51b3d6ccddab75ba33b7fded5ffaf75700770ccd
[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_QUICK,      /* don't try more than file page cache lookup */
82         SGP_READ,       /* don't exceed i_size, don't allocate page */
83         SGP_CACHE,      /* don't exceed i_size, may allocate page */
84         SGP_WRITE,      /* may exceed i_size, may allocate page */
85         SGP_FAULT,      /* same as SGP_CACHE, return with page locked */
86 };
87
88 static int shmem_getpage(struct inode *inode, unsigned long idx,
89                          struct page **pagep, enum sgp_type sgp, int *type);
90
91 static inline struct page *shmem_dir_alloc(gfp_t gfp_mask)
92 {
93         /*
94          * The above definition of ENTRIES_PER_PAGE, and the use of
95          * BLOCKS_PER_PAGE on indirect pages, assume PAGE_CACHE_SIZE:
96          * might be reconsidered if it ever diverges from PAGE_SIZE.
97          *
98          * Mobility flags are masked out as swap vectors cannot move
99          */
100         return alloc_pages((gfp_mask & ~GFP_MOVABLE_MASK) | __GFP_ZERO,
101                                 PAGE_CACHE_SHIFT-PAGE_SHIFT);
102 }
103
104 static inline void shmem_dir_free(struct page *page)
105 {
106         __free_pages(page, PAGE_CACHE_SHIFT-PAGE_SHIFT);
107 }
108
109 static struct page **shmem_dir_map(struct page *page)
110 {
111         return (struct page **)kmap_atomic(page, KM_USER0);
112 }
113
114 static inline void shmem_dir_unmap(struct page **dir)
115 {
116         kunmap_atomic(dir, KM_USER0);
117 }
118
119 static swp_entry_t *shmem_swp_map(struct page *page)
120 {
121         return (swp_entry_t *)kmap_atomic(page, KM_USER1);
122 }
123
124 static inline void shmem_swp_balance_unmap(void)
125 {
126         /*
127          * When passing a pointer to an i_direct entry, to code which
128          * also handles indirect entries and so will shmem_swp_unmap,
129          * we must arrange for the preempt count to remain in balance.
130          * What kmap_atomic of a lowmem page does depends on config
131          * and architecture, so pretend to kmap_atomic some lowmem page.
132          */
133         (void) kmap_atomic(ZERO_PAGE(0), KM_USER1);
134 }
135
136 static inline void shmem_swp_unmap(swp_entry_t *entry)
137 {
138         kunmap_atomic(entry, KM_USER1);
139 }
140
141 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
142 {
143         return sb->s_fs_info;
144 }
145
146 /*
147  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
148  * for shared memory and for shared anonymous (/dev/zero) mappings
149  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
150  * consistent with the pre-accounting of private mappings ...
151  */
152 static inline int shmem_acct_size(unsigned long flags, loff_t size)
153 {
154         return (flags & VM_ACCOUNT)?
155                 security_vm_enough_memory(VM_ACCT(size)): 0;
156 }
157
158 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
159 {
160         if (flags & VM_ACCOUNT)
161                 vm_unacct_memory(VM_ACCT(size));
162 }
163
164 /*
165  * ... whereas tmpfs objects are accounted incrementally as
166  * pages are allocated, in order to allow huge sparse files.
167  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
168  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
169  */
170 static inline int shmem_acct_block(unsigned long flags)
171 {
172         return (flags & VM_ACCOUNT)?
173                 0: security_vm_enough_memory(VM_ACCT(PAGE_CACHE_SIZE));
174 }
175
176 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
177 {
178         if (!(flags & VM_ACCOUNT))
179                 vm_unacct_memory(pages * VM_ACCT(PAGE_CACHE_SIZE));
180 }
181
182 static const struct super_operations shmem_ops;
183 static const struct address_space_operations shmem_aops;
184 static const struct file_operations shmem_file_operations;
185 static const struct inode_operations shmem_inode_operations;
186 static const struct inode_operations shmem_dir_inode_operations;
187 static const struct inode_operations shmem_special_inode_operations;
188 static struct vm_operations_struct shmem_vm_ops;
189
190 static struct backing_dev_info shmem_backing_dev_info  __read_mostly = {
191         .ra_pages       = 0,    /* No readahead */
192         .capabilities   = BDI_CAP_NO_ACCT_DIRTY | BDI_CAP_NO_WRITEBACK,
193         .unplug_io_fn   = default_unplug_io_fn,
194 };
195
196 static LIST_HEAD(shmem_swaplist);
197 static DEFINE_SPINLOCK(shmem_swaplist_lock);
198
199 static void shmem_free_blocks(struct inode *inode, long pages)
200 {
201         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
202         if (sbinfo->max_blocks) {
203                 spin_lock(&sbinfo->stat_lock);
204                 sbinfo->free_blocks += pages;
205                 inode->i_blocks -= pages*BLOCKS_PER_PAGE;
206                 spin_unlock(&sbinfo->stat_lock);
207         }
208 }
209
210 /*
211  * shmem_recalc_inode - recalculate the size of an inode
212  *
213  * @inode: inode to recalc
214  *
215  * We have to calculate the free blocks since the mm can drop
216  * undirtied hole pages behind our back.
217  *
218  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
219  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
220  *
221  * It has to be called with the spinlock held.
222  */
223 static void shmem_recalc_inode(struct inode *inode)
224 {
225         struct shmem_inode_info *info = SHMEM_I(inode);
226         long freed;
227
228         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
229         if (freed > 0) {
230                 info->alloced -= freed;
231                 shmem_unacct_blocks(info->flags, freed);
232                 shmem_free_blocks(inode, freed);
233         }
234 }
235
236 /*
237  * shmem_swp_entry - find the swap vector position in the info structure
238  *
239  * @info:  info structure for the inode
240  * @index: index of the page to find
241  * @page:  optional page to add to the structure. Has to be preset to
242  *         all zeros
243  *
244  * If there is no space allocated yet it will return NULL when
245  * page is NULL, else it will use the page for the needed block,
246  * setting it to NULL on return to indicate that it has been used.
247  *
248  * The swap vector is organized the following way:
249  *
250  * There are SHMEM_NR_DIRECT entries directly stored in the
251  * shmem_inode_info structure. So small files do not need an addional
252  * allocation.
