size, &dir, punch_lock);
710                         if (!dir)
711                                 dir = shmem_dir_map(middir);
712                         nr_swaps_freed += freed;
713                         if (offset || punch_lock) {
714                                 spin_lock(&info->lock);
715                                 set_page_private(subdir,
716                                         page_private(subdir) - freed);
717                                 spin_unlock(&info->lock);
718                         } else
719                                 BUG_ON(page_private(subdir) != freed);
720                 }
721                 offset = 0;
722         }
723 done1:
724         shmem_dir_unmap(dir);
725 done2:
726         if (inode->i_mapping->nrpages && (info->flags & SHMEM_PAGEIN)) {
727                 /*
728                  * Call truncate_inode_pages again: racing shmem_unuse_inode
729                  * may have swizzled a page in from swap since
730                  * truncate_pagecache or generic_delete_inode did it, before we
731                  * lowered next_index.  Also, though shmem_getpage checks
732                  * i_size before adding to cache, no recheck after: so fix the
733                  * narrow window there too.
734                  *
735                  * Recalling truncate_inode_pages_range and unmap_mapping_range
736                  * every time for punch_hole (which never got a chance to clear
737                  * SHMEM_PAGEIN at the start of vmtruncate_range) is expensive,
738                  * yet hardly ever necessary: try to optimize them out later.
739                  */
740                 truncate_inode_pages_range(inode->i_mapping, start, end);
741                 if (punch_hole)
742                         unmap_mapping_range(inode->i_mapping, start,
743                                                         end - start, 1);
744         }
745
746         spin_lock(&info->lock);
747         info->flags &= ~SHMEM_TRUNCATE;
748         info->swapped -= nr_swaps_freed;
749         if (nr_pages_to_free)
750                 shmem_free_blocks(inode, nr_pages_to_free);
751         shmem_recalc_inode(inode);
752         spin_unlock(&info->lock);
753
754         /*
755          * Empty swap vector directory pages to be freed?
756          */
757         if (!list_empty(&pages_to_free)) {
758                 pages_to_free.prev->next = NULL;
759                 shmem_free_pages(pages_to_free.next);
760         }
761 }
762
763 static int shmem_notify_change(struct dentry *dentry, struct iattr *attr)
764 {
765         struct inode *inode = dentry->d_inode;
766         loff_t newsize = attr->ia_size;
767         int error;
768
769         error = inode_change_ok(inode, attr);
770         if (error)
771                 return error;
772
773         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)
774                                         && newsize != inode->i_size) {
775                 struct page *page = NULL;
776
777                 if (newsize < inode->i_size) {
778                         /*
779                          * If truncating down to a partial page, then
780                          * if that page is already allocated, hold it
781                          * in memory until the truncation is over, so
782                          * truncate_partial_page cannot miss it were
783                          * it assigned to swap.
784                          */
785                         if (newsize & (PAGE_CACHE_SIZE-1)) {
786                                 (void) shmem_getpage(inode,
787                                         newsize >> PAGE_CACHE_SHIFT,
788                                                 &page, SGP_READ, NULL);
789                                 if (page)
790                                         unlock_page(page);
791                         }
792                         /*
793                          * Reset SHMEM_PAGEIN flag so that shmem_truncate can
794                          * detect if any pages might have been added to cache
795                          * after truncate_inode_pages.  But we needn't bother
796                          * if it's being fully truncated to zero-length: the
797                          * nrpages check is efficient enough in that case.
798                          */
799                         if (newsize) {
800                                 struct shmem_inode_info *info = SHMEM_I(inode);
801                                 spin_lock(&info->lock);
802                                 info->flags &= ~SHMEM_PAGEIN;
803                                 spin_unlock(&info->lock);
804                         }
805                 }
806
807                 /* XXX(truncate): truncate_setsize should be called last */
808                 truncate_setsize(inode, newsize);
809                 if (page)
810                         page_cache_release(page);
811                 shmem_truncate_range(inode, newsize, (loff_t)-1);
812         }
813
814         setattr_copy(inode, attr);
815 #ifdef CONFIG_TMPFS_POSIX_ACL
816         if (attr->ia_valid & ATTR_MODE)
817                 error = generic_acl_chmod(inode);
818 #endif
819         return error;
820 }
821
822 static void shmem_evict_inode(struct inode *inode)
823 {
824         struct shmem_inode_info *info = SHMEM_I(inode);
825
826         if (inode->i_mapping->a_ops == &shmem_aops) {
827                 truncate_inode_pages(inode->i_mapping, 0);
828                 shmem_unacct_size(info->flags, inode->i_size);
829                 inode->i_size = 0;
830                 shmem_truncate_range(inode, 0, (loff_t)-1);
831                 if (!list_empty(&info->swaplist)) {
832                         mutex_lock(&shmem_swaplist_mutex);
833                         list_del_init(&info->swaplist);
834                         mutex_unlock(&shmem_swaplist_mutex);
835                 }
836         }
837         BUG_ON(inode->i_blocks);
838         shmem_free_inode(inode->i_sb);
839         end_writeback(inode);
840 }
841
842 static inline int shmem_find_swp(swp_entry_t entry, swp_entry_t *dir, swp_entry_t *edir)
843 {
844         swp_entry_t *ptr;
845
846         for (ptr = dir; ptr < edir; ptr++) {
847                 if (ptr->val == entry.val)
848                         return ptr - dir;
849         }
850         return -1;
851 }
852
853 static int shmem_unuse_inode(struct shmem_inode_info *info, swp_entry_t entry, struct page *page)
854 {
855         struct address_space *mapping;
856         unsigned long idx;
857         unsigned long size;
858         unsigned long limit;
859         unsigned long stage;
860         struct page **dir;
861         struct page *subdir;
862         swp_entry_t *ptr;
863         int offset;
864         int error;
865
866         idx = 0;
867         ptr = info->i_direct;
868         spin_lock(&info->lock);
869         if (!info->swapped) {
870                 list_del_init(&info->swaplist);
871                 goto lost2;
872         }
873         limit = info->next_index;
874         size = limit;
875         if (size > SHMEM_NR_DIRECT)
876                 size = SHMEM_NR_DIRECT;
877         offset = shmem_find_swp(entry, ptr, ptr+size);
878         if (offset >= 0) {
879                 shmem_swp_balance_unmap();
880                 goto found;
881         }
882         if (!info->i_indirect)
883                 goto lost2;
884
885         dir = shmem_dir_map(info->i_indirect);
886         stage = SHMEM_NR_DIRECT + ENTRIES_PER_PAGEPAGE/2;
887
888         for (idx = SHMEM_NR_DIRECT; idx < limit; idx += ENTRIES_PER_PAGE, dir++) {
889                 if (unlikely(idx == stage)) {
890                         shmem_dir_unmap(dir-1);
891                         if (cond_resched_lock(&info->lock)) {
892                                 /* check it has not been truncated */
893                                 if (limit > info->next_index) {
894                                         limit = info->next_index;
895                                         if (idx >= limit)
896                                                 goto lost2;
897                                 }
898                         }
899                         dir = shmem_dir_map(info->i_indirect) +
900                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
901                         while (!*dir) {
902                                 dir++;
903                                 idx += ENTRIES_PER_PAGEPAGE;
904                                 if (idx >= limit)
905                                         goto lost1;
906                         }
907                         stage = idx + ENTRIES_PER_PAGEPAGE;
908                         subdir = *dir;
909                         shmem_dir_unmap(dir);
910                         dir = shmem_dir_map(subdir);
911                 }
912                 subdir = *dir;
913                 if (subdir && page_private(subdir)) {
914                         ptr = shmem_swp_map(subdir);
915                         size = limit - idx;
916                         if (size > ENTRIES_PER_PAGE)
917                                 size = ENTRIES_PER_PAGE;
918                         offset = shmem_find_swp(entry, ptr, ptr+size);
919                         if (offset >= 0) {
920                                 shmem_dir_unmap(dir);
921                                 goto found;
922                         }
923                         shmem_swp_unmap(ptr);
924                 }
925         }
926 lost1:
927         shmem_dir_unmap(dir-1);
928 lost2:
929         spin_unlock(&info->lock);
930         return 0;
931 found:
932         idx += offset;
933         ptr += offset;
934
935         /*
936          * Move _head_ to start search for next from here.
937          * But be careful: shmem_evict_inode checks list_empty without taking
938          * mutex, and there's an instant in list_move_tail when info->swaplist
939          * would appear empty, if it were the only one on shmem_swaplist.  We
940          * could avoid doing it if inode NULL; or use this minor optimization.
941          */
942         if (shmem_swaplist.next != &info->swaplist)
943                 list_move_tail(&shmem_swaplist, &info->swaplist);
944
945         /*
946          * We rely on shmem_swaplist_mutex, not only to protect the swaplist,
947          * but also to hold up shmem_evict_inode(): so inode cannot be freed
948          * beneath us (pagelock doesn't help until the page is in pagecache).
949          */
950         mapping = info->vfs_inode.i_mapping;
951         error = add_to_page_cache_locked(page, mapping, idx, GFP_NOWAIT);
952         /* which does mem_cgroup_uncharge_cache_page on error */
953
954         if (error == -EEXIST) {
955                 struct page *filepage = find_get_page(mapping, idx);
956                 error = 1;
957                 if (filepage) {
958                         /*
959                          * There might be a more uptodate page coming down
960                          * from a stacked writepage: forget our swappage if so.
