memcg: remove refcnt from page_cgroup
[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  * This file is released under the GPL.
18  */
19
20 /*
21  * This virtual memory filesystem is heavily based on the ramfs. It
22  * extends ramfs by the ability to use swap and honor resource limits
23  * which makes it a completely usable filesystem.
24  */
25
26 #include <linux/module.h>
27 #include <linux/init.h>
28 #include <linux/fs.h>
29 #include <linux/xattr.h>
30 #include <linux/exportfs.h>
31 #include <linux/generic_acl.h>
32 #include <linux/mm.h>
33 #include <linux/mman.h>
34 #include <linux/file.h>
35 #include <linux/swap.h>
36 #include <linux/pagemap.h>
37 #include <linux/string.h>
38 #include <linux/slab.h>
39 #include <linux/backing-dev.h>
40 #include <linux/shmem_fs.h>
41 #include <linux/mount.h>
42 #include <linux/writeback.h>
43 #include <linux/vfs.h>
44 #include <linux/blkdev.h>
45 #include <linux/security.h>
46 #include <linux/swapops.h>
47 #include <linux/mempolicy.h>
48 #include <linux/namei.h>
49 #include <linux/ctype.h>
50 #include <linux/migrate.h>
51 #include <linux/highmem.h>
52 #include <linux/seq_file.h>
53
54 #include <asm/uaccess.h>
55 #include <asm/div64.h>
56 #include <asm/pgtable.h>
57
58 /* This magic number is used in glibc for posix shared memory */
59 #define TMPFS_MAGIC     0x01021994
60
61 #define ENTRIES_PER_PAGE (PAGE_CACHE_SIZE/sizeof(unsigned long))
62 #define ENTRIES_PER_PAGEPAGE (ENTRIES_PER_PAGE*ENTRIES_PER_PAGE)
63 #define BLOCKS_PER_PAGE  (PAGE_CACHE_SIZE/512)
64
65 #define SHMEM_MAX_INDEX  (SHMEM_NR_DIRECT + (ENTRIES_PER_PAGEPAGE/2) * (ENTRIES_PER_PAGE+1))
66 #define SHMEM_MAX_BYTES  ((unsigned long long)SHMEM_MAX_INDEX << PAGE_CACHE_SHIFT)
67
68 #define VM_ACCT(size)    (PAGE_CACHE_ALIGN(size) >> PAGE_SHIFT)
69
70 /* info->flags needs VM_flags to handle pagein/truncate races efficiently */
71 #define SHMEM_PAGEIN     VM_READ
72 #define SHMEM_TRUNCATE   VM_WRITE
73
74 /* Definition to limit shmem_truncate's steps between cond_rescheds */
75 #define LATENCY_LIMIT    64
76
77 /* Pretend that each entry is of this size in directory's i_size */
78 #define BOGO_DIRENT_SIZE 20
79
80 /* Flag allocation requirements to shmem_getpage and shmem_swp_alloc */
81 enum sgp_type {
82         SGP_READ,       /* don't exceed i_size, don't allocate page */
83         SGP_CACHE,      /* don't exceed i_size, may allocate page */
84         SGP_DIRTY,      /* like SGP_CACHE, but set new page dirty */
85         SGP_WRITE,      /* may exceed i_size, may allocate page */
86 };
87
88 #ifdef CONFIG_TMPFS
89 static unsigned long shmem_default_max_blocks(void)
90 {
91         return totalram_pages / 2;
92 }
93
94 static unsigned long shmem_default_max_inodes(void)
95 {
96         return min(totalram_pages - totalhigh_pages, totalram_pages / 2);
97 }
98 #endif
99
100 static int shmem_getpage(struct inode *inode, unsigned long idx,
101                          struct page **pagep, enum sgp_type sgp, int *type);
102
103 static inline struct page *shmem_dir_alloc(gfp_t gfp_mask)
104 {
105         /*
106          * The above definition of ENTRIES_PER_PAGE, and the use of
107          * BLOCKS_PER_PAGE on indirect pages, assume PAGE_CACHE_SIZE:
108          * might be reconsidered if it ever diverges from PAGE_SIZE.
109          *
110          * Mobility flags are masked out as swap vectors cannot move
111          */
112         return alloc_pages((gfp_mask & ~GFP_MOVABLE_MASK) | __GFP_ZERO,
113                                 PAGE_CACHE_SHIFT-PAGE_SHIFT);
114 }
115
116 static inline void shmem_dir_free(struct page *page)
117 {
118         __free_pages(page, PAGE_CACHE_SHIFT-PAGE_SHIFT);
119 }
120
121 static struct page **shmem_dir_map(struct page *page)
122 {
123         return (struct page **)kmap_atomic(page, KM_USER0);
124 }
125
126 static inline void shmem_dir_unmap(struct page **dir)
127 {
128         kunmap_atomic(dir, KM_USER0);
129 }
130
131 static swp_entry_t *shmem_swp_map(struct page *page)
132 {
133         return (swp_entry_t *)kmap_atomic(page, KM_USER1);
134 }
135
136 static inline void shmem_swp_balance_unmap(void)
137 {
138         /*
139          * When passing a pointer to an i_direct entry, to code which
140          * also handles indirect entries and so will shmem_swp_unmap,
141          * we must arrange for the preempt count to remain in balance.
142          * What kmap_atomic of a lowmem page does depends on config
143          * and architecture, so pretend to kmap_atomic some lowmem page.
144          */
145         (void) kmap_atomic(ZERO_PAGE(0), KM_USER1);
146 }
147
148 static inline void shmem_swp_unmap(swp_entry_t *entry)
149 {
150         kunmap_atomic(entry, KM_USER1);
151 }
152
153 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
154 {
155         return sb->s_fs_info;
156 }
157
158 /*
159  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
160  * for shared memory and for shared anonymous (/dev/zero) mappings
161  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
162  * consistent with the pre-accounting of private mappings ...
163  */
164 static inline int shmem_acct_size(unsigned long flags, loff_t size)
165 {
166         return (flags & VM_ACCOUNT)?
167                 security_vm_enough_memory(VM_ACCT(size)): 0;
168 }
169
170 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
171 {
172         if (flags & VM_ACCOUNT)
173                 vm_unacct_memory(VM_ACCT(size));
174 }
175
176 /*
177  * ... whereas tmpfs objects are accounted incrementally as
178  * pages are allocated, in order to allow huge sparse files.
179  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
180  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
181  */
182 static inline int shmem_acct_block(unsigned long flags)
183 {
184         return (flags & VM_ACCOUNT)?
185                 0: security_vm_enough_memory(VM_ACCT(PAGE_CACHE_SIZE));
186 }
187
188 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
189 {
190         if (!(flags & VM_ACCOUNT))
191                 vm_unacct_memory(pages * VM_ACCT(PAGE_CACHE_SIZE));
192 }
193
194 static const struct super_operations shmem_ops;
195 static const struct address_space_operations shmem_aops;
196 static const struct file_operations shmem_file_operations;
197 static const struct inode_operations shmem_inode_operations;
198 static const struct inode_operations shmem_dir_inode_operations;
199 static const struct inode_operations shmem_special_inode_operations;
200 static struct vm_operations_struct shmem_vm_ops;
201
202 static struct backing_dev_info shmem_backing_dev_info  __read_mostly = {
203         .ra_pages       = 0,    /* No readahead */
204         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK,
205         .unplug_io_fn   = default_unplug_io_fn,
206 };
207
208 static LIST_HEAD(shmem_swaplist);
209 static DEFINE_MUTEX(shmem_swaplist_mutex);
210
211 static void shmem_free_blocks(struct inode *inode, long pages)
212 {
213         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
214         if (sbinfo->max_blocks) {
215                 spin_lock(&sbinfo->stat_lock);
216                 sbinfo->free_blocks += pages;
217                 inode->i_blocks -= pages*BLOCKS_PER_PAGE;
218                 spin_unlock(&sbinfo->stat_lock);
219         }
220 }
221
222 static int shmem_reserve_inode(struct super_block *sb)
223 {
224         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
225         if (sbinfo->max_inodes) {
226                 spin_lock(&sbinfo->stat_lock);
227                 if (!sbinfo->free_inodes) {
228                         spin_unlock(&sbinfo->stat_lock);
229                         return -ENOSPC;
230                 }
231                 sbinfo->free_inodes--;
232                 spin_unlock(&sbinfo->stat_lock);
233         }
234         return 0;
235 }
236
237 static void shmem_free_inode(struct super_block *sb)
238 {
239         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
240         if (sbinfo->max_inodes) {
241                 spin_lock(&sbinfo->stat_lock);
242                 sbinfo->free_inodes++;
243                 spin_unlock(&sbinfo->stat_lock);
244         }
245 }
246
247 /**
248  * shmem_recalc_inode - recalculate the size of an inode
249  * @inode: inode to recalc
250  *
251  * We have to calculate the free blocks since the mm can drop
252  * undirtied hole pages behind our back.
253  *
254  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
255  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
256  *
257  * It has to be called with the spinlock held.
258  */
259 static void shmem_recalc_inode(struct inode *inode)
260 {
261         struct shmem_inode_info *info = SHMEM_I(inode);
262         long freed;
263
264         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
265         if (freed > 0) {
266                 info->alloced -= freed;
267                 shmem_unacct_blocks(info->flags, freed);
268                 shmem_free_blocks(inode, freed);
269         }
270 }
271
272 /**
273  * shmem_swp_entry - find the swap vector position in the info structure
274  * @info:  info structure for the inode
275  * @index: index of the page to find
276  * @page:  optional page to add to the structure. Has to be preset to
277  *         all zeros
278  *
279  * If there is no space allocated yet it will return NULL when
280  * page is NULL, else it will use the page for the needed block,
281  * setting it to NULL on return to indicate that it has been used.
282  *
283  * The swap vector is organized the following way:
284  *
285  * There are SHMEM_NR_DIRECT entries directly stored in the
286  * shmem_inode_info structure. So small files do not need an addional
287  * allocation.
288  *
289  * For pages with index > SHMEM_NR_DIRECT there is the pointer
290  * i_indirect which points to a page which holds in the first half
291  * doubly indirect blocks, in the second half triple indirect blocks:
292  *
293  * For an artificial ENTRIES_PER_PAGE = 4 this would lead to the
294  * following layout (for SHMEM_NR_DIRECT == 16):
295  *
296  * i_indirect -> dir --> 16-19
297  *            |      +-> 20-23
298  *            |
299  *            +-->dir2 --> 24-27
300  *            |        +-> 28-31
301  *            |        +-> 32-35
302  *            |        +-> 36-39
303  *            |
304  *            +-->dir3 --> 40-43
305  *                     +-> 44-47
306  *                     +-> 48-51
307  *                     +-> 52-55
308  */
309 static swp_entry_t *shmem_swp_entry(struct shmem_inode_info *info, unsigned long index, struct page **page)
310 {
311         unsigned long offset;
312         struct page **dir;
313         struct page *subdir;
314
315         if (index < SHMEM_NR_DIRECT) {
316                 shmem_swp_balance_unmap();
317                 return info->i_direct+index;
318         }
319         if (!info->i_indirect) {
320                 if (page) {
321                         info->i_indirect = *page;
322                         *page = NULL;
323                 }
324                 return NULL;                    /* need another page */
325         }
326
327         index -= SHMEM_NR_DIRECT;
328         offset = index % ENTRIES_PER_PAGE;
329         index /= ENTRIES_PER_PAGE;
330         dir = shmem_dir_map(info->i_indirect);
331
332         if (index >= ENTRIES_PER_PAGE/2) {
333                 index -= ENTRIES_PER_PAGE/2;
334                 dir += ENTRIES_PER_PAGE/2 + index/ENTRIES_PER_PAGE;
335                 index %= ENTRIES_PER_PAGE;
336                 subdir = *dir;
337                 if (!subdir) {
338                         if (page) {
339                                 *dir = *page;
340                                 *page = NULL;
341                         }
342                         shmem_dir_unmap(dir);
343                         return NULL;            /* need another page */
344                 }
345                 shmem_dir_unmap(dir);
346                 dir = shmem_dir_map(subdir);
347         }
348
349         dir += index;
350         subdir = *dir;
351         if (!subdir) {
352                 if (!page || !(subdir = *page)) {
353                         shmem_dir_unmap(dir);
354                         return NULL;            /* need a page */
355                 }
356                 *dir = subdir;
357                 *page = NULL;
358         }
359         shmem_dir_unmap(dir);
360         return shmem_swp_map(subdir) + offset;
361 }
362
363 static void shmem_swp_set(struct shmem_inode_info *info, swp_entry_t *entry, unsigned long value)
364 {
365         long incdec = value? 1: -1;
366
367         entry->val = value;
368         info->swapped += incdec;
369         if ((unsigned long)(entry - info->i_direct) >= SHMEM_NR_DIRECT) {
370                 struct page *page = kmap_atomic_to_page(entry);
371                 set_page_private(page, page_private(page) + incdec);
372         }
373 }
374
375 /**
376  * shmem_swp_alloc - get the position of the swap entry for the page.
