5dfe79048f6d5e66f3379a2bfa66ebaf12e0074c
[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
53 #include <asm/uaccess.h>
54 #include <asm/div64.h>
55 #include <asm/pgtable.h>
56
57 /* This magic number is used in glibc for posix shared memory */
58 #define TMPFS_MAGIC     0x01021994
59
60 #define ENTRIES_PER_PAGE (PAGE_CACHE_SIZE/sizeof(unsigned long))
61 #define ENTRIES_PER_PAGEPAGE (ENTRIES_PER_PAGE*ENTRIES_PER_PAGE)
62 #define BLOCKS_PER_PAGE  (PAGE_CACHE_SIZE/512)
63
64 #define SHMEM_MAX_INDEX  (SHMEM_NR_DIRECT + (ENTRIES_PER_PAGEPAGE/2) * (ENTRIES_PER_PAGE+1))
65 #define SHMEM_MAX_BYTES  ((unsigned long long)SHMEM_MAX_INDEX << PAGE_CACHE_SHIFT)
66
67 #define VM_ACCT(size)    (PAGE_CACHE_ALIGN(size) >> PAGE_SHIFT)
68
69 /* info->flags needs VM_flags to handle pagein/truncate races efficiently */
70 #define SHMEM_PAGEIN     VM_READ
71 #define SHMEM_TRUNCATE   VM_WRITE
72
73 /* Definition to limit shmem_truncate's steps between cond_rescheds */
74 #define LATENCY_LIMIT    64
75
76 /* Pretend that each entry is of this size in directory's i_size */
77 #define BOGO_DIRENT_SIZE 20
78
79 /* Flag allocation requirements to shmem_getpage and shmem_swp_alloc */
80 enum sgp_type {
81         SGP_READ,       /* don't exceed i_size, don't allocate page */
82         SGP_CACHE,      /* don't exceed i_size, may allocate page */
83         SGP_WRITE,      /* may exceed i_size, may allocate page */
84 };
85
86 static int shmem_getpage(struct inode *inode, unsigned long idx,
87                          struct page **pagep, enum sgp_type sgp, int *type);
88
89 static inline struct page *shmem_dir_alloc(gfp_t gfp_mask)
90 {
91         /*
92          * The above definition of ENTRIES_PER_PAGE, and the use of
93          * BLOCKS_PER_PAGE on indirect pages, assume PAGE_CACHE_SIZE:
94          * might be reconsidered if it ever diverges from PAGE_SIZE.
95          *
96          * Mobility flags are masked out as swap vectors cannot move
97          */
98         return alloc_pages((gfp_mask & ~GFP_MOVABLE_MASK) | __GFP_ZERO,
99                                 PAGE_CACHE_SHIFT-PAGE_SHIFT);
100 }
101
102 static inline void shmem_dir_free(struct page *page)
103 {
104         __free_pages(page, PAGE_CACHE_SHIFT-PAGE_SHIFT);
105 }
106
107 static struct page **shmem_dir_map(struct page *page)
108 {
109         return (struct page **)kmap_atomic(page, KM_USER0);
110 }
111
112 static inline void shmem_dir_unmap(struct page **dir)
113 {
114         kunmap_atomic(dir, KM_USER0);
115 }
116
117 static swp_entry_t *shmem_swp_map(struct page *page)
118 {
119         return (swp_entry_t *)kmap_atomic(page, KM_USER1);
120 }
121
122 static inline void shmem_swp_balance_unmap(void)
123 {
124         /*
125          * When passing a pointer to an i_direct entry, to code which
126          * also handles indirect entries and so will shmem_swp_unmap,
127          * we must arrange for the preempt count to remain in balance.
128          * What kmap_atomic of a lowmem page does depends on config
129          * and architecture, so pretend to kmap_atomic some lowmem page.
130          */
131         (void) kmap_atomic(ZERO_PAGE(0), KM_USER1);
132 }
133
134 static inline void shmem_swp_unmap(swp_entry_t *entry)
135 {
136         kunmap_atomic(entry, KM_USER1);
137 }
138
139 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
140 {
141         return sb->s_fs_info;
142 }
143
144 /*
145  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
146  * for shared memory and for shared anonymous (/dev/zero) mappings
147  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
148  * consistent with the pre-accounting of private mappings ...
149  */
150 static inline int shmem_acct_size(unsigned long flags, loff_t size)
151 {
152         return (flags & VM_ACCOUNT)?
153                 security_vm_enough_memory(VM_ACCT(size)): 0;
154 }
155
156 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
157 {
158         if (flags & VM_ACCOUNT)
159                 vm_unacct_memory(VM_ACCT(size));
160 }
161
162 /*
163  * ... whereas tmpfs objects are accounted incrementally as
164  * pages are allocated, in order to allow huge sparse files.
165  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
166  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
167  */
168 static inline int shmem_acct_block(unsigned long flags)
169 {
170         return (flags & VM_ACCOUNT)?
171                 0: security_vm_enough_memory(VM_ACCT(PAGE_CACHE_SIZE));
172 }
173
174 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
175 {
176         if (!(flags & VM_ACCOUNT))
177                 vm_unacct_memory(pages * VM_ACCT(PAGE_CACHE_SIZE));
178 }
179
180 static const struct super_operations shmem_ops;
181 static const struct address_space_operations shmem_aops;
182 static const struct file_operations shmem_file_operations;
183 static const struct inode_operations shmem_inode_operations;
184 static const struct inode_operations shmem_dir_inode_operations;
185 static const struct inode_operations shmem_special_inode_operations;
186 static struct vm_operations_struct shmem_vm_ops;
187
188 static struct backing_dev_info shmem_backing_dev_info  __read_mostly = {
189         .ra_pages       = 0,    /* No readahead */
190         .capabilities   = BDI_CAP_NO_ACCT_DIRTY | BDI_CAP_NO_WRITEBACK,
191         .unplug_io_fn   = default_unplug_io_fn,
192 };
193
194 static LIST_HEAD(shmem_swaplist);
195 static DEFINE_SPINLOCK(shmem_swaplist_lock);
196
197 static void shmem_free_blocks(struct inode *inode, long pages)
198 {
199         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
200         if (sbinfo->max_blocks) {
201                 spin_lock(&sbinfo->stat_lock);
202                 sbinfo->free_blocks += pages;
203                 inode->i_blocks -= pages*BLOCKS_PER_PAGE;
204                 spin_unlock(&sbinfo->stat_lock);
205         }
206 }
207
208 /*
209  * shmem_recalc_inode - recalculate the size of an inode
210  *
211  * @inode: inode to recalc
212  *
213  * We have to calculate the free blocks since the mm can drop
214  * undirtied hole pages behind our back.
215  *
216  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
217  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
218  *
219  * It has to be called with the spinlock held.
220  */
221 static void shmem_recalc_inode(struct inode *inode)
222 {
223         struct shmem_inode_info *info = SHMEM_I(inode);
224         long freed;
225
226         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
227         if (freed > 0) {
228                 info->alloced -= freed;
229                 shmem_unacct_blocks(info->flags, freed);
230                 shmem_free_blocks(inode, freed);
231         }
232 }
233
234 /*
235  * shmem_swp_entry - find the swap vector position in the info structure
236  *
237  * @info:  info structure for the inode
238  * @index: index of the page to find
239  * @page:  optional page to add to the structure. Has to be preset to
240  *         all zeros
241  *
242  * If there is no space allocated yet it will return NULL when
243  * page is NULL, else it will use the page for the needed block,
244  * setting it to NULL on return to indicate that it has been used.
245  *
246  * The swap vector is organized the following way:
247  *
248  * There are SHMEM_NR_DIRECT entries directly stored in the
249  * shmem_inode_info structure. So small files do not need an addional
250  * allocation.
251  *
252  * For pages with index > SHMEM_NR_DIRECT there is the pointer
253  * i_indirect which points to a page which holds in the first half
254  * doubly indirect blocks, in the second half triple indirect blocks:
255  *
256  * For an artificial ENTRIES_PER_PAGE = 4 this would lead to the
257  * following layout (for SHMEM_NR_DIRECT == 16):
258  *
259  * i_indirect -> dir --> 16-19
260  *            |      +-> 20-23
261  *            |
262  *            +-->dir2 --> 24-27
263  *            |        +-> 28-31
264  *            |        +-> 32-35
265  *            |        +-> 36-39
266  *            |
267  *            +-->dir3 --> 40-43
268  *                     +-> 44-47
269  *                     +-> 48-51
270  *                     +-> 52-55
271  */
272 static swp_entry_t *shmem_swp_entry(struct shmem_inode_info *info, unsigned long index, struct page **page)
273 {
274         unsigned long offset;
275         struct page **dir;
276         struct page *subdir;
277
278         if (index < SHMEM_NR_DIRECT) {
279                 shmem_swp_balance_unmap();
280                 return info->i_direct+index;
281         }
282         if (!info->i_indirect) {
283                 if (page) {
284                         info->i_indirect = *page;
285                         *page = NULL;
286                 }
287                 return NULL;                    /* need another page */
288         }
289
290         index -= SHMEM_NR_DIRECT;
291         offset = index % ENTRIES_PER_PAGE;
292         index /= ENTRIES_PER_PAGE;
293         dir = shmem_dir_map(info->i_indirect);
294
295         if (index >= ENTRIES_PER_PAGE/2) {
296                 index -= ENTRIES_PER_PAGE/2;
297                 dir += ENTRIES_PER_PAGE/2 + index/ENTRIES_PER_PAGE;
298                 index %= ENTRIES_PER_PAGE;
299                 subdir = *dir;
300                 if (!subdir) {
301                         if (page) {
302                                 *dir = *page;
303                                 *page = NULL;
304                         }
305                         shmem_dir_unmap(dir);
306                         return NULL;            /* need another page */
307                 }
308                 shmem_dir_unmap(dir);
309                 dir = shmem_dir_map(subdir);
310         }
311
312         dir += index;
313         subdir = *dir;
314         if (!subdir) {
315                 if (!page || !(subdir = *page)) {
316                         shmem_dir_unmap(dir);
317                         return NULL;            /* need a page */
318                 }
319                 *dir = subdir;
320                 *page = NULL;
321         }
322         shmem_dir_unmap(dir);
323         return shmem_swp_map(subdir) + offset;
324 }
325
326 static void shmem_swp_set(struct shmem_inode_info *info, swp_entry_t *entry, unsigned long value)
327 {
328         long incdec = value? 1: -1;
329
330         entry->val = value;
331         info->swapped += incdec;
332         if ((unsigned long)(entry - info->i_direct) >= SHMEM_NR_DIRECT) {
333                 struct page *page = kmap_atomic_to_page(entry);
334                 set_page_private(page, page_private(page) + incdec);
335         }
336 }
337
338 /*
339  * shmem_swp_alloc - get the position of the swap entry for the page.
340  *                   If it does not exist allocate the entry.