961                          */
962                         if (PageUptodate(filepage))
963                                 error = 0;
964                         page_cache_release(filepage);
965                 }
966         }
967         if (!error) {
968                 delete_from_swap_cache(page);
969                 set_page_dirty(page);
970                 info->flags |= SHMEM_PAGEIN;
971                 shmem_swp_set(info, ptr, 0);
972                 swap_free(entry);
973                 error = 1;      /* not an error, but entry was found */
974         }
975         shmem_swp_unmap(ptr);
976         spin_unlock(&info->lock);
977         return error;
978 }
979
980 /*
981  * shmem_unuse() search for an eventually swapped out shmem page.
982  */
983 int shmem_unuse(swp_entry_t entry, struct page *page)
984 {
985         struct list_head *p, *next;
986         struct shmem_inode_info *info;
987         int found = 0;
988         int error;
989
990         /*
991          * Charge page using GFP_KERNEL while we can wait, before taking
992          * the shmem_swaplist_mutex which might hold up shmem_writepage().
993          * Charged back to the user (not to caller) when swap account is used.
994          * add_to_page_cache() will be called with GFP_NOWAIT.
995          */
996         error = mem_cgroup_cache_charge(page, current->mm, GFP_KERNEL);
997         if (error)
998                 goto out;
999         /*
1000          * Try to preload while we can wait, to not make a habit of
1001          * draining atomic reserves; but don't latch on to this cpu,
1002          * it's okay if sometimes we get rescheduled after this.
1003          */
1004         error = radix_tree_preload(GFP_KERNEL);
1005         if (error)
1006                 goto uncharge;
1007         radix_tree_preload_end();
1008
1009         mutex_lock(&shmem_swaplist_mutex);
1010         list_for_each_safe(p, next, &shmem_swaplist) {
1011                 info = list_entry(p, struct shmem_inode_info, swaplist);
1012                 found = shmem_unuse_inode(info, entry, page);
1013                 cond_resched();
1014                 if (found)
1015                         break;
1016         }
1017         mutex_unlock(&shmem_swaplist_mutex);
1018
1019 uncharge:
1020         if (!found)
1021                 mem_cgroup_uncharge_cache_page(page);
1022         if (found < 0)
1023                 error = found;
1024 out:
1025         unlock_page(page);
1026         page_cache_release(page);
1027         return error;
1028 }
1029
1030 /*
1031  * Move the page from the page cache to the swap cache.
1032  */
1033 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1034 {
1035         struct shmem_inode_info *info;
1036         swp_entry_t *entry, swap;
1037         struct address_space *mapping;
1038         unsigned long index;
1039         struct inode *inode;
1040         bool unlock_mutex = false;
1041
1042         BUG_ON(!PageLocked(page));
1043         mapping = page->mapping;
1044         index = page->index;
1045         inode = mapping->host;
1046         info = SHMEM_I(inode);
1047         if (info->flags & VM_LOCKED)
1048                 goto redirty;
1049         if (!total_swap_pages)
1050                 goto redirty;
1051
1052         /*
1053          * shmem_backing_dev_info's capabilities prevent regular writeback or
1054          * sync from ever calling shmem_writepage; but a stacking filesystem
1055          * may use the ->writepage of its underlying filesystem, in which case
1056          * tmpfs should write out to swap only in response to memory pressure,
1057          * and not for the writeback threads or sync.  However, in those cases,
1058          * we do still want to check if there's a redundant swappage to be
1059          * discarded.
1060          */
1061         if (wbc->for_reclaim)
1062                 swap = get_swap_page();
1063         else
1064                 swap.val = 0;
1065
1066         /*
1067          * Add inode to shmem_unuse()'s list of swapped-out inodes,
1068          * if it's not already there.  Do it now because we cannot take
1069          * mutex while holding spinlock, and must do so before the page
1070          * is moved to swap cache, when its pagelock no longer protects
1071          * the inode from eviction.  But don't unlock the mutex until
1072          * we've taken the spinlock, because shmem_unuse_inode() will
1073          * prune a !swapped inode from the swaplist under both locks.
1074          */
1075         if (swap.val && list_empty(&info->swaplist)) {
1076                 mutex_lock(&shmem_swaplist_mutex);
1077                 /* move instead of add in case we're racing */
1078                 list_move_tail(&info->swaplist, &shmem_swaplist);
1079                 unlock_mutex = true;
1080         }
1081
1082         spin_lock(&info->lock);
1083         if (unlock_mutex)
1084                 mutex_unlock(&shmem_swaplist_mutex);
1085
1086         if (index >= info->next_index) {
1087                 BUG_ON(!(info->flags & SHMEM_TRUNCATE));
1088                 goto unlock;
1089         }
1090         entry = shmem_swp_entry(info, index, NULL);
1091         if (entry->val) {
1092                 /*
1093                  * The more uptodate page coming down from a stacked
1094                  * writepage should replace our old swappage.
1095                  */
1096                 free_swap_and_cache(*entry);
1097                 shmem_swp_set(info, entry, 0);
1098         }
1099         shmem_recalc_inode(inode);
1100
1101         if (swap.val && add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
1102                 delete_from_page_cache(page);
1103                 shmem_swp_set(info, entry, swap.val);
1104                 shmem_swp_unmap(entry);
1105                 spin_unlock(&info->lock);
1106                 swap_shmem_alloc(swap);
1107                 BUG_ON(page_mapped(page));
1108                 swap_writepage(page, wbc);
1109                 return 0;
1110         }
1111
1112         shmem_swp_unmap(entry);
1113 unlock:
1114         spin_unlock(&info->lock);
1115         /*
1116          * add_to_swap_cache() doesn't return -EEXIST, so we can safely
1117          * clear SWAP_HAS_CACHE flag.
1118          */
1119         swapcache_free(swap, NULL);
1120 redirty:
1121         set_page_dirty(page);
1122         if (wbc->for_reclaim)
1123                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
1124         unlock_page(page);
1125         return 0;
1126 }
1127
1128 #ifdef CONFIG_NUMA
1129 #ifdef CONFIG_TMPFS
1130 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1131 {
1132         char buffer[64];
1133
1134         if (!mpol || mpol->mode == MPOL_DEFAULT)
1135                 return;         /* show nothing */
1136
1137         mpol_to_str(buffer, sizeof(buffer), mpol, 1);
1138
1139         seq_printf(seq, ",mpol=%s", buffer);
1140 }
1141
1142 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1143 {
1144         struct mempolicy *mpol = NULL;
1145         if (sbinfo->mpol) {
1146                 spin_lock(&sbinfo->stat_lock);  /* prevent replace/use races */
1147                 mpol = sbinfo->mpol;
1148                 mpol_get(mpol);
1149                 spin_unlock(&sbinfo->stat_lock);
1150         }
1151         return mpol;
1152 }
1153 #endif /* CONFIG_TMPFS */
1154
1155 static struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1156                         struct shmem_inode_info *info, unsigned long idx)
1157 {
1158         struct mempolicy mpol, *spol;
1159         struct vm_area_struct pvma;
1160         struct page *page;
1161
1162         spol = mpol_cond_copy(&mpol,
1163                                 mpol_shared_policy_lookup(&info->policy, idx));
1164
1165         /* Create a pseudo vma that just contains the policy */
1166         pvma.vm_start = 0;
1167         pvma.vm_pgoff = idx;
1168         pvma.vm_ops = NULL;
1169         pvma.vm_policy = spol;
1170         page = swapin_readahead(entry, gfp, &pvma, 0);
1171         return page;
1172 }
1173
1174 static struct page *shmem_alloc_page(gfp_t gfp,
1175                         struct shmem_inode_info *info, unsigned long idx)
1176 {
1177         struct vm_area_struct pvma;
1178
1179         /* Create a pseudo vma that just contains the policy */
1180         pvma.vm_start = 0;
1181         pvma.vm_pgoff = idx;
1182         pvma.vm_ops = NULL;
1183         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, idx);
1184
1185         /*
1186          * alloc_page_vma() will drop the shared policy reference
1187          */
1188         return alloc_page_vma(gfp, &pvma, 0);
1189 }
1190 #else /* !CONFIG_NUMA */
1191 #ifdef CONFIG_TMPFS
1192 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *p)
1193 {
1194 }
1195 #endif /* CONFIG_TMPFS */
1196
1197 static inline struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1198                         struct shmem_inode_info *info, unsigned long idx)
1199 {
1200         return swapin_readahead(entry, gfp, NULL, 0);
1201 }
1202
1203 static inline struct page *shmem_alloc_page(gfp_t gfp,
1204                         struct shmem_inode_info *info, unsigned long idx)
1205 {
1206         return alloc_page(gfp);
1207 }
1208 #endif /* CONFIG_NUMA */
1209
1210 #if !defined(CONFIG_NUMA) || !defined(CONFIG_TMPFS)
1211 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1212 {
1213         return NULL;
1214 }
1215 #endif
1216
1217 /*
1218  * shmem_getpage - either get the page from swap or allocate a new one
1219  *
1220  * If we allocate a new one we do not mark it dirty. That's up to the
1221  * vm. If we swap it in we mark it dirty since we also free the swap
1222  * entry since a page cannot live in both the swap and page cache
1223  */
1224 static int shmem_getpage(struct inode *inode, unsigned long idx,
1225                         struct page **pagep, enum sgp_type sgp, int *type)
1226 {
1227         struct address_space *mapping = inode->i_mapping;
1228         struct shmem_inode_info *info = SHMEM_I(inode);
1229         struct shmem_sb_info *sbinfo;
1230         struct page *filepage = *pagep;
1231         struct page *swappage;
1232         struct page *prealloc_page = NULL;
1233         swp_entry_t *entry;
1234         swp_entry_t swap;
1235         gfp_t gfp;
1236         int error;
1237
1238         if (idx >= SHMEM_MAX_INDEX)
1239                 return -EFBIG;
1240
1241         if (type)
1242                 *type = 0;
1243
1244         /*
1245          * Normally, filepage is NULL on entry, and either found
1246          * uptodate immediately, or allocated and zeroed, or read
1247          * in under swappage, which is then assigned to filepage.
