566f9a481e64afb3805d77d4a8465c85fb7e9ea9
[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         .unplug_io_fn   = default_unplug_io_fn,
228 };
229
230 static LIST_HEAD(shmem_swaplist);
231 static DEFINE_MUTEX(shmem_swaplist_mutex);
232
233 static void shmem_free_blocks(struct inode *inode, long pages)
234 {
235         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
236         if (sbinfo->max_blocks) {
237                 percpu_counter_add(&sbinfo->used_blocks, -pages);
238                 spin_lock(&inode->i_lock);
239                 inode->i_blocks -= pages*BLOCKS_PER_PAGE;
240                 spin_unlock(&inode->i_lock);
241         }
242 }
243
244 static int shmem_reserve_inode(struct super_block *sb)
245 {
246         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
247         if (sbinfo->max_inodes) {
248                 spin_lock(&sbinfo->stat_lock);
249                 if (!sbinfo->free_inodes) {
250                         spin_unlock(&sbinfo->stat_lock);
251                         return -ENOSPC;
252                 }
253                 sbinfo->free_inodes--;
254                 spin_unlock(&sbinfo->stat_lock);
255         }
256         return 0;
257 }
258
259 static void shmem_free_inode(struct super_block *sb)
260 {
261         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
262         if (sbinfo->max_inodes) {
263                 spin_lock(&sbinfo->stat_lock);
264                 sbinfo->free_inodes++;
265                 spin_unlock(&sbinfo->stat_lock);
266         }
267 }
268
269 /**
270  * shmem_recalc_inode - recalculate the size of an inode
271  * @inode: inode to recalc
272  *
273  * We have to calculate the free blocks since the mm can drop
274  * undirtied hole pages behind our back.
275  *
276  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
277  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
278  *
279  * It has to be called with the spinlock held.
280  */
281 static void shmem_recalc_inode(struct inode *inode)
282 {
283         struct shmem_inode_info *info = SHMEM_I(inode);
284         long freed;
285
286         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
287         if (freed > 0) {
288                 info->alloced -= freed;
289                 shmem_unacct_blocks(info->flags, freed);
290                 shmem_free_blocks(inode, freed);
291         }
292 }
293
294 /**
295  * shmem_swp_entry - find the swap vector position in the info structure
296  * @info:  info structure for the inode
297  * @index: index of the page to find
298  * @page:  optional page to add to the structure. Has to be preset to
299  *         all zeros
300  *
301  * If there is no space allocated yet it will return NULL when
302  * page is NULL, else it will use the page for the needed block,
303  * setting it to NULL on return to indicate that it has been used.
304  *
305  * The swap vector is organized the following way:
306  *
307  * There are SHMEM_NR_DIRECT entries directly stored in the
308  * shmem_inode_info structure. So small files do not need an addional
309  * allocation.
310  *
311  * For pages with index > SHMEM_NR_DIRECT there is the pointer
312  * i_indirect which points to a page which holds in the first half
313  * doubly indirect blocks, in the second half triple indirect blocks:
314  *
315  * For an artificial ENTRIES_PER_PAGE = 4 this would lead to the
316  * following layout (for SHMEM_NR_DIRECT == 16):
317  *
318  * i_indirect -> dir --> 16-19
319  *            |      +-> 20-23
320  *            |
321  *            +-->dir2 --> 24-27
322  *            |        +-> 28-31
323  *            |        +-> 32-35
324  *            |        +-> 36-39
325  *            |
326  *            +-->dir3 --> 40-43
327  *                     +-> 44-47
328  *                     +-> 48-51
329  *                     +-> 52-55
330  */
331 static swp_entry_t *shmem_swp_entry(struct shmem_inode_info *info, unsigned long index, struct page **page)
332 {
333         unsigned long offset;
334         struct page **dir;
335         struct page *subdir;
336
337         if (index < SHMEM_NR_DIRECT) {
338                 shmem_swp_balance_unmap();
339                 return info->i_direct+index;
340         }
341         if (!info->i_indirect) {
342                 if (page) {
343                         info->i_indirect = *page;
344                         *page = NULL;
345                 }
346                 return NULL;                    /* need another page */
347         }
348
349         index -= SHMEM_NR_DIRECT;
350         offset = index % ENTRIES_PER_PAGE;
351         index /= ENTRIES_PER_PAGE;
352         dir = shmem_dir_map(info->i_indirect);
353
354         if (index >= ENTRIES_PER_PAGE/2) {
355                 index -= ENTRIES_PER_PAGE/2;
356                 dir += ENTRIES_PER_PAGE/2 + index/ENTRIES_PER_PAGE;
357                 index %= ENTRIES_PER_PAGE;
358                 subdir = *dir;
359                 if (!subdir) {
360                         if (page) {
361                                 *dir = *page;
362                                 *page = NULL;
363                         }
364                         shmem_dir_unmap(dir);
365                         return NULL;            /* need another page */
366                 }
367                 shmem_dir_unmap(dir);
368                 dir = shmem_dir_map(subdir);
369         }
370
371         dir += index;
372         subdir = *dir;
373         if (!subdir) {
374                 if (!page || !(subdir = *page)) {
375                         shmem_dir_unmap(dir);
376                         return NULL;            /* need a page */
377                 }
378                 *dir = subdir;
379                 *page = NULL;
380         }
381         shmem_dir_unmap(dir);
382         return shmem_swp_map(subdir) + offset;
383 }
384
385 static void shmem_swp_set(struct shmem_inode_info *info, swp_entry_t *entry, unsigned long value)
386 {
387         long incdec = value? 1: -1;
388
389         entry->val = value;
390         info->swapped += incdec;
391         if ((unsigned long)(entry - info->i_direct) >= SHMEM_NR_DIRECT) {
392                 struct page *page = kmap_atomic_to_page(entry);
393                 set_page_private(page, page_private(page) + incdec);
394         }
395 }
396
397 /**
398  * shmem_swp_alloc - get the position of the swap entry for the page.
399  * @info:       info structure for the inode
400  * @index:      index of the page to find
401  * @sgp:        check and recheck i_size? skip allocation?
402  *
403  * If the entry does not exist, allocate it.
404  */
405 static swp_entry_t *shmem_swp_alloc(struct shmem_inode_info *info, unsigned long index, enum sgp_type sgp)
406 {
407         struct inode *inode = &info->vfs_inode;
408         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
409         struct page *page = NULL;
410         swp_entry_t *entry;
411
412         if (sgp != SGP_WRITE &&
413             ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode))
414                 return ERR_PTR(-EINVAL);
415
416         while (!(entry = shmem_swp_entry(info, index, &page))) {
417                 if (sgp == SGP_READ)
418                         return shmem_swp_map(ZERO_PAGE(0));
419                 /*
420                  * Test used_blocks against 1 less max_blocks, since we have 1 data
421                  * page (and perhaps indirect index pages) yet to allocate:
422                  * a waste to allocate index if we cannot allocate data.
423                  */
424                 if (sbinfo->max_blocks) {
425                         if (percpu_counter_compare(&sbinfo->used_blocks, (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         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)
770                                         && newsize != inode->i_size) {
771                 struct page *page = NULL;
772
773                 if (newsize < inode->i_size) {
774                         /*
775                          * If truncating down to a partial page, then
776                          * if that page is already allocated, hold it
777                          * in memory until the truncation is over, so
778                          * truncate_partial_page cannnot miss it were
779                          * it assigned to swap.
780                          */
781                         if (newsize & (PAGE_CACHE_SIZE-1)) {
782                                 (void) shmem_getpage(inode,
783                                         newsize >> PAGE_CACHE_SHIFT,
784                                                 &page, SGP_READ, NULL);
785                                 if (page)
786                                         unlock_page(page);
787                         }
788                         /*
789                          * Reset SHMEM_PAGEIN flag so that shmem_truncate can
790                          * detect if any pages might have been added to cache
791                          * after truncate_inode_pages.  But we needn't bother
792                          * if it's being fully truncated to zero-length: the
793                          * nrpages check is efficient enough in that case.