341  *
342  * @info:       info structure for the inode
343  * @index:      index of the page to find
344  * @sgp:        check and recheck i_size? skip allocation?
345  */
346 static swp_entry_t *shmem_swp_alloc(struct shmem_inode_info *info, unsigned long index, enum sgp_type sgp)
347 {
348         struct inode *inode = &info->vfs_inode;
349         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
350         struct page *page = NULL;
351         swp_entry_t *entry;
352
353         if (sgp != SGP_WRITE &&
354             ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode))
355                 return ERR_PTR(-EINVAL);
356
357         while (!(entry = shmem_swp_entry(info, index, &page))) {
358                 if (sgp == SGP_READ)
359                         return shmem_swp_map(ZERO_PAGE(0));
360                 /*
361                  * Test free_blocks against 1 not 0, since we have 1 data
362                  * page (and perhaps indirect index pages) yet to allocate:
363                  * a waste to allocate index if we cannot allocate data.
364                  */
365                 if (sbinfo->max_blocks) {
366                         spin_lock(&sbinfo->stat_lock);
367                         if (sbinfo->free_blocks <= 1) {
368                                 spin_unlock(&sbinfo->stat_lock);
369                                 return ERR_PTR(-ENOSPC);
370                         }
371                         sbinfo->free_blocks--;
372                         inode->i_blocks += BLOCKS_PER_PAGE;
373                         spin_unlock(&sbinfo->stat_lock);
374                 }
375
376                 spin_unlock(&info->lock);
377                 page = shmem_dir_alloc(mapping_gfp_mask(inode->i_mapping));
378                 if (page)
379                         set_page_private(page, 0);
380                 spin_lock(&info->lock);
381
382                 if (!page) {
383                         shmem_free_blocks(inode, 1);
384                         return ERR_PTR(-ENOMEM);
385                 }
386                 if (sgp != SGP_WRITE &&
387                     ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
388                         entry = ERR_PTR(-EINVAL);
389                         break;
390                 }
391                 if (info->next_index <= index)
392                         info->next_index = index + 1;
393         }
394         if (page) {
395                 /* another task gave its page, or truncated the file */
396                 shmem_free_blocks(inode, 1);
397                 shmem_dir_free(page);
398         }
399         if (info->next_index <= index && !IS_ERR(entry))
400                 info->next_index = index + 1;
401         return entry;
402 }
403
404 /*
405  * shmem_free_swp - free some swap entries in a directory
406  *
407  * @dir:        pointer to the directory
408  * @edir:       pointer after last entry of the directory
409  * @punch_lock: pointer to spinlock when needed for the holepunch case
410  */
411 static int shmem_free_swp(swp_entry_t *dir, swp_entry_t *edir,
412                                                 spinlock_t *punch_lock)
413 {
414         spinlock_t *punch_unlock = NULL;
415         swp_entry_t *ptr;
416         int freed = 0;
417
418         for (ptr = dir; ptr < edir; ptr++) {
419                 if (ptr->val) {
420                         if (unlikely(punch_lock)) {
421                                 punch_unlock = punch_lock;
422                                 punch_lock = NULL;
423                                 spin_lock(punch_unlock);
424                                 if (!ptr->val)
425                                         continue;
426                         }
427                         free_swap_and_cache(*ptr);
428                         *ptr = (swp_entry_t){0};
429                         freed++;
430                 }
431         }
432         if (punch_unlock)
433                 spin_unlock(punch_unlock);
434         return freed;
435 }
436
437 static int shmem_map_and_free_swp(struct page *subdir, int offset,
438                 int limit, struct page ***dir, spinlock_t *punch_lock)
439 {
440         swp_entry_t *ptr;
441         int freed = 0;
442
443         ptr = shmem_swp_map(subdir);
444         for (; offset < limit; offset += LATENCY_LIMIT) {
445                 int size = limit - offset;
446                 if (size > LATENCY_LIMIT)
447                         size = LATENCY_LIMIT;
448                 freed += shmem_free_swp(ptr+offset, ptr+offset+size,
449                                                         punch_lock);
450                 if (need_resched()) {
451                         shmem_swp_unmap(ptr);
452                         if (*dir) {
453                                 shmem_dir_unmap(*dir);
454                                 *dir = NULL;
455                         }
456                         cond_resched();
457                         ptr = shmem_swp_map(subdir);
458                 }
459         }
460         shmem_swp_unmap(ptr);
461         return freed;
462 }
463
464 static void shmem_free_pages(struct list_head *next)
465 {
466         struct page *page;
467         int freed = 0;
468
469         do {
470                 page = container_of(next, struct page, lru);
471                 next = next->next;
472                 shmem_dir_free(page);
473                 freed++;
474                 if (freed >= LATENCY_LIMIT) {
475                         cond_resched();
476                         freed = 0;
477                 }
478         } while (next);
479 }
480
481 static void shmem_truncate_range(struct inode *inode, loff_t start, loff_t end)
482 {
483         struct shmem_inode_info *info = SHMEM_I(inode);
484         unsigned long idx;
485         unsigned long size;
486         unsigned long limit;
487         unsigned long stage;
488         unsigned long diroff;
489         struct page **dir;
490         struct page *topdir;
491         struct page *middir;
492         struct page *subdir;
493         swp_entry_t *ptr;
494         LIST_HEAD(pages_to_free);
495         long nr_pages_to_free = 0;
496         long nr_swaps_freed = 0;
497         int offset;
498         int freed;
499         int punch_hole;
500         spinlock_t *needs_lock;
501         spinlock_t *punch_lock;
502         unsigned long upper_limit;
503
504         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
505         idx = (start + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
506         if (idx >= info->next_index)
507                 return;
508
509         spin_lock(&info->lock);
510         info->flags |= SHMEM_TRUNCATE;
511         if (likely(end == (loff_t) -1)) {
512                 limit = info->next_index;
513                 upper_limit = SHMEM_MAX_INDEX;
514                 info->next_index = idx;
515                 needs_lock = NULL;
516                 punch_hole = 0;
517         } else {
518                 if (end + 1 >= inode->i_size) { /* we may free a little more */
519                         limit = (inode->i_size + PAGE_CACHE_SIZE - 1) >>
520                                                         PAGE_CACHE_SHIFT;
521                         upper_limit = SHMEM_MAX_INDEX;
522                 } else {
523                         limit = (end + 1) >> PAGE_CACHE_SHIFT;
524                         upper_limit = limit;
525                 }
526                 needs_lock = &info->lock;
527                 punch_hole = 1;
528         }
529
530         topdir = info->i_indirect;
531         if (topdir && idx <= SHMEM_NR_DIRECT && !punch_hole) {
532                 info->i_indirect = NULL;
533                 nr_pages_to_free++;
534                 list_add(&topdir->lru, &pages_to_free);
535         }
536         spin_unlock(&info->lock);
537
538         if (info->swapped && idx < SHMEM_NR_DIRECT) {
539                 ptr = info->i_direct;
540                 size = limit;
541                 if (size > SHMEM_NR_DIRECT)
542                         size = SHMEM_NR_DIRECT;
543                 nr_swaps_freed = shmem_free_swp(ptr+idx, ptr+size, needs_lock);
544         }
545
546         /*
547          * If there are no indirect blocks or we are punching a hole
548          * below indirect blocks, nothing to be done.
549          */
550         if (!topdir || limit <= SHMEM_NR_DIRECT)
551                 goto done2;
552
553         /*
554          * The truncation case has already dropped info->lock, and we're safe
555          * because i_size and next_index have already been lowered, preventing
556          * access beyond.  But in the punch_hole case, we still need to take
557          * the lock when updating the swap directory, because there might be
558          * racing accesses by shmem_getpage(SGP_CACHE), shmem_unuse_inode or
559          * shmem_writepage.  However, whenever we find we can remove a whole
560          * directory page (not at the misaligned start or end of the range),
561          * we first NULLify its pointer in the level above, and then have no
562          * need to take the lock when updating its contents: needs_lock and
563          * punch_lock (either pointing to info->lock or NULL) manage this.
564          */
565
566         upper_limit -= SHMEM_NR_DIRECT;
567         limit -= SHMEM_NR_DIRECT;
568         idx = (idx > SHMEM_NR_DIRECT)? (idx - SHMEM_NR_DIRECT): 0;
569         offset = idx % ENTRIES_PER_PAGE;
570         idx -= offset;
571
572         dir = shmem_dir_map(topdir);
573         stage = ENTRIES_PER_PAGEPAGE/2;
574         if (idx < ENTRIES_PER_PAGEPAGE/2) {
575                 middir = topdir;
576                 diroff = idx/ENTRIES_PER_PAGE;
577         } else {
578                 dir += ENTRIES_PER_PAGE/2;
579                 dir += (idx - ENTRIES_PER_PAGEPAGE/2)/ENTRIES_PER_PAGEPAGE;
580                 while (stage <= idx)
581                         stage += ENTRIES_PER_PAGEPAGE;
582                 middir = *dir;
583                 if (*dir) {
584                         diroff = ((idx - ENTRIES_PER_PAGEPAGE/2) %
585                                 ENTRIES_PER_PAGEPAGE) / ENTRIES_PER_PAGE;
586                         if (!diroff && !offset && upper_limit >= stage) {
587                                 if (needs_lock) {
588                                         spin_lock(needs_lock);
589                                         *dir = NULL;
590                                         spin_unlock(needs_lock);
591                                         needs_lock = NULL;
592                                 } else
593                                         *dir = NULL;
594                                 nr_pages_to_free++;
595                                 list_add(&middir->lru, &pages_to_free);
596                         }
597                         shmem_dir_unmap(dir);
598                         dir = shmem_dir_map(middir);
599                 } else {
600                         diroff = 0;
601                         offset = 0;
602                         idx = stage;
603                 }
604         }
605
606         for (; idx < limit; idx += ENTRIES_PER_PAGE, diroff++) {
607                 if (unlikely(idx == stage)) {
608                         shmem_dir_unmap(dir);
609                         dir = shmem_dir_map(topdir) +
610                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
611                         while (!*dir) {
612                                 dir++;
613                                 idx += ENTRIES_PER_PAGEPAGE;
614                                 if (idx >= limit)
615                                         goto done1;
616                         }
617                         stage = idx + ENTRIES_PER_PAGEPAGE;
618                         middir = *dir;
619                         if (punch_hole)
620                                 needs_lock = &info->lock;
621                         if (upper_limit >= stage) {
622                                 if (needs_lock) {
623                                         spin_lock(needs_lock);
624                                         *dir = NULL;
625                                         spin_unlock(needs_lock);
626                                         needs_lock = NULL;
627                                 } else
628                                         *dir = NULL;
629                                 nr_pages_to_free++;
630                                 list_add(&middir->lru, &pages_to_free);
631                         }
632                         shmem_dir_unmap(dir);
633                         cond_resched();
634                         dir = shmem_dir_map(middir);
635                         diroff = 0;
636                 }
637                 punch_lock = needs_lock;
638                 subdir = dir[diroff];
639                 if (subdir && !offset && upper_limit-idx >= ENTRIES_PER_PAGE) {
640                         if (needs_lock) {
641                                 spin_lock(needs_lock);
642                                 dir[diroff] = NULL;
643                                 spin_unlock(needs_lock);
644                                 punch_lock = NULL;
645                         } else
646                                 dir[diroff] = NULL;
647                         nr_pages_to_free++;
648                         list_add(&subdir->lru, &pages_to_free);
649                 }
650                 if (subdir && page_private(subdir) /* has swap entries */) {
651                         size = limit - idx;
652                         if (size > ENTRIES_PER_PAGE)
653                                 size = ENTRIES_PER_PAGE;
654                         freed = shmem_map_and_free_swp(subdir,
655                                         offset, size, &dir, punch_lock);
656                         if (!dir)
657                                 dir = shmem_dir_map(middir);
658                         nr_swaps_freed += freed;
659                         if (offset || punch_lock) {
660                                 spin_lock(&info->lock);
661                                 set_page_private(subdir,
662                                         page_private(subdir) - freed);
663                                 spin_unlock(&info->lock);
664                         } else
665                                 BUG_ON(page_private(subdir) != freed);
666                 }
667                 offset = 0;
668         }
669 done1:
670         shmem_dir_unmap(dir);
671 done2:
672         if (inode->i_mapping->nrpages && (info->flags & SHMEM_PAGEIN)) {
673                 /*
674                  * Call truncate_inode_pages again: racing shmem_unuse_inode
675                  * may have swizzled a page in from swap since vmtruncate or
676                  * generic_delete_inode did it, before we lowered next_index.