253  *
254  * For pages with index > SHMEM_NR_DIRECT there is the pointer
255  * i_indirect which points to a page which holds in the first half
256  * doubly indirect blocks, in the second half triple indirect blocks:
257  *
258  * For an artificial ENTRIES_PER_PAGE = 4 this would lead to the
259  * following layout (for SHMEM_NR_DIRECT == 16):
260  *
261  * i_indirect -> dir --> 16-19
262  *            |      +-> 20-23
263  *            |
264  *            +-->dir2 --> 24-27
265  *            |        +-> 28-31
266  *            |        +-> 32-35
267  *            |        +-> 36-39
268  *            |
269  *            +-->dir3 --> 40-43
270  *                     +-> 44-47
271  *                     +-> 48-51
272  *                     +-> 52-55
273  */
274 static swp_entry_t *shmem_swp_entry(struct shmem_inode_info *info, unsigned long index, struct page **page)
275 {
276         unsigned long offset;
277         struct page **dir;
278         struct page *subdir;
279
280         if (index < SHMEM_NR_DIRECT) {
281                 shmem_swp_balance_unmap();
282                 return info->i_direct+index;
283         }
284         if (!info->i_indirect) {
285                 if (page) {
286                         info->i_indirect = *page;
287                         *page = NULL;
288                 }
289                 return NULL;                    /* need another page */
290         }
291
292         index -= SHMEM_NR_DIRECT;
293         offset = index % ENTRIES_PER_PAGE;
294         index /= ENTRIES_PER_PAGE;
295         dir = shmem_dir_map(info->i_indirect);
296
297         if (index >= ENTRIES_PER_PAGE/2) {
298                 index -= ENTRIES_PER_PAGE/2;
299                 dir += ENTRIES_PER_PAGE/2 + index/ENTRIES_PER_PAGE;
300                 index %= ENTRIES_PER_PAGE;
301                 subdir = *dir;
302                 if (!subdir) {
303                         if (page) {
304                                 *dir = *page;
305                                 *page = NULL;
306                         }
307                         shmem_dir_unmap(dir);
308                         return NULL;            /* need another page */
309                 }
310                 shmem_dir_unmap(dir);
311                 dir = shmem_dir_map(subdir);
312         }
313
314         dir += index;
315         subdir = *dir;
316         if (!subdir) {
317                 if (!page || !(subdir = *page)) {
318                         shmem_dir_unmap(dir);
319                         return NULL;            /* need a page */
320                 }
321                 *dir = subdir;
322                 *page = NULL;
323         }
324         shmem_dir_unmap(dir);
325         return shmem_swp_map(subdir) + offset;
326 }
327
328 static void shmem_swp_set(struct shmem_inode_info *info, swp_entry_t *entry, unsigned long value)
329 {
330         long incdec = value? 1: -1;
331
332         entry->val = value;
333         info->swapped += incdec;
334         if ((unsigned long)(entry - info->i_direct) >= SHMEM_NR_DIRECT) {
335                 struct page *page = kmap_atomic_to_page(entry);
336                 set_page_private(page, page_private(page) + incdec);
337         }
338 }
339
340 /*
341  * shmem_swp_alloc - get the position of the swap entry for the page.
342  *                   If it does not exist allocate the entry.
343  *
344  * @info:       info structure for the inode
345  * @index:      index of the page to find
346  * @sgp:        check and recheck i_size? skip allocation?
347  */
348 static swp_entry_t *shmem_swp_alloc(struct shmem_inode_info *info, unsigned long index, enum sgp_type sgp)
349 {
350         struct inode *inode = &info->vfs_inode;
351         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
352         struct page *page = NULL;
353         swp_entry_t *entry;
354
355         if (sgp != SGP_WRITE &&
356             ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode))
357                 return ERR_PTR(-EINVAL);
358
359         while (!(entry = shmem_swp_entry(info, index, &page))) {
360                 if (sgp == SGP_READ)
361                         return shmem_swp_map(ZERO_PAGE(0));
362                 /*
363                  * Test free_blocks against 1 not 0, since we have 1 data
364                  * page (and perhaps indirect index pages) yet to allocate:
365                  * a waste to allocate index if we cannot allocate data.
366                  */
367                 if (sbinfo->max_blocks) {
368                         spin_lock(&sbinfo->stat_lock);
369                         if (sbinfo->free_blocks <= 1) {
370                                 spin_unlock(&sbinfo->stat_lock);
371                                 return ERR_PTR(-ENOSPC);
372                         }
373                         sbinfo->free_blocks--;
374                         inode->i_blocks += BLOCKS_PER_PAGE;
375                         spin_unlock(&sbinfo->stat_lock);
376                 }
377
378                 spin_unlock(&info->lock);
379                 page = shmem_dir_alloc(mapping_gfp_mask(inode->i_mapping));
380                 if (page)
381                         set_page_private(page, 0);
382                 spin_lock(&info->lock);
383
384                 if (!page) {
385                         shmem_free_blocks(inode, 1);
386                         return ERR_PTR(-ENOMEM);
387                 }
388                 if (sgp != SGP_WRITE &&
389                     ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
390                         entry = ERR_PTR(-EINVAL);
391                         break;
392                 }
393                 if (info->next_index <= index)
394                         info->next_index = index + 1;
395         }
396         if (page) {
397                 /* another task gave its page, or truncated the file */
398                 shmem_free_blocks(inode, 1);
399                 shmem_dir_free(page);
400         }
401         if (info->next_index <= index && !IS_ERR(entry))
402                 info->next_index = index + 1;
403         return entry;
404 }
405
406 /*
407  * shmem_free_swp - free some swap entries in a directory
408  *
409  * @dir:        pointer to the directory
410  * @edir:       pointer after last entry of the directory
411  * @punch_lock: pointer to spinlock when needed for the holepunch case
412  */
413 static int shmem_free_swp(swp_entry_t *dir, swp_entry_t *edir,
414                                                 spinlock_t *punch_lock)
415 {
416         spinlock_t *punch_unlock = NULL;
417         swp_entry_t *ptr;
418         int freed = 0;
419
420         for (ptr = dir; ptr < edir; ptr++) {
421                 if (ptr->val) {
422                         if (unlikely(punch_lock)) {
423                                 punch_unlock = punch_lock;
424                                 punch_lock = NULL;
425                                 spin_lock(punch_unlock);
426                                 if (!ptr->val)
427                                         continue;
428                         }
429                         free_swap_and_cache(*ptr);
430                         *ptr = (swp_entry_t){0};
431                         freed++;
432                 }
433         }
434         if (punch_unlock)
435                 spin_unlock(punch_unlock);
436         return freed;
437 }
438
439 static int shmem_map_and_free_swp(struct page *subdir, int offset,
440                 int limit, struct page ***dir, spinlock_t *punch_lock)
441 {
442         swp_entry_t *ptr;
443         int freed = 0;
444
445         ptr = shmem_swp_map(subdir);
446         for (; offset < limit; offset += LATENCY_LIMIT) {
447                 int size = limit - offset;
448                 if (size > LATENCY_LIMIT)
449                         size = LATENCY_LIMIT;
450                 freed += shmem_free_swp(ptr+offset, ptr+offset+size,
451                                                         punch_lock);
452                 if (need_resched()) {
453                         shmem_swp_unmap(ptr);
454                         if (*dir) {
455                                 shmem_dir_unmap(*dir);
456                                 *dir = NULL;
457                         }
458                         cond_resched();
459                         ptr = shmem_swp_map(subdir);
460                 }
461         }
462         shmem_swp_unmap(ptr);
463         return freed;
464 }
465
466 static void shmem_free_pages(struct list_head *next)
467 {
468         struct page *page;
469         int freed = 0;
470
471         do {
472                 page = container_of(next, struct page, lru);
473                 next = next->next;
474                 shmem_dir_free(page);
475                 freed++;
476                 if (freed >= LATENCY_LIMIT) {
477                         cond_resched();
478                         freed = 0;
479                 }
480         } while (next);
481 }
482
483 static void shmem_truncate_range(struct inode *inode, loff_t start, loff_t end)
484 {
485         struct shmem_inode_info *info = SHMEM_I(inode);