794                          */
795                         if (newsize) {
796                                 struct shmem_inode_info *info = SHMEM_I(inode);
797                                 spin_lock(&info->lock);
798                                 info->flags &= ~SHMEM_PAGEIN;
799                                 spin_unlock(&info->lock);
800                         }
801                 }
802
803                 error = simple_setsize(inode, newsize);
804                 if (page)
805                         page_cache_release(page);
806                 if (error)
807                         return error;
808                 shmem_truncate_range(inode, newsize, (loff_t)-1);
809         }
810
811         error = inode_change_ok(inode, attr);
812         if (!error)
813                 generic_setattr(inode, attr);
814 #ifdef CONFIG_TMPFS_POSIX_ACL
815         if (!error && (attr->ia_valid & ATTR_MODE))
816                 error = generic_acl_chmod(inode);
817 #endif
818         return error;
819 }
820
821 static void shmem_delete_inode(struct inode *inode)
822 {
823         struct shmem_inode_info *info = SHMEM_I(inode);
824
825         if (inode->i_mapping->a_ops == &shmem_aops) {
826                 truncate_inode_pages(inode->i_mapping, 0);
827                 shmem_unacct_size(info->flags, inode->i_size);
828                 inode->i_size = 0;
829                 shmem_truncate_range(inode, 0, (loff_t)-1);
830                 if (!list_empty(&info->swaplist)) {
831                         mutex_lock(&shmem_swaplist_mutex);
832                         list_del_init(&info->swaplist);
833                         mutex_unlock(&shmem_swaplist_mutex);
834                 }
835         }
836         BUG_ON(inode->i_blocks);
837         shmem_free_inode(inode->i_sb);
838         clear_inode(inode);
839 }
840
841 static inline int shmem_find_swp(swp_entry_t entry, swp_entry_t *dir, swp_entry_t *edir)
842 {
843         swp_entry_t *ptr;
844
845         for (ptr = dir; ptr < edir; ptr++) {
846                 if (ptr->val == entry.val)
847                         return ptr - dir;
848         }
849         return -1;
850 }
851
852 static int shmem_unuse_inode(struct shmem_inode_info *info, swp_entry_t entry, struct page *page)
853 {
854         struct inode *inode;
855         unsigned long idx;
856         unsigned long size;
857         unsigned long limit;
858         unsigned long stage;
859         struct page **dir;
860         struct page *subdir;
861         swp_entry_t *ptr;
862         int offset;
863         int error;
864
865         idx = 0;
866         ptr = info->i_direct;
867         spin_lock(&info->lock);
868         if (!info->swapped) {
869                 list_del_init(&info->swaplist);
870                 goto lost2;
871         }
872         limit = info->next_index;
873         size = limit;
874         if (size > SHMEM_NR_DIRECT)
875                 size = SHMEM_NR_DIRECT;
876         offset = shmem_find_swp(entry, ptr, ptr+size);
877         if (offset >= 0)
878                 goto found;
879         if (!info->i_indirect)
880                 goto lost2;
881
882         dir = shmem_dir_map(info->i_indirect);
883         stage = SHMEM_NR_DIRECT + ENTRIES_PER_PAGEPAGE/2;
884
885         for (idx = SHMEM_NR_DIRECT; idx < limit; idx += ENTRIES_PER_PAGE, dir++) {
886                 if (unlikely(idx == stage)) {
887                         shmem_dir_unmap(dir-1);
888                         if (cond_resched_lock(&info->lock)) {
889                                 /* check it has not been truncated */
890                                 if (limit > info->next_index) {
891                                         limit = info->next_index;
892                                         if (idx >= limit)
893                                                 goto lost2;
894                                 }
895                         }
896                         dir = shmem_dir_map(info->i_indirect) +
897                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
898                         while (!*dir) {
899                                 dir++;
900                                 idx += ENTRIES_PER_PAGEPAGE;
901                                 if (idx >= limit)
902                                         goto lost1;
903                         }
904                         stage = idx + ENTRIES_PER_PAGEPAGE;
905                         subdir = *dir;
906                         shmem_dir_unmap(dir);
907                         dir = shmem_dir_map(subdir);
908                 }
909                 subdir = *dir;
910                 if (subdir && page_private(subdir)) {
911                         ptr = shmem_swp_map(subdir);
912                         size = limit - idx;
913                         if (size > ENTRIES_PER_PAGE)
914                                 size = ENTRIES_PER_PAGE;
915                         offset = shmem_find_swp(entry, ptr, ptr+size);
916                         shmem_swp_unmap(ptr);
917                         if (offset >= 0) {
918                                 shmem_dir_unmap(dir);
919                                 goto found;
920                         }
921                 }
922         }
923 lost1:
924         shmem_dir_unmap(dir-1);
925 lost2:
926         spin_unlock(&info->lock);
927         return 0;
928 found:
929         idx += offset;
930         inode = igrab(&info->vfs_inode);
931         spin_unlock(&info->lock);
932
933         /*
934          * Move _head_ to start search for next from here.
935          * But be careful: shmem_delete_inode checks list_empty without taking
936          * mutex, and there's an instant in list_move_tail when info->swaplist
937          * would appear empty, if it were the only one on shmem_swaplist.  We
938          * could avoid doing it if inode NULL; or use this minor optimization.
939          */
940         if (shmem_swaplist.next != &info->swaplist)
941                 list_move_tail(&shmem_swaplist, &info->swaplist);
942         mutex_unlock(&shmem_swaplist_mutex);
943
944         error = 1;
945         if (!inode)
946                 goto out;
947         /*
948          * Charge page using GFP_KERNEL while we can wait.
949          * Charged back to the user(not to caller) when swap account is used.
950          * add_to_page_cache() will be called with GFP_NOWAIT.
951          */
952         error = mem_cgroup_cache_charge(page, current->mm, GFP_KERNEL);
953         if (error)
954                 goto out;
955         error = radix_tree_preload(GFP_KERNEL);
956         if (error) {
957                 mem_cgroup_uncharge_cache_page(page);
958                 goto out;
959         }
960         error = 1;
961
962         spin_lock(&info->lock);
963         ptr = shmem_swp_entry(info, idx, NULL);
964         if (ptr && ptr->val == entry.val) {
965                 error = add_to_page_cache_locked(page, inode->i_mapping,
966                                                 idx, GFP_NOWAIT);
967                 /* does mem_cgroup_uncharge_cache_page on error */
968         } else  /* we must compensate for our precharge above */
969                 mem_cgroup_uncharge_cache_page(page);
970
971         if (error == -EEXIST) {
972                 struct page *filepage = find_get_page(inode->i_mapping, idx);
973                 error = 1;
974                 if (filepage) {
975                         /*
976                          * There might be a more uptodate page coming down
977                          * from a stacked writepage: forget our swappage if so.
978                          */
979                         if (PageUptodate(filepage))
980                                 error = 0;
981                         page_cache_release(filepage);
982                 }
983         }
984         if (!error) {
985                 delete_from_swap_cache(page);
986                 set_page_dirty(page);
987                 info->flags |= SHMEM_PAGEIN;
988                 shmem_swp_set(info, ptr, 0);
989                 swap_free(entry);
990                 error = 1;      /* not an error, but entry was found */
991         }
992         if (ptr)
993                 shmem_swp_unmap(ptr);
994         spin_unlock(&info->lock);
995         radix_tree_preload_end();
996 out:
997         unlock_page(page);
998         page_cache_release(page);
999         iput(inode);            /* allows for NULL */
1000         return error;
1001 }
1002
1003 /*
1004  * shmem_unuse() search for an eventually swapped out shmem page.
1005  */
1006 int shmem_unuse(swp_entry_t entry, struct page *page)
1007 {
1008         struct list_head *p, *next;
1009         struct shmem_inode_info *info;
1010         int found = 0;
1011
1012         mutex_lock(&shmem_swaplist_mutex);
1013         list_for_each_safe(p, next, &shmem_swaplist) {
1014                 info = list_entry(p, struct shmem_inode_info, swaplist);
1015                 found = shmem_unuse_inode(info, entry, page);
1016                 cond_resched();
1017                 if (found)
1018                         goto out;
1019         }
1020         mutex_unlock(&shmem_swaplist_mutex);
1021         /*
1022          * Can some race bring us here?  We've been holding page lock,
1023          * so I think not; but would rather try again later than BUG()
1024          */
1025         unlock_page(page);
1026         page_cache_release(page);
1027 out:
1028         return (found < 0) ? found : 0;
1029 }
1030
1031 /*
1032  * Move the page from the page cache to the swap cache.
1033  */
1034 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1035 {
1036         struct shmem_inode_info *info;
1037         swp_entry_t *entry, swap;
1038         struct address_space *mapping;
1039         unsigned long index;
1040         struct inode *inode;
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         spin_lock(&info->lock);
1067         if (index >= info->next_index) {
1068                 BUG_ON(!(info->flags & SHMEM_TRUNCATE));
1069                 goto unlock;
1070         }
1071         entry = shmem_swp_entry(info, index, NULL);
1072         if (entry->val) {
1073                 /*
1074                  * The more uptodate page coming down from a stacked
1075                  * writepage should replace our old swappage.
1076                  */
1077                 free_swap_and_cache(*entry);
1078                 shmem_swp_set(info, entry, 0);
1079         }
1080         shmem_recalc_inode(inode);
1081
1082         if (swap.val && add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
1083                 remove_from_page_cache(page);
1084                 shmem_swp_set(info, entry, swap.val);
1085                 shmem_swp_unmap(entry);
1086                 if (list_empty(&info->swaplist))
1087                         inode = igrab(inode);
1088                 else
1089                         inode = NULL;
1090                 spin_unlock(&info->lock);
1091                 swap_shmem_alloc(swap);
1092                 BUG_ON(page_mapped(page));
1093                 page_cache_release(page);       /* pagecache ref */
1094                 swap_writepage(page, wbc);
1095                 if (inode) {
1096                         mutex_lock(&shmem_swaplist_mutex);
1097                         /* move instead of add in case we're racing */
1098                         list_move_tail(&info->swaplist, &shmem_swaplist);
1099                         mutex_unlock(&shmem_swaplist_mutex);
1100                         iput(inode);
1101                 }
1102                 return 0;
1103         }
1104
1105         shmem_swp_unmap(entry);
1106 unlock:
1107         spin_unlock(&info->lock);
1108         /*
1109          * add_to_swap_cache() doesn't return -EEXIST, so we can safely
1110          * clear SWAP_HAS_CACHE flag.