677                  * Also, though shmem_getpage checks i_size before adding to
678                  * cache, no recheck after: so fix the narrow window there too.
679                  *
680                  * Recalling truncate_inode_pages_range and unmap_mapping_range
681                  * every time for punch_hole (which never got a chance to clear
682                  * SHMEM_PAGEIN at the start of vmtruncate_range) is expensive,
683                  * yet hardly ever necessary: try to optimize them out later.
684                  */
685                 truncate_inode_pages_range(inode->i_mapping, start, end);
686                 if (punch_hole)
687                         unmap_mapping_range(inode->i_mapping, start,
688                                                         end - start, 1);
689         }
690
691         spin_lock(&info->lock);
692         info->flags &= ~SHMEM_TRUNCATE;
693         info->swapped -= nr_swaps_freed;
694         if (nr_pages_to_free)
695                 shmem_free_blocks(inode, nr_pages_to_free);
696         shmem_recalc_inode(inode);
697         spin_unlock(&info->lock);
698
699         /*
700          * Empty swap vector directory pages to be freed?
701          */
702         if (!list_empty(&pages_to_free)) {
703                 pages_to_free.prev->next = NULL;
704                 shmem_free_pages(pages_to_free.next);
705         }
706 }
707
708 static void shmem_truncate(struct inode *inode)
709 {
710         shmem_truncate_range(inode, inode->i_size, (loff_t)-1);
711 }
712
713 static int shmem_notify_change(struct dentry *dentry, struct iattr *attr)
714 {
715         struct inode *inode = dentry->d_inode;
716         struct page *page = NULL;
717         int error;
718
719         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
720                 if (attr->ia_size < inode->i_size) {
721                         /*
722                          * If truncating down to a partial page, then
723                          * if that page is already allocated, hold it
724                          * in memory until the truncation is over, so
725                          * truncate_partial_page cannnot miss it were
726                          * it assigned to swap.
727                          */
728                         if (attr->ia_size & (PAGE_CACHE_SIZE-1)) {
729                                 (void) shmem_getpage(inode,
730                                         attr->ia_size>>PAGE_CACHE_SHIFT,
731                                                 &page, SGP_READ, NULL);
732                                 if (page)
733                                         unlock_page(page);
734                         }
735                         /*
736                          * Reset SHMEM_PAGEIN flag so that shmem_truncate can
737                          * detect if any pages might have been added to cache
738                          * after truncate_inode_pages.  But we needn't bother
739                          * if it's being fully truncated to zero-length: the
740                          * nrpages check is efficient enough in that case.
741                          */
742                         if (attr->ia_size) {
743                                 struct shmem_inode_info *info = SHMEM_I(inode);
744                                 spin_lock(&info->lock);
745                                 info->flags &= ~SHMEM_PAGEIN;
746                                 spin_unlock(&info->lock);
747                         }
748                 }
749         }
750
751         error = inode_change_ok(inode, attr);
752         if (!error)
753                 error = inode_setattr(inode, attr);
754 #ifdef CONFIG_TMPFS_POSIX_ACL
755         if (!error && (attr->ia_valid & ATTR_MODE))
756                 error = generic_acl_chmod(inode, &shmem_acl_ops);
757 #endif
758         if (page)
759                 page_cache_release(page);
760         return error;
761 }
762
763 static void shmem_delete_inode(struct inode *inode)
764 {
765         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
766         struct shmem_inode_info *info = SHMEM_I(inode);
767
768         if (inode->i_op->truncate == shmem_truncate) {
769                 truncate_inode_pages(inode->i_mapping, 0);
770                 shmem_unacct_size(info->flags, inode->i_size);
771                 inode->i_size = 0;
772                 shmem_truncate(inode);
773                 if (!list_empty(&info->swaplist)) {
774                         spin_lock(&shmem_swaplist_lock);
775                         list_del_init(&info->swaplist);
776                         spin_unlock(&shmem_swaplist_lock);
777                 }
778         }
779         BUG_ON(inode->i_blocks);
780         if (sbinfo->max_inodes) {
781                 spin_lock(&sbinfo->stat_lock);
782                 sbinfo->free_inodes++;
783                 spin_unlock(&sbinfo->stat_lock);
784         }
785         clear_inode(inode);
786 }
787
788 static inline int shmem_find_swp(swp_entry_t entry, swp_entry_t *dir, swp_entry_t *edir)
789 {
790         swp_entry_t *ptr;
791
792         for (ptr = dir; ptr < edir; ptr++) {
793                 if (ptr->val == entry.val)
794                         return ptr - dir;
795         }
796         return -1;
797 }
798
799 static int shmem_unuse_inode(struct shmem_inode_info *info, swp_entry_t entry, struct page *page)
800 {
801         struct inode *inode;
802         unsigned long idx;
803         unsigned long size;
804         unsigned long limit;
805         unsigned long stage;
806         struct page **dir;
807         struct page *subdir;
808         swp_entry_t *ptr;
809         int offset;
810
811         idx = 0;
812         ptr = info->i_direct;
813         spin_lock(&info->lock);
814         limit = info->next_index;
815         size = limit;
816         if (size > SHMEM_NR_DIRECT)
817                 size = SHMEM_NR_DIRECT;
818         offset = shmem_find_swp(entry, ptr, ptr+size);
819         if (offset >= 0) {
820                 shmem_swp_balance_unmap();
821                 goto found;
822         }
823         if (!info->i_indirect)
824                 goto lost2;
825
826         dir = shmem_dir_map(info->i_indirect);
827         stage = SHMEM_NR_DIRECT + ENTRIES_PER_PAGEPAGE/2;
828
829         for (idx = SHMEM_NR_DIRECT; idx < limit; idx += ENTRIES_PER_PAGE, dir++) {
830                 if (unlikely(idx == stage)) {
831                         shmem_dir_unmap(dir-1);
832                         dir = shmem_dir_map(info->i_indirect) +
833                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
834                         while (!*dir) {
835                                 dir++;
836                                 idx += ENTRIES_PER_PAGEPAGE;
837                                 if (idx >= limit)
838                                         goto lost1;
839                         }
840                         stage = idx + ENTRIES_PER_PAGEPAGE;
841                         subdir = *dir;
842                         shmem_dir_unmap(dir);
843                         dir = shmem_dir_map(subdir);
844                 }
845                 subdir = *dir;
846                 if (subdir && page_private(subdir)) {
847                         ptr = shmem_swp_map(subdir);
848                         size = limit - idx;
849                         if (size > ENTRIES_PER_PAGE)
850                                 size = ENTRIES_PER_PAGE;
851                         offset = shmem_find_swp(entry, ptr, ptr+size);
852                         if (offset >= 0) {
853                                 shmem_dir_unmap(dir);
854                                 goto found;
855                         }
856                         shmem_swp_unmap(ptr);
857                 }
858         }
859 lost1:
860         shmem_dir_unmap(dir-1);
861 lost2:
862         spin_unlock(&info->lock);
863         return 0;
864 found:
865         idx += offset;
866         inode = &info->vfs_inode;
867         if (move_from_swap_cache(page, idx, inode->i_mapping) == 0) {
868                 info->flags |= SHMEM_PAGEIN;
869                 shmem_swp_set(info, ptr + offset, 0);
870         }
871         shmem_swp_unmap(ptr);
872         spin_unlock(&info->lock);
873         /*
874          * Decrement swap count even when the entry is left behind:
875          * try_to_unuse will skip over mms, then reincrement count.
876          */
877         swap_free(entry);
878         return 1;
879 }
880
881 /*
882  * shmem_unuse() search for an eventually swapped out shmem page.
883  */
884 int shmem_unuse(swp_entry_t entry, struct page *page)
885 {
886         struct list_head *p, *next;
887         struct shmem_inode_info *info;
888         int found = 0;
889
890         spin_lock(&shmem_swaplist_lock);
891         list_for_each_safe(p, next, &shmem_swaplist) {
892                 info = list_entry(p, struct shmem_inode_info, swaplist);
893                 if (!info->swapped)
894                         list_del_init(&info->swaplist);
895                 else if (shmem_unuse_inode(info, entry, page)) {
896                         /* move head to start search for next from here */
897                         list_move_tail(&shmem_swaplist, &info->swaplist);
898                         found = 1;
899                         break;
900                 }
901         }
902         spin_unlock(&shmem_swaplist_lock);
903         return found;
904 }
905
906 /*
907  * Move the page from the page cache to the swap cache.
908  */
909 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
910 {
911         struct shmem_inode_info *info;
912         swp_entry_t *entry, swap;
913         struct address_space *mapping;
914         unsigned long index;
915         struct inode *inode;
916
917         BUG_ON(!PageLocked(page));
918         /*
919          * shmem_backing_dev_info's capabilities prevent regular writeback or
920          * sync from ever calling shmem_writepage; but a stacking filesystem
921          * may use the ->writepage of its underlying filesystem, in which case
922          * we want to do nothing when that underlying filesystem is tmpfs
923          * (writing out to swap is useful as a response to memory pressure, but
924          * of no use to stabilize the data) - just redirty the page, unlock it
925          * and claim success in this case.  AOP_WRITEPAGE_ACTIVATE, and the
926          * page_mapped check below, must be avoided unless we're in reclaim.