1248          * But shmem_readpage (required for splice) passes in a locked
1249          * filepage, which may be found not uptodate by other callers
1250          * too, and may need to be copied from the swappage read in.
1251          */
1252 repeat:
1253         if (!filepage)
1254                 filepage = find_lock_page(mapping, idx);
1255         if (filepage && PageUptodate(filepage))
1256                 goto done;
1257         gfp = mapping_gfp_mask(mapping);
1258         if (!filepage) {
1259                 /*
1260                  * Try to preload while we can wait, to not make a habit of
1261                  * draining atomic reserves; but don't latch on to this cpu.
1262                  */
1263                 error = radix_tree_preload(gfp & ~__GFP_HIGHMEM);
1264                 if (error)
1265                         goto failed;
1266                 radix_tree_preload_end();
1267                 if (sgp != SGP_READ && !prealloc_page) {
1268                         /* We don't care if this fails */
1269                         prealloc_page = shmem_alloc_page(gfp, info, idx);
1270                         if (prealloc_page) {
1271                                 if (mem_cgroup_cache_charge(prealloc_page,
1272                                                 current->mm, GFP_KERNEL)) {
1273                                         page_cache_release(prealloc_page);
1274                                         prealloc_page = NULL;
1275                                 }
1276                         }
1277                 }
1278         }
1279         error = 0;
1280
1281         spin_lock(&info->lock);
1282         shmem_recalc_inode(inode);
1283         entry = shmem_swp_alloc(info, idx, sgp);
1284         if (IS_ERR(entry)) {
1285                 spin_unlock(&info->lock);
1286                 error = PTR_ERR(entry);
1287                 goto failed;
1288         }
1289         swap = *entry;
1290
1291         if (swap.val) {
1292                 /* Look it up and read it in.. */
1293                 swappage = lookup_swap_cache(swap);
1294                 if (!swappage) {
1295                         shmem_swp_unmap(entry);
1296                         /* here we actually do the io */
1297                         if (type && !(*type & VM_FAULT_MAJOR)) {
1298                                 __count_vm_event(PGMAJFAULT);
1299                                 *type |= VM_FAULT_MAJOR;
1300                         }
1301                         spin_unlock(&info->lock);
1302                         swappage = shmem_swapin(swap, gfp, info, idx);
1303                         if (!swappage) {
1304                                 spin_lock(&info->lock);
1305                                 entry = shmem_swp_alloc(info, idx, sgp);
1306                                 if (IS_ERR(entry))
1307                                         error = PTR_ERR(entry);
1308                                 else {
1309                                         if (entry->val == swap.val)
1310                                                 error = -ENOMEM;
1311                                         shmem_swp_unmap(entry);
1312                                 }
1313                                 spin_unlock(&info->lock);
1314                                 if (error)
1315                                         goto failed;
1316                                 goto repeat;
1317                         }
1318                         wait_on_page_locked(swappage);
1319                         page_cache_release(swappage);
1320                         goto repeat;
1321                 }
1322
1323                 /* We have to do this with page locked to prevent races */
1324                 if (!trylock_page(swappage)) {
1325                         shmem_swp_unmap(entry);
1326                         spin_unlock(&info->lock);
1327                         wait_on_page_locked(swappage);
1328                         page_cache_release(swappage);
1329                         goto repeat;
1330                 }
1331                 if (PageWriteback(swappage)) {
1332                         shmem_swp_unmap(entry);
1333                         spin_unlock(&info->lock);
1334                         wait_on_page_writeback(swappage);
1335                         unlock_page(swappage);
1336                         page_cache_release(swappage);
1337                         goto repeat;
1338                 }
1339                 if (!PageUptodate(swappage)) {
1340                         shmem_swp_unmap(entry);
1341                         spin_unlock(&info->lock);
1342                         unlock_page(swappage);
1343                         page_cache_release(swappage);
1344                         error = -EIO;
1345                         goto failed;
1346                 }
1347
1348                 if (filepage) {
1349                         shmem_swp_set(info, entry, 0);
1350                         shmem_swp_unmap(entry);
1351                         delete_from_swap_cache(swappage);
1352                         spin_unlock(&info->lock);
1353                         copy_highpage(filepage, swappage);
1354                         unlock_page(swappage);
1355                         page_cache_release(swappage);
1356                         flush_dcache_page(filepage);
1357                         SetPageUptodate(filepage);
1358                         set_page_dirty(filepage);
1359                         swap_free(swap);
1360                 } else if (!(error = add_to_page_cache_locked(swappage, mapping,
1361                                         idx, GFP_NOWAIT))) {
1362                         info->flags |= SHMEM_PAGEIN;
1363                         shmem_swp_set(info, entry, 0);
1364                         shmem_swp_unmap(entry);
1365                         delete_from_swap_cache(swappage);
1366                         spin_unlock(&info->lock);
1367                         filepage = swappage;
1368                         set_page_dirty(filepage);
1369                         swap_free(swap);
1370                 } else {
1371                         shmem_swp_unmap(entry);
1372                         spin_unlock(&info->lock);
1373                         if (error == -ENOMEM) {
1374                                 /*
1375                                  * reclaim from proper memory cgroup and
1376                                  * call memcg's OOM if needed.
1377                                  */
1378                                 error = mem_cgroup_shmem_charge_fallback(
1379                                                                 swappage,
1380                                                                 current->mm,
1381                                                                 gfp);
1382                                 if (error) {
1383                                         unlock_page(swappage);
1384                                         page_cache_release(swappage);
1385                                         goto failed;
1386                                 }
1387                         }
1388                         unlock_page(swappage);
1389                         page_cache_release(swappage);
1390                         goto repeat;
1391                 }
1392         } else if (sgp == SGP_READ && !filepage) {
1393                 shmem_swp_unmap(entry);
1394                 filepage = find_get_page(mapping, idx);
1395                 if (filepage &&
1396                     (!PageUptodate(filepage) || !trylock_page(filepage))) {
1397                         spin_unlock(&info->lock);
1398                         wait_on_page_locked(filepage);
1399                         page_cache_release(filepage);
1400                         filepage = NULL;
1401                         goto repeat;
1402                 }
1403                 spin_unlock(&info->lock);
1404         } else {
1405                 shmem_swp_unmap(entry);
1406                 sbinfo = SHMEM_SB(inode->i_sb);
1407                 if (sbinfo->max_blocks) {
1408                         if (percpu_counter_compare(&sbinfo->used_blocks,
1409                                                 sbinfo->max_blocks) >= 0 ||
1410                             shmem_acct_block(info->flags)) {
1411                                 spin_unlock(&info->lock);
1412                                 error = -ENOSPC;
1413                                 goto failed;
1414                         }
1415                         percpu_counter_inc(&sbinfo->used_blocks);
1416                         spin_lock(&inode->i_lock);
1417                         inode->i_blocks += BLOCKS_PER_PAGE;
1418                         spin_unlock(&inode->i_lock);
1419                 } else if (shmem_acct_block(info->flags)) {
1420                         spin_unlock(&info->lock);
1421                         error = -ENOSPC;
1422                         goto failed;
1423                 }
1424
1425                 if (!filepage) {
1426                         int ret;
1427
1428                         if (!prealloc_page) {
1429                                 spin_unlock(&info->lock);
1430                                 filepage = shmem_alloc_page(gfp, info, idx);
1431                                 if (!filepage) {
1432                                         shmem_unacct_blocks(info->flags, 1);
1433                                         shmem_free_blocks(inode, 1);
1434                                         error = -ENOMEM;
1435                                         goto failed;
1436                                 }
1437                                 SetPageSwapBacked(filepage);
1438
1439                                 /*
1440                                  * Precharge page while we can wait, compensate
1441                                  * after
1442                                  */
1443                                 error = mem_cgroup_cache_charge(filepage,
1444                                         current->mm, GFP_KERNEL);
1445                                 if (error) {
1446                                         page_cache_release(filepage);
1447                                         shmem_unacct_blocks(info->flags, 1);
1448                                         shmem_free_blocks(inode, 1);
1449                                         filepage = NULL;
1450                                         goto failed;
1451                                 }
1452
1453                                 spin_lock(&info->lock);
1454                         } else {
1455                                 filepage = prealloc_page;
1456                                 prealloc_page = NULL;
1457                                 SetPageSwapBacked(filepage);
1458                         }
1459
1460                         entry = shmem_swp_alloc(info, idx, sgp);
1461                         if (IS_ERR(entry))
1462                                 error = PTR_ERR(entry);
1463                         else {
1464                                 swap = *entry;
1465                                 shmem_swp_unmap(entry);
1466                         }
1467                         ret = error || swap.val;
1468                         if (ret)
1469                                 mem_cgroup_uncharge_cache_page(filepage);
1470                         else
1471                                 ret = add_to_page_cache_lru(filepage, mapping,
1472                                                 idx, GFP_NOWAIT);
1473                         /*
1474                          * At add_to_page_cache_lru() failure, uncharge will
1475                          * be done automatically.