1111          */
1112         swapcache_free(swap, NULL);
1113 redirty:
1114         set_page_dirty(page);
1115         if (wbc->for_reclaim)
1116                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
1117         unlock_page(page);
1118         return 0;
1119 }
1120
1121 #ifdef CONFIG_NUMA
1122 #ifdef CONFIG_TMPFS
1123 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1124 {
1125         char buffer[64];
1126
1127         if (!mpol || mpol->mode == MPOL_DEFAULT)
1128                 return;         /* show nothing */
1129
1130         mpol_to_str(buffer, sizeof(buffer), mpol, 1);
1131
1132         seq_printf(seq, ",mpol=%s", buffer);
1133 }
1134
1135 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1136 {
1137         struct mempolicy *mpol = NULL;
1138         if (sbinfo->mpol) {
1139                 spin_lock(&sbinfo->stat_lock);  /* prevent replace/use races */
1140                 mpol = sbinfo->mpol;
1141                 mpol_get(mpol);
1142                 spin_unlock(&sbinfo->stat_lock);
1143         }
1144         return mpol;
1145 }
1146 #endif /* CONFIG_TMPFS */
1147
1148 static struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1149                         struct shmem_inode_info *info, unsigned long idx)
1150 {
1151         struct mempolicy mpol, *spol;
1152         struct vm_area_struct pvma;
1153         struct page *page;
1154
1155         spol = mpol_cond_copy(&mpol,
1156                                 mpol_shared_policy_lookup(&info->policy, idx));
1157
1158         /* Create a pseudo vma that just contains the policy */
1159         pvma.vm_start = 0;
1160         pvma.vm_pgoff = idx;
1161         pvma.vm_ops = NULL;
1162         pvma.vm_policy = spol;
1163         page = swapin_readahead(entry, gfp, &pvma, 0);
1164         return page;
1165 }
1166
1167 static struct page *shmem_alloc_page(gfp_t gfp,
1168                         struct shmem_inode_info *info, unsigned long idx)
1169 {
1170         struct vm_area_struct pvma;
1171
1172         /* Create a pseudo vma that just contains the policy */
1173         pvma.vm_start = 0;
1174         pvma.vm_pgoff = idx;
1175         pvma.vm_ops = NULL;
1176         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, idx);
1177
1178         /*
1179          * alloc_page_vma() will drop the shared policy reference
1180          */
1181         return alloc_page_vma(gfp, &pvma, 0);
1182 }
1183 #else /* !CONFIG_NUMA */
1184 #ifdef CONFIG_TMPFS
1185 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *p)
1186 {
1187 }
1188 #endif /* CONFIG_TMPFS */
1189
1190 static inline struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1191                         struct shmem_inode_info *info, unsigned long idx)
1192 {
1193         return swapin_readahead(entry, gfp, NULL, 0);
1194 }
1195
1196 static inline struct page *shmem_alloc_page(gfp_t gfp,
1197                         struct shmem_inode_info *info, unsigned long idx)
1198 {
1199         return alloc_page(gfp);
1200 }
1201 #endif /* CONFIG_NUMA */
1202
1203 #if !defined(CONFIG_NUMA) || !defined(CONFIG_TMPFS)
1204 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1205 {
1206         return NULL;
1207 }
1208 #endif
1209
1210 /*
1211  * shmem_getpage - either get the page from swap or allocate a new one
1212  *
1213  * If we allocate a new one we do not mark it dirty. That's up to the
1214  * vm. If we swap it in we mark it dirty since we also free the swap
1215  * entry since a page cannot live in both the swap and page cache
1216  */
1217 static int shmem_getpage(struct inode *inode, unsigned long idx,
1218                         struct page **pagep, enum sgp_type sgp, int *type)
1219 {
1220         struct address_space *mapping = inode->i_mapping;
1221         struct shmem_inode_info *info = SHMEM_I(inode);
1222         struct shmem_sb_info *sbinfo;
1223         struct page *filepage = *pagep;
1224         struct page *swappage;
1225         struct page *prealloc_page = NULL;
1226         swp_entry_t *entry;
1227         swp_entry_t swap;
1228         gfp_t gfp;
1229         int error;
1230
1231         if (idx >= SHMEM_MAX_INDEX)
1232                 return -EFBIG;
1233
1234         if (type)
1235                 *type = 0;
1236
1237         /*
1238          * Normally, filepage is NULL on entry, and either found
1239          * uptodate immediately, or allocated and zeroed, or read
1240          * in under swappage, which is then assigned to filepage.
1241          * But shmem_readpage (required for splice) passes in a locked
1242          * filepage, which may be found not uptodate by other callers
1243          * too, and may need to be copied from the swappage read in.
1244          */
1245 repeat:
1246         if (!filepage)
1247                 filepage = find_lock_page(mapping, idx);
1248         if (filepage && PageUptodate(filepage))
1249                 goto done;
1250         gfp = mapping_gfp_mask(mapping);
1251         if (!filepage) {
1252                 /*
1253                  * Try to preload while we can wait, to not make a habit of
1254                  * draining atomic reserves; but don't latch on to this cpu.
1255                  */
1256                 error = radix_tree_preload(gfp & ~__GFP_HIGHMEM);
1257                 if (error)
1258                         goto failed;
1259                 radix_tree_preload_end();
1260                 if (sgp != SGP_READ && !prealloc_page) {
1261                         /* We don't care if this fails */
1262                         prealloc_page = shmem_alloc_page(gfp, info, idx);
1263                         if (prealloc_page) {
1264                                 if (mem_cgroup_cache_charge(prealloc_page,
1265                                                 current->mm, GFP_KERNEL)) {
1266                                         page_cache_release(prealloc_page);
1267                                         prealloc_page = NULL;
1268                                 }
1269                         }
1270                 }
1271         }
1272         error = 0;
1273
1274         spin_lock(&info->lock);
1275         shmem_recalc_inode(inode);
1276         entry = shmem_swp_alloc(info, idx, sgp);
1277         if (IS_ERR(entry)) {
1278                 spin_unlock(&info->lock);
1279                 error = PTR_ERR(entry);
1280                 goto failed;
1281         }
1282         swap = *entry;
1283
1284         if (swap.val) {
1285                 /* Look it up and read it in.. */
1286                 swappage = lookup_swap_cache(swap);
1287                 if (!swappage) {
1288                         shmem_swp_unmap(entry);
1289                         /* here we actually do the io */
1290                         if (type && !(*type & VM_FAULT_MAJOR)) {
1291                                 __count_vm_event(PGMAJFAULT);
1292                                 *type |= VM_FAULT_MAJOR;
1293                         }
1294                         spin_unlock(&info->lock);
1295                         swappage = shmem_swapin(swap, gfp, info, idx);
1296                         if (!swappage) {
1297                                 spin_lock(&info->lock);
1298                                 entry = shmem_swp_alloc(info, idx, sgp);
1299                                 if (IS_ERR(entry))
1300                                         error = PTR_ERR(entry);
1301                                 else {
1302                                         if (entry->val == swap.val)
1303                                                 error = -ENOMEM;
1304                                         shmem_swp_unmap(entry);
1305                                 }
1306                                 spin_unlock(&info->lock);
1307                                 if (error)
1308                                         goto failed;
1309                                 goto repeat;
1310                         }
1311                         wait_on_page_locked(swappage);
1312                         page_cache_release(swappage);
1313                         goto repeat;
1314                 }
1315
1316                 /* We have to do this with page locked to prevent races */
1317                 if (!trylock_page(swappage)) {
1318                         shmem_swp_unmap(entry);
1319                         spin_unlock(&info->lock);
1320                         wait_on_page_locked(swappage);
1321                         page_cache_release(swappage);
1322                         goto repeat;
1323                 }
1324                 if (PageWriteback(swappage)) {
1325                         shmem_swp_unmap(entry);
1326                         spin_unlock(&info->lock);
1327                         wait_on_page_writeback(swappage);
1328                         unlock_page(swappage);
1329                         page_cache_release(swappage);
1330                         goto repeat;
1331                 }
1332                 if (!PageUptodate(swappage)) {
1333                         shmem_swp_unmap(entry);
1334                         spin_unlock(&info->lock);
1335                         unlock_page(swappage);
1336                         page_cache_release(swappage);
1337                         error = -EIO;
1338                         goto failed;
1339                 }
1340
1341                 if (filepage) {
1342                         shmem_swp_set(info, entry, 0);
1343                         shmem_swp_unmap(entry);
1344                         delete_from_swap_cache(swappage);
1345                         spin_unlock(&info->lock);
1346                         copy_highpage(filepage, swappage);
1347                         unlock_page(swappage);
1348                         page_cache_release(swappage);
1349                         flush_dcache_page(filepage);
1350                         SetPageUptodate(filepage);
1351                         set_page_dirty(filepage);
1352                         swap_free(swap);
1353                 } else if (!(error = add_to_page_cache_locked(swappage, mapping,
1354                                         idx, GFP_NOWAIT))) {
1355                         info->flags |= SHMEM_PAGEIN;
1356                         shmem_swp_set(info, entry, 0);
1357                         shmem_swp_unmap(entry);
1358                         delete_from_swap_cache(swappage);
1359                         spin_unlock(&info->lock);
1360                         filepage = swappage;
1361                         set_page_dirty(filepage);
1362                         swap_free(swap);
1363                 } else {
1364                         shmem_swp_unmap(entry);
1365                         spin_unlock(&info->lock);
1366                         if (error == -ENOMEM) {
1367                                 /*
1368                                  * reclaim from proper memory cgroup and
1369                                  * call memcg's OOM if needed.