927          */
928         if (!wbc->for_reclaim) {
929                 set_page_dirty(page);
930                 unlock_page(page);
931                 return 0;
932         }
933         BUG_ON(page_mapped(page));
934
935         mapping = page->mapping;
936         index = page->index;
937         inode = mapping->host;
938         info = SHMEM_I(inode);
939         if (info->flags & VM_LOCKED)
940                 goto redirty;
941         swap = get_swap_page();
942         if (!swap.val)
943                 goto redirty;
944
945         spin_lock(&info->lock);
946         shmem_recalc_inode(inode);
947         if (index >= info->next_index) {
948                 BUG_ON(!(info->flags & SHMEM_TRUNCATE));
949                 goto unlock;
950         }
951         entry = shmem_swp_entry(info, index, NULL);
952         BUG_ON(!entry);
953         BUG_ON(entry->val);
954
955         if (move_to_swap_cache(page, swap) == 0) {
956                 shmem_swp_set(info, entry, swap.val);
957                 shmem_swp_unmap(entry);
958                 spin_unlock(&info->lock);
959                 if (list_empty(&info->swaplist)) {
960                         spin_lock(&shmem_swaplist_lock);
961                         /* move instead of add in case we're racing */
962                         list_move_tail(&info->swaplist, &shmem_swaplist);
963                         spin_unlock(&shmem_swaplist_lock);
964                 }
965                 unlock_page(page);
966                 return 0;
967         }
968
969         shmem_swp_unmap(entry);
970 unlock:
971         spin_unlock(&info->lock);
972         swap_free(swap);
973 redirty:
974         set_page_dirty(page);
975         return AOP_WRITEPAGE_ACTIVATE;  /* Return with the page locked */
976 }
977
978 #ifdef CONFIG_NUMA
979 static inline int shmem_parse_mpol(char *value, int *policy, nodemask_t *policy_nodes)
980 {
981         char *nodelist = strchr(value, ':');
982         int err = 1;
983
984         if (nodelist) {
985                 /* NUL-terminate policy string */
986                 *nodelist++ = '\0';
987                 if (nodelist_parse(nodelist, *policy_nodes))
988                         goto out;
989                 if (!nodes_subset(*policy_nodes, node_states[N_HIGH_MEMORY]))
990                         goto out;
991         }
992         if (!strcmp(value, "default")) {
993                 *policy = MPOL_DEFAULT;
994                 /* Don't allow a nodelist */
995                 if (!nodelist)
996                         err = 0;
997         } else if (!strcmp(value, "prefer")) {
998                 *policy = MPOL_PREFERRED;
999                 /* Insist on a nodelist of one node only */
1000                 if (nodelist) {
1001                         char *rest = nodelist;
1002                         while (isdigit(*rest))
1003                                 rest++;
1004                         if (!*rest)
1005                                 err = 0;
1006                 }
1007         } else if (!strcmp(value, "bind")) {
1008                 *policy = MPOL_BIND;
1009                 /* Insist on a nodelist */
1010                 if (nodelist)
1011                         err = 0;
1012         } else if (!strcmp(value, "interleave")) {
1013                 *policy = MPOL_INTERLEAVE;
1014                 /*
1015                  * Default to online nodes with memory if no nodelist
1016                  */
1017                 if (!nodelist)
1018                         *policy_nodes = node_states[N_HIGH_MEMORY];
1019                 err = 0;
1020         }
1021 out:
1022         /* Restore string for error message */
1023         if (nodelist)
1024                 *--nodelist = ':';
1025         return err;
1026 }
1027
1028 static struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1029                         struct shmem_inode_info *info, unsigned long idx)
1030 {
1031         struct vm_area_struct pvma;
1032         struct page *page;
1033
1034         /* Create a pseudo vma that just contains the policy */
1035         pvma.vm_start = 0;
1036         pvma.vm_pgoff = idx;
1037         pvma.vm_ops = NULL;
1038         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, idx);
1039         page = swapin_readahead(entry, gfp, &pvma, 0);
1040         mpol_free(pvma.vm_policy);
1041         return page;
1042 }
1043
1044 static struct page *shmem_alloc_page(gfp_t gfp,
1045                         struct shmem_inode_info *info, unsigned long idx)
1046 {
1047         struct vm_area_struct pvma;
1048         struct page *page;
1049
1050         /* Create a pseudo vma that just contains the policy */
1051         pvma.vm_start = 0;
1052         pvma.vm_pgoff = idx;
1053         pvma.vm_ops = NULL;
1054         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, idx);
1055         page = alloc_page_vma(gfp, &pvma, 0);
1056         mpol_free(pvma.vm_policy);
1057         return page;
1058 }
1059 #else
1060 static inline int shmem_parse_mpol(char *value, int *policy,
1061                                                 nodemask_t *policy_nodes)
1062 {
1063         return 1;
1064 }
1065
1066 static inline struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1067                         struct shmem_inode_info *info, unsigned long idx)
1068 {
1069         return swapin_readahead(entry, gfp, NULL, 0);
1070 }
1071
1072 static inline struct page *shmem_alloc_page(gfp_t gfp,
1073                         struct shmem_inode_info *info, unsigned long idx)
1074 {
1075         return alloc_page(gfp);
1076 }
1077 #endif
1078
1079 /*
1080  * shmem_getpage - either get the page from swap or allocate a new one
1081  *
1082  * If we allocate a new one we do not mark it dirty. That's up to the
1083  * vm. If we swap it in we mark it dirty since we also free the swap
1084  * entry since a page cannot live in both the swap and page cache
1085  */
1086 static int shmem_getpage(struct inode *inode, unsigned long idx,
1087                         struct page **pagep, enum sgp_type sgp, int *type)
1088 {
1089         struct address_space *mapping = inode->i_mapping;
1090         struct shmem_inode_info *info = SHMEM_I(inode);
1091         struct shmem_sb_info *sbinfo;
1092         struct page *filepage = *pagep;
1093         struct page *swappage;
1094         swp_entry_t *entry;
1095         swp_entry_t swap;
1096         gfp_t gfp;
1097         int error;
1098
1099         if (idx >= SHMEM_MAX_INDEX)
1100                 return -EFBIG;
1101
1102         if (type)
1103                 *type = 0;
1104
1105         /*
1106          * Normally, filepage is NULL on entry, and either found
1107          * uptodate immediately, or allocated and zeroed, or read
1108          * in under swappage, which is then assigned to filepage.
1109          * But shmem_readpage (required for splice) passes in a locked
1110          * filepage, which may be found not uptodate by other callers
1111          * too, and may need to be copied from the swappage read in.
1112          */
1113 repeat:
1114         if (!filepage)
1115                 filepage = find_lock_page(mapping, idx);
1116         if (filepage && PageUptodate(filepage))
1117                 goto done;
1118         error = 0;
1119         gfp = mapping_gfp_mask(mapping);
1120
1121         spin_lock(&info->lock);
1122         shmem_recalc_inode(inode);
1123         entry = shmem_swp_alloc(info, idx, sgp);
1124         if (IS_ERR(entry)) {
1125                 spin_unlock(&info->lock);
1126                 error = PTR_ERR(entry);
1127                 goto failed;
1128         }
1129         swap = *entry;
1130
1131         if (swap.val) {
1132                 /* Look it up and read it in.. */
1133                 swappage = lookup_swap_cache(swap);
1134                 if (!swappage) {
1135                         shmem_swp_unmap(entry);
1136                         /* here we actually do the io */
1137                         if (type && !(*type & VM_FAULT_MAJOR)) {
1138                                 __count_vm_event(PGMAJFAULT);
1139                                 *type |= VM_FAULT_MAJOR;
1140                         }
1141                         spin_unlock(&info->lock);
1142                         swappage = shmem_swapin(swap, gfp, info, idx);
1143                         if (!swappage) {
1144                                 spin_lock(&info->lock);
1145                                 entry = shmem_swp_alloc(info, idx, sgp);
1146                                 if (IS_ERR(entry))
1147                                         error = PTR_ERR(entry);
1148                                 else {
1149                                         if (entry->val == swap.val)
1150                                                 error = -ENOMEM;
1151                                         shmem_swp_unmap(entry);
1152                                 }
1153                                 spin_unlock(&info->lock);
1154                                 if (error)
1155                                         goto failed;
1156                                 goto repeat;
1157                         }
1158                         wait_on_page_locked(swappage);
1159                         page_cache_release(swappage);
1160                         goto repeat;
1161                 }
1162
1163                 /* We have to do this with page locked to prevent races */
1164                 if (TestSetPageLocked(swappage)) {
1165                         shmem_swp_unmap(entry);
1166                         spin_unlock(&info->lock);
1167                         wait_on_page_locked(swappage);
1168                         page_cache_release(swappage);
1169                         goto repeat;
1170                 }
1171                 if (PageWriteback(swappage)) {
1172                         shmem_swp_unmap(entry);
1173                         spin_unlock(&info->lock);
1174                         wait_on_page_writeback(swappage);
1175                         unlock_page(swappage);
1176                         page_cache_release(swappage);
1177                         goto repeat;
1178                 }
1179                 if (!PageUptodate(swappage)) {
1180                         shmem_swp_unmap(entry);
1181                         spin_unlock(&info->lock);
1182                         unlock_page(swappage);
1183                         page_cache_release(swappage);
1184                         error = -EIO;
1185                         goto failed;
1186                 }
1187
1188                 if (filepage) {
1189                         shmem_swp_set(info, entry, 0);
1190                         shmem_swp_unmap(entry);
1191                         delete_from_swap_cache(swappage);
1192                         spin_unlock(&info->lock);
1193                         copy_highpage(filepage, swappage);
1194                         unlock_page(swappage);
1195                         page_cache_release(swappage);
1196                         flush_dcache_page(filepage);
1197                         SetPageUptodate(filepage);
1198                         set_page_dirty(filepage);
1199                         swap_free(swap);
1200                 } else if (!(error = move_from_swap_cache(
1201                                 swappage, idx, mapping))) {
1202                         info->flags |= SHMEM_PAGEIN;
1203                         shmem_swp_set(info, entry, 0);
1204                         shmem_swp_unmap(entry);
1205                         spin_unlock(&info->lock);
1206                         filepage = swappage;
1207                         swap_free(swap);
1208                 } else {
1209                         shmem_swp_unmap(entry);
1210                         spin_unlock(&info->lock);
1211                         unlock_page(swappage);
1212                         page_cache_release(swappage);
1213                         if (error == -ENOMEM) {
1214                                 /* let kswapd refresh zone for GFP_ATOMICs */
1215                                 congestion_wait(WRITE, HZ/50);
1216                         }
1217                         goto repeat;
1218                 }
1219         } else if (sgp == SGP_READ && !filepage) {
1220                 shmem_swp_unmap(entry);
1221                 filepage = find_get_page(mapping, idx);
1222                 if (filepage &&
1223                     (!PageUptodate(filepage) || TestSetPageLocked(filepage))) {
1224                         spin_unlock(&info->lock);
1225                         wait_on_page_locked(filepage);
1226                         page_cache_release(filepage);
1227                         filepage = NULL;
1228                         goto repeat;
1229                 }
1230                 spin_unlock(&info->lock);