1476                          */
1477                         if (ret) {
1478                                 spin_unlock(&info->lock);
1479                                 page_cache_release(filepage);
1480                                 shmem_unacct_blocks(info->flags, 1);
1481                                 shmem_free_blocks(inode, 1);
1482                                 filepage = NULL;
1483                                 if (error)
1484                                         goto failed;
1485                                 goto repeat;
1486                         }
1487                         info->flags |= SHMEM_PAGEIN;
1488                 }
1489
1490                 info->alloced++;
1491                 spin_unlock(&info->lock);
1492                 clear_highpage(filepage);
1493                 flush_dcache_page(filepage);
1494                 SetPageUptodate(filepage);
1495                 if (sgp == SGP_DIRTY)
1496                         set_page_dirty(filepage);
1497         }
1498 done:
1499         *pagep = filepage;
1500         error = 0;
1501         goto out;
1502
1503 failed:
1504         if (*pagep != filepage) {
1505                 unlock_page(filepage);
1506                 page_cache_release(filepage);
1507         }
1508 out:
1509         if (prealloc_page) {
1510                 mem_cgroup_uncharge_cache_page(prealloc_page);
1511                 page_cache_release(prealloc_page);
1512         }
1513         return error;
1514 }
1515
1516 static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1517 {
1518         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1519         int error;
1520         int ret;
1521
1522         if (((loff_t)vmf->pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
1523                 return VM_FAULT_SIGBUS;
1524
1525         error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_CACHE, &ret);
1526         if (error)
1527                 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1528
1529         return ret | VM_FAULT_LOCKED;
1530 }
1531
1532 #ifdef CONFIG_NUMA
1533 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *new)
1534 {
1535         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1536         return mpol_set_shared_policy(&SHMEM_I(i)->policy, vma, new);
1537 }
1538
1539 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
1540                                           unsigned long addr)
1541 {
1542         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1543         unsigned long idx;
1544
1545         idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1546         return mpol_shared_policy_lookup(&SHMEM_I(i)->policy, idx);
1547 }
1548 #endif
1549
1550 int shmem_lock(struct file *file, int lock, struct user_struct *user)
1551 {
1552         struct inode *inode = file->f_path.dentry->d_inode;
1553         struct shmem_inode_info *info = SHMEM_I(inode);
1554         int retval = -ENOMEM;
1555
1556         spin_lock(&info->lock);
1557         if (lock && !(info->flags & VM_LOCKED)) {
1558                 if (!user_shm_lock(inode->i_size, user))
1559                         goto out_nomem;
1560                 info->flags |= VM_LOCKED;
1561                 mapping_set_unevictable(file->f_mapping);
1562         }
1563         if (!lock && (info->flags & VM_LOCKED) && user) {
1564                 user_shm_unlock(inode->i_size, user);
1565                 info->flags &= ~VM_LOCKED;
1566                 mapping_clear_unevictable(file->f_mapping);
1567                 scan_mapping_unevictable_pages(file->f_mapping);
1568         }
1569         retval = 0;
1570
1571 out_nomem:
1572         spin_unlock(&info->lock);
1573         return retval;
1574 }
1575
1576 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1577 {
1578         file_accessed(file);
1579         vma->vm_ops = &shmem_vm_ops;
1580         vma->vm_flags |= VM_CAN_NONLINEAR;
1581         return 0;
1582 }
1583
1584 static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
1585                                      int mode, dev_t dev, unsigned long flags)
1586 {
1587         struct inode *inode;
1588         struct shmem_inode_info *info;
1589         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1590
1591         if (shmem_reserve_inode(sb))
1592                 return NULL;
1593
1594         inode = new_inode(sb);
1595         if (inode) {
1596                 inode->i_ino = get_next_ino();
1597                 inode_init_owner(inode, dir, mode);
1598                 inode->i_blocks = 0;
1599                 inode->i_mapping->backing_dev_info = &shmem_backing_dev_info;
1600                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1601                 inode->i_generation = get_seconds();
1602                 info = SHMEM_I(inode);
1603                 memset(info, 0, (char *)inode - (char *)info);
1604                 spin_lock_init(&info->lock);
1605                 info->flags = flags & VM_NORESERVE;
1606                 INIT_LIST_HEAD(&info->swaplist);
1607                 cache_no_acl(inode);
1608
1609                 switch (mode & S_IFMT) {
1610                 default:
1611                         inode->i_op = &shmem_special_inode_operations;
1612                         init_special_inode(inode, mode, dev);
1613                         break;
1614                 case S_IFREG:
1615                         inode->i_mapping->a_ops = &shmem_aops;
1616                         inode->i_op = &shmem_inode_operations;
1617                         inode->i_fop = &shmem_file_operations;
1618                         mpol_shared_policy_init(&info->policy,
1619                                                  shmem_get_sbmpol(sbinfo));
1620                         break;
1621                 case S_IFDIR:
1622                         inc_nlink(inode);
1623                         /* Some things misbehave if size == 0 on a directory */
1624                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
1625                         inode->i_op = &shmem_dir_inode_operations;
1626                         inode->i_fop = &simple_dir_operations;
1627                         break;
1628                 case S_IFLNK:
1629                         /*
1630                          * Must not load anything in the rbtree,
1631                          * mpol_free_shared_policy will not be called.
1632                          */
1633                         mpol_shared_policy_init(&info->policy, NULL);
1634                         break;
1635                 }
1636         } else
1637                 shmem_free_inode(sb);
1638         return inode;
1639 }
1640
1641 #ifdef CONFIG_TMPFS
1642 static const struct inode_operations shmem_symlink_inode_operations;
1643 static const struct inode_operations shmem_symlink_inline_operations;
1644
1645 /*
1646  * Normally tmpfs avoids the use of shmem_readpage and shmem_write_begin;
1647  * but providing them allows a tmpfs file to be used for splice, sendfile, and
1648  * below the loop driver, in the generic fashion that many filesystems support.
1649  */
1650 static int shmem_readpage(struct file *file, struct page *page)
1651 {
1652         struct inode *inode = page->mapping->host;
1653         int error = shmem_getpage(inode, page->index, &page, SGP_CACHE, NULL);
1654         unlock_page(page);
1655         return error;
1656 }
1657
1658 static int
1659 shmem_write_begin(struct file *file, struct address_space *mapping,
1660                         loff_t pos, unsigned len, unsigned flags,
1661                         struct page **pagep, void **fsdata)
1662 {
1663         struct inode *inode = mapping->host;
1664         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1665         *pagep = NULL;
1666         return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1667 }
1668
1669 static int
1670 shmem_write_end(struct file *file, struct address_space *mapping,
1671                         loff_t pos, unsigned len, unsigned copied,
1672                         struct page *page, void *fsdata)
1673 {
1674         struct inode *inode = mapping->host;
1675
1676         if (pos + copied > inode->i_size)
1677                 i_size_write(inode, pos + copied);
1678
1679         set_page_dirty(page);
1680         unlock_page(page);
1681         page_cache_release(page);
1682
1683         return copied;
1684 }
1685
1686 static void do_shmem_file_read(struct file *filp, loff_t *ppos, read_descriptor_t *desc, read_actor_t actor)
1687 {
1688         struct inode *inode = filp->f_path.dentry->d_inode;
1689         struct address_space *mapping = inode->i_mapping;
1690         unsigned long index, offset;
1691         enum sgp_type sgp = SGP_READ;
1692
1693         /*
1694          * Might this read be for a stacking filesystem?  Then when reading
1695          * holes of a sparse file, we actually need to allocate those pages,
1696          * and even mark them dirty, so it cannot exceed the max_blocks limit.
1697          */
1698         if (segment_eq(get_fs(), KERNEL_DS))
1699                 sgp = SGP_DIRTY;
1700
1701         index = *ppos >> PAGE_CACHE_SHIFT;
1702         offset = *ppos & ~PAGE_CACHE_MASK;
1703
1704         for (;;) {
1705                 struct page *page = NULL;
1706                 unsigned long end_index, nr, ret;
1707                 loff_t i_size = i_size_read(inode);
1708
1709                 end_index = i_size >> PAGE_CACHE_SHIFT;
1710                 if (index > end_index)
1711                         break;
1712                 if (index == end_index) {
1713                         nr = i_size & ~PAGE_CACHE_MASK;
1714                         if (nr <= offset)
1715                                 break;
1716                 }
1717
1718                 desc->error = shmem_getpage(inode, index, &page, sgp, NULL);
1719                 if (desc->error) {
1720                         if (desc->error == -EINVAL)
1721                                 desc->error = 0;
1722                         break;
1723                 }
1724                 if (page)
1725                         unlock_page(page);
1726
1727                 /*
1728                  * We must evaluate after, since reads (unlike writes)
1729                  * are called without i_mutex protection against truncate
1730                  */
1731                 nr = PAGE_CACHE_SIZE;
1732                 i_size = i_size_read(inode);
1733                 end_index = i_size >> PAGE_CACHE_SHIFT;
1734                 if (index == end_index) {
1735                         nr = i_size & ~PAGE_CACHE_MASK;
1736                         if (nr <= offset) {
1737                                 if (page)
1738                                         page_cache_release(page);
1739                                 break;
1740                         }
1741                 }
1742                 nr -= offset;
1743
1744                 if (page) {
1745                         /*
1746                          * If users can be writing to this page using arbitrary
1747                          * virtual addresses, take care about potential aliasing
1748                          * before reading the page on the kernel side.