1370                                  */
1371                                 error = mem_cgroup_shmem_charge_fallback(
1372                                                                 swappage,
1373                                                                 current->mm,
1374                                                                 gfp);
1375                                 if (error) {
1376                                         unlock_page(swappage);
1377                                         page_cache_release(swappage);
1378                                         goto failed;
1379                                 }
1380                         }
1381                         unlock_page(swappage);
1382                         page_cache_release(swappage);
1383                         goto repeat;
1384                 }
1385         } else if (sgp == SGP_READ && !filepage) {
1386                 shmem_swp_unmap(entry);
1387                 filepage = find_get_page(mapping, idx);
1388                 if (filepage &&
1389                     (!PageUptodate(filepage) || !trylock_page(filepage))) {
1390                         spin_unlock(&info->lock);
1391                         wait_on_page_locked(filepage);
1392                         page_cache_release(filepage);
1393                         filepage = NULL;
1394                         goto repeat;
1395                 }
1396                 spin_unlock(&info->lock);
1397         } else {
1398                 shmem_swp_unmap(entry);
1399                 sbinfo = SHMEM_SB(inode->i_sb);
1400                 if (sbinfo->max_blocks) {
1401                         if ((percpu_counter_compare(&sbinfo->used_blocks, sbinfo->max_blocks) > 0) ||
1402                             shmem_acct_block(info->flags)) {
1403                                 spin_unlock(&info->lock);
1404                                 error = -ENOSPC;
1405                                 goto failed;
1406                         }
1407                         percpu_counter_inc(&sbinfo->used_blocks);
1408                         spin_lock(&inode->i_lock);
1409                         inode->i_blocks += BLOCKS_PER_PAGE;
1410                         spin_unlock(&inode->i_lock);
1411                 } else if (shmem_acct_block(info->flags)) {
1412                         spin_unlock(&info->lock);
1413                         error = -ENOSPC;
1414                         goto failed;
1415                 }
1416
1417                 if (!filepage) {
1418                         int ret;
1419
1420                         if (!prealloc_page) {
1421                                 spin_unlock(&info->lock);
1422                                 filepage = shmem_alloc_page(gfp, info, idx);
1423                                 if (!filepage) {
1424                                         shmem_unacct_blocks(info->flags, 1);
1425                                         shmem_free_blocks(inode, 1);
1426                                         error = -ENOMEM;
1427                                         goto failed;
1428                                 }
1429                                 SetPageSwapBacked(filepage);
1430
1431                                 /*
1432                                  * Precharge page while we can wait, compensate
1433                                  * after
1434                                  */
1435                                 error = mem_cgroup_cache_charge(filepage,
1436                                         current->mm, GFP_KERNEL);
1437                                 if (error) {
1438                                         page_cache_release(filepage);
1439                                         shmem_unacct_blocks(info->flags, 1);
1440                                         shmem_free_blocks(inode, 1);
1441                                         filepage = NULL;
1442                                         goto failed;
1443                                 }
1444
1445                                 spin_lock(&info->lock);
1446                         } else {
1447                                 filepage = prealloc_page;
1448                                 prealloc_page = NULL;
1449                                 SetPageSwapBacked(filepage);
1450                         }
1451
1452                         entry = shmem_swp_alloc(info, idx, sgp);
1453                         if (IS_ERR(entry))
1454                                 error = PTR_ERR(entry);
1455                         else {
1456                                 swap = *entry;
1457                                 shmem_swp_unmap(entry);
1458                         }
1459                         ret = error || swap.val;
1460                         if (ret)
1461                                 mem_cgroup_uncharge_cache_page(filepage);
1462                         else
1463                                 ret = add_to_page_cache_lru(filepage, mapping,
1464                                                 idx, GFP_NOWAIT);
1465                         /*
1466                          * At add_to_page_cache_lru() failure, uncharge will
1467                          * be done automatically.
1468                          */
1469                         if (ret) {
1470                                 spin_unlock(&info->lock);
1471                                 page_cache_release(filepage);
1472                                 shmem_unacct_blocks(info->flags, 1);
1473                                 shmem_free_blocks(inode, 1);
1474                                 filepage = NULL;
1475                                 if (error)
1476                                         goto failed;
1477                                 goto repeat;
1478                         }
1479                         info->flags |= SHMEM_PAGEIN;
1480                 }
1481
1482                 info->alloced++;
1483                 spin_unlock(&info->lock);
1484                 clear_highpage(filepage);
1485                 flush_dcache_page(filepage);
1486                 SetPageUptodate(filepage);
1487                 if (sgp == SGP_DIRTY)
1488                         set_page_dirty(filepage);
1489         }
1490 done:
1491         *pagep = filepage;
1492         error = 0;
1493         goto out;
1494
1495 failed:
1496         if (*pagep != filepage) {
1497                 unlock_page(filepage);
1498                 page_cache_release(filepage);
1499         }
1500 out:
1501         if (prealloc_page) {
1502                 mem_cgroup_uncharge_cache_page(prealloc_page);
1503                 page_cache_release(prealloc_page);
1504         }
1505         return error;
1506 }
1507
1508 static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1509 {
1510         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1511         int error;
1512         int ret;
1513
1514         if (((loff_t)vmf->pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
1515                 return VM_FAULT_SIGBUS;
1516
1517         error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_CACHE, &ret);
1518         if (error)
1519                 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1520
1521         return ret | VM_FAULT_LOCKED;
1522 }
1523
1524 #ifdef CONFIG_NUMA
1525 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *new)
1526 {
1527         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1528         return mpol_set_shared_policy(&SHMEM_I(i)->policy, vma, new);
1529 }
1530
1531 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
1532                                           unsigned long addr)
1533 {
1534         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1535         unsigned long idx;
1536
1537         idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1538         return mpol_shared_policy_lookup(&SHMEM_I(i)->policy, idx);
1539 }
1540 #endif
1541
1542 int shmem_lock(struct file *file, int lock, struct user_struct *user)
1543 {
1544         struct inode *inode = file->f_path.dentry->d_inode;
1545         struct shmem_inode_info *info = SHMEM_I(inode);
1546         int retval = -ENOMEM;
1547
1548         spin_lock(&info->lock);
1549         if (lock && !(info->flags & VM_LOCKED)) {
1550                 if (!user_shm_lock(inode->i_size, user))
1551                         goto out_nomem;
1552                 info->flags |= VM_LOCKED;
1553                 mapping_set_unevictable(file->f_mapping);
1554         }
1555         if (!lock && (info->flags & VM_LOCKED) && user) {
1556                 user_shm_unlock(inode->i_size, user);
1557                 info->flags &= ~VM_LOCKED;
1558                 mapping_clear_unevictable(file->f_mapping);
1559                 scan_mapping_unevictable_pages(file->f_mapping);
1560         }
1561         retval = 0;
1562
1563 out_nomem:
1564         spin_unlock(&info->lock);
1565         return retval;
1566 }
1567
1568 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1569 {
1570         file_accessed(file);
1571         vma->vm_ops = &shmem_vm_ops;
1572         vma->vm_flags |= VM_CAN_NONLINEAR;
1573         return 0;
1574 }
1575
1576 static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
1577                                      int mode, dev_t dev, unsigned long flags)
1578 {
1579         struct inode *inode;
1580         struct shmem_inode_info *info;
1581         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1582
1583         if (shmem_reserve_inode(sb))
1584                 return NULL;
1585
1586         inode = new_inode(sb);
1587         if (inode) {
1588                 inode_init_owner(inode, dir, mode);
1589                 inode->i_blocks = 0;
1590                 inode->i_mapping->backing_dev_info = &shmem_backing_dev_info;
1591                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1592                 inode->i_generation = get_seconds();
1593                 info = SHMEM_I(inode);
1594                 memset(info, 0, (char *)inode - (char *)info);
1595                 spin_lock_init(&info->lock);
1596                 info->flags = flags & VM_NORESERVE;
1597                 INIT_LIST_HEAD(&info->swaplist);
1598                 cache_no_acl(inode);
1599
1600                 switch (mode & S_IFMT) {
1601                 default:
1602                         inode->i_op = &shmem_special_inode_operations;
1603                         init_special_inode(inode, mode, dev);
1604                         break;
1605                 case S_IFREG:
1606                         inode->i_mapping->a_ops = &shmem_aops;
1607                         inode->i_op = &shmem_inode_operations;
1608                         inode->i_fop = &shmem_file_operations;
1609                         mpol_shared_policy_init(&info->policy,
1610                                                  shmem_get_sbmpol(sbinfo));
1611                         break;
1612                 case S_IFDIR:
1613                         inc_nlink(inode);
1614                         /* Some things misbehave if size == 0 on a directory */
1615                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
1616                         inode->i_op = &shmem_dir_inode_operations;
1617                         inode->i_fop = &simple_dir_operations;
1618                         break;
1619                 case S_IFLNK:
1620                         /*
1621                          * Must not load anything in the rbtree,
1622                          * mpol_free_shared_policy will not be called.
1623                          */
1624                         mpol_shared_policy_init(&info->policy, NULL);
1625                         break;
1626                 }
1627         } else
1628                 shmem_free_inode(sb);
1629         return inode;
1630 }
1631
1632 #ifdef CONFIG_TMPFS
1633 static const struct inode_operations shmem_symlink_inode_operations;
1634 static const struct inode_operations shmem_symlink_inline_operations;
1635
1636 /*
1637  * Normally tmpfs avoids the use of shmem_readpage and shmem_write_begin;
1638  * but providing them allows a tmpfs file to be used for splice, sendfile, and
1639  * below the loop driver, in the generic fashion that many filesystems support.
1640  */
1641 static int shmem_readpage(struct file *file, struct page *page)
1642 {
1643         struct inode *inode = page->mapping->host;
1644         int error = shmem_getpage(inode, page->index, &page, SGP_CACHE, NULL);
1645         unlock_page(page);
1646         return error;
1647 }
1648
1649 static int
1650 shmem_write_begin(struct file *file, struct address_space *mapping,
1651                         loff_t pos, unsigned len, unsigned flags,
1652                         struct page **pagep, void **fsdata)
1653 {
1654         struct inode *inode = mapping->host;
1655         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1656         *pagep = NULL;
1657         return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1658 }
1659
1660 static int
1661 shmem_write_end(struct file *file, struct address_space *mapping,
1662                         loff_t pos, unsigned len, unsigned copied,
1663                         struct page *page, void *fsdata)
1664 {
1665         struct inode *inode = mapping->host;
1666
1667         if (pos + copied > inode->i_size)
1668                 i_size_write(inode, pos + copied);
1669
1670         set_page_dirty(page);
1671         unlock_page(page);
1672         page_cache_release(page);
1673
1674         return copied;
1675 }
1676
1677 static void do_shmem_file_read(struct file *filp, loff_t *ppos, read_descriptor_t *desc, read_actor_t actor)
1678 {
1679         struct inode *inode = filp->f_path.dentry->d_inode;
1680         struct address_space *mapping = inode->i_mapping;
1681         unsigned long index, offset;
1682         enum sgp_type sgp = SGP_READ;
1683
1684         /*
1685          * Might this read be for a stacking filesystem?  Then when reading
1686          * holes of a sparse file, we actually need to allocate those pages,
1687          * and even mark them dirty, so it cannot exceed the max_blocks limit.