1231         } else {
1232                 shmem_swp_unmap(entry);
1233                 sbinfo = SHMEM_SB(inode->i_sb);
1234                 if (sbinfo->max_blocks) {
1235                         spin_lock(&sbinfo->stat_lock);
1236                         if (sbinfo->free_blocks == 0 ||
1237                             shmem_acct_block(info->flags)) {
1238                                 spin_unlock(&sbinfo->stat_lock);
1239                                 spin_unlock(&info->lock);
1240                                 error = -ENOSPC;
1241                                 goto failed;
1242                         }
1243                         sbinfo->free_blocks--;
1244                         inode->i_blocks += BLOCKS_PER_PAGE;
1245                         spin_unlock(&sbinfo->stat_lock);
1246                 } else if (shmem_acct_block(info->flags)) {
1247                         spin_unlock(&info->lock);
1248                         error = -ENOSPC;
1249                         goto failed;
1250                 }
1251
1252                 if (!filepage) {
1253                         spin_unlock(&info->lock);
1254                         filepage = shmem_alloc_page(gfp, info, idx);
1255                         if (!filepage) {
1256                                 shmem_unacct_blocks(info->flags, 1);
1257                                 shmem_free_blocks(inode, 1);
1258                                 error = -ENOMEM;
1259                                 goto failed;
1260                         }
1261
1262                         spin_lock(&info->lock);
1263                         entry = shmem_swp_alloc(info, idx, sgp);
1264                         if (IS_ERR(entry))
1265                                 error = PTR_ERR(entry);
1266                         else {
1267                                 swap = *entry;
1268                                 shmem_swp_unmap(entry);
1269                         }
1270                         if (error || swap.val || 0 != add_to_page_cache_lru(
1271                                         filepage, mapping, idx, GFP_ATOMIC)) {
1272                                 spin_unlock(&info->lock);
1273                                 page_cache_release(filepage);
1274                                 shmem_unacct_blocks(info->flags, 1);
1275                                 shmem_free_blocks(inode, 1);
1276                                 filepage = NULL;
1277                                 if (error)
1278                                         goto failed;
1279                                 goto repeat;
1280                         }
1281                         info->flags |= SHMEM_PAGEIN;
1282                 }
1283
1284                 info->alloced++;
1285                 spin_unlock(&info->lock);
1286                 clear_highpage(filepage);
1287                 flush_dcache_page(filepage);
1288                 SetPageUptodate(filepage);
1289         }
1290 done:
1291         *pagep = filepage;
1292         return 0;
1293
1294 failed:
1295         if (*pagep != filepage) {
1296                 unlock_page(filepage);
1297                 page_cache_release(filepage);
1298         }
1299         return error;
1300 }
1301
1302 static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1303 {
1304         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1305         int error;
1306         int ret;
1307
1308         if (((loff_t)vmf->pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
1309                 return VM_FAULT_SIGBUS;
1310
1311         error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_CACHE, &ret);
1312         if (error)
1313                 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1314
1315         mark_page_accessed(vmf->page);
1316         return ret | VM_FAULT_LOCKED;
1317 }
1318
1319 #ifdef CONFIG_NUMA
1320 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *new)
1321 {
1322         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1323         return mpol_set_shared_policy(&SHMEM_I(i)->policy, vma, new);
1324 }
1325
1326 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
1327                                           unsigned long addr)
1328 {
1329         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1330         unsigned long idx;
1331
1332         idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1333         return mpol_shared_policy_lookup(&SHMEM_I(i)->policy, idx);
1334 }
1335 #endif
1336
1337 int shmem_lock(struct file *file, int lock, struct user_struct *user)
1338 {
1339         struct inode *inode = file->f_path.dentry->d_inode;
1340         struct shmem_inode_info *info = SHMEM_I(inode);
1341         int retval = -ENOMEM;
1342
1343         spin_lock(&info->lock);
1344         if (lock && !(info->flags & VM_LOCKED)) {
1345                 if (!user_shm_lock(inode->i_size, user))
1346                         goto out_nomem;
1347                 info->flags |= VM_LOCKED;
1348         }
1349         if (!lock && (info->flags & VM_LOCKED) && user) {
1350                 user_shm_unlock(inode->i_size, user);
1351                 info->flags &= ~VM_LOCKED;
1352         }
1353         retval = 0;
1354 out_nomem:
1355         spin_unlock(&info->lock);
1356         return retval;
1357 }
1358
1359 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1360 {
1361         file_accessed(file);
1362         vma->vm_ops = &shmem_vm_ops;
1363         vma->vm_flags |= VM_CAN_NONLINEAR;
1364         return 0;
1365 }
1366
1367 static struct inode *
1368 shmem_get_inode(struct super_block *sb, int mode, dev_t dev)
1369 {
1370         struct inode *inode;
1371         struct shmem_inode_info *info;
1372         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1373
1374         if (sbinfo->max_inodes) {
1375                 spin_lock(&sbinfo->stat_lock);
1376                 if (!sbinfo->free_inodes) {
1377                         spin_unlock(&sbinfo->stat_lock);
1378                         return NULL;
1379                 }
1380                 sbinfo->free_inodes--;
1381                 spin_unlock(&sbinfo->stat_lock);
1382         }
1383
1384         inode = new_inode(sb);
1385         if (inode) {
1386                 inode->i_mode = mode;
1387                 inode->i_uid = current->fsuid;
1388                 inode->i_gid = current->fsgid;
1389                 inode->i_blocks = 0;
1390                 inode->i_mapping->a_ops = &shmem_aops;
1391                 inode->i_mapping->backing_dev_info = &shmem_backing_dev_info;
1392                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1393                 inode->i_generation = get_seconds();
1394                 info = SHMEM_I(inode);
1395                 memset(info, 0, (char *)inode - (char *)info);
1396                 spin_lock_init(&info->lock);
1397                 INIT_LIST_HEAD(&info->swaplist);
1398
1399                 switch (mode & S_IFMT) {
1400                 default:
1401                         inode->i_op = &shmem_special_inode_operations;
1402                         init_special_inode(inode, mode, dev);
1403                         break;
1404                 case S_IFREG:
1405                         inode->i_op = &shmem_inode_operations;
1406                         inode->i_fop = &shmem_file_operations;
1407                         mpol_shared_policy_init(&info->policy, sbinfo->policy,
1408                                                         &sbinfo->policy_nodes);
1409                         break;
1410                 case S_IFDIR:
1411                         inc_nlink(inode);
1412                         /* Some things misbehave if size == 0 on a directory */
1413                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
1414                         inode->i_op = &shmem_dir_inode_operations;
1415                         inode->i_fop = &simple_dir_operations;
1416                         break;
1417                 case S_IFLNK:
1418                         /*
1419                          * Must not load anything in the rbtree,
1420                          * mpol_free_shared_policy will not be called.
1421                          */
1422                         mpol_shared_policy_init(&info->policy, MPOL_DEFAULT,
1423                                                 NULL);
1424                         break;
1425                 }
1426         } else if (sbinfo->max_inodes) {
1427                 spin_lock(&sbinfo->stat_lock);
1428                 sbinfo->free_inodes++;
1429                 spin_unlock(&sbinfo->stat_lock);
1430         }
1431         return inode;
1432 }
1433
1434 #ifdef CONFIG_TMPFS
1435 static const struct inode_operations shmem_symlink_inode_operations;
1436 static const struct inode_operations shmem_symlink_inline_operations;
1437
1438 /*
1439  * Normally tmpfs avoids the use of shmem_readpage and shmem_write_begin;
1440  * but providing them allows a tmpfs file to be used for splice, sendfile, and
1441  * below the loop driver, in the generic fashion that many filesystems support.
1442  */
1443 static int shmem_readpage(struct file *file, struct page *page)
1444 {
1445         struct inode *inode = page->mapping->host;
1446         int error = shmem_getpage(inode, page->index, &page, SGP_CACHE, NULL);
1447         unlock_page(page);
1448         return error;
1449 }
1450
1451 static int
1452 shmem_write_begin(struct file *file, struct address_space *mapping,
1453                         loff_t pos, unsigned len, unsigned flags,
1454                         struct page **pagep, void **fsdata)
1455 {
1456         struct inode *inode = mapping->host;
1457         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1458         *pagep = NULL;
1459         return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1460 }
1461
1462 static int
1463 shmem_write_end(struct file *file, struct address_space *mapping,
1464                         loff_t pos, unsigned len, unsigned copied,
1465                         struct page *page, void *fsdata)
1466 {
1467         struct inode *inode = mapping->host;
1468
1469         if (pos + copied > inode->i_size)
1470                 i_size_write(inode, pos + copied);
1471
1472         unlock_page(page);
1473         set_page_dirty(page);
1474         page_cache_release(page);
1475
1476         return copied;
1477 }
1478
1479 static void do_shmem_file_read(struct file *filp, loff_t *ppos, read_descriptor_t *desc, read_actor_t actor)
1480 {
1481         struct inode *inode = filp->f_path.dentry->d_inode;
1482         struct address_space *mapping = inode->i_mapping;
1483         unsigned long index, offset;
1484
1485         index = *ppos >> PAGE_CACHE_SHIFT;
1486         offset = *ppos & ~PAGE_CACHE_MASK;
1487
1488         for (;;) {
1489                 struct page *page = NULL;
1490                 unsigned long end_index, nr, ret;
1491                 loff_t i_size = i_size_read(inode);
1492
1493                 end_index = i_size >> PAGE_CACHE_SHIFT;
1494                 if (index > end_index)
1495                         break;
1496                 if (index == end_index) {
1497                         nr = i_size & ~PAGE_CACHE_MASK;
1498                         if (nr <= offset)
1499                                 break;
1500                 }
1501
1502                 desc->error = shmem_getpage(inode, index, &page, SGP_READ, NULL);
1503                 if (desc->error) {
1504                         if (desc->error == -EINVAL)
1505                                 desc->error = 0;
1506                         break;
1507                 }
1508                 if (page)
1509                         unlock_page(page);
1510
1511                 /*
1512                  * We must evaluate after, since reads (unlike writes)
1513                  * are called without i_mutex protection against truncate
1514                  */
1515                 nr = PAGE_CACHE_SIZE;
1516                 i_size = i_size_read(inode);
1517                 end_index = i_size >> PAGE_CACHE_SHIFT;
1518                 if (index == end_index) {
1519                         nr = i_size & ~PAGE_CACHE_MASK;
1520                         if (nr <= offset) {
1521                                 if (page)
1522                                         page_cache_release(page);
1523                                 break;
1524                         }
1525                 }
1526                 nr -= offset;
1527
1528                 if (page) {
1529                         /*
1530                          * If users can be writing to this page using arbitrary
1531                          * virtual addresses, take care about potential aliasing
1532                          * before reading the page on the kernel side.