377  * @info:       info structure for the inode
378  * @index:      index of the page to find
379  * @sgp:        check and recheck i_size? skip allocation?
380  *
381  * If the entry does not exist, allocate it.
382  */
383 static swp_entry_t *shmem_swp_alloc(struct shmem_inode_info *info, unsigned long index, enum sgp_type sgp)
384 {
385         struct inode *inode = &info->vfs_inode;
386         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
387         struct page *page = NULL;
388         swp_entry_t *entry;
389
390         if (sgp != SGP_WRITE &&
391             ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode))
392                 return ERR_PTR(-EINVAL);
393
394         while (!(entry = shmem_swp_entry(info, index, &page))) {
395                 if (sgp == SGP_READ)
396                         return shmem_swp_map(ZERO_PAGE(0));
397                 /*
398                  * Test free_blocks against 1 not 0, since we have 1 data
399                  * page (and perhaps indirect index pages) yet to allocate:
400                  * a waste to allocate index if we cannot allocate data.
401                  */
402                 if (sbinfo->max_blocks) {
403                         spin_lock(&sbinfo->stat_lock);
404                         if (sbinfo->free_blocks <= 1) {
405                                 spin_unlock(&sbinfo->stat_lock);
406                                 return ERR_PTR(-ENOSPC);
407                         }
408                         sbinfo->free_blocks--;
409                         inode->i_blocks += BLOCKS_PER_PAGE;
410                         spin_unlock(&sbinfo->stat_lock);
411                 }
412
413                 spin_unlock(&info->lock);
414                 page = shmem_dir_alloc(mapping_gfp_mask(inode->i_mapping));
415                 if (page)
416                         set_page_private(page, 0);
417                 spin_lock(&info->lock);
418
419                 if (!page) {
420                         shmem_free_blocks(inode, 1);
421                         return ERR_PTR(-ENOMEM);
422                 }
423                 if (sgp != SGP_WRITE &&
424                     ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
425                         entry = ERR_PTR(-EINVAL);
426                         break;
427                 }
428                 if (info->next_index <= index)
429                         info->next_index = index + 1;
430         }
431         if (page) {
432                 /* another task gave its page, or truncated the file */
433                 shmem_free_blocks(inode, 1);
434                 shmem_dir_free(page);
435         }
436         if (info->next_index <= index && !IS_ERR(entry))
437                 info->next_index = index + 1;
438         return entry;
439 }
440
441 /**
442  * shmem_free_swp - free some swap entries in a directory
443  * @dir:        pointer to the directory
444  * @edir:       pointer after last entry of the directory
445  * @punch_lock: pointer to spinlock when needed for the holepunch case
446  */
447 static int shmem_free_swp(swp_entry_t *dir, swp_entry_t *edir,
448                                                 spinlock_t *punch_lock)
449 {
450         spinlock_t *punch_unlock = NULL;
451         swp_entry_t *ptr;
452         int freed = 0;
453
454         for (ptr = dir; ptr < edir; ptr++) {
455                 if (ptr->val) {
456                         if (unlikely(punch_lock)) {
457                                 punch_unlock = punch_lock;
458                                 punch_lock = NULL;
459                                 spin_lock(punch_unlock);
460                                 if (!ptr->val)
461                                         continue;
462                         }
463                         free_swap_and_cache(*ptr);
464                         *ptr = (swp_entry_t){0};
465                         freed++;
466                 }
467         }
468         if (punch_unlock)
469                 spin_unlock(punch_unlock);
470         return freed;
471 }
472
473 static int shmem_map_and_free_swp(struct page *subdir, int offset,
474                 int limit, struct page ***dir, spinlock_t *punch_lock)
475 {
476         swp_entry_t *ptr;
477         int freed = 0;
478
479         ptr = shmem_swp_map(subdir);
480         for (; offset < limit; offset += LATENCY_LIMIT) {
481                 int size = limit - offset;
482                 if (size > LATENCY_LIMIT)
483                         size = LATENCY_LIMIT;
484                 freed += shmem_free_swp(ptr+offset, ptr+offset+size,
485                                                         punch_lock);
486                 if (need_resched()) {
487                         shmem_swp_unmap(ptr);
488                         if (*dir) {
489                                 shmem_dir_unmap(*dir);
490                                 *dir = NULL;
491                         }
492                         cond_resched();
493                         ptr = shmem_swp_map(subdir);
494                 }
495         }
496         shmem_swp_unmap(ptr);
497         return freed;
498 }
499
500 static void shmem_free_pages(struct list_head *next)
501 {
502         struct page *page;
503         int freed = 0;
504
505         do {
506                 page = container_of(next, struct page, lru);
507                 next = next->next;
508                 shmem_dir_free(page);
509                 freed++;
510                 if (freed >= LATENCY_LIMIT) {
511                         cond_resched();
512                         freed = 0;
513                 }
514         } while (next);
515 }
516
517 static void shmem_truncate_range(struct inode *inode, loff_t start, loff_t end)
518 {
519         struct shmem_inode_info *info = SHMEM_I(inode);
520         unsigned long idx;
521         unsigned long size;
522         unsigned long limit;
523         unsigned long stage;
524         unsigned long diroff;
525         struct page **dir;
526         struct page *topdir;
527         struct page *middir;
528         struct page *subdir;
529         swp_entry_t *ptr;
530         LIST_HEAD(pages_to_free);
531         long nr_pages_to_free = 0;
532         long nr_swaps_freed = 0;
533         int offset;
534         int freed;
535         int punch_hole;
536         spinlock_t *needs_lock;
537         spinlock_t *punch_lock;
538         unsigned long upper_limit;
539
540         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
541         idx = (start + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
542         if (idx >= info->next_index)
543                 return;
544
545         spin_lock(&info->lock);
546         info->flags |= SHMEM_TRUNCATE;
547         if (likely(end == (loff_t) -1)) {
548                 limit = info->next_index;
549                 upper_limit = SHMEM_MAX_INDEX;
550                 info->next_index = idx;
551                 needs_lock = NULL;
552                 punch_hole = 0;
553         } else {
554                 if (end + 1 >= inode->i_size) { /* we may free a little more */
555                         limit = (inode->i_size + PAGE_CACHE_SIZE - 1) >>
556                                                         PAGE_CACHE_SHIFT;
557                         upper_limit = SHMEM_MAX_INDEX;
558                 } else {
559                         limit = (end + 1) >> PAGE_CACHE_SHIFT;
560                         upper_limit = limit;
561                 }
562                 needs_lock = &info->lock;
563                 punch_hole = 1;
564         }
565
566         topdir = info->i_indirect;
567         if (topdir && idx <= SHMEM_NR_DIRECT && !punch_hole) {
568                 info->i_indirect = NULL;
569                 nr_pages_to_free++;
570                 list_add(&topdir->lru, &pages_to_free);
571         }
572         spin_unlock(&info->lock);
573
574         if (info->swapped && idx < SHMEM_NR_DIRECT) {
575                 ptr = info->i_direct;
576                 size = limit;
577                 if (size > SHMEM_NR_DIRECT)
578                         size = SHMEM_NR_DIRECT;
579                 nr_swaps_freed = shmem_free_swp(ptr+idx, ptr+size, needs_lock);
580         }
581
582         /*
583          * If there are no indirect blocks or we are punching a hole
584          * below indirect blocks, nothing to be done.
585          */
586         if (!topdir || limit <= SHMEM_NR_DIRECT)
587                 goto done2;
588
589         /*
590          * The truncation case has already dropped info->lock, and we're safe
591          * because i_size and next_index have already been lowered, preventing
592          * access beyond.  But in the punch_hole case, we still need to take
593          * the lock when updating the swap directory, because there might be
594          * racing accesses by shmem_getpage(SGP_CACHE), shmem_unuse_inode or
595          * shmem_writepage.  However, whenever we find we can remove a whole
596          * directory page (not at the misaligned start or end of the range),
597          * we first NULLify its pointer in the level above, and then have no
598          * need to take the lock when updating its contents: needs_lock and
599          * punch_lock (either pointing to info->lock or NULL) manage this.