1749                          */
1750                         if (mapping_writably_mapped(mapping))
1751                                 flush_dcache_page(page);
1752                         /*
1753                          * Mark the page accessed if we read the beginning.
1754                          */
1755                         if (!offset)
1756                                 mark_page_accessed(page);
1757                 } else {
1758                         page = ZERO_PAGE(0);
1759                         page_cache_get(page);
1760                 }
1761
1762                 /*
1763                  * Ok, we have the page, and it's up-to-date, so
1764                  * now we can copy it to user space...
1765                  *
1766                  * The actor routine returns how many bytes were actually used..
1767                  * NOTE! This may not be the same as how much of a user buffer
1768                  * we filled up (we may be padding etc), so we can only update
1769                  * "pos" here (the actor routine has to update the user buffer
1770                  * pointers and the remaining count).
1771                  */
1772                 ret = actor(desc, page, offset, nr);
1773                 offset += ret;
1774                 index += offset >> PAGE_CACHE_SHIFT;
1775                 offset &= ~PAGE_CACHE_MASK;
1776
1777                 page_cache_release(page);
1778                 if (ret != nr || !desc->count)
1779                         break;
1780
1781                 cond_resched();
1782         }
1783
1784         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1785         file_accessed(filp);
1786 }
1787
1788 static ssize_t shmem_file_aio_read(struct kiocb *iocb,
1789                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
1790 {
1791         struct file *filp = iocb->ki_filp;
1792         ssize_t retval;
1793         unsigned long seg;
1794         size_t count;
1795         loff_t *ppos = &iocb->ki_pos;
1796
1797         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1798         if (retval)
1799                 return retval;
1800
1801         for (seg = 0; seg < nr_segs; seg++) {
1802                 read_descriptor_t desc;
1803
1804                 desc.written = 0;
1805                 desc.arg.buf = iov[seg].iov_base;
1806                 desc.count = iov[seg].iov_len;
1807                 if (desc.count == 0)
1808                         continue;
1809                 desc.error = 0;
1810                 do_shmem_file_read(filp, ppos, &desc, file_read_actor);
1811                 retval += desc.written;
1812                 if (desc.error) {
1813                         retval = retval ?: desc.error;
1814                         break;
1815                 }
1816                 if (desc.count > 0)
1817                         break;
1818         }
1819         return retval;
1820 }
1821
1822 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
1823 {
1824         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
1825
1826         buf->f_type = TMPFS_MAGIC;
1827         buf->f_bsize = PAGE_CACHE_SIZE;
1828         buf->f_namelen = NAME_MAX;
1829         if (sbinfo->max_blocks) {
1830                 buf->f_blocks = sbinfo->max_blocks;
1831                 buf->f_bavail = buf->f_bfree =
1832                                 sbinfo->max_blocks - percpu_counter_sum(&sbinfo->used_blocks);
1833         }
1834         if (sbinfo->max_inodes) {
1835                 buf->f_files = sbinfo->max_inodes;
1836                 buf->f_ffree = sbinfo->free_inodes;
1837         }
1838         /* else leave those fields 0 like simple_statfs */
1839         return 0;
1840 }
1841
1842 /*
1843  * File creation. Allocate an inode, and we're done..
1844  */
1845 static int
1846 shmem_mknod(struct inode *dir, struct dentry *dentry, int mode, dev_t dev)
1847 {
1848         struct inode *inode;
1849         int error = -ENOSPC;
1850
1851         inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
1852         if (inode) {
1853                 error = security_inode_init_security(inode, dir,
1854                                                      &dentry->d_name, NULL,
1855                                                      NULL, NULL);
1856                 if (error) {
1857                         if (error != -EOPNOTSUPP) {
1858                                 iput(inode);
1859                                 return error;
1860                         }
1861                 }
1862 #ifdef CONFIG_TMPFS_POSIX_ACL
1863                 error = generic_acl_init(inode, dir);
1864                 if (error) {
1865                         iput(inode);
1866                         return error;
1867                 }
1868 #else
1869                 error = 0;
1870 #endif
1871                 dir->i_size += BOGO_DIRENT_SIZE;
1872                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1873                 d_instantiate(dentry, inode);
1874                 dget(dentry); /* Extra count - pin the dentry in core */
1875         }
1876         return error;
1877 }
1878
1879 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, int mode)
1880 {
1881         int error;
1882
1883         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
1884                 return error;
1885         inc_nlink(dir);
1886         return 0;
1887 }
1888
1889 static int shmem_create(struct inode *dir, struct dentry *dentry, int mode,
1890                 struct nameidata *nd)
1891 {
1892         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
1893 }
1894
1895 /*
1896  * Link a file..
1897  */
1898 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
1899 {
1900         struct inode *inode = old_dentry->d_inode;
1901         int ret;
1902
1903         /*
1904          * No ordinary (disk based) filesystem counts links as inodes;
1905          * but each new link needs a new dentry, pinning lowmem, and
1906          * tmpfs dentries cannot be pruned until they are unlinked.
1907          */
1908         ret = shmem_reserve_inode(inode->i_sb);
1909         if (ret)
1910                 goto out;
1911
1912         dir->i_size += BOGO_DIRENT_SIZE;
1913         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1914         inc_nlink(inode);
1915         ihold(inode);   /* New dentry reference */
1916         dget(dentry);           /* Extra pinning count for the created dentry */
1917         d_instantiate(dentry, inode);
1918 out:
1919         return ret;
1920 }
1921
1922 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
1923 {
1924         struct inode *inode = dentry->d_inode;
1925
1926         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
1927                 shmem_free_inode(inode->i_sb);
1928
1929         dir->i_size -= BOGO_DIRENT_SIZE;
1930         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1931         drop_nlink(inode);
1932         dput(dentry);   /* Undo the count from "create" - this does all the work */
1933         return 0;
1934 }
1935
1936 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
1937 {
1938         if (!simple_empty(dentry))
1939                 return -ENOTEMPTY;
1940
1941         drop_nlink(dentry->d_inode);
1942         drop_nlink(dir);
1943         return shmem_unlink(dir, dentry);
1944 }
1945
1946 /*
1947  * The VFS layer already does all the dentry stuff for rename,
1948  * we just have to decrement the usage count for the target if
1949  * it exists so that the VFS layer correctly free's it when it
1950  * gets overwritten.
1951  */
1952 static int shmem_rename(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
1953 {
1954         struct inode *inode = old_dentry->d_inode;
1955         int they_are_dirs = S_ISDIR(inode->i_mode);
1956
1957         if (!simple_empty(new_dentry))
1958                 return -ENOTEMPTY;
1959
1960         if (new_dentry->d_inode) {
1961                 (void) shmem_unlink(new_dir, new_dentry);
1962                 if (they_are_dirs)
1963                         drop_nlink(old_dir);
1964         } else if (they_are_dirs) {
1965                 drop_nlink(old_dir);
1966                 inc_nlink(new_dir);
1967         }
1968
1969         old_dir->i_size -= BOGO_DIRENT_SIZE;
1970         new_dir->i_size += BOGO_DIRENT_SIZE;
1971         old_dir->i_ctime = old_dir->i_mtime =
1972         new_dir->i_ctime = new_dir->i_mtime =
1973         inode->i_ctime = CURRENT_TIME;
1974         return 0;
1975 }
1976
1977 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
1978 {
1979         int error;
1980         int len;
1981         struct inode *inode;
1982         struct page *page = NULL;
1983         char *kaddr;
1984         struct shmem_inode_info *info;
1985
1986         len = strlen(symname) + 1;
1987         if (len > PAGE_CACHE_SIZE)
1988                 return -ENAMETOOLONG;
1989
1990         inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0, VM_NORESERVE);
1991         if (!inode)
1992                 return -ENOSPC;
1993
1994         error = security_inode_init_security(inode, dir, &dentry->d_name, NULL,
1995                                              NULL, NULL);
1996         if (error) {
1997                 if (error != -EOPNOTSUPP) {
1998                         iput(inode);
1999                         return error;
2000                 }
2001                 error = 0;
2002         }
2003
2004         info = SHMEM_I(inode);
2005         inode->i_size = len-1;
2006         if (len <= (char *)inode - (char *)info) {
2007                 /* do it inline */
2008                 memcpy(info, symname, len);
2009                 inode->i_op = &shmem_symlink_inline_operations;
2010         } else {
2011                 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
2012                 if (error) {
2013                         iput(inode);
2014                         return error;
2015                 }
2016                 inode->i_mapping->a_ops = &shmem_aops;
2017                 inode->i_op = &shmem_symlink_inode_operations;
2018                 kaddr = kmap_atomic(page, KM_USER0);
2019                 memcpy(kaddr, symname, len);
2020                 kunmap_atomic(kaddr, KM_USER0);
2021                 set_page_dirty(page);
2022                 unlock_page(page);
2023                 page_cache_release(page);
2024         }
2025         dir->i_size += BOGO_DIRENT_SIZE;
2026         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2027         d_instantiate(dentry, inode);
2028         dget(dentry);
2029         return 0;
2030 }
2031
2032 static void *shmem_follow_link_inline(struct dentry *dentry, struct nameidata *nd)
2033 {
2034         nd_set_link(nd, (char *)SHMEM_I(dentry->d_inode));
2035         return NULL;
2036 }
2037
2038 static void *shmem_follow_link(struct dentry *dentry, struct nameidata *nd)
2039 {
2040         struct page *page = NULL;
2041         int res = shmem_getpage(dentry->d_inode, 0, &page, SGP_READ, NULL);
2042         nd_set_link(nd, res ? ERR_PTR(res) : kmap(page));
2043         if (page)
2044                 unlock_page(page);
2045         return page;
2046 }
2047
2048 static void shmem_put_link(struct dentry *dentry, struct nameidata *nd, void *cookie)
2049 {
2050         if (!IS_ERR(nd_get_link(nd))) {
2051                 struct page *page = cookie;
2052                 kunmap(page);
2053                 mark_page_accessed(page);
2054                 page_cache_release(page);
2055         }
2056 }
2057
2058 static const struct inode_operations shmem_symlink_inline_operations = {
2059         .readlink       = generic_readlink,
2060         .follow_link    = shmem_follow_link_inline,
2061 };
2062
2063 static const struct inode_operations shmem_symlink_inode_operations = {
2064         .readlink       = generic_readlink,
2065         .follow_link    = shmem_follow_link,
2066         .put_link       = shmem_put_link,
2067 };
2068
2069 #ifdef CONFIG_TMPFS_POSIX_ACL
2070 /*
2071  * Superblocks without xattr inode operations will get security.* xattr
2072  * support from the VFS "for free". As soon as we have any other xattrs
2073  * like ACLs, we also need to implement the security.* handlers at
2074  * filesystem level, though.