486         unsigned long idx;
487         unsigned long size;
488         unsigned long limit;
489         unsigned long stage;
490         unsigned long diroff;
491         struct page **dir;
492         struct page *topdir;
493         struct page *middir;
494         struct page *subdir;
495         swp_entry_t *ptr;
496         LIST_HEAD(pages_to_free);
497         long nr_pages_to_free = 0;
498         long nr_swaps_freed = 0;
499         int offset;
500         int freed;
501         int punch_hole;
502         spinlock_t *needs_lock;
503         spinlock_t *punch_lock;
504         unsigned long upper_limit;
505
506         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
507         idx = (start + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
508         if (idx >= info->next_index)
509                 return;
510
511         spin_lock(&info->lock);
512         info->flags |= SHMEM_TRUNCATE;
513         if (likely(end == (loff_t) -1)) {
514                 limit = info->next_index;
515                 upper_limit = SHMEM_MAX_INDEX;
516                 info->next_index = idx;
517                 needs_lock = NULL;
518                 punch_hole = 0;
519         } else {
520                 if (end + 1 >= inode->i_size) { /* we may free a little more */
521                         limit = (inode->i_size + PAGE_CACHE_SIZE - 1) >>
522                                                         PAGE_CACHE_SHIFT;
523                         upper_limit = SHMEM_MAX_INDEX;
524                 } else {
525                         limit = (end + 1) >> PAGE_CACHE_SHIFT;
526                         upper_limit = limit;
527                 }
528                 needs_lock = &info->lock;
529                 punch_hole = 1;
530         }
531
532         topdir = info->i_indirect;
533         if (topdir && idx <= SHMEM_NR_DIRECT && !punch_hole) {
534                 info->i_indirect = NULL;
535                 nr_pages_to_free++;
536                 list_add(&topdir->lru, &pages_to_free);
537         }
538         spin_unlock(&info->lock);
539
540         if (info->swapped && idx < SHMEM_NR_DIRECT) {
541                 ptr = info->i_direct;
542                 size = limit;
543                 if (size > SHMEM_NR_DIRECT)
544                         size = SHMEM_NR_DIRECT;
545                 nr_swaps_freed = shmem_free_swp(ptr+idx, ptr+size, needs_lock);
546         }
547
548         /*
549          * If there are no indirect blocks or we are punching a hole
550          * below indirect blocks, nothing to be done.
551          */
552         if (!topdir || limit <= SHMEM_NR_DIRECT)
553                 goto done2;
554
555         /*
556          * The truncation case has already dropped info->lock, and we're safe
557          * because i_size and next_index have already been lowered, preventing
558          * access beyond.  But in the punch_hole case, we still need to take
559          * the lock when updating the swap directory, because there might be
560          * racing accesses by shmem_getpage(SGP_CACHE), shmem_unuse_inode or
561          * shmem_writepage.  However, whenever we find we can remove a whole
562          * directory page (not at the misaligned start or end of the range),
563          * we first NULLify its pointer in the level above, and then have no
564          * need to take the lock when updating its contents: needs_lock and
565          * punch_lock (either pointing to info->lock or NULL) manage this.
566          */
567
568         upper_limit -= SHMEM_NR_DIRECT;
569         limit -= SHMEM_NR_DIRECT;
570         idx = (idx > SHMEM_NR_DIRECT)? (idx - SHMEM_NR_DIRECT): 0;
571         offset = idx % ENTRIES_PER_PAGE;
572         idx -= offset;
573
574         dir = shmem_dir_map(topdir);
575         stage = ENTRIES_PER_PAGEPAGE/2;
576         if (idx < ENTRIES_PER_PAGEPAGE/2) {
577                 middir = topdir;
578                 diroff = idx/ENTRIES_PER_PAGE;
579         } else {
580                 dir += ENTRIES_PER_PAGE/2;
581                 dir += (idx - ENTRIES_PER_PAGEPAGE/2)/ENTRIES_PER_PAGEPAGE;
582                 while (stage <= idx)
583                         stage += ENTRIES_PER_PAGEPAGE;
584                 middir = *dir;
585                 if (*dir) {
586                         diroff = ((idx - ENTRIES_PER_PAGEPAGE/2) %
587                                 ENTRIES_PER_PAGEPAGE) / ENTRIES_PER_PAGE;
588                         if (!diroff && !offset && upper_limit >= stage) {
589                                 if (needs_lock) {
590                                         spin_lock(needs_lock);
591                                         *dir = NULL;
592                                         spin_unlock(needs_lock);
593                                         needs_lock = NULL;
594                                 } else
595                                         *dir = NULL;
596                                 nr_pages_to_free++;
597                                 list_add(&middir->lru, &pages_to_free);
598                         }
599                         shmem_dir_unmap(dir);
600                         dir = shmem_dir_map(middir);
601                 } else {
602                         diroff = 0;
603                         offset = 0;
604                         idx = stage;
605                 }
606         }
607
608         for (; idx < limit; idx += ENTRIES_PER_PAGE, diroff++) {
609                 if (unlikely(idx == stage)) {
610                         shmem_dir_unmap(dir);
611                         dir = shmem_dir_map(topdir) +
612                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
613                         while (!*dir) {
614                                 dir++;
615                                 idx += ENTRIES_PER_PAGEPAGE;
616                                 if (idx >= limit)
617                                         goto done1;
618                         }
619                         stage = idx + ENTRIES_PER_PAGEPAGE;
620                         middir = *dir;
621                         if (punch_hole)
622                                 needs_lock = &info->lock;
623                         if (upper_limit >= stage) {
624                                 if (needs_lock) {
625                                         spin_lock(needs_lock);
626                                         *dir = NULL;
627                                         spin_unlock(needs_lock);
628                                         needs_lock = NULL;
629                                 } else
630                                         *dir = NULL;
631                                 nr_pages_to_free++;
632                                 list_add(&middir->lru, &pages_to_free);
633                         }
634                         shmem_dir_unmap(dir);
635                         cond_resched();
636                         dir = shmem_dir_map(middir);
637                         diroff = 0;
638                 }
639                 punch_lock = needs_lock;
640                 subdir = dir[diroff];
641                 if (subdir && !offset && upper_limit-idx >= ENTRIES_PER_PAGE) {
642                         if (needs_lock) {
643                                 spin_lock(needs_lock);
644                                 dir[diroff] = NULL;
645                                 spin_unlock(needs_lock);
646                                 punch_lock = NULL;
647                         } else
648                                 dir[diroff] = NULL;
649                         nr_pages_to_free++;
650                         list_add(&subdir->lru, &pages_to_free);
651                 }
652                 if (subdir && page_private(subdir) /* has swap entries */) {
653                         size = limit - idx;
654                         if (size > ENTRIES_PER_PAGE)
655                                 size = ENTRIES_PER_PAGE;
656                         freed = shmem_map_and_free_swp(subdir,
657                                         offset, size, &dir, punch_lock);
658                         if (!dir)
659                                 dir = shmem_dir_map(middir);
660                         nr_swaps_freed += freed;
661                         if (offset || punch_lock) {
662                                 spin_lock(&info->lock);
663                                 set_page_private(subdir,
664                                         page_private(subdir) - freed);
665                                 spin_unlock(&info->lock);
666                         } else
667                                 BUG_ON(page_private(subdir) != freed);
668                 }
669                 offset = 0;
670         }
671 done1:
672         shmem_dir_unmap(dir);
673 done2:
674         if (inode->i_mapping->nrpages && (info->flags & SHMEM_PAGEIN)) {
675                 /*
676                  * Call truncate_inode_pages again: racing shmem_unuse_inode
677                  * may have swizzled a page in from swap since vmtruncate or
678                  * generic_delete_inode did it, before we lowered next_index.