1533                          */
1534                         if (mapping_writably_mapped(mapping))
1535                                 flush_dcache_page(page);
1536                         /*
1537                          * Mark the page accessed if we read the beginning.
1538                          */
1539                         if (!offset)
1540                                 mark_page_accessed(page);
1541                 } else {
1542                         page = ZERO_PAGE(0);
1543                         page_cache_get(page);
1544                 }
1545
1546                 /*
1547                  * Ok, we have the page, and it's up-to-date, so
1548                  * now we can copy it to user space...
1549                  *
1550                  * The actor routine returns how many bytes were actually used..
1551                  * NOTE! This may not be the same as how much of a user buffer
1552                  * we filled up (we may be padding etc), so we can only update
1553                  * "pos" here (the actor routine has to update the user buffer
1554                  * pointers and the remaining count).
1555                  */
1556                 ret = actor(desc, page, offset, nr);
1557                 offset += ret;
1558                 index += offset >> PAGE_CACHE_SHIFT;
1559                 offset &= ~PAGE_CACHE_MASK;
1560
1561                 page_cache_release(page);
1562                 if (ret != nr || !desc->count)
1563                         break;
1564
1565                 cond_resched();
1566         }
1567
1568         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1569         file_accessed(filp);
1570 }
1571
1572 static ssize_t shmem_file_read(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
1573 {
1574         read_descriptor_t desc;
1575
1576         if ((ssize_t) count < 0)
1577                 return -EINVAL;
1578         if (!access_ok(VERIFY_WRITE, buf, count))
1579                 return -EFAULT;
1580         if (!count)
1581                 return 0;
1582
1583         desc.written = 0;
1584         desc.count = count;
1585         desc.arg.buf = buf;
1586         desc.error = 0;
1587
1588         do_shmem_file_read(filp, ppos, &desc, file_read_actor);
1589         if (desc.written)
1590                 return desc.written;
1591         return desc.error;
1592 }
1593
1594 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
1595 {
1596         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
1597
1598         buf->f_type = TMPFS_MAGIC;
1599         buf->f_bsize = PAGE_CACHE_SIZE;
1600         buf->f_namelen = NAME_MAX;
1601         spin_lock(&sbinfo->stat_lock);
1602         if (sbinfo->max_blocks) {
1603                 buf->f_blocks = sbinfo->max_blocks;
1604                 buf->f_bavail = buf->f_bfree = sbinfo->free_blocks;
1605         }
1606         if (sbinfo->max_inodes) {
1607                 buf->f_files = sbinfo->max_inodes;
1608                 buf->f_ffree = sbinfo->free_inodes;
1609         }
1610         /* else leave those fields 0 like simple_statfs */
1611         spin_unlock(&sbinfo->stat_lock);
1612         return 0;
1613 }
1614
1615 /*
1616  * File creation. Allocate an inode, and we're done..
1617  */
1618 static int
1619 shmem_mknod(struct inode *dir, struct dentry *dentry, int mode, dev_t dev)
1620 {
1621         struct inode *inode = shmem_get_inode(dir->i_sb, mode, dev);
1622         int error = -ENOSPC;
1623
1624         if (inode) {
1625                 error = security_inode_init_security(inode, dir, NULL, NULL,
1626                                                      NULL);
1627                 if (error) {
1628                         if (error != -EOPNOTSUPP) {
1629                                 iput(inode);
1630                                 return error;
1631                         }
1632                 }
1633                 error = shmem_acl_init(inode, dir);
1634                 if (error) {
1635                         iput(inode);
1636                         return error;
1637                 }
1638                 if (dir->i_mode & S_ISGID) {
1639                         inode->i_gid = dir->i_gid;
1640                         if (S_ISDIR(mode))
1641                                 inode->i_mode |= S_ISGID;
1642                 }
1643                 dir->i_size += BOGO_DIRENT_SIZE;
1644                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1645                 d_instantiate(dentry, inode);
1646                 dget(dentry); /* Extra count - pin the dentry in core */
1647         }
1648         return error;
1649 }
1650
1651 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, int mode)
1652 {
1653         int error;
1654
1655         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
1656                 return error;
1657         inc_nlink(dir);
1658         return 0;
1659 }
1660
1661 static int shmem_create(struct inode *dir, struct dentry *dentry, int mode,
1662                 struct nameidata *nd)
1663 {
1664         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
1665 }
1666
1667 /*
1668  * Link a file..
1669  */
1670 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
1671 {
1672         struct inode *inode = old_dentry->d_inode;
1673         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1674
1675         /*
1676          * No ordinary (disk based) filesystem counts links as inodes;
1677          * but each new link needs a new dentry, pinning lowmem, and
1678          * tmpfs dentries cannot be pruned until they are unlinked.
1679          */
1680         if (sbinfo->max_inodes) {
1681                 spin_lock(&sbinfo->stat_lock);
1682                 if (!sbinfo->free_inodes) {
1683                         spin_unlock(&sbinfo->stat_lock);
1684                         return -ENOSPC;
1685                 }
1686                 sbinfo->free_inodes--;
1687                 spin_unlock(&sbinfo->stat_lock);
1688         }
1689
1690         dir->i_size += BOGO_DIRENT_SIZE;
1691         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1692         inc_nlink(inode);
1693         atomic_inc(&inode->i_count);    /* New dentry reference */
1694         dget(dentry);           /* Extra pinning count for the created dentry */
1695         d_instantiate(dentry, inode);
1696         return 0;
1697 }
1698
1699 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
1700 {
1701         struct inode *inode = dentry->d_inode;
1702
1703         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode)) {
1704                 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1705                 if (sbinfo->max_inodes) {
1706                         spin_lock(&sbinfo->stat_lock);
1707                         sbinfo->free_inodes++;
1708                         spin_unlock(&sbinfo->stat_lock);
1709                 }
1710         }
1711
1712         dir->i_size -= BOGO_DIRENT_SIZE;
1713         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1714         drop_nlink(inode);
1715         dput(dentry);   /* Undo the count from "create" - this does all the work */
1716         return 0;
1717 }
1718
1719 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
1720 {
1721         if (!simple_empty(dentry))
1722                 return -ENOTEMPTY;
1723
1724         drop_nlink(dentry->d_inode);
1725         drop_nlink(dir);
1726         return shmem_unlink(dir, dentry);
1727 }
1728
1729 /*
1730  * The VFS layer already does all the dentry stuff for rename,
1731  * we just have to decrement the usage count for the target if
1732  * it exists so that the VFS layer correctly free's it when it
1733  * gets overwritten.
1734  */
1735 static int shmem_rename(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
1736 {
1737         struct inode *inode = old_dentry->d_inode;
1738         int they_are_dirs = S_ISDIR(inode->i_mode);
1739
1740         if (!simple_empty(new_dentry))
1741                 return -ENOTEMPTY;
1742
1743         if (new_dentry->d_inode) {
1744                 (void) shmem_unlink(new_dir, new_dentry);
1745                 if (they_are_dirs)
1746                         drop_nlink(old_dir);
1747         } else if (they_are_dirs) {
1748                 drop_nlink(old_dir);
1749                 inc_nlink(new_dir);
1750         }
1751
1752         old_dir->i_size -= BOGO_DIRENT_SIZE;
1753         new_dir->i_size += BOGO_DIRENT_SIZE;
1754         old_dir->i_ctime = old_dir->i_mtime =
1755         new_dir->i_ctime = new_dir->i_mtime =
1756         inode->i_ctime = CURRENT_TIME;
1757         return 0;
1758 }
1759
1760 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
1761 {
1762         int error;
1763         int len;
1764         struct inode *inode;
1765         struct page *page = NULL;
1766         char *kaddr;
1767         struct shmem_inode_info *info;
1768
1769         len = strlen(symname) + 1;
1770         if (len > PAGE_CACHE_SIZE)
1771                 return -ENAMETOOLONG;
1772
1773         inode = shmem_get_inode(dir->i_sb, S_IFLNK|S_IRWXUGO, 0);
1774         if (!inode)
1775                 return -ENOSPC;
1776
1777         error = security_inode_init_security(inode, dir, NULL, NULL,
1778                                              NULL);
1779         if (error) {
1780                 if (error != -EOPNOTSUPP) {
1781                         iput(inode);
1782                         return error;
1783                 }
1784                 error = 0;
1785         }
1786
1787         info = SHMEM_I(inode);
1788         inode->i_size = len-1;
1789         if (len <= (char *)inode - (char *)info) {
1790                 /* do it inline */
1791                 memcpy(info, symname, len);
1792                 inode->i_op = &shmem_symlink_inline_operations;
1793         } else {
1794                 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
1795                 if (error) {
1796                         iput(inode);
1797                         return error;
1798                 }
1799                 unlock_page(page);
1800                 inode->i_op = &shmem_symlink_inode_operations;
1801                 kaddr = kmap_atomic(page, KM_USER0);
1802                 memcpy(kaddr, symname, len);
1803                 kunmap_atomic(kaddr, KM_USER0);
1804                 set_page_dirty(page);
1805                 page_cache_release(page);
1806         }
1807         if (dir->i_mode & S_ISGID)
1808                 inode->i_gid = dir->i_gid;
1809         dir->i_size += BOGO_DIRENT_SIZE;
1810         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1811         d_instantiate(dentry, inode);
1812         dget(dentry);
1813         return 0;
1814 }
1815
1816 static void *shmem_follow_link_inline(struct dentry *dentry, struct nameidata *nd)
1817 {
1818         nd_set_link(nd, (char *)SHMEM_I(dentry->d_inode));
1819         return NULL;
1820 }
1821
1822 static void *shmem_follow_link(struct dentry *dentry, struct nameidata *nd)
1823 {
1824         struct page *page = NULL;
1825         int res = shmem_getpage(dentry->d_inode, 0, &page, SGP_READ, NULL);
1826         nd_set_link(nd, res ? ERR_PTR(res) : kmap(page));
1827         if (page)
1828                 unlock_page(page);
1829         return page;
1830 }
1831
1832 static void shmem_put_link(struct dentry *dentry, struct nameidata *nd, void *cookie)
1833 {
1834         if (!IS_ERR(nd_get_link(nd))) {
1835                 struct page *page = cookie;
1836                 kunmap(page);
1837                 mark_page_accessed(page);
1838                 page_cache_release(page);
1839         }
1840 }
1841
1842 static const struct inode_operations shmem_symlink_inline_operations = {
1843         .readlink       = generic_readlink,
1844         .follow_link    = shmem_follow_link_inline,
1845 };
1846
1847 static const struct inode_operations shmem_symlink_inode_operations = {
1848         .truncate       = shmem_truncate,
1849         .readlink       = generic_readlink,
1850         .follow_link    = shmem_follow_link,
1851         .put_link       = shmem_put_link,
1852 };
1853
1854 #ifdef CONFIG_TMPFS_POSIX_ACL
1855 /**
1856  * Superblocks without xattr inode operations will get security.* xattr
1857  * support from the VFS "for free". As soon as we have any other xattrs
1858  * like ACLs, we also need to implement the security.* handlers at
1859  * filesystem level, though.