1688          */
1689         if (segment_eq(get_fs(), KERNEL_DS))
1690                 sgp = SGP_DIRTY;
1691
1692         index = *ppos >> PAGE_CACHE_SHIFT;
1693         offset = *ppos & ~PAGE_CACHE_MASK;
1694
1695         for (;;) {
1696                 struct page *page = NULL;
1697                 unsigned long end_index, nr, ret;
1698                 loff_t i_size = i_size_read(inode);
1699
1700                 end_index = i_size >> PAGE_CACHE_SHIFT;
1701                 if (index > end_index)
1702                         break;
1703                 if (index == end_index) {
1704                         nr = i_size & ~PAGE_CACHE_MASK;
1705                         if (nr <= offset)
1706                                 break;
1707                 }
1708
1709                 desc->error = shmem_getpage(inode, index, &page, sgp, NULL);
1710                 if (desc->error) {
1711                         if (desc->error == -EINVAL)
1712                                 desc->error = 0;
1713                         break;
1714                 }
1715                 if (page)
1716                         unlock_page(page);
1717
1718                 /*
1719                  * We must evaluate after, since reads (unlike writes)
1720                  * are called without i_mutex protection against truncate
1721                  */
1722                 nr = PAGE_CACHE_SIZE;
1723                 i_size = i_size_read(inode);
1724                 end_index = i_size >> PAGE_CACHE_SHIFT;
1725                 if (index == end_index) {
1726                         nr = i_size & ~PAGE_CACHE_MASK;
1727                         if (nr <= offset) {
1728                                 if (page)
1729                                         page_cache_release(page);
1730                                 break;
1731                         }
1732                 }
1733                 nr -= offset;
1734
1735                 if (page) {
1736                         /*
1737                          * If users can be writing to this page using arbitrary
1738                          * virtual addresses, take care about potential aliasing
1739                          * before reading the page on the kernel side.
1740                          */
1741                         if (mapping_writably_mapped(mapping))
1742                                 flush_dcache_page(page);
1743                         /*
1744                          * Mark the page accessed if we read the beginning.
1745                          */
1746                         if (!offset)
1747                                 mark_page_accessed(page);
1748                 } else {
1749                         page = ZERO_PAGE(0);
1750                         page_cache_get(page);
1751                 }
1752
1753                 /*
1754                  * Ok, we have the page, and it's up-to-date, so
1755                  * now we can copy it to user space...
1756                  *
1757                  * The actor routine returns how many bytes were actually used..
1758                  * NOTE! This may not be the same as how much of a user buffer
1759                  * we filled up (we may be padding etc), so we can only update
1760                  * "pos" here (the actor routine has to update the user buffer
1761                  * pointers and the remaining count).
1762                  */
1763                 ret = actor(desc, page, offset, nr);
1764                 offset += ret;
1765                 index += offset >> PAGE_CACHE_SHIFT;
1766                 offset &= ~PAGE_CACHE_MASK;
1767
1768                 page_cache_release(page);
1769                 if (ret != nr || !desc->count)
1770                         break;
1771
1772                 cond_resched();
1773         }
1774
1775         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1776         file_accessed(filp);
1777 }
1778
1779 static ssize_t shmem_file_aio_read(struct kiocb *iocb,
1780                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
1781 {
1782         struct file *filp = iocb->ki_filp;
1783         ssize_t retval;
1784         unsigned long seg;
1785         size_t count;
1786         loff_t *ppos = &iocb->ki_pos;
1787
1788         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1789         if (retval)
1790                 return retval;
1791
1792         for (seg = 0; seg < nr_segs; seg++) {
1793                 read_descriptor_t desc;
1794
1795                 desc.written = 0;
1796                 desc.arg.buf = iov[seg].iov_base;
1797                 desc.count = iov[seg].iov_len;
1798                 if (desc.count == 0)
1799                         continue;
1800                 desc.error = 0;
1801                 do_shmem_file_read(filp, ppos, &desc, file_read_actor);
1802                 retval += desc.written;
1803                 if (desc.error) {
1804                         retval = retval ?: desc.error;
1805                         break;
1806                 }
1807                 if (desc.count > 0)
1808                         break;
1809         }
1810         return retval;
1811 }
1812
1813 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
1814 {
1815         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
1816
1817         buf->f_type = TMPFS_MAGIC;
1818         buf->f_bsize = PAGE_CACHE_SIZE;
1819         buf->f_namelen = NAME_MAX;
1820         if (sbinfo->max_blocks) {
1821                 buf->f_blocks = sbinfo->max_blocks;
1822                 buf->f_bavail = buf->f_bfree =
1823                                 sbinfo->max_blocks - percpu_counter_sum(&sbinfo->used_blocks);
1824         }
1825         if (sbinfo->max_inodes) {
1826                 buf->f_files = sbinfo->max_inodes;
1827                 buf->f_ffree = sbinfo->free_inodes;
1828         }
1829         /* else leave those fields 0 like simple_statfs */
1830         return 0;
1831 }
1832
1833 /*
1834  * File creation. Allocate an inode, and we're done..
1835  */
1836 static int
1837 shmem_mknod(struct inode *dir, struct dentry *dentry, int mode, dev_t dev)
1838 {
1839         struct inode *inode;
1840         int error = -ENOSPC;
1841
1842         inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
1843         if (inode) {
1844                 error = security_inode_init_security(inode, dir, NULL, NULL,
1845                                                      NULL);
1846                 if (error) {
1847                         if (error != -EOPNOTSUPP) {
1848                                 iput(inode);
1849                                 return error;
1850                         }
1851                 }
1852 #ifdef CONFIG_TMPFS_POSIX_ACL
1853                 error = generic_acl_init(inode, dir);
1854                 if (error) {
1855                         iput(inode);
1856                         return error;
1857                 }
1858 #else
1859                 error = 0;
1860 #endif
1861                 dir->i_size += BOGO_DIRENT_SIZE;
1862                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1863                 d_instantiate(dentry, inode);
1864                 dget(dentry); /* Extra count - pin the dentry in core */
1865         }
1866         return error;
1867 }
1868
1869 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, int mode)
1870 {
1871         int error;
1872
1873         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
1874                 return error;
1875         inc_nlink(dir);
1876         return 0;
1877 }
1878
1879 static int shmem_create(struct inode *dir, struct dentry *dentry, int mode,
1880                 struct nameidata *nd)
1881 {
1882         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
1883 }
1884
1885 /*
1886  * Link a file..
1887  */
1888 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
1889 {
1890         struct inode *inode = old_dentry->d_inode;
1891         int ret;
1892
1893         /*
1894          * No ordinary (disk based) filesystem counts links as inodes;
1895          * but each new link needs a new dentry, pinning lowmem, and
1896          * tmpfs dentries cannot be pruned until they are unlinked.
1897          */
1898         ret = shmem_reserve_inode(inode->i_sb);
1899         if (ret)
1900                 goto out;
1901
1902         dir->i_size += BOGO_DIRENT_SIZE;
1903         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1904         inc_nlink(inode);
1905         atomic_inc(&inode->i_count);    /* New dentry reference */
1906         dget(dentry);           /* Extra pinning count for the created dentry */
1907         d_instantiate(dentry, inode);
1908 out:
1909         return ret;
1910 }
1911
1912 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
1913 {
1914         struct inode *inode = dentry->d_inode;
1915
1916         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
1917                 shmem_free_inode(inode->i_sb);
1918
1919         dir->i_size -= BOGO_DIRENT_SIZE;
1920         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1921         drop_nlink(inode);
1922         dput(dentry);   /* Undo the count from "create" - this does all the work */
1923         return 0;
1924 }
1925
1926 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
1927 {
1928         if (!simple_empty(dentry))
1929                 return -ENOTEMPTY;
1930
1931         drop_nlink(dentry->d_inode);
1932         drop_nlink(dir);
1933         return shmem_unlink(dir, dentry);
1934 }
1935
1936 /*
1937  * The VFS layer already does all the dentry stuff for rename,
1938  * we just have to decrement the usage count for the target if
1939  * it exists so that the VFS layer correctly free's it when it
1940  * gets overwritten.