679                  * Also, though shmem_getpage checks i_size before adding to
680                  * cache, no recheck after: so fix the narrow window there too.
681                  *
682                  * Recalling truncate_inode_pages_range and unmap_mapping_range
683                  * every time for punch_hole (which never got a chance to clear
684                  * SHMEM_PAGEIN at the start of vmtruncate_range) is expensive,
685                  * yet hardly ever necessary: try to optimize them out later.
686                  */
687                 truncate_inode_pages_range(inode->i_mapping, start, end);
688                 if (punch_hole)
689                         unmap_mapping_range(inode->i_mapping, start,
690                                                         end - start, 1);
691         }
692
693         spin_lock(&info->lock);
694         info->flags &= ~SHMEM_TRUNCATE;
695         info->swapped -= nr_swaps_freed;
696         if (nr_pages_to_free)
697                 shmem_free_blocks(inode, nr_pages_to_free);
698         shmem_recalc_inode(inode);
699         spin_unlock(&info->lock);
700
701         /*
702          * Empty swap vector directory pages to be freed?
703          */
704         if (!list_empty(&pages_to_free)) {
705                 pages_to_free.prev->next = NULL;
706                 shmem_free_pages(pages_to_free.next);
707         }
708 }
709
710 static void shmem_truncate(struct inode *inode)
711 {
712         shmem_truncate_range(inode, inode->i_size, (loff_t)-1);
713 }
714
715 static int shmem_notify_change(struct dentry *dentry, struct iattr *attr)
716 {
717         struct inode *inode = dentry->d_inode;
718         struct page *page = NULL;
719         int error;
720
721         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
722                 if (attr->ia_size < inode->i_size) {
723                         /*
724                          * If truncating down to a partial page, then
725                          * if that page is already allocated, hold it
726                          * in memory until the truncation is over, so
727                          * truncate_partial_page cannnot miss it were
728                          * it assigned to swap.
729                          */
730                         if (attr->ia_size & (PAGE_CACHE_SIZE-1)) {
731                                 (void) shmem_getpage(inode,
732                                         attr->ia_size>>PAGE_CACHE_SHIFT,
733                                                 &page, SGP_READ, NULL);
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_async(struct shared_policy *p,
1029                                        swp_entry_t entry, unsigned long idx)
1030 {
1031         struct page *page;
1032         struct vm_area_struct pvma;
1033
1034         /* Create a pseudo vma that just contains the policy */
1035         memset(&pvma, 0, sizeof(struct vm_area_struct));
1036         pvma.vm_end = PAGE_SIZE;
1037         pvma.vm_pgoff = idx;
1038         pvma.vm_policy = mpol_shared_policy_lookup(p, idx);
1039         page = read_swap_cache_async(entry, &pvma, 0);
1040         mpol_free(pvma.vm_policy);
1041         return page;
1042 }
1043
1044 static struct page *shmem_swapin(struct shmem_inode_info *info,
1045                                  swp_entry_t entry, unsigned long idx)
1046 {
1047         struct shared_policy *p = &info->policy;
1048         int i, num;
1049         struct page *page;
1050         unsigned long offset;
1051
1052         num = valid_swaphandles(entry, &offset);
1053         for (i = 0; i < num; offset++, i++) {
1054                 page = shmem_swapin_async(p,
1055                                 swp_entry(swp_type(entry), offset), idx);
1056                 if (!page)
1057                         break;
1058                 page_cache_release(page);
1059         }
1060         lru_add_drain();        /* Push any new pages onto the LRU now */
1061         return shmem_swapin_async(p, entry, idx);
1062 }
1063
1064 static struct page *
1065 shmem_alloc_page(gfp_t gfp, struct shmem_inode_info *info,
1066                  unsigned long idx)
1067 {
1068         struct vm_area_struct pvma;
1069         struct page *page;
1070
1071         memset(&pvma, 0, sizeof(struct vm_area_struct));
1072         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, idx);
1073         pvma.vm_pgoff = idx;
1074         pvma.vm_end = PAGE_SIZE;
1075         page = alloc_page_vma(gfp, &pvma, 0);
1076         mpol_free(pvma.vm_policy);
1077         return page;
1078 }
1079 #else
1080 static inline int shmem_parse_mpol(char *value, int *policy,
1081                                                 nodemask_t *policy_nodes)
1082 {
1083         return 1;
1084 }
1085
1086 static inline struct page *
1087 shmem_swapin(struct shmem_inode_info *info,swp_entry_t entry,unsigned long idx)
1088 {
1089         swapin_readahead(entry, 0, NULL);
1090         return read_swap_cache_async(entry, NULL, 0);
1091 }
1092
1093 static inline struct page *
1094 shmem_alloc_page(gfp_t gfp,struct shmem_inode_info *info, unsigned long idx)
1095 {
1096         return alloc_page(gfp);
1097 }
1098 #endif
1099
1100 /*
1101  * shmem_getpage - either get the page from swap or allocate a new one
1102  *
1103  * If we allocate a new one we do not mark it dirty. That's up to the
1104  * vm. If we swap it in we mark it dirty since we also free the swap
1105  * entry since a page cannot live in both the swap and page cache
1106  */
1107 static int shmem_getpage(struct inode *inode, unsigned long idx,
1108                         struct page **pagep, enum sgp_type sgp, int *type)
1109 {
1110         struct address_space *mapping = inode->i_mapping;
1111         struct shmem_inode_info *info = SHMEM_I(inode);
1112         struct shmem_sb_info *sbinfo;
1113         struct page *filepage = *pagep;
1114         struct page *swappage;
1115         swp_entry_t *entry;
1116         swp_entry_t swap;
1117         int error;
1118
1119         if (idx >= SHMEM_MAX_INDEX)
1120                 return -EFBIG;
1121
1122         if (type)
1123                 *type = 0;
1124
1125         /*
1126          * Normally, filepage is NULL on entry, and either found
1127          * uptodate immediately, or allocated and zeroed, or read
1128          * in under swappage, which is then assigned to filepage.