600          */
601
602         upper_limit -= SHMEM_NR_DIRECT;
603         limit -= SHMEM_NR_DIRECT;
604         idx = (idx > SHMEM_NR_DIRECT)? (idx - SHMEM_NR_DIRECT): 0;
605         offset = idx % ENTRIES_PER_PAGE;
606         idx -= offset;
607
608         dir = shmem_dir_map(topdir);
609         stage = ENTRIES_PER_PAGEPAGE/2;
610         if (idx < ENTRIES_PER_PAGEPAGE/2) {
611                 middir = topdir;
612                 diroff = idx/ENTRIES_PER_PAGE;
613         } else {
614                 dir += ENTRIES_PER_PAGE/2;
615                 dir += (idx - ENTRIES_PER_PAGEPAGE/2)/ENTRIES_PER_PAGEPAGE;
616                 while (stage <= idx)
617                         stage += ENTRIES_PER_PAGEPAGE;
618                 middir = *dir;
619                 if (*dir) {
620                         diroff = ((idx - ENTRIES_PER_PAGEPAGE/2) %
621                                 ENTRIES_PER_PAGEPAGE) / ENTRIES_PER_PAGE;
622                         if (!diroff && !offset && upper_limit >= stage) {
623                                 if (needs_lock) {
624                                         spin_lock(needs_lock);
625                                         *dir = NULL;
626                                         spin_unlock(needs_lock);
627                                         needs_lock = NULL;
628                                 } else
629                                         *dir = NULL;
630                                 nr_pages_to_free++;
631                                 list_add(&middir->lru, &pages_to_free);
632                         }
633                         shmem_dir_unmap(dir);
634                         dir = shmem_dir_map(middir);
635                 } else {
636                         diroff = 0;
637                         offset = 0;
638                         idx = stage;
639                 }
640         }
641
642         for (; idx < limit; idx += ENTRIES_PER_PAGE, diroff++) {
643                 if (unlikely(idx == stage)) {
644                         shmem_dir_unmap(dir);
645                         dir = shmem_dir_map(topdir) +
646                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
647                         while (!*dir) {
648                                 dir++;
649                                 idx += ENTRIES_PER_PAGEPAGE;
650                                 if (idx >= limit)
651                                         goto done1;
652                         }
653                         stage = idx + ENTRIES_PER_PAGEPAGE;
654                         middir = *dir;
655                         if (punch_hole)
656                                 needs_lock = &info->lock;
657                         if (upper_limit >= stage) {
658                                 if (needs_lock) {
659                                         spin_lock(needs_lock);
660                                         *dir = NULL;
661                                         spin_unlock(needs_lock);
662                                         needs_lock = NULL;
663                                 } else
664                                         *dir = NULL;
665                                 nr_pages_to_free++;
666                                 list_add(&middir->lru, &pages_to_free);
667                         }
668                         shmem_dir_unmap(dir);
669                         cond_resched();
670                         dir = shmem_dir_map(middir);
671                         diroff = 0;
672                 }
673                 punch_lock = needs_lock;
674                 subdir = dir[diroff];
675                 if (subdir && !offset && upper_limit-idx >= ENTRIES_PER_PAGE) {
676                         if (needs_lock) {
677                                 spin_lock(needs_lock);
678                                 dir[diroff] = NULL;
679                                 spin_unlock(needs_lock);
680                                 punch_lock = NULL;
681                         } else
682                                 dir[diroff] = NULL;
683                         nr_pages_to_free++;
684                         list_add(&subdir->lru, &pages_to_free);
685                 }
686                 if (subdir && page_private(subdir) /* has swap entries */) {
687                         size = limit - idx;
688                         if (size > ENTRIES_PER_PAGE)
689                                 size = ENTRIES_PER_PAGE;
690                         freed = shmem_map_and_free_swp(subdir,
691                                         offset, size, &dir, punch_lock);
692                         if (!dir)
693                                 dir = shmem_dir_map(middir);
694                         nr_swaps_freed += freed;
695                         if (offset || punch_lock) {
696                                 spin_lock(&info->lock);
697                                 set_page_private(subdir,
698                                         page_private(subdir) - freed);
699                                 spin_unlock(&info->lock);
700                         } else
701                                 BUG_ON(page_private(subdir) != freed);
702                 }
703                 offset = 0;
704         }
705 done1:
706         shmem_dir_unmap(dir);
707 done2:
708         if (inode->i_mapping->nrpages && (info->flags & SHMEM_PAGEIN)) {
709                 /*
710                  * Call truncate_inode_pages again: racing shmem_unuse_inode
711                  * may have swizzled a page in from swap since vmtruncate or
712                  * generic_delete_inode did it, before we lowered next_index.
713                  * Also, though shmem_getpage checks i_size before adding to
714                  * cache, no recheck after: so fix the narrow window there too.
715                  *
716                  * Recalling truncate_inode_pages_range and unmap_mapping_range
717                  * every time for punch_hole (which never got a chance to clear
718                  * SHMEM_PAGEIN at the start of vmtruncate_range) is expensive,
719                  * yet hardly ever necessary: try to optimize them out later.
720                  */
721                 truncate_inode_pages_range(inode->i_mapping, start, end);
722                 if (punch_hole)
723                         unmap_mapping_range(inode->i_mapping, start,
724                                                         end - start, 1);
725         }
726
727         spin_lock(&info->lock);
728         info->flags &= ~SHMEM_TRUNCATE;
729         info->swapped -= nr_swaps_freed;
730         if (nr_pages_to_free)
731                 shmem_free_blocks(inode, nr_pages_to_free);
732         shmem_recalc_inode(inode);
733         spin_unlock(&info->lock);
734
735         /*
736          * Empty swap vector directory pages to be freed?
737          */
738         if (!list_empty(&pages_to_free)) {
739                 pages_to_free.prev->next = NULL;
740                 shmem_free_pages(pages_to_free.next);
741         }
742 }
743
744 static void shmem_truncate(struct inode *inode)
745 {
746         shmem_truncate_range(inode, inode->i_size, (loff_t)-1);
747 }
748
749 static int shmem_notify_change(struct dentry *dentry, struct iattr *attr)
750 {
751         struct inode *inode = dentry->d_inode;
752         struct page *page = NULL;
753         int error;
754
755         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
756                 if (attr->ia_size < inode->i_size) {
757                         /*
758                          * If truncating down to a partial page, then
759                          * if that page is already allocated, hold it
760                          * in memory until the truncation is over, so
761                          * truncate_partial_page cannnot miss it were
762                          * it assigned to swap.
763                          */
764                         if (attr->ia_size & (PAGE_CACHE_SIZE-1)) {
765                                 (void) shmem_getpage(inode,
766                                         attr->ia_size>>PAGE_CACHE_SHIFT,
767                                                 &page, SGP_READ, NULL);
768                                 if (page)
769                                         unlock_page(page);
770                         }
771                         /*
772                          * Reset SHMEM_PAGEIN flag so that shmem_truncate can
773                          * detect if any pages might have been added to cache
774                          * after truncate_inode_pages.  But we needn't bother
775                          * if it's being fully truncated to zero-length: the
776                          * nrpages check is efficient enough in that case.
777                          */
778                         if (attr->ia_size) {
779                                 struct shmem_inode_info *info = SHMEM_I(inode);
780                                 spin_lock(&info->lock);
781                                 info->flags &= ~SHMEM_PAGEIN;
782                                 spin_unlock(&info->lock);
783                         }
784                 }
785         }
786
787         error = inode_change_ok(inode, attr);
788         if (!error)
789                 error = inode_setattr(inode, attr);
790 #ifdef CONFIG_TMPFS_POSIX_ACL
791         if (!error && (attr->ia_valid & ATTR_MODE))
792                 error = generic_acl_chmod(inode, &shmem_acl_ops);
793 #endif
794         if (page)
795                 page_cache_release(page);
796         return error;
797 }
798
799 static void shmem_delete_inode(struct inode *inode)
800 {
801         struct shmem_inode_info *info = SHMEM_I(inode);
802
803         if (inode->i_op->truncate == shmem_truncate) {
804                 truncate_inode_pages(inode->i_mapping, 0);
805                 shmem_unacct_size(info->flags, inode->i_size);
806                 inode->i_size = 0;
807                 shmem_truncate(inode);
808                 if (!list_empty(&info->swaplist)) {
809                         mutex_lock(&shmem_swaplist_mutex);
810                         list_del_init(&info->swaplist);
811                         mutex_unlock(&shmem_swaplist_mutex);
812                 }
813         }
814         BUG_ON(inode->i_blocks);
815         shmem_free_inode(inode->i_sb);
816         clear_inode(inode);
817 }
818
819 static inline int shmem_find_swp(swp_entry_t entry, swp_entry_t *dir, swp_entry_t *edir)
820 {
821         swp_entry_t *ptr;
822
823         for (ptr = dir; ptr < edir; ptr++) {
824                 if (ptr->val == entry.val)
825                         return ptr - dir;
826         }
827         return -1;
828 }
829
830 static int shmem_unuse_inode(struct shmem_inode_info *info, swp_entry_t entry, struct page *page)
831 {
832         struct inode *inode;
833         unsigned long idx;
834         unsigned long size;
835         unsigned long limit;
836         unsigned long stage;
837         struct page **dir;
838         struct page *subdir;
839         swp_entry_t *ptr;
840         int offset;
841         int error;
842
843         idx = 0;
844         ptr = info->i_direct;
845         spin_lock(&info->lock);
846         if (!info->swapped) {
847                 list_del_init(&info->swaplist);
848                 goto lost2;
849         }
850         limit = info->next_index;
851         size = limit;
852         if (size > SHMEM_NR_DIRECT)
853                 size = SHMEM_NR_DIRECT;
854         offset = shmem_find_swp(entry, ptr, ptr+size);
855         if (offset >= 0)
856                 goto found;
857         if (!info->i_indirect)
858                 goto lost2;
859
860         dir = shmem_dir_map(info->i_indirect);
861         stage = SHMEM_NR_DIRECT + ENTRIES_PER_PAGEPAGE/2;
862
863         for (idx = SHMEM_NR_DIRECT; idx < limit; idx += ENTRIES_PER_PAGE, dir++) {
864                 if (unlikely(idx == stage)) {
865                         shmem_dir_unmap(dir-1);
866                         if (cond_resched_lock(&info->lock)) {
867                                 /* check it has not been truncated */
868                                 if (limit > info->next_index) {
869                                         limit = info->next_index;
870                                         if (idx >= limit)
871                                                 goto lost2;
872                                 }
873                         }
874                         dir = shmem_dir_map(info->i_indirect) +
875                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
876                         while (!*dir) {
877                                 dir++;
878                                 idx += ENTRIES_PER_PAGEPAGE;
879                                 if (idx >= limit)
880                                         goto lost1;
881                         }
882                         stage = idx + ENTRIES_PER_PAGEPAGE;
883                         subdir = *dir;
884                         shmem_dir_unmap(dir);
885                         dir = shmem_dir_map(subdir);
886                 }
887                 subdir = *dir;
888                 if (subdir && page_private(subdir)) {
889                         ptr = shmem_swp_map(subdir);
890                         size = limit - idx;
891                         if (size > ENTRIES_PER_PAGE)
892                                 size = ENTRIES_PER_PAGE;
893                         offset = shmem_find_swp(entry, ptr, ptr+size);
894                         shmem_swp_unmap(ptr);
895                         if (offset >= 0) {
896                                 shmem_dir_unmap(dir);
897                                 goto found;
898                         }
899                 }
900         }
901 lost1:
902         shmem_dir_unmap(dir-1);
903 lost2:
904         spin_unlock(&info->lock);
905         return 0;
906 found:
907         idx += offset;
908         inode = igrab(&info->vfs_inode);
909         spin_unlock(&info->lock);
910
911         /*
912          * Move _head_ to start search for next from here.
913          * But be careful: shmem_delete_inode checks list_empty without taking
914          * mutex, and there's an instant in list_move_tail when info->swaplist
915          * would appear empty, if it were the only one on shmem_swaplist.  We
916          * could avoid doing it if inode NULL; or use this minor optimization.
917          */
918         if (shmem_swaplist.next != &info->swaplist)
919                 list_move_tail(&shmem_swaplist, &info->swaplist);
920         mutex_unlock(&shmem_swaplist_mutex);
921
922         error = 1;
923         if (!inode)
924                 goto out;
925         /* Precharge page using GFP_KERNEL while we can wait */
926         error = mem_cgroup_cache_charge(page, current->mm, GFP_KERNEL);
927         if (error)
928                 goto out;
929         error = radix_tree_preload(GFP_KERNEL);
930         if (error) {
931                 mem_cgroup_uncharge_cache_page(page);
932                 goto out;
933         }
934         error = 1;
935
936         spin_lock(&info->lock);
937         ptr = shmem_swp_entry(info, idx, NULL);
938         if (ptr && ptr->val == entry.val) {
939                 error = add_to_page_cache(page, inode->i_mapping,
940                                                 idx, GFP_NOWAIT);
941                 /* does mem_cgroup_uncharge_cache_page on error */
942         } else  /* we must compensate for our precharge above */
943                 mem_cgroup_uncharge_cache_page(page);
944
945         if (error == -EEXIST) {
946                 struct page *filepage = find_get_page(inode->i_mapping, idx);
947                 error = 1;
948                 if (filepage) {
949                         /*
950                          * There might be a more uptodate page coming down
951                          * from a stacked writepage: forget our swappage if so.
952                          */
953                         if (PageUptodate(filepage))
954                                 error = 0;
955                         page_cache_release(filepage);
956                 }
957         }
958         if (!error) {
959                 delete_from_swap_cache(page);
960                 set_page_dirty(page);
961                 info->flags |= SHMEM_PAGEIN;
962                 shmem_swp_set(info, ptr, 0);
963                 swap_free(entry);
964                 error = 1;      /* not an error, but entry was found */
965         }
966         if (ptr)
967                 shmem_swp_unmap(ptr);
968         spin_unlock(&info->lock);
969         radix_tree_preload_end();
970 out:
971         unlock_page(page);
972         page_cache_release(page);
973         iput(inode);            /* allows for NULL */
974         return error;
975 }
976
977 /*
978  * shmem_unuse() search for an eventually swapped out shmem page.