1941  */
1942 static int shmem_rename(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
1943 {
1944         struct inode *inode = old_dentry->d_inode;
1945         int they_are_dirs = S_ISDIR(inode->i_mode);
1946
1947         if (!simple_empty(new_dentry))
1948                 return -ENOTEMPTY;
1949
1950         if (new_dentry->d_inode) {
1951                 (void) shmem_unlink(new_dir, new_dentry);
1952                 if (they_are_dirs)
1953                         drop_nlink(old_dir);
1954         } else if (they_are_dirs) {
1955                 drop_nlink(old_dir);
1956                 inc_nlink(new_dir);
1957         }
1958
1959         old_dir->i_size -= BOGO_DIRENT_SIZE;
1960         new_dir->i_size += BOGO_DIRENT_SIZE;
1961         old_dir->i_ctime = old_dir->i_mtime =
1962         new_dir->i_ctime = new_dir->i_mtime =
1963         inode->i_ctime = CURRENT_TIME;
1964         return 0;
1965 }
1966
1967 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
1968 {
1969         int error;
1970         int len;
1971         struct inode *inode;
1972         struct page *page = NULL;
1973         char *kaddr;
1974         struct shmem_inode_info *info;
1975
1976         len = strlen(symname) + 1;
1977         if (len > PAGE_CACHE_SIZE)
1978                 return -ENAMETOOLONG;
1979
1980         inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0, VM_NORESERVE);
1981         if (!inode)
1982                 return -ENOSPC;
1983
1984         error = security_inode_init_security(inode, dir, NULL, NULL,
1985                                              NULL);
1986         if (error) {
1987                 if (error != -EOPNOTSUPP) {
1988                         iput(inode);
1989                         return error;
1990                 }
1991                 error = 0;
1992         }
1993
1994         info = SHMEM_I(inode);
1995         inode->i_size = len-1;
1996         if (len <= (char *)inode - (char *)info) {
1997                 /* do it inline */
1998                 memcpy(info, symname, len);
1999                 inode->i_op = &shmem_symlink_inline_operations;
2000         } else {
2001                 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
2002                 if (error) {
2003                         iput(inode);
2004                         return error;
2005                 }
2006                 inode->i_mapping->a_ops = &shmem_aops;
2007                 inode->i_op = &shmem_symlink_inode_operations;
2008                 kaddr = kmap_atomic(page, KM_USER0);
2009                 memcpy(kaddr, symname, len);
2010                 kunmap_atomic(kaddr, KM_USER0);
2011                 set_page_dirty(page);
2012                 unlock_page(page);
2013                 page_cache_release(page);
2014         }
2015         dir->i_size += BOGO_DIRENT_SIZE;
2016         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2017         d_instantiate(dentry, inode);
2018         dget(dentry);
2019         return 0;
2020 }
2021
2022 static void *shmem_follow_link_inline(struct dentry *dentry, struct nameidata *nd)
2023 {
2024         nd_set_link(nd, (char *)SHMEM_I(dentry->d_inode));
2025         return NULL;
2026 }
2027
2028 static void *shmem_follow_link(struct dentry *dentry, struct nameidata *nd)
2029 {
2030         struct page *page = NULL;
2031         int res = shmem_getpage(dentry->d_inode, 0, &page, SGP_READ, NULL);
2032         nd_set_link(nd, res ? ERR_PTR(res) : kmap(page));
2033         if (page)
2034                 unlock_page(page);
2035         return page;
2036 }
2037
2038 static void shmem_put_link(struct dentry *dentry, struct nameidata *nd, void *cookie)
2039 {
2040         if (!IS_ERR(nd_get_link(nd))) {
2041                 struct page *page = cookie;
2042                 kunmap(page);
2043                 mark_page_accessed(page);
2044                 page_cache_release(page);
2045         }
2046 }
2047
2048 static const struct inode_operations shmem_symlink_inline_operations = {
2049         .readlink       = generic_readlink,
2050         .follow_link    = shmem_follow_link_inline,
2051 };
2052
2053 static const struct inode_operations shmem_symlink_inode_operations = {
2054         .readlink       = generic_readlink,
2055         .follow_link    = shmem_follow_link,
2056         .put_link       = shmem_put_link,
2057 };
2058
2059 #ifdef CONFIG_TMPFS_POSIX_ACL
2060 /*
2061  * Superblocks without xattr inode operations will get security.* xattr
2062  * support from the VFS "for free". As soon as we have any other xattrs
2063  * like ACLs, we also need to implement the security.* handlers at
2064  * filesystem level, though.
2065  */
2066
2067 static size_t shmem_xattr_security_list(struct dentry *dentry, char *list,
2068                                         size_t list_len, const char *name,
2069                                         size_t name_len, int handler_flags)
2070 {
2071         return security_inode_listsecurity(dentry->d_inode, list, list_len);
2072 }
2073
2074 static int shmem_xattr_security_get(struct dentry *dentry, const char *name,
2075                 void *buffer, size_t size, int handler_flags)
2076 {
2077         if (strcmp(name, "") == 0)
2078                 return -EINVAL;
2079         return xattr_getsecurity(dentry->d_inode, name, buffer, size);
2080 }
2081
2082 static int shmem_xattr_security_set(struct dentry *dentry, const char *name,
2083                 const void *value, size_t size, int flags, int handler_flags)
2084 {
2085         if (strcmp(name, "") == 0)
2086                 return -EINVAL;
2087         return security_inode_setsecurity(dentry->d_inode, name, value,
2088                                           size, flags);
2089 }
2090
2091 static const struct xattr_handler shmem_xattr_security_handler = {
2092         .prefix = XATTR_SECURITY_PREFIX,
2093         .list   = shmem_xattr_security_list,
2094         .get    = shmem_xattr_security_get,
2095         .set    = shmem_xattr_security_set,
2096 };
2097
2098 static const struct xattr_handler *shmem_xattr_handlers[] = {
2099         &generic_acl_access_handler,
2100         &generic_acl_default_handler,
2101         &shmem_xattr_security_handler,
2102         NULL
2103 };
2104 #endif
2105
2106 static struct dentry *shmem_get_parent(struct dentry *child)
2107 {
2108         return ERR_PTR(-ESTALE);
2109 }
2110
2111 static int shmem_match(struct inode *ino, void *vfh)
2112 {
2113         __u32 *fh = vfh;
2114         __u64 inum = fh[2];
2115         inum = (inum << 32) | fh[1];
2116         return ino->i_ino == inum && fh[0] == ino->i_generation;
2117 }
2118
2119 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
2120                 struct fid *fid, int fh_len, int fh_type)
2121 {
2122         struct inode *inode;
2123         struct dentry *dentry = NULL;
2124         u64 inum = fid->raw[2];
2125         inum = (inum << 32) | fid->raw[1];
2126
2127         if (fh_len < 3)
2128                 return NULL;
2129
2130         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
2131                         shmem_match, fid->raw);
2132         if (inode) {
2133                 dentry = d_find_alias(inode);
2134                 iput(inode);
2135         }
2136
2137         return dentry;
2138 }
2139
2140 static int shmem_encode_fh(struct dentry *dentry, __u32 *fh, int *len,
2141                                 int connectable)
2142 {
2143         struct inode *inode = dentry->d_inode;
2144
2145         if (*len < 3)
2146                 return 255;
2147
2148         if (hlist_unhashed(&inode->i_hash)) {
2149                 /* Unfortunately insert_inode_hash is not idempotent,
2150                  * so as we hash inodes here rather than at creation
2151                  * time, we need a lock to ensure we only try
2152                  * to do it once
2153                  */
2154                 static DEFINE_SPINLOCK(lock);
2155                 spin_lock(&lock);
2156                 if (hlist_unhashed(&inode->i_hash))
2157                         __insert_inode_hash(inode,
2158                                             inode->i_ino + inode->i_generation);
2159                 spin_unlock(&lock);
2160         }
2161
2162         fh[0] = inode->i_generation;
2163         fh[1] = inode->i_ino;
2164         fh[2] = ((__u64)inode->i_ino) >> 32;
2165
2166         *len = 3;
2167         return 1;
2168 }
2169
2170 static const struct export_operations shmem_export_ops = {
2171         .get_parent     = shmem_get_parent,
2172         .encode_fh      = shmem_encode_fh,
2173         .fh_to_dentry   = shmem_fh_to_dentry,
2174 };
2175
2176 static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
2177                                bool remount)
2178 {
2179         char *this_char, *value, *rest;
2180
2181         while (options != NULL) {
2182                 this_char = options;
2183                 for (;;) {
2184                         /*
2185                          * NUL-terminate this option: unfortunately,
2186                          * mount options form a comma-separated list,
2187                          * but mpol's nodelist may also contain commas.
2188                          */
2189                         options = strchr(options, ',');
2190                         if (options == NULL)
2191                                 break;
2192                         options++;
2193                         if (!isdigit(*options)) {
2194                                 options[-1] = '\0';
2195                                 break;
2196                         }
2197                 }
2198                 if (!*this_char)
2199                         continue;
2200                 if ((value = strchr(this_char,'=')) != NULL) {
2201                         *value++ = 0;
2202                 } else {
2203                         printk(KERN_ERR
2204                             "tmpfs: No value for mount option '%s'\n",
2205                             this_char);
2206                         return 1;
2207                 }
2208
2209                 if (!strcmp(this_char,"size")) {
2210                         unsigned long long size;
2211                         size = memparse(value,&rest);
2212                         if (*rest == '%') {
2213                                 size <<= PAGE_SHIFT;
2214                                 size *= totalram_pages;
2215                                 do_div(size, 100);
2216                                 rest++;
2217                         }
2218                         if (*rest)
2219                                 goto bad_val;
2220                         sbinfo->max_blocks =
2221                                 DIV_ROUND_UP(size, PAGE_CACHE_SIZE);
2222                 } else if (!strcmp(this_char,"nr_blocks")) {
2223                         sbinfo->max_blocks = memparse(value, &rest);
2224                         if (*rest)
2225                                 goto bad_val;
2226                 } else if (!strcmp(this_char,"nr_inodes")) {
2227                         sbinfo->max_inodes = memparse(value, &rest);
2228                         if (*rest)
2229                                 goto bad_val;
2230                 } else if (!strcmp(this_char,"mode")) {
2231                         if (remount)
2232                                 continue;
2233                         sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
2234                         if (*rest)
2235                                 goto bad_val;
2236                 } else if (!strcmp(this_char,"uid")) {
2237                         if (remount)
2238                                 continue;
2239                         sbinfo->uid = simple_strtoul(value, &rest, 0);
2240                         if (*rest)
2241                                 goto bad_val;
2242                 } else if (!strcmp(this_char,"gid")) {
2243                         if (remount)
2244                                 continue;
2245                         sbinfo->gid = simple_strtoul(value, &rest, 0);
2246                         if (*rest)
2247                                 goto bad_val;
2248                 } else if (!strcmp(this_char,"mpol")) {
2249                         if (mpol_parse_str(value, &sbinfo->mpol, 1))
2250                                 goto bad_val;
2251                 } else {
2252                         printk(KERN_ERR "tmpfs: Bad mount option %s\n",
2253                                this_char);
2254                         return 1;
2255                 }
2256         }
2257         return 0;
2258
2259 bad_val:
2260         printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
2261                value, this_char);
2262         return 1;
2263
2264 }
2265
2266 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2267 {
2268         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2269         struct shmem_sb_info config = *sbinfo;
2270         unsigned long inodes;
2271         int error = -EINVAL;
2272
2273         if (shmem_parse_options(data, &config, true))
2274                 return error;
2275
2276         spin_lock(&sbinfo->stat_lock);
2277         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2278         if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
2279                 goto out;
2280         if (config.max_inodes < inodes)
2281                 goto out;
2282         /*
2283          * Those tests also disallow limited->unlimited while any are in
2284          * use, so i_blocks will always be zero when max_blocks is zero;
2285          * but we must separately disallow unlimited->limited, because
2286          * in that case we have no record of how much is already in use.