2075  */
2076
2077 static size_t shmem_xattr_security_list(struct dentry *dentry, char *list,
2078                                         size_t list_len, const char *name,
2079                                         size_t name_len, int handler_flags)
2080 {
2081         return security_inode_listsecurity(dentry->d_inode, list, list_len);
2082 }
2083
2084 static int shmem_xattr_security_get(struct dentry *dentry, const char *name,
2085                 void *buffer, size_t size, int handler_flags)
2086 {
2087         if (strcmp(name, "") == 0)
2088                 return -EINVAL;
2089         return xattr_getsecurity(dentry->d_inode, name, buffer, size);
2090 }
2091
2092 static int shmem_xattr_security_set(struct dentry *dentry, const char *name,
2093                 const void *value, size_t size, int flags, int handler_flags)
2094 {
2095         if (strcmp(name, "") == 0)
2096                 return -EINVAL;
2097         return security_inode_setsecurity(dentry->d_inode, name, value,
2098                                           size, flags);
2099 }
2100
2101 static const struct xattr_handler shmem_xattr_security_handler = {
2102         .prefix = XATTR_SECURITY_PREFIX,
2103         .list   = shmem_xattr_security_list,
2104         .get    = shmem_xattr_security_get,
2105         .set    = shmem_xattr_security_set,
2106 };
2107
2108 static const struct xattr_handler *shmem_xattr_handlers[] = {
2109         &generic_acl_access_handler,
2110         &generic_acl_default_handler,
2111         &shmem_xattr_security_handler,
2112         NULL
2113 };
2114 #endif
2115
2116 static struct dentry *shmem_get_parent(struct dentry *child)
2117 {
2118         return ERR_PTR(-ESTALE);
2119 }
2120
2121 static int shmem_match(struct inode *ino, void *vfh)
2122 {
2123         __u32 *fh = vfh;
2124         __u64 inum = fh[2];
2125         inum = (inum << 32) | fh[1];
2126         return ino->i_ino == inum && fh[0] == ino->i_generation;
2127 }
2128
2129 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
2130                 struct fid *fid, int fh_len, int fh_type)
2131 {
2132         struct inode *inode;
2133         struct dentry *dentry = NULL;
2134         u64 inum = fid->raw[2];
2135         inum = (inum << 32) | fid->raw[1];
2136
2137         if (fh_len < 3)
2138                 return NULL;
2139
2140         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
2141                         shmem_match, fid->raw);
2142         if (inode) {
2143                 dentry = d_find_alias(inode);
2144                 iput(inode);
2145         }
2146
2147         return dentry;
2148 }
2149
2150 static int shmem_encode_fh(struct dentry *dentry, __u32 *fh, int *len,
2151                                 int connectable)
2152 {
2153         struct inode *inode = dentry->d_inode;
2154
2155         if (*len < 3) {
2156                 *len = 3;
2157                 return 255;
2158         }
2159
2160         if (inode_unhashed(inode)) {
2161                 /* Unfortunately insert_inode_hash is not idempotent,
2162                  * so as we hash inodes here rather than at creation
2163                  * time, we need a lock to ensure we only try
2164                  * to do it once
2165                  */
2166                 static DEFINE_SPINLOCK(lock);
2167                 spin_lock(&lock);
2168                 if (inode_unhashed(inode))
2169                         __insert_inode_hash(inode,
2170                                             inode->i_ino + inode->i_generation);
2171                 spin_unlock(&lock);
2172         }
2173
2174         fh[0] = inode->i_generation;
2175         fh[1] = inode->i_ino;
2176         fh[2] = ((__u64)inode->i_ino) >> 32;
2177
2178         *len = 3;
2179         return 1;
2180 }
2181
2182 static const struct export_operations shmem_export_ops = {
2183         .get_parent     = shmem_get_parent,
2184         .encode_fh      = shmem_encode_fh,
2185         .fh_to_dentry   = shmem_fh_to_dentry,
2186 };
2187
2188 static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
2189                                bool remount)
2190 {
2191         char *this_char, *value, *rest;
2192
2193         while (options != NULL) {
2194                 this_char = options;
2195                 for (;;) {
2196                         /*
2197                          * NUL-terminate this option: unfortunately,
2198                          * mount options form a comma-separated list,
2199                          * but mpol's nodelist may also contain commas.
2200                          */
2201                         options = strchr(options, ',');
2202                         if (options == NULL)
2203                                 break;
2204                         options++;
2205                         if (!isdigit(*options)) {
2206                                 options[-1] = '\0';
2207                                 break;
2208                         }
2209                 }
2210                 if (!*this_char)
2211                         continue;
2212                 if ((value = strchr(this_char,'=')) != NULL) {
2213                         *value++ = 0;
2214                 } else {
2215                         printk(KERN_ERR
2216                             "tmpfs: No value for mount option '%s'\n",
2217                             this_char);
2218                         return 1;
2219                 }
2220
2221                 if (!strcmp(this_char,"size")) {
2222                         unsigned long long size;
2223                         size = memparse(value,&rest);
2224                         if (*rest == '%') {
2225                                 size <<= PAGE_SHIFT;
2226                                 size *= totalram_pages;
2227                                 do_div(size, 100);
2228                                 rest++;
2229                         }
2230                         if (*rest)
2231                                 goto bad_val;
2232                         sbinfo->max_blocks =
2233                                 DIV_ROUND_UP(size, PAGE_CACHE_SIZE);
2234                 } else if (!strcmp(this_char,"nr_blocks")) {
2235                         sbinfo->max_blocks = memparse(value, &rest);
2236                         if (*rest)
2237                                 goto bad_val;
2238                 } else if (!strcmp(this_char,"nr_inodes")) {
2239                         sbinfo->max_inodes = memparse(value, &rest);
2240                         if (*rest)
2241                                 goto bad_val;
2242                 } else if (!strcmp(this_char,"mode")) {
2243                         if (remount)
2244                                 continue;
2245                         sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
2246                         if (*rest)
2247                                 goto bad_val;
2248                 } else if (!strcmp(this_char,"uid")) {
2249                         if (remount)
2250                                 continue;
2251                         sbinfo->uid = simple_strtoul(value, &rest, 0);
2252                         if (*rest)
2253                                 goto bad_val;
2254                 } else if (!strcmp(this_char,"gid")) {
2255                         if (remount)
2256                                 continue;
2257                         sbinfo->gid = simple_strtoul(value, &rest, 0);
2258                         if (*rest)
2259                                 goto bad_val;
2260                 } else if (!strcmp(this_char,"mpol")) {
2261                         if (mpol_parse_str(value, &sbinfo->mpol, 1))
2262                                 goto bad_val;
2263                 } else {
2264                         printk(KERN_ERR "tmpfs: Bad mount option %s\n",
2265                                this_char);
2266                         return 1;
2267                 }
2268         }
2269         return 0;
2270
2271 bad_val:
2272         printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
2273                value, this_char);
2274         return 1;
2275
2276 }
2277
2278 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2279 {
2280         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2281         struct shmem_sb_info config = *sbinfo;
2282         unsigned long inodes;
2283         int error = -EINVAL;
2284
2285         if (shmem_parse_options(data, &config, true))
2286                 return error;
2287
2288         spin_lock(&sbinfo->stat_lock);
2289         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2290         if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
2291                 goto out;
2292         if (config.max_inodes < inodes)
2293                 goto out;
2294         /*
2295          * Those tests also disallow limited->unlimited while any are in
2296          * use, so i_blocks will always be zero when max_blocks is zero;
2297          * but we must separately disallow unlimited->limited, because
2298          * in that case we have no record of how much is already in use.