979  */
980 int shmem_unuse(swp_entry_t entry, struct page *page)
981 {
982         struct list_head *p, *next;
983         struct shmem_inode_info *info;
984         int found = 0;
985
986         mutex_lock(&shmem_swaplist_mutex);
987         list_for_each_safe(p, next, &shmem_swaplist) {
988                 info = list_entry(p, struct shmem_inode_info, swaplist);
989                 found = shmem_unuse_inode(info, entry, page);
990                 cond_resched();
991                 if (found)
992                         goto out;
993         }
994         mutex_unlock(&shmem_swaplist_mutex);
995 out:    return found;   /* 0 or 1 or -ENOMEM */
996 }
997
998 /*
999  * Move the page from the page cache to the swap cache.
1000  */
1001 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1002 {
1003         struct shmem_inode_info *info;
1004         swp_entry_t *entry, swap;
1005         struct address_space *mapping;
1006         unsigned long index;
1007         struct inode *inode;
1008
1009         BUG_ON(!PageLocked(page));
1010         mapping = page->mapping;
1011         index = page->index;
1012         inode = mapping->host;
1013         info = SHMEM_I(inode);
1014         if (info->flags & VM_LOCKED)
1015                 goto redirty;
1016         if (!total_swap_pages)
1017                 goto redirty;
1018
1019         /*
1020          * shmem_backing_dev_info's capabilities prevent regular writeback or
1021          * sync from ever calling shmem_writepage; but a stacking filesystem
1022          * may use the ->writepage of its underlying filesystem, in which case
1023          * tmpfs should write out to swap only in response to memory pressure,
1024          * and not for pdflush or sync.  However, in those cases, we do still
1025          * want to check if there's a redundant swappage to be discarded.
1026          */
1027         if (wbc->for_reclaim)
1028                 swap = get_swap_page();
1029         else
1030                 swap.val = 0;
1031
1032         spin_lock(&info->lock);
1033         if (index >= info->next_index) {
1034                 BUG_ON(!(info->flags & SHMEM_TRUNCATE));
1035                 goto unlock;
1036         }
1037         entry = shmem_swp_entry(info, index, NULL);
1038         if (entry->val) {
1039                 /*
1040                  * The more uptodate page coming down from a stacked
1041                  * writepage should replace our old swappage.
1042                  */
1043                 free_swap_and_cache(*entry);
1044                 shmem_swp_set(info, entry, 0);
1045         }
1046         shmem_recalc_inode(inode);
1047
1048         if (swap.val && add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
1049                 remove_from_page_cache(page);
1050                 shmem_swp_set(info, entry, swap.val);
1051                 shmem_swp_unmap(entry);
1052                 if (list_empty(&info->swaplist))
1053                         inode = igrab(inode);
1054                 else
1055                         inode = NULL;
1056                 spin_unlock(&info->lock);
1057                 swap_duplicate(swap);
1058                 BUG_ON(page_mapped(page));
1059                 page_cache_release(page);       /* pagecache ref */
1060                 set_page_dirty(page);
1061                 unlock_page(page);
1062                 if (inode) {
1063                         mutex_lock(&shmem_swaplist_mutex);
1064                         /* move instead of add in case we're racing */
1065                         list_move_tail(&info->swaplist, &shmem_swaplist);
1066                         mutex_unlock(&shmem_swaplist_mutex);
1067                         iput(inode);
1068                 }
1069                 return 0;
1070         }
1071
1072         shmem_swp_unmap(entry);
1073 unlock:
1074         spin_unlock(&info->lock);
1075         swap_free(swap);
1076 redirty:
1077         set_page_dirty(page);
1078         if (wbc->for_reclaim)
1079                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
1080         unlock_page(page);
1081         return 0;
1082 }
1083
1084 #ifdef CONFIG_NUMA
1085 #ifdef CONFIG_TMPFS
1086 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1087 {
1088         char buffer[64];
1089
1090         if (!mpol || mpol->mode == MPOL_DEFAULT)
1091                 return;         /* show nothing */
1092
1093         mpol_to_str(buffer, sizeof(buffer), mpol, 1);
1094
1095         seq_printf(seq, ",mpol=%s", buffer);
1096 }
1097
1098 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1099 {
1100         struct mempolicy *mpol = NULL;
1101         if (sbinfo->mpol) {
1102                 spin_lock(&sbinfo->stat_lock);  /* prevent replace/use races */
1103                 mpol = sbinfo->mpol;
1104                 mpol_get(mpol);
1105                 spin_unlock(&sbinfo->stat_lock);
1106         }
1107         return mpol;
1108 }
1109 #endif /* CONFIG_TMPFS */
1110
1111 static struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1112                         struct shmem_inode_info *info, unsigned long idx)
1113 {
1114         struct mempolicy mpol, *spol;
1115         struct vm_area_struct pvma;
1116         struct page *page;
1117
1118         spol = mpol_cond_copy(&mpol,
1119                                 mpol_shared_policy_lookup(&info->policy, idx));
1120
1121         /* Create a pseudo vma that just contains the policy */
1122         pvma.vm_start = 0;
1123         pvma.vm_pgoff = idx;
1124         pvma.vm_ops = NULL;
1125         pvma.vm_policy = spol;
1126         page = swapin_readahead(entry, gfp, &pvma, 0);
1127         return page;
1128 }
1129
1130 static struct page *shmem_alloc_page(gfp_t gfp,
1131                         struct shmem_inode_info *info, unsigned long idx)
1132 {
1133         struct vm_area_struct pvma;
1134
1135         /* Create a pseudo vma that just contains the policy */
1136         pvma.vm_start = 0;
1137         pvma.vm_pgoff = idx;
1138         pvma.vm_ops = NULL;
1139         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, idx);
1140
1141         /*
1142          * alloc_page_vma() will drop the shared policy reference
1143          */
1144         return alloc_page_vma(gfp, &pvma, 0);
1145 }
1146 #else /* !CONFIG_NUMA */
1147 #ifdef CONFIG_TMPFS
1148 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *p)
1149 {
1150 }
1151 #endif /* CONFIG_TMPFS */
1152
1153 static inline struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1154                         struct shmem_inode_info *info, unsigned long idx)
1155 {
1156         return swapin_readahead(entry, gfp, NULL, 0);
1157 }
1158
1159 static inline struct page *shmem_alloc_page(gfp_t gfp,
1160                         struct shmem_inode_info *info, unsigned long idx)
1161 {
1162         return alloc_page(gfp);
1163 }
1164 #endif /* CONFIG_NUMA */
1165
1166 #if !defined(CONFIG_NUMA) || !defined(CONFIG_TMPFS)
1167 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1168 {
1169         return NULL;
1170 }
1171 #endif
1172
1173 /*
1174  * shmem_getpage - either get the page from swap or allocate a new one
1175  *
1176  * If we allocate a new one we do not mark it dirty. That's up to the
1177  * vm. If we swap it in we mark it dirty since we also free the swap
1178  * entry since a page cannot live in both the swap and page cache
1179  */
1180 static int shmem_getpage(struct inode *inode, unsigned long idx,
1181                         struct page **pagep, enum sgp_type sgp, int *type)
1182 {
1183         struct address_space *mapping = inode->i_mapping;
1184         struct shmem_inode_info *info = SHMEM_I(inode);
1185         struct shmem_sb_info *sbinfo;
1186         struct page *filepage = *pagep;
1187         struct page *swappage;
1188         swp_entry_t *entry;
1189         swp_entry_t swap;
1190         gfp_t gfp;
1191         int error;
1192
1193         if (idx >= SHMEM_MAX_INDEX)
1194                 return -EFBIG;
1195
1196         if (type)
1197                 *type = 0;
1198
1199         /*
1200          * Normally, filepage is NULL on entry, and either found
1201          * uptodate immediately, or allocated and zeroed, or read
1202          * in under swappage, which is then assigned to filepage.
1203          * But shmem_readpage (required for splice) passes in a locked
1204          * filepage, which may be found not uptodate by other callers
1205          * too, and may need to be copied from the swappage read in.
1206          */
1207 repeat:
1208         if (!filepage)
1209                 filepage = find_lock_page(mapping, idx);
1210         if (filepage && PageUptodate(filepage))
1211                 goto done;
1212         error = 0;
1213         gfp = mapping_gfp_mask(mapping);
1214         if (!filepage) {
1215                 /*
1216                  * Try to preload while we can wait, to not make a habit of
1217                  * draining atomic reserves; but don't latch on to this cpu.