2287          */
2288         if (config.max_blocks && !sbinfo->max_blocks)
2289                 goto out;
2290         if (config.max_inodes && !sbinfo->max_inodes)
2291                 goto out;
2292
2293         error = 0;
2294         sbinfo->max_blocks  = config.max_blocks;
2295         sbinfo->max_inodes  = config.max_inodes;
2296         sbinfo->free_inodes = config.max_inodes - inodes;
2297
2298         mpol_put(sbinfo->mpol);
2299         sbinfo->mpol        = config.mpol;      /* transfers initial ref */
2300 out:
2301         spin_unlock(&sbinfo->stat_lock);
2302         return error;
2303 }
2304
2305 static int shmem_show_options(struct seq_file *seq, struct vfsmount *vfs)
2306 {
2307         struct shmem_sb_info *sbinfo = SHMEM_SB(vfs->mnt_sb);
2308
2309         if (sbinfo->max_blocks != shmem_default_max_blocks())
2310                 seq_printf(seq, ",size=%luk",
2311                         sbinfo->max_blocks << (PAGE_CACHE_SHIFT - 10));
2312         if (sbinfo->max_inodes != shmem_default_max_inodes())
2313                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
2314         if (sbinfo->mode != (S_IRWXUGO | S_ISVTX))
2315                 seq_printf(seq, ",mode=%03o", sbinfo->mode);
2316         if (sbinfo->uid != 0)
2317                 seq_printf(seq, ",uid=%u", sbinfo->uid);
2318         if (sbinfo->gid != 0)
2319                 seq_printf(seq, ",gid=%u", sbinfo->gid);
2320         shmem_show_mpol(seq, sbinfo->mpol);
2321         return 0;
2322 }
2323 #endif /* CONFIG_TMPFS */
2324
2325 static void shmem_put_super(struct super_block *sb)
2326 {
2327         kfree(sb->s_fs_info);
2328         sb->s_fs_info = NULL;
2329 }
2330
2331 int shmem_fill_super(struct super_block *sb, void *data, int silent)
2332 {
2333         struct inode *inode;
2334         struct dentry *root;
2335         struct shmem_sb_info *sbinfo;
2336         int err = -ENOMEM;
2337
2338         /* Round up to L1_CACHE_BYTES to resist false sharing */
2339         sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
2340                                 L1_CACHE_BYTES), GFP_KERNEL);
2341         if (!sbinfo)
2342                 return -ENOMEM;
2343
2344         sbinfo->mode = S_IRWXUGO | S_ISVTX;
2345         sbinfo->uid = current_fsuid();
2346         sbinfo->gid = current_fsgid();
2347         sb->s_fs_info = sbinfo;
2348
2349 #ifdef CONFIG_TMPFS
2350         /*
2351          * Per default we only allow half of the physical ram per
2352          * tmpfs instance, limiting inodes to one per page of lowmem;
2353          * but the internal instance is left unlimited.
2354          */
2355         if (!(sb->s_flags & MS_NOUSER)) {
2356                 sbinfo->max_blocks = shmem_default_max_blocks();
2357                 sbinfo->max_inodes = shmem_default_max_inodes();
2358                 if (shmem_parse_options(data, sbinfo, false)) {
2359                         err = -EINVAL;
2360                         goto failed;
2361                 }
2362         }
2363         sb->s_export_op = &shmem_export_ops;
2364 #else
2365         sb->s_flags |= MS_NOUSER;
2366 #endif
2367
2368         spin_lock_init(&sbinfo->stat_lock);
2369         percpu_counter_init(&sbinfo->used_blocks, 0);
2370         sbinfo->free_inodes = sbinfo->max_inodes;
2371
2372         sb->s_maxbytes = SHMEM_MAX_BYTES;
2373         sb->s_blocksize = PAGE_CACHE_SIZE;
2374         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
2375         sb->s_magic = TMPFS_MAGIC;
2376         sb->s_op = &shmem_ops;
2377         sb->s_time_gran = 1;
2378 #ifdef CONFIG_TMPFS_POSIX_ACL
2379         sb->s_xattr = shmem_xattr_handlers;
2380         sb->s_flags |= MS_POSIXACL;
2381 #endif
2382
2383         inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
2384         if (!inode)
2385                 goto failed;
2386         inode->i_uid = sbinfo->uid;
2387         inode->i_gid = sbinfo->gid;
2388         root = d_alloc_root(inode);
2389         if (!root)
2390                 goto failed_iput;
2391         sb->s_root = root;
2392         return 0;
2393
2394 failed_iput:
2395         iput(inode);
2396 failed:
2397         shmem_put_super(sb);
2398         return err;
2399 }
2400
2401 static struct kmem_cache *shmem_inode_cachep;
2402
2403 static struct inode *shmem_alloc_inode(struct super_block *sb)
2404 {
2405         struct shmem_inode_info *p;
2406         p = (struct shmem_inode_info *)kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
2407         if (!p)
2408                 return NULL;
2409         return &p->vfs_inode;
2410 }
2411
2412 static void shmem_destroy_inode(struct inode *inode)
2413 {
2414         if ((inode->i_mode & S_IFMT) == S_IFREG) {
2415                 /* only struct inode is valid if it's an inline symlink */
2416                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
2417         }
2418         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
2419 }
2420
2421 static void init_once(void *foo)
2422 {
2423         struct shmem_inode_info *p = (struct shmem_inode_info *) foo;
2424
2425         inode_init_once(&p->vfs_inode);
2426 }
2427
2428 static int init_inodecache(void)
2429 {
2430         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
2431                                 sizeof(struct shmem_inode_info),
2432                                 0, SLAB_PANIC, init_once);
2433         return 0;
2434 }
2435
2436 static void destroy_inodecache(void)
2437 {
2438         kmem_cache_destroy(shmem_inode_cachep);
2439 }
2440
2441 static const struct address_space_operations shmem_aops = {
2442         .writepage      = shmem_writepage,
2443         .set_page_dirty = __set_page_dirty_no_writeback,
2444 #ifdef CONFIG_TMPFS
2445         .readpage       = shmem_readpage,
2446         .write_begin    = shmem_write_begin,
2447         .write_end      = shmem_write_end,
2448 #endif
2449         .migratepage    = migrate_page,
2450         .error_remove_page = generic_error_remove_page,
2451 };
2452
2453 static const struct file_operations shmem_file_operations = {
2454         .mmap           = shmem_mmap,
2455 #ifdef CONFIG_TMPFS
2456         .llseek         = generic_file_llseek,
2457         .read           = do_sync_read,
2458         .write          = do_sync_write,
2459         .aio_read       = shmem_file_aio_read,
2460         .aio_write      = generic_file_aio_write,
2461         .fsync          = noop_fsync,
2462         .splice_read    = generic_file_splice_read,
2463         .splice_write   = generic_file_splice_write,
2464 #endif
2465 };
2466
2467 static const struct inode_operations shmem_inode_operations = {
2468         .setattr        = shmem_notify_change,
2469         .truncate_range = shmem_truncate_range,
2470 #ifdef CONFIG_TMPFS_POSIX_ACL
2471         .setxattr       = generic_setxattr,
2472         .getxattr       = generic_getxattr,
2473         .listxattr      = generic_listxattr,
2474         .removexattr    = generic_removexattr,
2475         .check_acl      = generic_check_acl,
2476 #endif
2477
2478 };
2479
2480 static const struct inode_operations shmem_dir_inode_operations = {
2481 #ifdef CONFIG_TMPFS
2482         .create         = shmem_create,
2483         .lookup         = simple_lookup,
2484         .link           = shmem_link,
2485         .unlink         = shmem_unlink,
2486         .symlink        = shmem_symlink,
2487         .mkdir          = shmem_mkdir,
2488         .rmdir          = shmem_rmdir,
2489         .mknod          = shmem_mknod,
2490         .rename         = shmem_rename,
2491 #endif
2492 #ifdef CONFIG_TMPFS_POSIX_ACL
2493         .setattr        = shmem_notify_change,
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 static const struct inode_operations shmem_special_inode_operations = {
2503 #ifdef CONFIG_TMPFS_POSIX_ACL
2504         .setattr        = shmem_notify_change,
2505         .setxattr       = generic_setxattr,
2506         .getxattr       = generic_getxattr,
2507         .listxattr      = generic_listxattr,
2508         .removexattr    = generic_removexattr,
2509         .check_acl      = generic_check_acl,
2510 #endif
2511 };
2512
2513 static const struct super_operations shmem_ops = {
2514         .alloc_inode    = shmem_alloc_inode,
2515         .destroy_inode  = shmem_destroy_inode,
2516 #ifdef CONFIG_TMPFS
2517         .statfs         = shmem_statfs,
2518         .remount_fs     = shmem_remount_fs,
2519         .show_options   = shmem_show_options,
2520 #endif
2521         .delete_inode   = shmem_delete_inode,
2522         .drop_inode     = generic_delete_inode,