2299          */
2300         if (config.max_blocks && !sbinfo->max_blocks)
2301                 goto out;
2302         if (config.max_inodes && !sbinfo->max_inodes)
2303                 goto out;
2304
2305         error = 0;
2306         sbinfo->max_blocks  = config.max_blocks;
2307         sbinfo->max_inodes  = config.max_inodes;
2308         sbinfo->free_inodes = config.max_inodes - inodes;
2309
2310         mpol_put(sbinfo->mpol);
2311         sbinfo->mpol        = config.mpol;      /* transfers initial ref */
2312 out:
2313         spin_unlock(&sbinfo->stat_lock);
2314         return error;
2315 }
2316
2317 static int shmem_show_options(struct seq_file *seq, struct vfsmount *vfs)
2318 {
2319         struct shmem_sb_info *sbinfo = SHMEM_SB(vfs->mnt_sb);
2320
2321         if (sbinfo->max_blocks != shmem_default_max_blocks())
2322                 seq_printf(seq, ",size=%luk",
2323                         sbinfo->max_blocks << (PAGE_CACHE_SHIFT - 10));
2324         if (sbinfo->max_inodes != shmem_default_max_inodes())
2325                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
2326         if (sbinfo->mode != (S_IRWXUGO | S_ISVTX))
2327                 seq_printf(seq, ",mode=%03o", sbinfo->mode);
2328         if (sbinfo->uid != 0)
2329                 seq_printf(seq, ",uid=%u", sbinfo->uid);
2330         if (sbinfo->gid != 0)
2331                 seq_printf(seq, ",gid=%u", sbinfo->gid);
2332         shmem_show_mpol(seq, sbinfo->mpol);
2333         return 0;
2334 }
2335 #endif /* CONFIG_TMPFS */
2336
2337 static void shmem_put_super(struct super_block *sb)
2338 {
2339         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2340
2341         percpu_counter_destroy(&sbinfo->used_blocks);
2342         kfree(sbinfo);
2343         sb->s_fs_info = NULL;
2344 }
2345
2346 int shmem_fill_super(struct super_block *sb, void *data, int silent)
2347 {
2348         struct inode *inode;
2349         struct dentry *root;
2350         struct shmem_sb_info *sbinfo;
2351         int err = -ENOMEM;
2352
2353         /* Round up to L1_CACHE_BYTES to resist false sharing */
2354         sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
2355                                 L1_CACHE_BYTES), GFP_KERNEL);
2356         if (!sbinfo)
2357                 return -ENOMEM;
2358
2359         sbinfo->mode = S_IRWXUGO | S_ISVTX;
2360         sbinfo->uid = current_fsuid();
2361         sbinfo->gid = current_fsgid();
2362         sb->s_fs_info = sbinfo;
2363
2364 #ifdef CONFIG_TMPFS
2365         /*
2366          * Per default we only allow half of the physical ram per
2367          * tmpfs instance, limiting inodes to one per page of lowmem;
2368          * but the internal instance is left unlimited.
2369          */
2370         if (!(sb->s_flags & MS_NOUSER)) {
2371                 sbinfo->max_blocks = shmem_default_max_blocks();
2372                 sbinfo->max_inodes = shmem_default_max_inodes();
2373                 if (shmem_parse_options(data, sbinfo, false)) {
2374                         err = -EINVAL;
2375                         goto failed;
2376                 }
2377         }
2378         sb->s_export_op = &shmem_export_ops;
2379 #else
2380         sb->s_flags |= MS_NOUSER;
2381 #endif
2382
2383         spin_lock_init(&sbinfo->stat_lock);
2384         if (percpu_counter_init(&sbinfo->used_blocks, 0))
2385                 goto failed;
2386         sbinfo->free_inodes = sbinfo->max_inodes;
2387
2388         sb->s_maxbytes = SHMEM_MAX_BYTES;
2389         sb->s_blocksize = PAGE_CACHE_SIZE;
2390         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
2391         sb->s_magic = TMPFS_MAGIC;
2392         sb->s_op = &shmem_ops;
2393         sb->s_time_gran = 1;
2394 #ifdef CONFIG_TMPFS_POSIX_ACL
2395         sb->s_xattr = shmem_xattr_handlers;
2396         sb->s_flags |= MS_POSIXACL;
2397 #endif
2398
2399         inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
2400         if (!inode)
2401                 goto failed;
2402         inode->i_uid = sbinfo->uid;
2403         inode->i_gid = sbinfo->gid;
2404         root = d_alloc_root(inode);
2405         if (!root)
2406                 goto failed_iput;
2407         sb->s_root = root;
2408         return 0;
2409
2410 failed_iput:
2411         iput(inode);
2412 failed:
2413         shmem_put_super(sb);
2414         return err;
2415 }
2416
2417 static struct kmem_cache *shmem_inode_cachep;
2418
2419 static struct inode *shmem_alloc_inode(struct super_block *sb)
2420 {
2421         struct shmem_inode_info *p;
2422         p = (struct shmem_inode_info *)kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
2423         if (!p)
2424                 return NULL;
2425         return &p->vfs_inode;
2426 }
2427
2428 static void shmem_i_callback(struct rcu_head *head)
2429 {
2430         struct inode *inode = container_of(head, struct inode, i_rcu);
2431         INIT_LIST_HEAD(&inode->i_dentry);
2432         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
2433 }
2434
2435 static void shmem_destroy_inode(struct inode *inode)
2436 {
2437         if ((inode->i_mode & S_IFMT) == S_IFREG) {
2438                 /* only struct inode is valid if it's an inline symlink */
2439                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
2440         }
2441         call_rcu(&inode->i_rcu, shmem_i_callback);
2442 }
2443
2444 static void init_once(void *foo)
2445 {
2446         struct shmem_inode_info *p = (struct shmem_inode_info *) foo;
2447
2448         inode_init_once(&p->vfs_inode);
2449 }
2450
2451 static int init_inodecache(void)
2452 {
2453         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
2454                                 sizeof(struct shmem_inode_info),
2455                                 0, SLAB_PANIC, init_once);
2456         return 0;
2457 }
2458
2459 static void destroy_inodecache(void)
2460 {
2461         kmem_cache_destroy(shmem_inode_cachep);
2462 }
2463
2464 static const struct address_space_operations shmem_aops = {
2465         .writepage      = shmem_writepage,
2466         .set_page_dirty = __set_page_dirty_no_writeback,
2467 #ifdef CONFIG_TMPFS
2468         .readpage       = shmem_readpage,
2469         .write_begin    = shmem_write_begin,
2470         .write_end      = shmem_write_end,
2471 #endif
2472         .migratepage    = migrate_page,
2473         .error_remove_page = generic_error_remove_page,
2474 };
2475
2476 static const struct file_operations shmem_file_operations = {
2477         .mmap           = shmem_mmap,
2478 #ifdef CONFIG_TMPFS
2479         .llseek         = generic_file_llseek,
2480         .read           = do_sync_read,
2481         .write          = do_sync_write,
2482         .aio_read       = shmem_file_aio_read,
2483         .aio_write      = generic_file_aio_write,
2484         .fsync          = noop_fsync,
2485         .splice_read    = generic_file_splice_read,
2486         .splice_write   = generic_file_splice_write,
2487 #endif
2488 };
2489
2490 static const struct inode_operations shmem_inode_operations = {
2491         .setattr        = shmem_notify_change,
2492         .truncate_range = shmem_truncate_range,
2493 #ifdef CONFIG_TMPFS_POSIX_ACL
2494         .setxattr       = generic_setxattr,
2495         .getxattr       = generic_getxattr,
2496         .listxattr      = generic_listxattr,
2497         .removexattr    = generic_removexattr,
2498         .check_acl      = generic_check_acl,
2499 #endif
2500
2501 };
2502
2503 static const struct inode_operations shmem_dir_inode_operations = {
2504 #ifdef CONFIG_TMPFS
2505         .create         = shmem_create,
2506         .lookup         = simple_lookup,
2507         .link           = shmem_link,
2508         .unlink         = shmem_unlink,
2509         .symlink        = shmem_symlink,
2510         .mkdir          = shmem_mkdir,
2511         .rmdir          = shmem_rmdir,
2512         .mknod          = shmem_mknod,
2513         .rename         = shmem_rename,
2514 #endif
2515 #ifdef CONFIG_TMPFS_POSIX_ACL
2516         .setattr        = shmem_notify_change,
2517         .setxattr       = generic_setxattr,
2518         .getxattr       = generic_getxattr,
2519         .listxattr      = generic_listxattr,
2520         .removexattr    = generic_removexattr,
2521         .check_acl      = generic_check_acl,
2522 #endif
2523 };
2524
2525 static const struct inode_operations shmem_special_inode_operations = {
2526 #ifdef CONFIG_TMPFS_POSIX_ACL
2527         .setattr        = shmem_notify_change,
2528         .setxattr       = generic_setxattr,
2529         .getxattr       = generic_getxattr,
2530         .listxattr      = generic_listxattr,
2531         .removexattr    = generic_removexattr,
2532         .check_acl      = generic_check_acl,
2533 #endif
2534 };
2535
2536 static const struct super_operations shmem_ops = {
2537         .alloc_inode    = shmem_alloc_inode,
2538         .destroy_inode  = shmem_destroy_inode,
2539 #ifdef CONFIG_TMPFS
2540         .statfs         = shmem_statfs,
2541         .remount_fs     = shmem_remount_fs,
2542         .show_options   = shmem_show_options,
2543 #endif
2544         .evict_inode    = shmem_evict_inode,
2545         .drop_inode     = generic_delete_inode,
2546         .put_super      = shmem_put_super,
2547 };
2548
2549 static const struct vm_operations_struct shmem_vm_ops = {
2550         .fault          = shmem_fault,
2551 #ifdef CONFIG_NUMA
2552         .set_policy     = shmem_set_policy,
2553         .get_policy     = shmem_get_policy,
2554 #endif
2555 };
2556
2557
2558 static struct dentry *shmem_mount(struct file_system_type *fs_type,
2559         int flags, const char *dev_name, void *data)
2560 {
2561         return mount_nodev(fs_type, flags, data, shmem_fill_super);
2562 }
2563
2564 static struct file_system_type tmpfs_fs_type = {
2565         .owner          = THIS_MODULE,
2566         .name           = "tmpfs",
2567         .mount          = shmem_mount,
2568         .kill_sb        = kill_litter_super,
2569 };
2570
2571 int __init init_tmpfs(void)
2572 {
2573         int error;
2574
2575         error = bdi_init(&shmem_backing_dev_info);
2576         if (error)
2577                 goto out4;
2578
2579         error = init_inodecache();
2580         if (error)
2581                 goto out3;
2582
2583         error = register_filesystem(&tmpfs_fs_type);
2584         if (error) {
2585                 printk(KERN_ERR "Could not register tmpfs\n");
2586                 goto out2;
2587         }
2588
2589         shm_mnt = vfs_kern_mount(&tmpfs_fs_type, MS_NOUSER,
2590                                 tmpfs_fs_type.name, NULL);
2591         if (IS_ERR(shm_mnt)) {
2592                 error = PTR_ERR(shm_mnt);
2593                 printk(KERN_ERR "Could not kern_mount tmpfs\n");
2594                 goto out1;
2595         }
2596         return 0;
2597
2598 out1:
2599         unregister_filesystem(&tmpfs_fs_type);
2600 out2:
2601         destroy_inodecache();
2602 out3:
2603         bdi_destroy(&shmem_backing_dev_info);
2604 out4:
2605         shm_mnt = ERR_PTR(error);
2606         return error;
2607 }
2608
2609 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2610 /**
2611  * mem_cgroup_get_shmem_target - find a page or entry assigned to the shmem file
2612  * @inode: the inode to be searched
2613  * @pgoff: the offset to be searched
2614  * @pagep: the pointer for the found page to be stored
2615  * @ent: the pointer for the found swap entry to be stored
2616  *
2617  * If a page is found, refcount of it is incremented. Callers should handle
2618  * these refcount.