1218                  */
1219                 error = radix_tree_preload(gfp & ~__GFP_HIGHMEM);
1220                 if (error)
1221                         goto failed;
1222                 radix_tree_preload_end();
1223         }
1224
1225         spin_lock(&info->lock);
1226         shmem_recalc_inode(inode);
1227         entry = shmem_swp_alloc(info, idx, sgp);
1228         if (IS_ERR(entry)) {
1229                 spin_unlock(&info->lock);
1230                 error = PTR_ERR(entry);
1231                 goto failed;
1232         }
1233         swap = *entry;
1234
1235         if (swap.val) {
1236                 /* Look it up and read it in.. */
1237                 swappage = lookup_swap_cache(swap);
1238                 if (!swappage) {
1239                         shmem_swp_unmap(entry);
1240                         /* here we actually do the io */
1241                         if (type && !(*type & VM_FAULT_MAJOR)) {
1242                                 __count_vm_event(PGMAJFAULT);
1243                                 *type |= VM_FAULT_MAJOR;
1244                         }
1245                         spin_unlock(&info->lock);
1246                         swappage = shmem_swapin(swap, gfp, info, idx);
1247                         if (!swappage) {
1248                                 spin_lock(&info->lock);
1249                                 entry = shmem_swp_alloc(info, idx, sgp);
1250                                 if (IS_ERR(entry))
1251                                         error = PTR_ERR(entry);
1252                                 else {
1253                                         if (entry->val == swap.val)
1254                                                 error = -ENOMEM;
1255                                         shmem_swp_unmap(entry);
1256                                 }
1257                                 spin_unlock(&info->lock);
1258                                 if (error)
1259                                         goto failed;
1260                                 goto repeat;
1261                         }
1262                         wait_on_page_locked(swappage);
1263                         page_cache_release(swappage);
1264                         goto repeat;
1265                 }
1266
1267                 /* We have to do this with page locked to prevent races */
1268                 if (TestSetPageLocked(swappage)) {
1269                         shmem_swp_unmap(entry);
1270                         spin_unlock(&info->lock);
1271                         wait_on_page_locked(swappage);
1272                         page_cache_release(swappage);
1273                         goto repeat;
1274                 }
1275                 if (PageWriteback(swappage)) {
1276                         shmem_swp_unmap(entry);
1277                         spin_unlock(&info->lock);
1278                         wait_on_page_writeback(swappage);
1279                         unlock_page(swappage);
1280                         page_cache_release(swappage);
1281                         goto repeat;
1282                 }
1283                 if (!PageUptodate(swappage)) {
1284                         shmem_swp_unmap(entry);
1285                         spin_unlock(&info->lock);
1286                         unlock_page(swappage);
1287                         page_cache_release(swappage);
1288                         error = -EIO;
1289                         goto failed;
1290                 }
1291
1292                 if (filepage) {
1293                         shmem_swp_set(info, entry, 0);
1294                         shmem_swp_unmap(entry);
1295                         delete_from_swap_cache(swappage);
1296                         spin_unlock(&info->lock);
1297                         copy_highpage(filepage, swappage);
1298                         unlock_page(swappage);
1299                         page_cache_release(swappage);
1300                         flush_dcache_page(filepage);
1301                         SetPageUptodate(filepage);
1302                         set_page_dirty(filepage);
1303                         swap_free(swap);
1304                 } else if (!(error = add_to_page_cache(
1305                                 swappage, mapping, idx, GFP_NOWAIT))) {
1306                         info->flags |= SHMEM_PAGEIN;
1307                         shmem_swp_set(info, entry, 0);
1308                         shmem_swp_unmap(entry);
1309                         delete_from_swap_cache(swappage);
1310                         spin_unlock(&info->lock);
1311                         filepage = swappage;
1312                         set_page_dirty(filepage);
1313                         swap_free(swap);
1314                 } else {
1315                         shmem_swp_unmap(entry);
1316                         spin_unlock(&info->lock);
1317                         unlock_page(swappage);
1318                         if (error == -ENOMEM) {
1319                                 /* allow reclaim from this memory cgroup */
1320                                 error = mem_cgroup_cache_charge(swappage,
1321                                         current->mm, gfp & ~__GFP_HIGHMEM);
1322                                 if (error) {
1323                                         page_cache_release(swappage);
1324                                         goto failed;
1325                                 }
1326                                 mem_cgroup_uncharge_cache_page(swappage);
1327                         }
1328                         page_cache_release(swappage);
1329                         goto repeat;
1330                 }
1331         } else if (sgp == SGP_READ && !filepage) {
1332                 shmem_swp_unmap(entry);
1333                 filepage = find_get_page(mapping, idx);
1334                 if (filepage &&
1335                     (!PageUptodate(filepage) || TestSetPageLocked(filepage))) {
1336                         spin_unlock(&info->lock);
1337                         wait_on_page_locked(filepage);
1338                         page_cache_release(filepage);
1339                         filepage = NULL;
1340                         goto repeat;
1341                 }
1342                 spin_unlock(&info->lock);
1343         } else {
1344                 shmem_swp_unmap(entry);
1345                 sbinfo = SHMEM_SB(inode->i_sb);
1346                 if (sbinfo->max_blocks) {
1347                         spin_lock(&sbinfo->stat_lock);
1348                         if (sbinfo->free_blocks == 0 ||
1349                             shmem_acct_block(info->flags)) {
1350                                 spin_unlock(&sbinfo->stat_lock);
1351                                 spin_unlock(&info->lock);
1352                                 error = -ENOSPC;
1353                                 goto failed;
1354                         }
1355                         sbinfo->free_blocks--;
1356                         inode->i_blocks += BLOCKS_PER_PAGE;
1357                         spin_unlock(&sbinfo->stat_lock);
1358                 } else if (shmem_acct_block(info->flags)) {
1359                         spin_unlock(&info->lock);
1360                         error = -ENOSPC;
1361                         goto failed;
1362                 }
1363
1364                 if (!filepage) {
1365                         int ret;
1366
1367                         spin_unlock(&info->lock);
1368                         filepage = shmem_alloc_page(gfp, info, idx);
1369                         if (!filepage) {
1370                                 shmem_unacct_blocks(info->flags, 1);
1371                                 shmem_free_blocks(inode, 1);
1372                                 error = -ENOMEM;
1373                                 goto failed;
1374                         }
1375
1376                         /* Precharge page while we can wait, compensate after */
1377                         error = mem_cgroup_cache_charge(filepage, current->mm,
1378                                                         gfp & ~__GFP_HIGHMEM);
1379                         if (error) {
1380                                 page_cache_release(filepage);
1381                                 shmem_unacct_blocks(info->flags, 1);
1382                                 shmem_free_blocks(inode, 1);
1383                                 filepage = NULL;
1384                                 goto failed;
1385                         }
1386
1387                         spin_lock(&info->lock);
1388                         entry = shmem_swp_alloc(info, idx, sgp);
1389                         if (IS_ERR(entry))
1390                                 error = PTR_ERR(entry);
1391                         else {
1392                                 swap = *entry;
1393                                 shmem_swp_unmap(entry);
1394                         }
1395                         ret = error || swap.val;
1396                         if (ret)
1397                                 mem_cgroup_uncharge_cache_page(filepage);
1398                         else
1399                                 ret = add_to_page_cache_lru(filepage, mapping,
1400                                                 idx, GFP_NOWAIT);
1401                         /*
1402                          * At add_to_page_cache_lru() failure, uncharge will
1403                          * be done automatically.
1404                          */
1405                         if (ret) {
1406                                 spin_unlock(&info->lock);
1407                                 page_cache_release(filepage);
1408                                 shmem_unacct_blocks(info->flags, 1);
1409                                 shmem_free_blocks(inode, 1);
1410                                 filepage = NULL;
1411                                 if (error)
1412                                         goto failed;
1413                                 goto repeat;
1414                         }
1415                         info->flags |= SHMEM_PAGEIN;
1416                 }
1417
1418                 info->alloced++;
1419                 spin_unlock(&info->lock);
1420                 clear_highpage(filepage);
1421                 flush_dcache_page(filepage);
1422                 SetPageUptodate(filepage);
1423                 if (sgp == SGP_DIRTY)
1424                         set_page_dirty(filepage);
1425         }
1426 done:
1427         *pagep = filepage;
1428         return 0;
1429
1430 failed:
1431         if (*pagep != filepage) {
1432                 unlock_page(filepage);
1433                 page_cache_release(filepage);
1434         }
1435         return error;
1436 }
1437
1438 static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1439 {
1440         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1441         int error;
1442         int ret;
1443
1444         if (((loff_t)vmf->pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
1445                 return VM_FAULT_SIGBUS;
1446
1447         error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_CACHE, &ret);
1448         if (error)
1449                 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1450
1451         mark_page_accessed(vmf->page);
1452         return ret | VM_FAULT_LOCKED;
1453 }
1454
1455 #ifdef CONFIG_NUMA
1456 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *new)
1457 {
1458         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1459         return mpol_set_shared_policy(&SHMEM_I(i)->policy, vma, new);
1460 }
1461
1462 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
1463                                           unsigned long addr)
1464 {
1465         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1466         unsigned long idx;
1467
1468         idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1469         return mpol_shared_policy_lookup(&SHMEM_I(i)->policy, idx);
1470 }
1471 #endif
1472
1473 int shmem_lock(struct file *file, int lock, struct user_struct *user)
1474 {
1475         struct inode *inode = file->f_path.dentry->d_inode;
1476         struct shmem_inode_info *info = SHMEM_I(inode);
1477         int retval = -ENOMEM;
1478
1479         spin_lock(&info->lock);
1480         if (lock && !(info->flags & VM_LOCKED)) {
1481                 if (!user_shm_lock(inode->i_size, user))
1482                         goto out_nomem;
1483                 info->flags |= VM_LOCKED;
1484         }
1485         if (!lock && (info->flags & VM_LOCKED) && user) {
1486                 user_shm_unlock(inode->i_size, user);
1487                 info->flags &= ~VM_LOCKED;
1488         }
1489         retval = 0;
1490 out_nomem:
1491         spin_unlock(&info->lock);
1492         return retval;
1493 }
1494
1495 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1496 {
1497         file_accessed(file);
1498         vma->vm_ops = &shmem_vm_ops;
1499         vma->vm_flags |= VM_CAN_NONLINEAR;
1500         return 0;
1501 }
1502
1503 static struct inode *
1504 shmem_get_inode(struct super_block *sb, int mode, dev_t dev)
1505 {
1506         struct inode *inode;
1507         struct shmem_inode_info *info;
1508         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1509
1510         if (shmem_reserve_inode(sb))
1511                 return NULL;
1512
1513         inode = new_inode(sb);
1514         if (inode) {
1515                 inode->i_mode = mode;
1516                 inode->i_uid = current->fsuid;
1517                 inode->i_gid = current->fsgid;
1518                 inode->i_blocks = 0;
1519                 inode->i_mapping->a_ops = &shmem_aops;
1520                 inode->i_mapping->backing_dev_info = &shmem_backing_dev_info;
1521                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1522                 inode->i_generation = get_seconds();
1523                 info = SHMEM_I(inode);
1524                 memset(info, 0, (char *)inode - (char *)info);
1525                 spin_lock_init(&info->lock);
1526                 INIT_LIST_HEAD(&info->swaplist);
1527
1528                 switch (mode & S_IFMT) {
1529                 default:
1530                         inode->i_op = &shmem_special_inode_operations;
1531                         init_special_inode(inode, mode, dev);
1532                         break;
1533                 case S_IFREG:
1534                         inode->i_op = &shmem_inode_operations;
1535                         inode->i_fop = &shmem_file_operations;
1536                         mpol_shared_policy_init(&info->policy,
1537                                                  shmem_get_sbmpol(sbinfo));
1538                         break;
1539                 case S_IFDIR:
1540                         inc_nlink(inode);
1541                         /* Some things misbehave if size == 0 on a directory */
1542                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
1543                         inode->i_op = &shmem_dir_inode_operations;
1544                         inode->i_fop = &simple_dir_operations;
1545                         break;
1546                 case S_IFLNK:
1547                         /*
1548                          * Must not load anything in the rbtree,
1549                          * mpol_free_shared_policy will not be called.
1550                          */
1551                         mpol_shared_policy_init(&info->policy, NULL);
1552                         break;
1553                 }
1554         } else
1555                 shmem_free_inode(sb);
1556         return inode;
1557 }
1558
1559 #ifdef CONFIG_TMPFS
1560 static const struct inode_operations shmem_symlink_inode_operations;
1561 static const struct inode_operations shmem_symlink_inline_operations;
1562
1563 /*
1564  * Normally tmpfs avoids the use of shmem_readpage and shmem_write_begin;
1565  * but providing them allows a tmpfs file to be used for splice, sendfile, and
1566  * below the loop driver, in the generic fashion that many filesystems support.
1567  */
1568 static int shmem_readpage(struct file *file, struct page *page)
1569 {
1570         struct inode *inode = page->mapping->host;
1571         int error = shmem_getpage(inode, page->index, &page, SGP_CACHE, NULL);
1572         unlock_page(page);
1573         return error;
1574 }
1575
1576 static int
1577 shmem_write_begin(struct file *file, struct address_space *mapping,
1578                         loff_t pos, unsigned len, unsigned flags,
1579                         struct page **pagep, void **fsdata)
1580 {
1581         struct inode *inode = mapping->host;
1582         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1583         *pagep = NULL;
1584         return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1585 }
1586
1587 static int
1588 shmem_write_end(struct file *file, struct address_space *mapping,
1589                         loff_t pos, unsigned len, unsigned copied,
1590                         struct page *page, void *fsdata)
1591 {
1592         struct inode *inode = mapping->host;
1593
1594         if (pos + copied > inode->i_size)
1595                 i_size_write(inode, pos + copied);
1596
1597         unlock_page(page);
1598         set_page_dirty(page);
1599         page_cache_release(page);
1600
1601         return copied;
1602 }
1603
1604 static void do_shmem_file_read(struct file *filp, loff_t *ppos, read_descriptor_t *desc, read_actor_t actor)
1605 {
1606         struct inode *inode = filp->f_path.dentry->d_inode;
1607         struct address_space *mapping = inode->i_mapping;
1608         unsigned long index, offset;
1609         enum sgp_type sgp = SGP_READ;
1610
1611         /*
1612          * Might this read be for a stacking filesystem?  Then when reading
1613          * holes of a sparse file, we actually need to allocate those pages,
1614          * and even mark them dirty, so it cannot exceed the max_blocks limit.