1129          * But shmem_readpage and shmem_write_begin pass in a locked
1130          * filepage, which may be found not uptodate by other callers
1131          * too, and may need to be copied from the swappage read in.
1132          */
1133 repeat:
1134         if (!filepage)
1135                 filepage = find_lock_page(mapping, idx);
1136         if (filepage && PageUptodate(filepage))
1137                 goto done;
1138         error = 0;
1139         if (sgp == SGP_QUICK)
1140                 goto failed;
1141
1142         spin_lock(&info->lock);
1143         shmem_recalc_inode(inode);
1144         entry = shmem_swp_alloc(info, idx, sgp);
1145         if (IS_ERR(entry)) {
1146                 spin_unlock(&info->lock);
1147                 error = PTR_ERR(entry);
1148                 goto failed;
1149         }
1150         swap = *entry;
1151
1152         if (swap.val) {
1153                 /* Look it up and read it in.. */
1154                 swappage = lookup_swap_cache(swap);
1155                 if (!swappage) {
1156                         shmem_swp_unmap(entry);
1157                         /* here we actually do the io */
1158                         if (type && !(*type & VM_FAULT_MAJOR)) {
1159                                 __count_vm_event(PGMAJFAULT);
1160                                 *type |= VM_FAULT_MAJOR;
1161                         }
1162                         spin_unlock(&info->lock);
1163                         swappage = shmem_swapin(info, swap, idx);
1164                         if (!swappage) {
1165                                 spin_lock(&info->lock);
1166                                 entry = shmem_swp_alloc(info, idx, sgp);
1167                                 if (IS_ERR(entry))
1168                                         error = PTR_ERR(entry);
1169                                 else {
1170                                         if (entry->val == swap.val)
1171                                                 error = -ENOMEM;
1172                                         shmem_swp_unmap(entry);
1173                                 }
1174                                 spin_unlock(&info->lock);
1175                                 if (error)
1176                                         goto failed;
1177                                 goto repeat;
1178                         }
1179                         wait_on_page_locked(swappage);
1180                         page_cache_release(swappage);
1181                         goto repeat;
1182                 }
1183
1184                 /* We have to do this with page locked to prevent races */
1185                 if (TestSetPageLocked(swappage)) {
1186                         shmem_swp_unmap(entry);
1187                         spin_unlock(&info->lock);
1188                         wait_on_page_locked(swappage);
1189                         page_cache_release(swappage);
1190                         goto repeat;
1191                 }
1192                 if (PageWriteback(swappage)) {
1193                         shmem_swp_unmap(entry);
1194                         spin_unlock(&info->lock);
1195                         wait_on_page_writeback(swappage);
1196                         unlock_page(swappage);
1197                         page_cache_release(swappage);
1198                         goto repeat;
1199                 }
1200                 if (!PageUptodate(swappage)) {
1201                         shmem_swp_unmap(entry);
1202                         spin_unlock(&info->lock);
1203                         unlock_page(swappage);
1204                         page_cache_release(swappage);
1205                         error = -EIO;
1206                         goto failed;
1207                 }
1208
1209                 if (filepage) {
1210                         shmem_swp_set(info, entry, 0);
1211                         shmem_swp_unmap(entry);
1212                         delete_from_swap_cache(swappage);
1213                         spin_unlock(&info->lock);
1214                         copy_highpage(filepage, swappage);
1215                         unlock_page(swappage);
1216                         page_cache_release(swappage);
1217                         flush_dcache_page(filepage);
1218                         SetPageUptodate(filepage);
1219                         set_page_dirty(filepage);
1220                         swap_free(swap);
1221                 } else if (!(error = move_from_swap_cache(
1222                                 swappage, idx, mapping))) {
1223                         info->flags |= SHMEM_PAGEIN;
1224                         shmem_swp_set(info, entry, 0);
1225                         shmem_swp_unmap(entry);
1226                         spin_unlock(&info->lock);
1227                         filepage = swappage;
1228                         swap_free(swap);
1229                 } else {
1230                         shmem_swp_unmap(entry);
1231                         spin_unlock(&info->lock);
1232                         unlock_page(swappage);
1233                         page_cache_release(swappage);
1234                         if (error == -ENOMEM) {
1235                                 /* let kswapd refresh zone for GFP_ATOMICs */
1236                                 congestion_wait(WRITE, HZ/50);
1237                         }
1238                         goto repeat;
1239                 }
1240         } else if (sgp == SGP_READ && !filepage) {
1241                 shmem_swp_unmap(entry);
1242                 filepage = find_get_page(mapping, idx);
1243                 if (filepage &&
1244                     (!PageUptodate(filepage) || TestSetPageLocked(filepage))) {
1245                         spin_unlock(&info->lock);
1246                         wait_on_page_locked(filepage);
1247                         page_cache_release(filepage);
1248                         filepage = NULL;
1249                         goto repeat;
1250                 }
1251                 spin_unlock(&info->lock);
1252         } else {
1253                 shmem_swp_unmap(entry);
1254                 sbinfo = SHMEM_SB(inode->i_sb);
1255                 if (sbinfo->max_blocks) {
1256                         spin_lock(&sbinfo->stat_lock);
1257                         if (sbinfo->free_blocks == 0 ||
1258                             shmem_acct_block(info->flags)) {
1259                                 spin_unlock(&sbinfo->stat_lock);
1260                                 spin_unlock(&info->lock);
1261                                 error = -ENOSPC;
1262                                 goto failed;
1263                         }
1264                         sbinfo->free_blocks--;
1265                         inode->i_blocks += BLOCKS_PER_PAGE;
1266                         spin_unlock(&sbinfo->stat_lock);
1267                 } else if (shmem_acct_block(info->flags)) {
1268                         spin_unlock(&info->lock);
1269                         error = -ENOSPC;
1270                         goto failed;
1271                 }
1272
1273                 if (!filepage) {
1274                         spin_unlock(&info->lock);
1275                         filepage = shmem_alloc_page(mapping_gfp_mask(mapping),
1276                                                     info,
1277                                                     idx);
1278                         if (!filepage) {
1279                                 shmem_unacct_blocks(info->flags, 1);
1280                                 shmem_free_blocks(inode, 1);
1281                                 error = -ENOMEM;
1282                                 goto failed;
1283                         }
1284