2619  */
2620 void mem_cgroup_get_shmem_target(struct inode *inode, pgoff_t pgoff,
2621                                         struct page **pagep, swp_entry_t *ent)
2622 {
2623         swp_entry_t entry = { .val = 0 }, *ptr;
2624         struct page *page = NULL;
2625         struct shmem_inode_info *info = SHMEM_I(inode);
2626
2627         if ((pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
2628                 goto out;
2629
2630         spin_lock(&info->lock);
2631         ptr = shmem_swp_entry(info, pgoff, NULL);
2632 #ifdef CONFIG_SWAP
2633         if (ptr && ptr->val) {
2634                 entry.val = ptr->val;
2635                 page = find_get_page(&swapper_space, entry.val);
2636         } else
2637 #endif
2638                 page = find_get_page(inode->i_mapping, pgoff);
2639         if (ptr)
2640                 shmem_swp_unmap(ptr);
2641         spin_unlock(&info->lock);
2642 out:
2643         *pagep = page;
2644         *ent = entry;
2645 }
2646 #endif
2647
2648 #else /* !CONFIG_SHMEM */
2649
2650 /*
2651  * tiny-shmem: simple shmemfs and tmpfs using ramfs code
2652  *
2653  * This is intended for small system where the benefits of the full
2654  * shmem code (swap-backed and resource-limited) are outweighed by
2655  * their complexity. On systems without swap this code should be
2656  * effectively equivalent, but much lighter weight.
2657  */
2658
2659 #include <linux/ramfs.h>
2660
2661 static struct file_system_type tmpfs_fs_type = {
2662         .name           = "tmpfs",
2663         .mount          = ramfs_mount,
2664         .kill_sb        = kill_litter_super,
2665 };
2666
2667 int __init init_tmpfs(void)
2668 {
2669         BUG_ON(register_filesystem(&tmpfs_fs_type) != 0);
2670
2671         shm_mnt = kern_mount(&tmpfs_fs_type);
2672         BUG_ON(IS_ERR(shm_mnt));
2673
2674         return 0;
2675 }
2676
2677 int shmem_unuse(swp_entry_t entry, struct page *page)
2678 {
2679         return 0;
2680 }
2681
2682 int shmem_lock(struct file *file, int lock, struct user_struct *user)
2683 {
2684         return 0;
2685 }
2686
2687 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2688 /**
2689  * mem_cgroup_get_shmem_target - find a page or entry assigned to the shmem file
2690  * @inode: the inode to be searched
2691  * @pgoff: the offset to be searched
2692  * @pagep: the pointer for the found page to be stored
2693  * @ent: the pointer for the found swap entry to be stored
2694  *
2695  * If a page is found, refcount of it is incremented. Callers should handle
2696  * these refcount.
2697  */
2698 void mem_cgroup_get_shmem_target(struct inode *inode, pgoff_t pgoff,
2699                                         struct page **pagep, swp_entry_t *ent)
2700 {
2701         struct page *page = NULL;
2702
2703         if ((pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
2704                 goto out;
2705         page = find_get_page(inode->i_mapping, pgoff);
2706 out:
2707         *pagep = page;
2708         *ent = (swp_entry_t){ .val = 0 };
2709 }
2710 #endif
2711
2712 #define shmem_vm_ops                            generic_file_vm_ops
2713 #define shmem_file_operations                   ramfs_file_operations
2714 #define shmem_get_inode(sb, dir, mode, dev, flags)      ramfs_get_inode(sb, dir, mode, dev)
2715 #define shmem_acct_size(flags, size)            0
2716 #define shmem_unacct_size(flags, size)          do {} while (0)
2717 #define SHMEM_MAX_BYTES                         MAX_LFS_FILESIZE
2718
2719 #endif /* CONFIG_SHMEM */
2720
2721 /* common code */
2722
2723 /**
2724  * shmem_file_setup - get an unlinked file living in tmpfs
2725  * @name: name for dentry (to be seen in /proc/<pid>/maps
2726  * @size: size to be set for the file
2727  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
2728  */
2729 struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
2730 {
2731         int error;
2732         struct file *file;
2733         struct inode *inode;
2734         struct path path;
2735         struct dentry *root;
2736         struct qstr this;
2737
2738         if (IS_ERR(shm_mnt))
2739                 return (void *)shm_mnt;
2740
2741         if (size < 0 || size > SHMEM_MAX_BYTES)
2742                 return ERR_PTR(-EINVAL);
2743
2744         if (shmem_acct_size(flags, size))
2745                 return ERR_PTR(-ENOMEM);
2746
2747         error = -ENOMEM;
2748         this.name = name;
2749         this.len = strlen(name);
2750         this.hash = 0; /* will go */
2751         root = shm_mnt->mnt_root;
2752         path.dentry = d_alloc(root, &this);
2753         if (!path.dentry)
2754                 goto put_memory;
2755         path.mnt = mntget(shm_mnt);
2756
2757         error = -ENOSPC;
2758         inode = shmem_get_inode(root->d_sb, NULL, S_IFREG | S_IRWXUGO, 0, flags);
2759         if (!inode)
2760                 goto put_dentry;
2761
2762         d_instantiate(path.dentry, inode);
2763         inode->i_size = size;
2764         inode->i_nlink = 0;     /* It is unlinked */
2765 #ifndef CONFIG_MMU
2766         error = ramfs_nommu_expand_for_mapping(inode, size);
2767         if (error)
2768                 goto put_dentry;
2769 #endif
2770
2771         error = -ENFILE;
2772         file = alloc_file(&path, FMODE_WRITE | FMODE_READ,
2773                   &shmem_file_operations);
2774         if (!file)
2775                 goto put_dentry;
2776
2777         return file;
2778
2779 put_dentry:
2780         path_put(&path);
2781 put_memory:
2782         shmem_unacct_size(flags, size);
2783         return ERR_PTR(error);
2784 }
2785 EXPORT_SYMBOL_GPL(shmem_file_setup);
2786
2787 /**
2788  * shmem_zero_setup - setup a shared anonymous mapping
2789  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
2790  */
2791 int shmem_zero_setup(struct vm_area_struct *vma)
2792 {
2793         struct file *file;
2794         loff_t size = vma->vm_end - vma->vm_start;
2795
2796         file = shmem_file_setup("dev/zero", size, vma->vm_flags);
2797         if (IS_ERR(file))
2798                 return PTR_ERR(file);
2799
2800         if (vma->vm_file)
2801                 fput(vma->vm_file);
2802         vma->vm_file = file;
2803         vma->vm_ops = &shmem_vm_ops;
2804         vma->vm_flags |= VM_CAN_NONLINEAR;
2805         return 0;
2806 }