1615          */
1616         if (segment_eq(get_fs(), KERNEL_DS))
1617                 sgp = SGP_DIRTY;
1618
1619         index = *ppos >> PAGE_CACHE_SHIFT;
1620         offset = *ppos & ~PAGE_CACHE_MASK;
1621
1622         for (;;) {
1623                 struct page *page = NULL;
1624                 unsigned long end_index, nr, ret;
1625                 loff_t i_size = i_size_read(inode);
1626
1627                 end_index = i_size >> PAGE_CACHE_SHIFT;
1628                 if (index > end_index)
1629                         break;
1630                 if (index == end_index) {
1631                         nr = i_size & ~PAGE_CACHE_MASK;
1632                         if (nr <= offset)
1633                                 break;
1634                 }
1635
1636                 desc->error = shmem_getpage(inode, index, &page, sgp, NULL);
1637                 if (desc->error) {
1638                         if (desc->error == -EINVAL)
1639                                 desc->error = 0;
1640                         break;
1641                 }
1642                 if (page)
1643                         unlock_page(page);
1644
1645                 /*
1646                  * We must evaluate after, since reads (unlike writes)
1647                  * are called without i_mutex protection against truncate
1648                  */
1649                 nr = PAGE_CACHE_SIZE;
1650                 i_size = i_size_read(inode);
1651                 end_index = i_size >> PAGE_CACHE_SHIFT;
1652                 if (index == end_index) {
1653                         nr = i_size & ~PAGE_CACHE_MASK;
1654                         if (nr <= offset) {
1655                                 if (page)
1656                                         page_cache_release(page);
1657                                 break;
1658                         }
1659                 }
1660                 nr -= offset;
1661
1662                 if (page) {
1663                         /*
1664                          * If users can be writing to this page using arbitrary
1665                          * virtual addresses, take care about potential aliasing
1666                          * before reading the page on the kernel side.
1667                          */
1668                         if (mapping_writably_mapped(mapping))
1669                                 flush_dcache_page(page);
1670                         /*
1671                          * Mark the page accessed if we read the beginning.
1672                          */
1673                         if (!offset)
1674                                 mark_page_accessed(page);
1675                 } else {
1676                         page = ZERO_PAGE(0);
1677                         page_cache_get(page);
1678                 }
1679
1680                 /*
1681                  * Ok, we have the page, and it's up-to-date, so
1682                  * now we can copy it to user space...
1683                  *
1684                  * The actor routine returns how many bytes were actually used..
1685                  * NOTE! This may not be the same as how much of a user buffer
1686                  * we filled up (we may be padding etc), so we can only update
1687                  * "pos" here (the actor routine has to update the user buffer
1688                  * pointers and the remaining count).
1689                  */
1690                 ret = actor(desc, page, offset, nr);
1691                 offset += ret;
1692                 index += offset >> PAGE_CACHE_SHIFT;
1693                 offset &= ~PAGE_CACHE_MASK;
1694
1695                 page_cache_release(page);
1696                 if (ret != nr || !desc->count)
1697                         break;
1698
1699                 cond_resched();
1700         }
1701
1702         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1703         file_accessed(filp);
1704 }
1705
1706 static ssize_t shmem_file_aio_read(struct kiocb *iocb,
1707                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
1708 {
1709         struct file *filp = iocb->ki_filp;
1710         ssize_t retval;
1711         unsigned long seg;
1712         size_t count;
1713         loff_t *ppos = &iocb->ki_pos;
1714
1715         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1716         if (retval)
1717                 return retval;
1718
1719         for (seg = 0; seg < nr_segs; seg++) {
1720                 read_descriptor_t desc;
1721
1722                 desc.written = 0;
1723                 desc.arg.buf = iov[seg].iov_base;
1724                 desc.count = iov[seg].iov_len;
1725                 if (desc.count == 0)
1726                         continue;
1727                 desc.error = 0;
1728                 do_shmem_file_read(filp, ppos, &desc, file_read_actor);
1729                 retval += desc.written;
1730                 if (desc.error) {
1731                         retval = retval ?: desc.error;
1732                         break;
1733                 }
1734                 if (desc.count > 0)
1735                         break;
1736         }
1737         return retval;
1738 }
1739
1740 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
1741 {
1742         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
1743
1744         buf->f_type = TMPFS_MAGIC;
1745         buf->f_bsize = PAGE_CACHE_SIZE;
1746         buf->f_namelen = NAME_MAX;
1747         spin_lock(&sbinfo->stat_lock);
1748         if (sbinfo->max_blocks) {
1749                 buf->f_blocks = sbinfo->max_blocks;
1750                 buf->f_bavail = buf->f_bfree = sbinfo->free_blocks;
1751         }
1752         if (sbinfo->max_inodes) {
1753                 buf->f_files = sbinfo->max_inodes;
1754                 buf->f_ffree = sbinfo->free_inodes;
1755         }
1756         /* else leave those fields 0 like simple_statfs */
1757         spin_unlock(&sbinfo->stat_lock);
1758         return 0;
1759 }
1760
1761 /*
1762  * File creation. Allocate an inode, and we're done..
1763  */
1764 static int
1765 shmem_mknod(struct inode *dir, struct dentry *dentry, int mode, dev_t dev)
1766 {
1767         struct inode *inode = shmem_get_inode(dir->i_sb, mode, dev);
1768         int error = -ENOSPC;
1769
1770         if (inode) {
1771                 error = security_inode_init_security(inode, dir, NULL, NULL,
1772                                                      NULL);
1773                 if (error) {
1774                         if (error != -EOPNOTSUPP) {
1775                                 iput(inode);
1776                                 return error;
1777                         }
1778                 }
1779                 error = shmem_acl_init(inode, dir);
1780                 if (error) {
1781                         iput(inode);
1782                         return error;
1783                 }
1784                 if (dir->i_mode & S_ISGID) {
1785                         inode->i_gid = dir->i_gid;
1786                         if (S_ISDIR(mode))
1787                                 inode->i_mode |= S_ISGID;
1788                 }
1789                 dir->i_size += BOGO_DIRENT_SIZE;
1790                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1791                 d_instantiate(dentry, inode);
1792                 dget(dentry); /* Extra count - pin the dentry in core */
1793         }
1794         return error;
1795 }
1796
1797 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, int mode)
1798 {
1799         int error;
1800
1801         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
1802                 return error;
1803         inc_nlink(dir);
1804         return 0;
1805 }
1806
1807 static int shmem_create(struct inode *dir, struct dentry *dentry, int mode,
1808                 struct nameidata *nd)
1809 {
1810         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
1811 }
1812
1813 /*
1814  * Link a file..
1815  */
1816 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
1817 {
1818         struct inode *inode = old_dentry->d_inode;
1819         int ret;
1820
1821         /*
1822          * No ordinary (disk based) filesystem counts links as inodes;
1823          * but each new link needs a new dentry, pinning lowmem, and
1824          * tmpfs dentries cannot be pruned until they are unlinked.
1825          */
1826         ret = shmem_reserve_inode(inode->i_sb);
1827         if (ret)
1828                 goto out;
1829
1830         dir->i_size += BOGO_DIRENT_SIZE;
1831         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1832         inc_nlink(inode);
1833         atomic_inc(&inode->i_count);    /* New dentry reference */
1834         dget(dentry);           /* Extra pinning count for the created dentry */
1835         d_instantiate(dentry, inode);
1836 out:
1837         return ret;
1838 }
1839
1840 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
1841 {
1842         struct inode *inode = dentry->d_inode;
1843
1844         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
1845                 shmem_free_inode(inode->i_sb);
1846
1847         dir->i_size -= BOGO_DIRENT_SIZE;
1848         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1849         drop_nlink(inode);
1850         dput(dentry);   /* Undo the count from "create" - this does all the work */
1851         return 0;
1852 }
1853
1854 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
1855 {
1856         if (!simple_empty(dentry))
1857                 return -ENOTEMPTY;
1858
1859         drop_nlink(dentry->d_inode);
1860         drop_nlink(dir);
1861         return shmem_unlink(dir, dentry);
1862 }
1863
1864 /*
1865  * The VFS layer already does all the dentry stuff for rename,
1866  * we just have to decrement the usage count for the target if
1867  * it exists so that the VFS layer correctly free's it when it
1868  * gets overwritten.
1869  */
1870 static int shmem_rename(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
1871 {
1872         struct inode *inode = old_dentry->d_inode;
1873         int they_are_dirs = S_ISDIR(inode->i_mode);
1874
1875         if (!simple_empty(new_dentry))
1876                 return -ENOTEMPTY;
1877
1878         if (new_dentry->d_inode) {
1879                 (void) shmem_unlink(new_dir, new_dentry);
1880                 if (they_are_dirs)
1881                         drop_nlink(old_dir);
1882         } else if (they_are_dirs) {
1883                 drop_nlink(old_dir);
1884                 inc_nlink(new_dir);
1885         }
1886
1887         old_dir->i_size -= BOGO_DIRENT_SIZE;
1888         new_dir->i_size += BOGO_DIRENT_SIZE;
1889         old_dir->i_ctime = old_dir->i_mtime =
1890         new_dir->i_ctime = new_dir->i_mtime =
1891         inode->i_ctime = CURRENT_TIME;
1892         return 0;
1893 }
1894
1895 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
1896 {
1897         int error;
1898         int len;
1899         struct inode *inode;
1900         struct page *page = NULL;
1901         char *kaddr;
1902         struct shmem_inode_info *info;
1903
1904         len = strlen(symname) + 1;
1905         if (len > PAGE_CACHE_SIZE)
1906                 return -ENAMETOOLONG;
1907
1908         inode = shmem_get_inode(dir->i_sb, S_IFLNK|S_IRWXUGO, 0);
1909         if (!inode)
1910                 return -ENOSPC;
1911
1912         error = security_inode_init_security(inode, dir, NULL, NULL,
1913                                              NULL);
1914         if (error) {
1915                 if (error != -EOPNOTSUPP) {
1916                         iput(inode);
1917                         return error;
1918                 }
1919                 error = 0;
1920         }
1921
1922         info = SHMEM_I(inode);
1923         inode->i_size = len-1;
1924         if (len <= (char *)inode - (char *)info) {
1925                 /* do it inline */
1926                 memcpy(info, symname, len);
1927                 inode->i_op = &shmem_symlink_inline_operations;
1928         } else {
1929                 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
1930                 if (error) {
1931                         iput(inode);
1932                         return error;
1933                 }
1934                 unlock_page(page);
1935                 inode->i_op = &shmem_symlink_inode_operations;
1936                 kaddr = kmap_atomic(page, KM_USER0);
1937                 memcpy(kaddr, symname, len);
1938                 kunmap_atomic(kaddr, KM_USER0);
1939                 set_page_dirty(page);
1940                 page_cache_release(page);
1941         }
1942         if (dir->i_mode & S_ISGID)
1943                 inode->i_gid = dir->i_gid;
1944         dir->i_size += BOGO_DIRENT_SIZE;
1945         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1946         d_instantiate(dentry, inode);
1947         dget(dentry);
1948         return 0;
1949 }
1950
1951 static void *shmem_follow_link_inline(struct dentry *dentry, struct nameidata *nd)
1952 {
1953         nd_set_link(nd, (char *)SHMEM_I(dentry->d_inode));
1954         return NULL;
1955 }
1956
1957 static void *shmem_follow_link(struct dentry *dentry, struct nameidata *nd)
1958 {
1959         struct page *page = NULL;
1960         int res = shmem_getpage(dentry->d_inode, 0, &page, SGP_READ, NULL);
1961         nd_set_link(nd, res ? ERR_PTR(res) : kmap(page));
1962         if (page)
1963                 unlock_page(page);
1964         return page;
1965 }
1966
1967 static void shmem_put_link(struct dentry *dentry, struct nameidata *nd, void *cookie)
1968 {
1969         if (!IS_ERR(nd_get_link(nd))) {
1970                 struct page *page = cookie;
1971                 kunmap(page);
1972                 mark_page_accessed(page);
1973                 page_cache_release(page);
1974         }
1975 }
1976
1977 static const struct inode_operations shmem_symlink_inline_operations = {
1978         .readlink       = generic_readlink,
1979         .follow_link    = shmem_follow_link_inline,
1980 };
1981
1982 static const struct inode_operations shmem_symlink_inode_operations = {
1983         .truncate       = shmem_truncate,
1984         .readlink       = generic_readlink,
1985         .follow_link    = shmem_follow_link,
1986         .put_link       = shmem_put_link,
1987 };
1988
1989 #ifdef CONFIG_TMPFS_POSIX_ACL
1990 /*
1991  * Superblocks without xattr inode operations will get security.* xattr
1992  * support from the VFS "for free". As soon as we have any other xattrs
1993  * like ACLs, we also need to implement the security.* handlers at
1994  * filesystem level, though.