1860  */
1861
1862 static size_t shmem_xattr_security_list(struct inode *inode, char *list,
1863                                         size_t list_len, const char *name,
1864                                         size_t name_len)
1865 {
1866         return security_inode_listsecurity(inode, list, list_len);
1867 }
1868
1869 static int shmem_xattr_security_get(struct inode *inode, const char *name,
1870                                     void *buffer, size_t size)
1871 {
1872         if (strcmp(name, "") == 0)
1873                 return -EINVAL;
1874         return security_inode_getsecurity(inode, name, buffer, size,
1875                                           -EOPNOTSUPP);
1876 }
1877
1878 static int shmem_xattr_security_set(struct inode *inode, const char *name,
1879                                     const void *value, size_t size, int flags)
1880 {
1881         if (strcmp(name, "") == 0)
1882                 return -EINVAL;
1883         return security_inode_setsecurity(inode, name, value, size, flags);
1884 }
1885
1886 static struct xattr_handler shmem_xattr_security_handler = {
1887         .prefix = XATTR_SECURITY_PREFIX,
1888         .list   = shmem_xattr_security_list,
1889         .get    = shmem_xattr_security_get,
1890         .set    = shmem_xattr_security_set,
1891 };
1892
1893 static struct xattr_handler *shmem_xattr_handlers[] = {
1894         &shmem_xattr_acl_access_handler,
1895         &shmem_xattr_acl_default_handler,
1896         &shmem_xattr_security_handler,
1897         NULL
1898 };
1899 #endif
1900
1901 static struct dentry *shmem_get_parent(struct dentry *child)
1902 {
1903         return ERR_PTR(-ESTALE);
1904 }
1905
1906 static int shmem_match(struct inode *ino, void *vfh)
1907 {
1908         __u32 *fh = vfh;
1909         __u64 inum = fh[2];
1910         inum = (inum << 32) | fh[1];
1911         return ino->i_ino == inum && fh[0] == ino->i_generation;
1912 }
1913
1914 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
1915                 struct fid *fid, int fh_len, int fh_type)
1916 {
1917         struct inode *inode;
1918         struct dentry *dentry = NULL;
1919         u64 inum = fid->raw[2];
1920         inum = (inum << 32) | fid->raw[1];
1921
1922         if (fh_len < 3)
1923                 return NULL;
1924
1925         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
1926                         shmem_match, fid->raw);
1927         if (inode) {
1928                 dentry = d_find_alias(inode);
1929                 iput(inode);
1930         }
1931
1932         return dentry;
1933 }
1934
1935 static int shmem_encode_fh(struct dentry *dentry, __u32 *fh, int *len,
1936                                 int connectable)
1937 {
1938         struct inode *inode = dentry->d_inode;
1939
1940         if (*len < 3)
1941                 return 255;
1942
1943         if (hlist_unhashed(&inode->i_hash)) {
1944                 /* Unfortunately insert_inode_hash is not idempotent,
1945                  * so as we hash inodes here rather than at creation
1946                  * time, we need a lock to ensure we only try
1947                  * to do it once
1948                  */
1949                 static DEFINE_SPINLOCK(lock);
1950                 spin_lock(&lock);
1951                 if (hlist_unhashed(&inode->i_hash))
1952                         __insert_inode_hash(inode,
1953                                             inode->i_ino + inode->i_generation);
1954                 spin_unlock(&lock);
1955         }
1956
1957         fh[0] = inode->i_generation;
1958         fh[1] = inode->i_ino;
1959         fh[2] = ((__u64)inode->i_ino) >> 32;
1960
1961         *len = 3;
1962         return 1;
1963 }
1964
1965 static const struct export_operations shmem_export_ops = {
1966         .get_parent     = shmem_get_parent,
1967         .encode_fh      = shmem_encode_fh,
1968         .fh_to_dentry   = shmem_fh_to_dentry,
1969 };
1970
1971 static int shmem_parse_options(char *options, int *mode, uid_t *uid,
1972         gid_t *gid, unsigned long *blocks, unsigned long *inodes,
1973         int *policy, nodemask_t *policy_nodes)
1974 {
1975         char *this_char, *value, *rest;
1976
1977         while (options != NULL) {
1978                 this_char = options;
1979                 for (;;) {
1980                         /*
1981                          * NUL-terminate this option: unfortunately,
1982                          * mount options form a comma-separated list,
1983                          * but mpol's nodelist may also contain commas.
1984                          */
1985                         options = strchr(options, ',');
1986                         if (options == NULL)
1987                                 break;
1988                         options++;
1989                         if (!isdigit(*options)) {
1990                                 options[-1] = '\0';
1991                                 break;
1992                         }
1993                 }
1994                 if (!*this_char)
1995                         continue;
1996                 if ((value = strchr(this_char,'=')) != NULL) {
1997                         *value++ = 0;
1998                 } else {
1999                         printk(KERN_ERR
2000                             "tmpfs: No value for mount option '%s'\n",
2001                             this_char);
2002                         return 1;
2003                 }
2004
2005                 if (!strcmp(this_char,"size")) {
2006                         unsigned long long size;
2007                         size = memparse(value,&rest);
2008                         if (*rest == '%') {
2009                                 size <<= PAGE_SHIFT;
2010                                 size *= totalram_pages;
2011                                 do_div(size, 100);
2012                                 rest++;
2013                         }
2014                         if (*rest)
2015                                 goto bad_val;
2016                         *blocks = size >> PAGE_CACHE_SHIFT;
2017                 } else if (!strcmp(this_char,"nr_blocks")) {
2018                         *blocks = memparse(value,&rest);
2019                         if (*rest)
2020                                 goto bad_val;
2021                 } else if (!strcmp(this_char,"nr_inodes")) {
2022                         *inodes = memparse(value,&rest);
2023                         if (*rest)
2024                                 goto bad_val;
2025                 } else if (!strcmp(this_char,"mode")) {
2026                         if (!mode)
2027                                 continue;
2028                         *mode = simple_strtoul(value,&rest,8);
2029                         if (*rest)
2030                                 goto bad_val;
2031                 } else if (!strcmp(this_char,"uid")) {
2032                         if (!uid)
2033                                 continue;
2034                         *uid = simple_strtoul(value,&rest,0);
2035                         if (*rest)
2036                                 goto bad_val;
2037                 } else if (!strcmp(this_char,"gid")) {
2038                         if (!gid)
2039                                 continue;
2040                         *gid = simple_strtoul(value,&rest,0);
2041                         if (*rest)
2042                                 goto bad_val;
2043                 } else if (!strcmp(this_char,"mpol")) {
2044                         if (shmem_parse_mpol(value,policy,policy_nodes))
2045                                 goto bad_val;
2046                 } else {
2047                         printk(KERN_ERR "tmpfs: Bad mount option %s\n",
2048                                this_char);
2049                         return 1;
2050                 }
2051         }
2052         return 0;
2053
2054 bad_val:
2055         printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
2056                value, this_char);
2057         return 1;
2058
2059 }
2060
2061 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2062 {
2063         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2064         unsigned long max_blocks = sbinfo->max_blocks;
2065         unsigned long max_inodes = sbinfo->max_inodes;
2066         int policy = sbinfo->policy;
2067         nodemask_t policy_nodes = sbinfo->policy_nodes;
2068         unsigned long blocks;
2069         unsigned long inodes;
2070         int error = -EINVAL;
2071
2072         if (shmem_parse_options(data, NULL, NULL, NULL, &max_blocks,
2073                                 &max_inodes, &policy, &policy_nodes))
2074                 return error;
2075
2076         spin_lock(&sbinfo->stat_lock);
2077         blocks = sbinfo->max_blocks - sbinfo->free_blocks;
2078         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2079         if (max_blocks < blocks)
2080                 goto out;
2081         if (max_inodes < inodes)
2082                 goto out;
2083         /*
2084          * Those tests also disallow limited->unlimited while any are in
2085          * use, so i_blocks will always be zero when max_blocks is zero;
2086          * but we must separately disallow unlimited->limited, because
2087          * in that case we have no record of how much is already in use.
2088          */
2089         if (max_blocks && !sbinfo->max_blocks)
2090                 goto out;
2091         if (max_inodes && !sbinfo->max_inodes)
2092                 goto out;
2093
2094         error = 0;
2095         sbinfo->max_blocks  = max_blocks;
2096         sbinfo->free_blocks = max_blocks - blocks;
2097         sbinfo->max_inodes  = max_inodes;
2098         sbinfo->free_inodes = max_inodes - inodes;
2099         sbinfo->policy = policy;
2100         sbinfo->policy_nodes = policy_nodes;
2101 out:
2102         spin_unlock(&sbinfo->stat_lock);
2103         return error;
2104 }
2105 #endif
2106
2107 static void shmem_put_super(struct super_block *sb)
2108 {
2109         kfree(sb->s_fs_info);
2110         sb->s_fs_info = NULL;
2111 }
2112
2113 static int shmem_fill_super(struct super_block *sb,
2114                             void *data, int silent)
2115 {
2116         struct inode *inode;
2117         struct dentry *root;
2118         int mode   = S_IRWXUGO | S_ISVTX;
2119         uid_t uid = current->fsuid;
2120         gid_t gid = current->fsgid;
2121         int err = -ENOMEM;
2122         struct shmem_sb_info *sbinfo;
2123         unsigned long blocks = 0;
2124         unsigned long inodes = 0;
2125         int policy = MPOL_DEFAULT;
2126         nodemask_t policy_nodes = node_states[N_HIGH_MEMORY];
2127
2128 #ifdef CONFIG_TMPFS
2129         /*
2130          * Per default we only allow half of the physical ram per
2131          * tmpfs instance, limiting inodes to one per page of lowmem;
2132          * but the internal instance is left unlimited.