2523         .put_super      = shmem_put_super,
2524 };
2525
2526 static const struct vm_operations_struct shmem_vm_ops = {
2527         .fault          = shmem_fault,
2528 #ifdef CONFIG_NUMA
2529         .set_policy     = shmem_set_policy,
2530         .get_policy     = shmem_get_policy,
2531 #endif
2532 };
2533
2534
2535 static int shmem_get_sb(struct file_system_type *fs_type,
2536         int flags, const char *dev_name, void *data, struct vfsmount *mnt)
2537 {
2538         return get_sb_nodev(fs_type, flags, data, shmem_fill_super, mnt);
2539 }
2540
2541 static struct file_system_type tmpfs_fs_type = {
2542         .owner          = THIS_MODULE,
2543         .name           = "tmpfs",
2544         .get_sb         = shmem_get_sb,
2545         .kill_sb        = kill_litter_super,
2546 };
2547
2548 int __init init_tmpfs(void)
2549 {
2550         int error;
2551
2552         error = bdi_init(&shmem_backing_dev_info);
2553         if (error)
2554                 goto out4;
2555
2556         error = init_inodecache();
2557         if (error)
2558                 goto out3;
2559
2560         error = register_filesystem(&tmpfs_fs_type);
2561         if (error) {
2562                 printk(KERN_ERR "Could not register tmpfs\n");
2563                 goto out2;
2564         }
2565
2566         shm_mnt = vfs_kern_mount(&tmpfs_fs_type, MS_NOUSER,
2567                                 tmpfs_fs_type.name, NULL);
2568         if (IS_ERR(shm_mnt)) {
2569                 error = PTR_ERR(shm_mnt);
2570                 printk(KERN_ERR "Could not kern_mount tmpfs\n");
2571                 goto out1;
2572         }
2573         return 0;
2574
2575 out1:
2576         unregister_filesystem(&tmpfs_fs_type);
2577 out2:
2578         destroy_inodecache();
2579 out3:
2580         bdi_destroy(&shmem_backing_dev_info);
2581 out4:
2582         shm_mnt = ERR_PTR(error);
2583         return error;
2584 }
2585
2586 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2587 /**
2588  * mem_cgroup_get_shmem_target - find a page or entry assigned to the shmem file
2589  * @inode: the inode to be searched
2590  * @pgoff: the offset to be searched
2591  * @pagep: the pointer for the found page to be stored
2592  * @ent: the pointer for the found swap entry to be stored
2593  *
2594  * If a page is found, refcount of it is incremented. Callers should handle
2595  * these refcount.
2596  */
2597 void mem_cgroup_get_shmem_target(struct inode *inode, pgoff_t pgoff,
2598                                         struct page **pagep, swp_entry_t *ent)
2599 {
2600         swp_entry_t entry = { .val = 0 }, *ptr;
2601         struct page *page = NULL;
2602         struct shmem_inode_info *info = SHMEM_I(inode);
2603
2604         if ((pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
2605                 goto out;
2606
2607         spin_lock(&info->lock);
2608         ptr = shmem_swp_entry(info, pgoff, NULL);
2609 #ifdef CONFIG_SWAP
2610         if (ptr && ptr->val) {
2611                 entry.val = ptr->val;
2612                 page = find_get_page(&swapper_space, entry.val);
2613         } else
2614 #endif
2615                 page = find_get_page(inode->i_mapping, pgoff);
2616         if (ptr)
2617                 shmem_swp_unmap(ptr);
2618         spin_unlock(&info->lock);
2619 out:
2620         *pagep = page;
2621         *ent = entry;
2622 }
2623 #endif
2624
2625 #else /* !CONFIG_SHMEM */
2626
2627 /*
2628  * tiny-shmem: simple shmemfs and tmpfs using ramfs code
2629  *
2630  * This is intended for small system where the benefits of the full
2631  * shmem code (swap-backed and resource-limited) are outweighed by
2632  * their complexity. On systems without swap this code should be
2633  * effectively equivalent, but much lighter weight.
2634  */
2635
2636 #include <linux/ramfs.h>
2637
2638 static struct file_system_type tmpfs_fs_type = {
2639         .name           = "tmpfs",
2640         .get_sb         = ramfs_get_sb,
2641         .kill_sb        = kill_litter_super,
2642 };
2643
2644 int __init init_tmpfs(void)
2645 {
2646         BUG_ON(register_filesystem(&tmpfs_fs_type) != 0);
2647
2648         shm_mnt = kern_mount(&tmpfs_fs_type);
2649         BUG_ON(IS_ERR(shm_mnt));
2650
2651         return 0;
2652 }
2653
2654 int shmem_unuse(swp_entry_t entry, struct page *page)
2655 {
2656         return 0;
2657 }
2658
2659 int shmem_lock(struct file *file, int lock, struct user_struct *user)
2660 {
2661         return 0;
2662 }
2663
2664 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2665 /**
2666  * mem_cgroup_get_shmem_target - find a page or entry assigned to the shmem file
2667  * @inode: the inode to be searched
2668  * @pgoff: the offset to be searched
2669  * @pagep: the pointer for the found page to be stored
2670  * @ent: the pointer for the found swap entry to be stored
2671  *
2672  * If a page is found, refcount of it is incremented. Callers should handle
2673  * these refcount.
2674  */
2675 void mem_cgroup_get_shmem_target(struct inode *inode, pgoff_t pgoff,
2676                                         struct page **pagep, swp_entry_t *ent)
2677 {
2678         struct page *page = NULL;
2679
2680         if ((pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
2681                 goto out;
2682         page = find_get_page(inode->i_mapping, pgoff);
2683 out:
2684         *pagep = page;
2685         *ent = (swp_entry_t){ .val = 0 };
2686 }
2687 #endif
2688
2689 #define shmem_vm_ops                            generic_file_vm_ops
2690 #define shmem_file_operations                   ramfs_file_operations
2691 #define shmem_get_inode(sb, dir, mode, dev, flags)      ramfs_get_inode(sb, dir, mode, dev)
2692 #define shmem_acct_size(flags, size)            0
2693 #define shmem_unacct_size(flags, size)          do {} while (0)
2694 #define SHMEM_MAX_BYTES                         MAX_LFS_FILESIZE
2695
2696 #endif /* CONFIG_SHMEM */
2697
2698 /* common code */
2699
2700 /**
2701  * shmem_file_setup - get an unlinked file living in tmpfs
2702  * @name: name for dentry (to be seen in /proc/<pid>/maps
2703  * @size: size to be set for the file
2704  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
2705  */
2706 struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
2707 {
2708         int error;
2709         struct file *file;
2710         struct inode *inode;
2711         struct path path;
2712         struct dentry *root;
2713         struct qstr this;
2714
2715         if (IS_ERR(shm_mnt))
2716                 return (void *)shm_mnt;
2717
2718         if (size < 0 || size > SHMEM_MAX_BYTES)
2719                 return ERR_PTR(-EINVAL);
2720
2721         if (shmem_acct_size(flags, size))
2722                 return ERR_PTR(-ENOMEM);
2723
2724         error = -ENOMEM;
2725         this.name = name;
2726         this.len = strlen(name);
2727         this.hash = 0; /* will go */
2728         root = shm_mnt->mnt_root;
2729         path.dentry = d_alloc(root, &this);
2730         if (!path.dentry)
2731                 goto put_memory;
2732         path.mnt = mntget(shm_mnt);
2733
2734         error = -ENOSPC;
2735         inode = shmem_get_inode(root->d_sb, NULL, S_IFREG | S_IRWXUGO, 0, flags);
2736         if (!inode)
2737                 goto put_dentry;
2738
2739         d_instantiate(path.dentry, inode);
2740         inode->i_size = size;
2741         inode->i_nlink = 0;     /* It is unlinked */
2742 #ifndef CONFIG_MMU
2743         error = ramfs_nommu_expand_for_mapping(inode, size);
2744         if (error)
2745                 goto put_dentry;
2746 #endif
2747
2748         error = -ENFILE;
2749         file = alloc_file(&path, FMODE_WRITE | FMODE_READ,
2750                   &shmem_file_operations);
2751         if (!file)
2752                 goto put_dentry;
2753
2754         return file;
2755
2756 put_dentry:
2757         path_put(&path);
2758 put_memory:
2759         shmem_unacct_size(flags, size);
2760         return ERR_PTR(error);
2761 }
2762 EXPORT_SYMBOL_GPL(shmem_file_setup);
2763
2764 /**
2765  * shmem_zero_setup - setup a shared anonymous mapping
2766  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
2767  */
2768 int shmem_zero_setup(struct vm_area_struct *vma)
2769 {
2770         struct file *file;
2771         loff_t size = vma->vm_end - vma->vm_start;
2772
2773         file = shmem_file_setup("dev/zero", size, vma->vm_flags);
2774         if (IS_ERR(file))
2775                 return PTR_ERR(file);
2776
2777         if (vma->vm_file)
2778                 fput(vma->vm_file);
2779         vma->vm_file = file;
2780         vma->vm_ops = &shmem_vm_ops;
2781         return 0;
2782 }