1995  */
1996
1997 static size_t shmem_xattr_security_list(struct inode *inode, char *list,
1998                                         size_t list_len, const char *name,
1999                                         size_t name_len)
2000 {
2001         return security_inode_listsecurity(inode, list, list_len);
2002 }
2003
2004 static int shmem_xattr_security_get(struct inode *inode, const char *name,
2005                                     void *buffer, size_t size)
2006 {
2007         if (strcmp(name, "") == 0)
2008                 return -EINVAL;
2009         return xattr_getsecurity(inode, name, buffer, size);
2010 }
2011
2012 static int shmem_xattr_security_set(struct inode *inode, const char *name,
2013                                     const void *value, size_t size, int flags)
2014 {
2015         if (strcmp(name, "") == 0)
2016                 return -EINVAL;
2017         return security_inode_setsecurity(inode, name, value, size, flags);
2018 }
2019
2020 static struct xattr_handler shmem_xattr_security_handler = {
2021         .prefix = XATTR_SECURITY_PREFIX,
2022         .list   = shmem_xattr_security_list,
2023         .get    = shmem_xattr_security_get,
2024         .set    = shmem_xattr_security_set,
2025 };
2026
2027 static struct xattr_handler *shmem_xattr_handlers[] = {
2028         &shmem_xattr_acl_access_handler,
2029         &shmem_xattr_acl_default_handler,
2030         &shmem_xattr_security_handler,
2031         NULL
2032 };
2033 #endif
2034
2035 static struct dentry *shmem_get_parent(struct dentry *child)
2036 {
2037         return ERR_PTR(-ESTALE);
2038 }
2039
2040 static int shmem_match(struct inode *ino, void *vfh)
2041 {
2042         __u32 *fh = vfh;
2043         __u64 inum = fh[2];
2044         inum = (inum << 32) | fh[1];
2045         return ino->i_ino == inum && fh[0] == ino->i_generation;
2046 }
2047
2048 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
2049                 struct fid *fid, int fh_len, int fh_type)
2050 {
2051         struct inode *inode;
2052         struct dentry *dentry = NULL;
2053         u64 inum = fid->raw[2];
2054         inum = (inum << 32) | fid->raw[1];
2055
2056         if (fh_len < 3)
2057                 return NULL;
2058
2059         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
2060                         shmem_match, fid->raw);
2061         if (inode) {
2062                 dentry = d_find_alias(inode);
2063                 iput(inode);
2064         }
2065
2066         return dentry;
2067 }
2068
2069 static int shmem_encode_fh(struct dentry *dentry, __u32 *fh, int *len,
2070                                 int connectable)
2071 {
2072         struct inode *inode = dentry->d_inode;
2073
2074         if (*len < 3)
2075                 return 255;
2076
2077         if (hlist_unhashed(&inode->i_hash)) {
2078                 /* Unfortunately insert_inode_hash is not idempotent,
2079                  * so as we hash inodes here rather than at creation
2080                  * time, we need a lock to ensure we only try
2081                  * to do it once
2082                  */
2083                 static DEFINE_SPINLOCK(lock);
2084                 spin_lock(&lock);
2085                 if (hlist_unhashed(&inode->i_hash))
2086                         __insert_inode_hash(inode,
2087                                             inode->i_ino + inode->i_generation);
2088                 spin_unlock(&lock);
2089         }
2090
2091         fh[0] = inode->i_generation;
2092         fh[1] = inode->i_ino;
2093         fh[2] = ((__u64)inode->i_ino) >> 32;
2094
2095         *len = 3;
2096         return 1;
2097 }
2098
2099 static const struct export_operations shmem_export_ops = {
2100         .get_parent     = shmem_get_parent,
2101         .encode_fh      = shmem_encode_fh,
2102         .fh_to_dentry   = shmem_fh_to_dentry,
2103 };
2104
2105 static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
2106                                bool remount)
2107 {
2108         char *this_char, *value, *rest;
2109
2110         while (options != NULL) {
2111                 this_char = options;
2112                 for (;;) {
2113                         /*
2114                          * NUL-terminate this option: unfortunately,
2115                          * mount options form a comma-separated list,
2116                          * but mpol's nodelist may also contain commas.
2117                          */
2118                         options = strchr(options, ',');
2119                         if (options == NULL)
2120                                 break;
2121                         options++;
2122                         if (!isdigit(*options)) {
2123                                 options[-1] = '\0';
2124                                 break;
2125                         }
2126                 }
2127                 if (!*this_char)
2128                         continue;
2129                 if ((value = strchr(this_char,'=')) != NULL) {
2130                         *value++ = 0;
2131                 } else {
2132                         printk(KERN_ERR
2133                             "tmpfs: No value for mount option '%s'\n",
2134                             this_char);
2135                         return 1;
2136                 }
2137
2138                 if (!strcmp(this_char,"size")) {
2139                         unsigned long long size;
2140                         size = memparse(value,&rest);
2141                         if (*rest == '%') {
2142                                 size <<= PAGE_SHIFT;
2143                                 size *= totalram_pages;
2144                                 do_div(size, 100);
2145                                 rest++;
2146                         }
2147                         if (*rest)
2148                                 goto bad_val;
2149                         sbinfo->max_blocks =
2150                                 DIV_ROUND_UP(size, PAGE_CACHE_SIZE);
2151                 } else if (!strcmp(this_char,"nr_blocks")) {
2152                         sbinfo->max_blocks = memparse(value, &rest);
2153                         if (*rest)
2154                                 goto bad_val;
2155                 } else if (!strcmp(this_char,"nr_inodes")) {
2156                         sbinfo->max_inodes = memparse(value, &rest);
2157                         if (*rest)
2158                                 goto bad_val;
2159                 } else if (!strcmp(this_char,"mode")) {
2160                         if (remount)
2161                                 continue;
2162                         sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
2163                         if (*rest)
2164                                 goto bad_val;
2165                 } else if (!strcmp(this_char,"uid")) {
2166                         if (remount)
2167                                 continue;
2168                         sbinfo->uid = simple_strtoul(value, &rest, 0);
2169                         if (*rest)
2170                                 goto bad_val;
2171                 } else if (!strcmp(this_char,"gid")) {
2172                         if (remount)
2173                                 continue;
2174                         sbinfo->gid = simple_strtoul(value, &rest, 0);
2175                         if (*rest)
2176                                 goto bad_val;
2177                 } else if (!strcmp(this_char,"mpol")) {
2178                         if (mpol_parse_str(value, &sbinfo->mpol, 1))
2179                                 goto bad_val;
2180                 } else {
2181                         printk(KERN_ERR "tmpfs: Bad mount option %s\n",
2182                                this_char);
2183                         return 1;
2184                 }
2185         }
2186         return 0;
2187
2188 bad_val:
2189         printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
2190                value, this_char);
2191         return 1;
2192
2193 }
2194
2195 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2196 {
2197         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2198         struct shmem_sb_info config = *sbinfo;
2199         unsigned long blocks;
2200         unsigned long inodes;
2201         int error = -EINVAL;
2202
2203         if (shmem_parse_options(data, &config, true))
2204                 return error;
2205
2206         spin_lock(&sbinfo->stat_lock);
2207         blocks = sbinfo->max_blocks - sbinfo->free_blocks;
2208         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2209         if (config.max_blocks < blocks)
2210                 goto out;
2211         if (config.max_inodes < inodes)
2212                 goto out;
2213         /*
2214          * Those tests also disallow limited->unlimited while any are in
2215          * use, so i_blocks will always be zero when max_blocks is zero;
2216          * but we must separately disallow unlimited->limited, because
2217          * in that case we have no record of how much is already in use.
2218          */
2219         if (config.max_blocks && !sbinfo->max_blocks)
2220                 goto out;
2221         if (config.max_inodes && !sbinfo->max_inodes)
2222                 goto out;
2223
2224         error = 0;
2225         sbinfo->max_blocks  = config.max_blocks;
2226         sbinfo->free_blocks = config.max_blocks - blocks;
2227         sbinfo->max_inodes  = config.max_inodes;
2228         sbinfo->free_inodes = config.max_inodes - inodes;
2229
2230         mpol_put(sbinfo->mpol);
2231         sbinfo->mpol        = config.mpol;      /* transfers initial ref */
2232 out:
2233         spin_unlock(&sbinfo->stat_lock);
2234         return error;
2235 }
2236
2237 static int shmem_show_options(struct seq_file *seq, struct vfsmount *vfs)
2238 {
2239         struct shmem_sb_info *sbinfo = SHMEM_SB(vfs->mnt_sb);
2240
2241         if (sbinfo->max_blocks != shmem_default_max_blocks())
2242                 seq_printf(seq, ",size=%luk",
2243                         sbinfo->max_blocks << (PAGE_CACHE_SHIFT - 10));
2244         if (sbinfo->max_inodes != shmem_default_max_inodes())
2245                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
2246         if (sbinfo->mode != (S_IRWXUGO | S_ISVTX))
2247                 seq_printf(seq, ",mode=%03o", sbinfo->mode);
2248         if (sbinfo->uid != 0)
2249                 seq_printf(seq, ",uid=%u", sbinfo->uid);
2250         if (sbinfo->gid != 0)
2251                 seq_printf(seq, ",gid=%u", sbinfo->gid);
2252         shmem_show_mpol(seq, sbinfo->mpol);
2253         return 0;
2254 }
2255 #endif /* CONFIG_TMPFS */
2256
2257 static void shmem_put_super(struct super_block *sb)
2258 {
2259         kfree(sb->s_fs_info);
2260         sb->s_fs_info = NULL;
2261 }
2262
2263 static int shmem_fill_super(struct super_block *sb,
2264                             void *data, int silent)
2265 {
2266         struct inode *inode;
2267         struct dentry *root;
2268         struct shmem_sb_info *sbinfo;
2269         int err = -ENOMEM;
2270
2271         /* Round up to L1_CACHE_BYTES to resist false sharing */
2272         sbinfo = kmalloc(max((int)sizeof(struct shmem_sb_info),
2273                                 L1_CACHE_BYTES), GFP_KERNEL);
2274         if (!sbinfo)
2275                 return -ENOMEM;
2276
2277         sbinfo->max_blocks = 0;
2278         sbinfo->max_inodes = 0;
2279         sbinfo->mode = S_IRWXUGO | S_ISVTX;
2280         sbinfo->uid = current->fsuid;
2281         sbinfo->gid = current->fsgid;
2282         sbinfo->mpol = NULL;
2283         sb->s_fs_info = sbinfo;
2284
2285 #ifdef CONFIG_TMPFS
2286         /*
2287          * Per default we only allow half of the physical ram per
2288          * tmpfs instance, limiting inodes to one per page of lowmem;
2289          * but the internal instance is left unlimited.