1285                         spin_lock(&info->lock);
1286                         entry = shmem_swp_alloc(info, idx, sgp);
1287                         if (IS_ERR(entry))
1288                                 error = PTR_ERR(entry);
1289                         else {
1290                                 swap = *entry;
1291                                 shmem_swp_unmap(entry);
1292                         }
1293                         if (error || swap.val || 0 != add_to_page_cache_lru(
1294                                         filepage, mapping, idx, GFP_ATOMIC)) {
1295                                 spin_unlock(&info->lock);
1296                                 page_cache_release(filepage);
1297                                 shmem_unacct_blocks(info->flags, 1);
1298                                 shmem_free_blocks(inode, 1);
1299                                 filepage = NULL;
1300                                 if (error)
1301                                         goto failed;
1302                                 goto repeat;
1303                         }
1304                         info->flags |= SHMEM_PAGEIN;
1305                 }
1306
1307                 info->alloced++;
1308                 spin_unlock(&info->lock);
1309                 clear_highpage(filepage);
1310                 flush_dcache_page(filepage);
1311                 SetPageUptodate(filepage);
1312         }
1313 done:
1314         if (*pagep != filepage) {
1315                 *pagep = filepage;
1316                 if (sgp != SGP_FAULT)
1317                         unlock_page(filepage);
1318
1319         }
1320         return 0;
1321
1322 failed:
1323         if (*pagep != filepage) {
1324                 unlock_page(filepage);
1325                 page_cache_release(filepage);
1326         }
1327         return error;
1328 }
1329
1330 static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1331 {
1332         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1333         int error;
1334         int ret;
1335
1336         if (((loff_t)vmf->pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
1337                 return VM_FAULT_SIGBUS;
1338
1339         error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_FAULT, &ret);
1340         if (error)
1341                 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1342
1343         mark_page_accessed(vmf->page);
1344         return ret | VM_FAULT_LOCKED;
1345 }
1346
1347 #ifdef CONFIG_NUMA
1348 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *new)
1349 {
1350         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1351         return mpol_set_shared_policy(&SHMEM_I(i)->policy, vma, new);
1352 }
1353
1354 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
1355                                           unsigned long addr)
1356 {
1357         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1358         unsigned long idx;
1359
1360         idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1361         return mpol_shared_policy_lookup(&SHMEM_I(i)->policy, idx);
1362 }
1363 #endif
1364
1365 int shmem_lock(struct file *file, int lock, struct user_struct *user)
1366 {
1367         struct inode *inode = file->f_path.dentry->d_inode;
1368         struct shmem_inode_info *info = SHMEM_I(inode);
1369         int retval = -ENOMEM;
1370
1371         spin_lock(&info->lock);
1372         if (lock && !(info->flags & VM_LOCKED)) {
1373                 if (!user_shm_lock(inode->i_size, user))
1374                         goto out_nomem;
1375                 info->flags |= VM_LOCKED;
1376         }
1377         if (!lock && (info->flags & VM_LOCKED) && user) {
1378                 user_shm_unlock(inode->i_size, user);
1379                 info->flags &= ~VM_LOCKED;
1380         }
1381         retval = 0;
1382 out_nomem:
1383         spin_unlock(&info->lock);
1384         return retval;
1385 }
1386
1387 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1388 {
1389         file_accessed(file);
1390         vma->vm_ops = &shmem_vm_ops;
1391         vma->vm_flags |= VM_CAN_NONLINEAR;
1392         return 0;
1393 }
1394
1395 static struct inode *
1396 shmem_get_inode(struct super_block *sb, int mode, dev_t dev)
1397 {
1398         struct inode *inode;
1399         struct shmem_inode_info *info;
1400         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1401
1402         if (sbinfo->max_inodes) {
1403                 spin_lock(&sbinfo->stat_lock);
1404                 if (!sbinfo->free_inodes) {
1405                         spin_unlock(&sbinfo->stat_lock);
1406                         return NULL;
1407                 }
1408                 sbinfo->free_inodes--;
1409                 spin_unlock(&sbinfo->stat_lock);
1410         }
1411
1412         inode = new_inode(sb);
1413         if (inode) {
1414                 inode->i_mode = mode;
1415                 inode->i_uid = current->fsuid;
1416                 inode->i_gid = current->fsgid;
1417                 inode->i_blocks = 0;
1418                 inode->i_mapping->a_ops = &shmem_aops;
1419                 inode->i_mapping->backing_dev_info = &shmem_backing_dev_info;
1420                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1421                 inode->i_generation = get_seconds();
1422                 info = SHMEM_I(inode);
1423                 memset(info, 0, (char *)inode - (char *)info);
1424                 spin_lock_init(&info->lock);
1425                 INIT_LIST_HEAD(&info->swaplist);
1426
1427                 switch (mode & S_IFMT) {
1428                 default:
1429                         inode->i_op = &shmem_special_inode_operations;
1430                         init_special_inode(inode, mode, dev);
1431                         break;
1432                 case S_IFREG:
1433                         inode->i_op = &shmem_inode_operations;
1434                         inode->i_fop = &shmem_file_operations;
1435                         mpol_shared_policy_init(&info->policy, sbinfo->policy,
1436                                                         &sbinfo->policy_nodes);
1437                         break;
1438                 case S_IFDIR:
1439                         inc_nlink(inode);
1440                         /* Some things misbehave if size == 0 on a directory */
1441                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
1442                         inode->i_op = &shmem_dir_inode_operations;
1443                         inode->i_fop = &simple_dir_operations;
1444                         break;
1445                 case S_IFLNK:
1446                         /*
1447                          * Must not load anything in the rbtree,
1448                          * mpol_free_shared_policy will not be called.
1449                          */
1450                         mpol_shared_policy_init(&info->policy, MPOL_DEFAULT,
1451                                                 NULL);
1452                         break;
1453                 }
1454         } else if (sbinfo->max_inodes) {
1455                 spin_lock(&sbinfo->stat_lock);
1456                 sbinfo->free_inodes++;
1457                 spin_unlock(&sbinfo->stat_lock);
1458         }
1459         return inode;
1460 }
1461
1462 #ifdef CONFIG_TMPFS
1463 static const struct inode_operations shmem_symlink_inode_operations;
1464 static const struct inode_operations shmem_symlink_inline_operations;
1465
1466 /*
1467  * Normally tmpfs avoids the use of shmem_readpage and shmem_write_begin;
1468  * but providing them allows a tmpfs file to be used for splice, sendfile, and
1469  * below the loop driver, in the generic fashion that many filesystems support.