2133          */
2134         if (!(sb->s_flags & MS_NOUSER)) {
2135                 blocks = totalram_pages / 2;
2136                 inodes = totalram_pages - totalhigh_pages;
2137                 if (inodes > blocks)
2138                         inodes = blocks;
2139                 if (shmem_parse_options(data, &mode, &uid, &gid, &blocks,
2140                                         &inodes, &policy, &policy_nodes))
2141                         return -EINVAL;
2142         }
2143         sb->s_export_op = &shmem_export_ops;
2144 #else
2145         sb->s_flags |= MS_NOUSER;
2146 #endif
2147
2148         /* Round up to L1_CACHE_BYTES to resist false sharing */
2149         sbinfo = kmalloc(max((int)sizeof(struct shmem_sb_info),
2150                                 L1_CACHE_BYTES), GFP_KERNEL);
2151         if (!sbinfo)
2152                 return -ENOMEM;
2153
2154         spin_lock_init(&sbinfo->stat_lock);
2155         sbinfo->max_blocks = blocks;
2156         sbinfo->free_blocks = blocks;
2157         sbinfo->max_inodes = inodes;
2158         sbinfo->free_inodes = inodes;
2159         sbinfo->policy = policy;
2160         sbinfo->policy_nodes = policy_nodes;
2161
2162         sb->s_fs_info = sbinfo;
2163         sb->s_maxbytes = SHMEM_MAX_BYTES;
2164         sb->s_blocksize = PAGE_CACHE_SIZE;
2165         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
2166         sb->s_magic = TMPFS_MAGIC;
2167         sb->s_op = &shmem_ops;
2168         sb->s_time_gran = 1;
2169 #ifdef CONFIG_TMPFS_POSIX_ACL
2170         sb->s_xattr = shmem_xattr_handlers;
2171         sb->s_flags |= MS_POSIXACL;
2172 #endif
2173
2174         inode = shmem_get_inode(sb, S_IFDIR | mode, 0);
2175         if (!inode)
2176                 goto failed;
2177         inode->i_uid = uid;
2178         inode->i_gid = gid;
2179         root = d_alloc_root(inode);
2180         if (!root)
2181                 goto failed_iput;
2182         sb->s_root = root;
2183         return 0;
2184
2185 failed_iput:
2186         iput(inode);
2187 failed:
2188         shmem_put_super(sb);
2189         return err;
2190 }
2191
2192 static struct kmem_cache *shmem_inode_cachep;
2193
2194 static struct inode *shmem_alloc_inode(struct super_block *sb)
2195 {
2196         struct shmem_inode_info *p;
2197         p = (struct shmem_inode_info *)kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
2198         if (!p)
2199                 return NULL;
2200         return &p->vfs_inode;
2201 }
2202
2203 static void shmem_destroy_inode(struct inode *inode)
2204 {
2205         if ((inode->i_mode & S_IFMT) == S_IFREG) {
2206                 /* only struct inode is valid if it's an inline symlink */
2207                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
2208         }
2209         shmem_acl_destroy_inode(inode);
2210         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
2211 }
2212
2213 static void init_once(struct kmem_cache *cachep, void *foo)
2214 {
2215         struct shmem_inode_info *p = (struct shmem_inode_info *) foo;
2216
2217         inode_init_once(&p->vfs_inode);
2218 #ifdef CONFIG_TMPFS_POSIX_ACL
2219         p->i_acl = NULL;
2220         p->i_default_acl = NULL;
2221 #endif
2222 }
2223
2224 static int init_inodecache(void)
2225 {
2226         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
2227                                 sizeof(struct shmem_inode_info),
2228                                 0, SLAB_PANIC, init_once);
2229         return 0;
2230 }
2231
2232 static void destroy_inodecache(void)
2233 {
2234         kmem_cache_destroy(shmem_inode_cachep);
2235 }
2236
2237 static const struct address_space_operations shmem_aops = {
2238         .writepage      = shmem_writepage,
2239         .set_page_dirty = __set_page_dirty_no_writeback,
2240 #ifdef CONFIG_TMPFS
2241         .readpage       = shmem_readpage,
2242         .write_begin    = shmem_write_begin,
2243         .write_end      = shmem_write_end,
2244 #endif
2245         .migratepage    = migrate_page,
2246 };
2247
2248 static const struct file_operations shmem_file_operations = {
2249         .mmap           = shmem_mmap,
2250 #ifdef CONFIG_TMPFS
2251         .llseek         = generic_file_llseek,
2252         .read           = shmem_file_read,
2253         .write          = do_sync_write,
2254         .aio_write      = generic_file_aio_write,
2255         .fsync          = simple_sync_file,
2256         .splice_read    = generic_file_splice_read,
2257         .splice_write   = generic_file_splice_write,
2258 #endif
2259 };
2260
2261 static const struct inode_operations shmem_inode_operations = {
2262         .truncate       = shmem_truncate,
2263         .setattr        = shmem_notify_change,
2264         .truncate_range = shmem_truncate_range,
2265 #ifdef CONFIG_TMPFS_POSIX_ACL
2266         .setxattr       = generic_setxattr,
2267         .getxattr       = generic_getxattr,
2268         .listxattr      = generic_listxattr,
2269         .removexattr    = generic_removexattr,
2270         .permission     = shmem_permission,
2271 #endif
2272
2273 };
2274
2275 static const struct inode_operations shmem_dir_inode_operations = {
2276 #ifdef CONFIG_TMPFS
2277         .create         = shmem_create,
2278         .lookup         = simple_lookup,
2279         .link           = shmem_link,
2280         .unlink         = shmem_unlink,
2281         .symlink        = shmem_symlink,
2282         .mkdir          = shmem_mkdir,
2283         .rmdir          = shmem_rmdir,
2284         .mknod          = shmem_mknod,
2285         .rename         = shmem_rename,
2286 #endif
2287 #ifdef CONFIG_TMPFS_POSIX_ACL
2288         .setattr        = shmem_notify_change,
2289         .setxattr       = generic_setxattr,
2290         .getxattr       = generic_getxattr,
2291         .listxattr      = generic_listxattr,
2292         .removexattr    = generic_removexattr,
2293         .permission     = shmem_permission,
2294 #endif
2295 };
2296
2297 static const struct inode_operations shmem_special_inode_operations = {
2298 #ifdef CONFIG_TMPFS_POSIX_ACL
2299         .setattr        = shmem_notify_change,
2300         .setxattr       = generic_setxattr,
2301         .getxattr       = generic_getxattr,
2302         .listxattr      = generic_listxattr,
2303         .removexattr    = generic_removexattr,
2304         .permission     = shmem_permission,
2305 #endif
2306 };
2307
2308 static const struct super_operations shmem_ops = {
2309         .alloc_inode    = shmem_alloc_inode,
2310         .destroy_inode  = shmem_destroy_inode,
2311 #ifdef CONFIG_TMPFS
2312         .statfs         = shmem_statfs,
2313         .remount_fs     = shmem_remount_fs,
2314 #endif
2315         .delete_inode   = shmem_delete_inode,
2316         .drop_inode     = generic_delete_inode,
2317         .put_super      = shmem_put_super,
2318 };
2319
2320 static struct vm_operations_struct shmem_vm_ops = {
2321         .fault          = shmem_fault,
2322 #ifdef CONFIG_NUMA
2323         .set_policy     = shmem_set_policy,
2324         .get_policy     = shmem_get_policy,
2325 #endif
2326 };
2327
2328
2329 static int shmem_get_sb(struct file_system_type *fs_type,
2330         int flags, const char *dev_name, void *data, struct vfsmount *mnt)
2331 {
2332         return get_sb_nodev(fs_type, flags, data, shmem_fill_super, mnt);
2333 }
2334
2335 static struct file_system_type tmpfs_fs_type = {
2336         .owner          = THIS_MODULE,
2337         .name           = "tmpfs",
2338         .get_sb         = shmem_get_sb,
2339         .kill_sb        = kill_litter_super,
2340 };
2341 static struct vfsmount *shm_mnt;
2342
2343 static int __init init_tmpfs(void)
2344 {
2345         int error;
2346
2347         error = bdi_init(&shmem_backing_dev_info);
2348         if (error)
2349                 goto out4;
2350
2351         error = init_inodecache();
2352         if (error)
2353                 goto out3;
2354
2355         error = register_filesystem(&tmpfs_fs_type);
2356         if (error) {
2357                 printk(KERN_ERR "Could not register tmpfs\n");
2358                 goto out2;
2359         }
2360
2361         shm_mnt = vfs_kern_mount(&tmpfs_fs_type, MS_NOUSER,
2362                                 tmpfs_fs_type.name, NULL);
2363         if (IS_ERR(shm_mnt)) {
2364                 error = PTR_ERR(shm_mnt);
2365                 printk(KERN_ERR "Could not kern_mount tmpfs\n");
2366                 goto out1;
2367         }
2368         return 0;
2369
2370 out1:
2371         unregister_filesystem(&tmpfs_fs_type);
2372 out2:
2373         destroy_inodecache();
2374 out3:
2375         bdi_destroy(&shmem_backing_dev_info);
2376 out4:
2377         shm_mnt = ERR_PTR(error);
2378         return error;
2379 }
2380 module_init(init_tmpfs)
2381
2382 /*
2383  * shmem_file_setup - get an unlinked file living in tmpfs
2384  *
2385  * @name: name for dentry (to be seen in /proc/<pid>/maps
2386  * @size: size to be set for the file
2387  *
2388  */
2389 struct file *shmem_file_setup(char *name, loff_t size, unsigned long flags)
2390 {
2391         int error;
2392         struct file *file;
2393         struct inode *inode;
2394         struct dentry *dentry, *root;
2395         struct qstr this;
2396
2397         if (IS_ERR(shm_mnt))
2398                 return (void *)shm_mnt;
2399
2400         if (size < 0 || size > SHMEM_MAX_BYTES)
2401                 return ERR_PTR(-EINVAL);
2402
2403         if (shmem_acct_size(flags, size))
2404                 return ERR_PTR(-ENOMEM);
2405
2406         error = -ENOMEM;
2407         this.name = name;
2408         this.len = strlen(name);
2409         this.hash = 0; /* will go */
2410         root = shm_mnt->mnt_root;
2411         dentry = d_alloc(root, &this);
2412         if (!dentry)
2413                 goto put_memory;
2414
2415         error = -ENFILE;
2416         file = get_empty_filp();
2417         if (!file)
2418                 goto put_dentry;
2419
2420         error = -ENOSPC;
2421         inode = shmem_get_inode(root->d_sb, S_IFREG | S_IRWXUGO, 0);
2422         if (!inode)
2423                 goto close_file;
2424
2425         SHMEM_I(inode)->flags = flags & VM_ACCOUNT;
2426         d_instantiate(dentry, inode);
2427         inode->i_size = size;
2428         inode->i_nlink = 0;     /* It is unlinked */
2429         init_file(file, shm_mnt, dentry, FMODE_WRITE | FMODE_READ,
2430                         &shmem_file_operations);
2431         return file;
2432
2433 close_file:
2434         put_filp(file);
2435 put_dentry:
2436         dput(dentry);
2437 put_memory:
2438         shmem_unacct_size(flags, size);
2439         return ERR_PTR(error);
2440 }
2441
2442 /*
2443  * shmem_zero_setup - setup a shared anonymous mapping
2444  *
2445  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
2446  */
2447 int shmem_zero_setup(struct vm_area_struct *vma)
2448 {
2449         struct file *file;
2450         loff_t size = vma->vm_end - vma->vm_start;
2451
2452         file = shmem_file_setup("dev/zero", size, vma->vm_flags);
2453         if (IS_ERR(file))
2454                 return PTR_ERR(file);
2455
2456         if (vma->vm_file)
2457                 fput(vma->vm_file);
2458         vma->vm_file = file;
2459         vma->vm_ops = &shmem_vm_ops;
2460         return 0;
2461 }