2290          */
2291         if (!(sb->s_flags & MS_NOUSER)) {
2292                 sbinfo->max_blocks = shmem_default_max_blocks();
2293                 sbinfo->max_inodes = shmem_default_max_inodes();
2294                 if (shmem_parse_options(data, sbinfo, false)) {
2295                         err = -EINVAL;
2296                         goto failed;
2297                 }
2298         }
2299         sb->s_export_op = &shmem_export_ops;
2300 #else
2301         sb->s_flags |= MS_NOUSER;
2302 #endif
2303
2304         spin_lock_init(&sbinfo->stat_lock);
2305         sbinfo->free_blocks = sbinfo->max_blocks;
2306         sbinfo->free_inodes = sbinfo->max_inodes;
2307
2308         sb->s_maxbytes = SHMEM_MAX_BYTES;
2309         sb->s_blocksize = PAGE_CACHE_SIZE;
2310         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
2311         sb->s_magic = TMPFS_MAGIC;
2312         sb->s_op = &shmem_ops;
2313         sb->s_time_gran = 1;
2314 #ifdef CONFIG_TMPFS_POSIX_ACL
2315         sb->s_xattr = shmem_xattr_handlers;
2316         sb->s_flags |= MS_POSIXACL;
2317 #endif
2318
2319         inode = shmem_get_inode(sb, S_IFDIR | sbinfo->mode, 0);
2320         if (!inode)
2321                 goto failed;
2322         inode->i_uid = sbinfo->uid;
2323         inode->i_gid = sbinfo->gid;
2324         root = d_alloc_root(inode);
2325         if (!root)
2326                 goto failed_iput;
2327         sb->s_root = root;
2328         return 0;
2329
2330 failed_iput:
2331         iput(inode);
2332 failed:
2333         shmem_put_super(sb);
2334         return err;
2335 }
2336
2337 static struct kmem_cache *shmem_inode_cachep;
2338
2339 static struct inode *shmem_alloc_inode(struct super_block *sb)
2340 {
2341         struct shmem_inode_info *p;
2342         p = (struct shmem_inode_info *)kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
2343         if (!p)
2344                 return NULL;
2345         return &p->vfs_inode;
2346 }
2347
2348 static void shmem_destroy_inode(struct inode *inode)
2349 {
2350         if ((inode->i_mode & S_IFMT) == S_IFREG) {
2351                 /* only struct inode is valid if it's an inline symlink */
2352                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
2353         }
2354         shmem_acl_destroy_inode(inode);
2355         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
2356 }
2357
2358 static void init_once(struct kmem_cache *cachep, void *foo)
2359 {
2360         struct shmem_inode_info *p = (struct shmem_inode_info *) foo;
2361
2362         inode_init_once(&p->vfs_inode);
2363 #ifdef CONFIG_TMPFS_POSIX_ACL
2364         p->i_acl = NULL;
2365         p->i_default_acl = NULL;
2366 #endif
2367 }
2368
2369 static int init_inodecache(void)
2370 {
2371         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
2372                                 sizeof(struct shmem_inode_info),
2373                                 0, SLAB_PANIC, init_once);
2374         return 0;
2375 }
2376
2377 static void destroy_inodecache(void)
2378 {
2379         kmem_cache_destroy(shmem_inode_cachep);
2380 }
2381
2382 static const struct address_space_operations shmem_aops = {
2383         .writepage      = shmem_writepage,
2384         .set_page_dirty = __set_page_dirty_no_writeback,
2385 #ifdef CONFIG_TMPFS
2386         .readpage       = shmem_readpage,
2387         .write_begin    = shmem_write_begin,
2388         .write_end      = shmem_write_end,
2389 #endif
2390         .migratepage    = migrate_page,
2391 };
2392
2393 static const struct file_operations shmem_file_operations = {
2394         .mmap           = shmem_mmap,
2395 #ifdef CONFIG_TMPFS
2396         .llseek         = generic_file_llseek,
2397         .read           = do_sync_read,
2398         .write          = do_sync_write,
2399         .aio_read       = shmem_file_aio_read,
2400         .aio_write      = generic_file_aio_write,
2401         .fsync          = simple_sync_file,
2402         .splice_read    = generic_file_splice_read,
2403         .splice_write   = generic_file_splice_write,
2404 #endif
2405 };
2406
2407 static const struct inode_operations shmem_inode_operations = {
2408         .truncate       = shmem_truncate,
2409         .setattr        = shmem_notify_change,
2410         .truncate_range = shmem_truncate_range,
2411 #ifdef CONFIG_TMPFS_POSIX_ACL
2412         .setxattr       = generic_setxattr,
2413         .getxattr       = generic_getxattr,
2414         .listxattr      = generic_listxattr,
2415         .removexattr    = generic_removexattr,
2416         .permission     = shmem_permission,
2417 #endif
2418
2419 };
2420
2421 static const struct inode_operations shmem_dir_inode_operations = {
2422 #ifdef CONFIG_TMPFS
2423         .create         = shmem_create,
2424         .lookup         = simple_lookup,
2425         .link           = shmem_link,
2426         .unlink         = shmem_unlink,
2427         .symlink        = shmem_symlink,
2428         .mkdir          = shmem_mkdir,
2429         .rmdir          = shmem_rmdir,
2430         .mknod          = shmem_mknod,
2431         .rename         = shmem_rename,
2432 #endif
2433 #ifdef CONFIG_TMPFS_POSIX_ACL
2434         .setattr        = shmem_notify_change,
2435         .setxattr       = generic_setxattr,
2436         .getxattr       = generic_getxattr,
2437         .listxattr      = generic_listxattr,
2438         .removexattr    = generic_removexattr,
2439         .permission     = shmem_permission,
2440 #endif
2441 };
2442
2443 static const struct inode_operations shmem_special_inode_operations = {
2444 #ifdef CONFIG_TMPFS_POSIX_ACL
2445         .setattr        = shmem_notify_change,
2446         .setxattr       = generic_setxattr,
2447         .getxattr       = generic_getxattr,
2448         .listxattr      = generic_listxattr,
2449         .removexattr    = generic_removexattr,
2450         .permission     = shmem_permission,
2451 #endif
2452 };
2453
2454 static const struct super_operations shmem_ops = {
2455         .alloc_inode    = shmem_alloc_inode,
2456         .destroy_inode  = shmem_destroy_inode,
2457 #ifdef CONFIG_TMPFS
2458         .statfs         = shmem_statfs,
2459         .remount_fs     = shmem_remount_fs,
2460         .show_options   = shmem_show_options,
2461 #endif
2462         .delete_inode   = shmem_delete_inode,
2463         .drop_inode     = generic_delete_inode,
2464         .put_super      = shmem_put_super,
2465 };
2466
2467 static struct vm_operations_struct shmem_vm_ops = {
2468         .fault          = shmem_fault,
2469 #ifdef CONFIG_NUMA
2470         .set_policy     = shmem_set_policy,
2471         .get_policy     = shmem_get_policy,
2472 #endif
2473 };
2474
2475
2476 static int shmem_get_sb(struct file_system_type *fs_type,
2477         int flags, const char *dev_name, void *data, struct vfsmount *mnt)
2478 {
2479         return get_sb_nodev(fs_type, flags, data, shmem_fill_super, mnt);
2480 }
2481
2482 static struct file_system_type tmpfs_fs_type = {
2483         .owner          = THIS_MODULE,
2484         .name           = "tmpfs",
2485         .get_sb         = shmem_get_sb,
2486         .kill_sb        = kill_litter_super,
2487 };
2488 static struct vfsmount *shm_mnt;
2489
2490 static int __init init_tmpfs(void)
2491 {
2492         int error;
2493
2494         error = bdi_init(&shmem_backing_dev_info);
2495         if (error)
2496                 goto out4;
2497
2498         error = init_inodecache();
2499         if (error)
2500                 goto out3;
2501
2502         error = register_filesystem(&tmpfs_fs_type);
2503         if (error) {
2504                 printk(KERN_ERR "Could not register tmpfs\n");
2505                 goto out2;
2506         }
2507
2508         shm_mnt = vfs_kern_mount(&tmpfs_fs_type, MS_NOUSER,
2509                                 tmpfs_fs_type.name, NULL);
2510         if (IS_ERR(shm_mnt)) {
2511                 error = PTR_ERR(shm_mnt);
2512                 printk(KERN_ERR "Could not kern_mount tmpfs\n");
2513                 goto out1;
2514         }
2515         return 0;
2516
2517 out1:
2518         unregister_filesystem(&tmpfs_fs_type);
2519 out2:
2520         destroy_inodecache();
2521 out3:
2522         bdi_destroy(&shmem_backing_dev_info);
2523 out4:
2524         shm_mnt = ERR_PTR(error);
2525         return error;
2526 }
2527 module_init(init_tmpfs)
2528
2529 /**
2530  * shmem_file_setup - get an unlinked file living in tmpfs
2531  * @name: name for dentry (to be seen in /proc/<pid>/maps
2532  * @size: size to be set for the file
2533  * @flags: vm_flags
2534  */
2535 struct file *shmem_file_setup(char *name, loff_t size, unsigned long flags)
2536 {
2537         int error;
2538         struct file *file;
2539         struct inode *inode;
2540         struct dentry *dentry, *root;
2541         struct qstr this;
2542
2543         if (IS_ERR(shm_mnt))
2544                 return (void *)shm_mnt;
2545
2546         if (size < 0 || size > SHMEM_MAX_BYTES)
2547                 return ERR_PTR(-EINVAL);
2548
2549         if (shmem_acct_size(flags, size))
2550                 return ERR_PTR(-ENOMEM);
2551
2552         error = -ENOMEM;
2553         this.name = name;
2554         this.len = strlen(name);
2555         this.hash = 0; /* will go */
2556         root = shm_mnt->mnt_root;
2557         dentry = d_alloc(root, &this);
2558         if (!dentry)
2559                 goto put_memory;
2560
2561         error = -ENFILE;
2562         file = get_empty_filp();
2563         if (!file)
2564                 goto put_dentry;
2565
2566         error = -ENOSPC;
2567         inode = shmem_get_inode(root->d_sb, S_IFREG | S_IRWXUGO, 0);
2568         if (!inode)
2569                 goto close_file;
2570
2571         SHMEM_I(inode)->flags = flags & VM_ACCOUNT;
2572         d_instantiate(dentry, inode);
2573         inode->i_size = size;
2574         inode->i_nlink = 0;     /* It is unlinked */
2575         init_file(file, shm_mnt, dentry, FMODE_WRITE | FMODE_READ,
2576                         &shmem_file_operations);
2577         return file;
2578
2579 close_file:
2580         put_filp(file);
2581 put_dentry:
2582         dput(dentry);
2583 put_memory:
2584         shmem_unacct_size(flags, size);
2585         return ERR_PTR(error);
2586 }
2587
2588 /**
2589  * shmem_zero_setup - setup a shared anonymous mapping
2590  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
2591  */
2592 int shmem_zero_setup(struct vm_area_struct *vma)
2593 {
2594         struct file *file;
2595         loff_t size = vma->vm_end - vma->vm_start;
2596
2597         file = shmem_file_setup("dev/zero", size, vma->vm_flags);
2598         if (IS_ERR(file))
2599                 return PTR_ERR(file);
2600
2601         if (vma->vm_file)
2602                 fput(vma->vm_file);
2603         vma->vm_file = file;
2604         vma->vm_ops = &shmem_vm_ops;
2605         return 0;
2606 }