1470  */
1471 static int shmem_readpage(struct file *file, struct page *page)
1472 {
1473         struct inode *inode = page->mapping->host;
1474         int error = shmem_getpage(inode, page->index, &page, SGP_CACHE, NULL);
1475         unlock_page(page);
1476         return error;
1477 }
1478
1479 static int
1480 shmem_write_begin(struct file *file, struct address_space *mapping,
1481                         loff_t pos, unsigned len, unsigned flags,
1482                         struct page **pagep, void **fsdata)
1483 {
1484         struct inode *inode = mapping->host;
1485         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1486         *pagep = NULL;
1487         return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1488 }
1489
1490 static int
1491 shmem_write_end(struct file *file, struct address_space *mapping,
1492                         loff_t pos, unsigned len, unsigned copied,
1493                         struct page *page, void *fsdata)
1494 {
1495         struct inode *inode = mapping->host;
1496
1497         set_page_dirty(page);
1498         page_cache_release(page);
1499
1500         if (pos+copied > inode->i_size)
1501                 i_size_write(inode, pos+copied);
1502
1503         return copied;
1504 }
1505
1506 static ssize_t
1507 shmem_file_write(struct file *file, const char __user *buf, size_t count, loff_t *ppos)
1508 {
1509         struct inode    *inode = file->f_path.dentry->d_inode;
1510         loff_t          pos;
1511         unsigned long   written;
1512         ssize_t         err;
1513
1514         if ((ssize_t) count < 0)
1515                 return -EINVAL;
1516
1517         if (!access_ok(VERIFY_READ, buf, count))
1518                 return -EFAULT;
1519
1520         mutex_lock(&inode->i_mutex);
1521
1522         pos = *ppos;
1523         written = 0;
1524
1525         err = generic_write_checks(file, &pos, &count, 0);
1526         if (err || !count)
1527                 goto out;
1528
1529         err = remove_suid(file->f_path.dentry);
1530         if (err)
1531                 goto out;
1532
1533         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1534
1535         do {
1536                 struct page *page = NULL;
1537                 unsigned long bytes, index, offset;
1538                 char *kaddr;
1539                 int left;
1540
1541                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
1542                 index = pos >> PAGE_CACHE_SHIFT;
1543                 bytes = PAGE_CACHE_SIZE - offset;
1544                 if (bytes > count)
1545                         bytes = count;
1546
1547                 /*
1548                  * We don't hold page lock across copy from user -
1549                  * what would it guard against? - so no deadlock here.
1550                  * But it still may be a good idea to prefault below.
1551                  */
1552
1553                 err = shmem_getpage(inode, index, &page, SGP_WRITE, NULL);
1554                 if (err)
1555                         break;
1556
1557                 left = bytes;
1558                 if (PageHighMem(page)) {
1559                         volatile unsigned char dummy;
1560                         __get_user(dummy, buf);
1561                         __get_user(dummy, buf + bytes - 1);
1562
1563                         kaddr = kmap_atomic(page, KM_USER0);
1564                         left = __copy_from_user_inatomic(kaddr + offset,
1565                                                         buf, bytes);
1566                         kunmap_atomic(kaddr, KM_USER0);
1567                 }
1568                 if (left) {
1569                         kaddr = kmap(page);
1570                         left = __copy_from_user(kaddr + offset, buf, bytes);
1571                         kunmap(page);
1572                 }
1573
1574                 written += bytes;
1575                 count -= bytes;
1576                 pos += bytes;
1577                 buf += bytes;
1578                 if (pos > inode->i_size)
1579                         i_size_write(inode, pos);
1580
1581                 flush_dcache_page(page);
1582                 set_page_dirty(page);
1583                 mark_page_accessed(page);
1584                 page_cache_release(page);
1585
1586                 if (left) {
1587                         pos -= left;
1588                         written -= left;
1589                         err = -EFAULT;
1590                         break;
1591                 }
1592
1593                 /*
1594                  * Our dirty pages are not counted in nr_dirty,
1595                  * and we do not attempt to balance dirty pages.
1596                  */
1597
1598                 cond_resched();
1599         } while (count);
1600
1601         *ppos = pos;
1602         if (written)
1603                 err = written;
1604 out:
1605         mutex_unlock(&inode->i_mutex);
1606         return err;
1607 }
1608
1609 static void do_shmem_file_read(struct file *filp, loff_t *ppos, read_descriptor_t *desc, read_actor_t actor)
1610 {
1611         struct inode *inode = filp->f_path.dentry->d_inode;
1612         struct address_space *mapping = inode->i_mapping;
1613         unsigned long index, offset;
1614
1615         index = *ppos >> PAGE_CACHE_SHIFT;
1616         offset = *ppos & ~PAGE_CACHE_MASK;
1617
1618         for (;;) {
1619                 struct page *page = NULL;
1620                 unsigned long end_index, nr, ret;
1621                 loff_t i_size = i_size_read(inode);
1622
1623                 end_index = i_size >> PAGE_CACHE_SHIFT;
1624                 if (index > end_index)
1625                         break;
1626                 if (index == end_index) {
1627                         nr = i_size & ~PAGE_CACHE_MASK;
1628                         if (nr <= offset)
1629                                 break;
1630                 }
1631
1632                 desc->error = shmem_getpage(inode, index, &page, SGP_READ, NULL);
1633                 if (desc->error) {
1634                         if (desc->error == -EINVAL)
1635                                 desc->error = 0;
1636                         break;
1637                 }
1638
1639                 /*
1640                  * We must evaluate after, since reads (unlike writes)
1641                  * are called without i_mutex protection against truncate
1642                  */
1643                 nr = PAGE_CACHE_SIZE;
1644                 i_size = i_size_read(inode);
1645                 end_index = i_size >> PAGE_CACHE_SHIFT;
1646                 if (index == end_index) {
1647                         nr = i_size & ~PAGE_CACHE_MASK;
1648                         if (nr <= offset) {
1649                                 if (page)
1650                                         page_cache_release(page);
1651                                 break;
1652                         }
1653                 }
1654                 nr -= offset;
1655
1656                 if (page) {
1657                         /*
1658                          * If users can be writing to this page using arbitrary
1659                          * virtual addresses, take care about potential aliasing
1660                          * before reading the page on the kernel side.
1661                          */
1662                         if (mapping_writably_mapped(mapping))
1663                                 flush_dcache_page(page);
1664                         /*
1665                          * Mark the page accessed if we read the beginning.
1666                          */
1667                         if (!offset)
1668                                 mark_page_accessed(page);
1669                 } else {
1670                         page = ZERO_PAGE(0);
1671                         page_cache_get(page);
1672                 }
1673
1674                 /*
1675                  * Ok, we have the page, and it's up-to-date, so
1676                  * now we can copy it to user space...
1677                  *
1678                  * The actor routine returns how many bytes were actually used..
1679                  * NOTE! This may not be the same as how much of a user buffer
1680                  * we filled up (we may be padding etc), so we can only update
1681                  * "pos" here (the actor routine has to update the user buffer
1682                  * pointers and the remaining count).
1683                  */
1684                 ret = actor(desc, page, offset, nr);
1685                 offset += ret;
1686                 index += offset >> PAGE_CACHE_SHIFT;
1687                 offset &= ~PAGE_CACHE_MASK;
1688
1689                 page_cache_release(page);
1690                 if (ret != nr || !desc->count)
1691                         break;
1692
1693                 cond_resched();
1694         }
1695
1696         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1697         file_accessed(filp);
1698 }
1699
1700 static ssize_t shmem_file_read(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
1701 {
1702         read_descriptor_t desc;
1703
1704         if ((ssize_t) count < 0)
1705                 return -EINVAL;
1706         if (!access_ok(VERIFY_WRITE, buf, count))
1707                 return -EFAULT;
1708         if (!count)
1709                 return 0;
1710
1711         desc.written = 0;
1712         desc.count = count;
1713         desc.arg.buf = buf;
1714         desc.error = 0;
1715
1716         do_shmem_file_read(filp, ppos, &desc, file_read_actor);
1717         if (desc.written)
1718                 return desc.written;
1719         return desc.error;
1720 }
1721
1722 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
1723 {
1724         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
1725
1726         buf->f_type = TMPFS_MAGIC;
1727         buf->f_bsize = PAGE_CACHE_SIZE;
1728         buf->f_namelen = NAME_MAX;
1729         spin_lock(&sbinfo->stat_lock);
1730         if (sbinfo->max_blocks) {
1731                 buf->f_blocks = sbinfo->max_blocks;
1732                 buf->f_bavail = buf->f_bfree = sbinfo->free_blocks;
1733         }
1734         if (sbinfo->max_inodes) {
1735                 buf->f_files = sbinfo->max_inodes;
1736                 buf->f_ffree = sbinfo->free_inodes;
1737         }
1738         /* else leave those fields 0 like simple_statfs */
1739         spin_unlock(&sbinfo->stat_lock);
1740         return 0;
1741 }
1742
1743 /*
1744  * File creation. Allocate an inode, and we're done..
1745  */
1746 static int
1747 shmem_mknod(struct inode *dir, struct dentry *dentry, int mode, dev_t dev)
1748 {
1749         struct inode *inode = shmem_get_inode(dir->i_sb, mode, dev);
1750         int error = -ENOSPC;
1751
1752         if (inode) {
1753                 error = security_inode_init_security(inode, dir, NULL, NULL,
1754                                                      NULL);
1755                 if (error) {
1756                         if (error != -EOPNOTSUPP) {
1757                                 iput(inode);
1758                                 return error;
1759                         }
1760                 }
1761                 error = shmem_acl_init(inode, dir);
1762                 if (error) {
1763                         iput(inode);
1764                         return error;
1765                 }
1766                 if (dir->i_mode & S_ISGID) {
1767                         inode->i_gid = dir->i_gid;
1768                         if (S_ISDIR(mode))
1769                                 inode->i_mode |= S_ISGID;
1770                 }
1771                 dir->i_size += BOGO_DIRENT_SIZE;
1772                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1773                 d_instantiate(dentry, inode);
1774                 dget(dentry); /* Extra count - pin the dentry in core */
1775         }
1776         return error;
1777 }
1778
1779 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, int mode)
1780 {
1781         int error;
1782
1783         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
1784                 return error;
1785         inc_nlink(dir);
1786         return 0;
1787 }
1788
1789 static int shmem_create(struct inode *dir, struct dentry *dentry, int mode,
1790                 struct nameidata *nd)
1791 {
1792         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
1793 }
1794
1795 /*
1796  * Link a file..
1797  */
1798 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
1799 {
1800         struct inode *inode = old_dentry->d_inode;
1801         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1802
1803         /*
1804          * No ordinary (disk based) filesystem counts links as inodes;
1805          * but each new link needs a new dentry, pinning lowmem, and
1806          * tmpfs dentries cannot be pruned until they are unlinked.
1807          */
1808         if (sbinfo->max_inodes) {
1809                 spin_lock(&sbinfo->stat_lock);
1810                 if (!sbinfo->free_inodes) {
1811                         spin_unlock(&sbinfo->stat_lock);
1812                         return -ENOSPC;
1813                 }
1814                 sbinfo->free_inodes--;
1815                 spin_unlock(&sbinfo->stat_lock);
1816         }
1817
1818         dir->i_size += BOGO_DIRENT_SIZE;
1819         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1820         inc_nlink(inode);
1821         atomic_inc(&inode->i_count);    /* New dentry reference */
1822         dget(dentry);           /* Extra pinning count for the created dentry */
1823         d_instantiate(dentry, inode);
1824         return 0;
1825 }
1826
1827 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
1828 {
1829         struct inode *inode = dentry->d_inode;
1830
1831         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode)) {
1832                 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1833                 if (sbinfo->max_inodes) {
1834                         spin_lock(&sbinfo->stat_lock);
1835                         sbinfo->free_inodes++;
1836                         spin_unlock(&sbinfo->stat_lock);
1837                 }
1838         }
1839
1840         dir->i_size -= BOGO_DIRENT_SIZE;
1841         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1842         drop_nlink(inode);
1843         dput(dentry);   /* Undo the count from "create" - this does all the work */
1844         return 0;
1845 }
1846
1847 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
1848 {
1849         if (!simple_empty(dentry))
1850                 return -ENOTEMPTY;
1851
1852         drop_nlink(dentry->d_inode);
1853         drop_nlink(dir);
1854         return shmem_unlink(dir, dentry);
1855 }
1856
1857 /*
1858  * The VFS layer already does all the dentry stuff for rename,
1859  * we just have to decrement the usage count for the target if
1860  * it exists so that the VFS layer correctly free's it when it
1861  * gets overwritten.
1862  */
1863 static int shmem_rename(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
1864 {
1865         struct inode *inode = old_dentry->d_inode;
1866         int they_are_dirs = S_ISDIR(inode->i_mode);
1867
1868         if (!simple_empty(new_dentry))
1869                 return -ENOTEMPTY;
1870
1871         if (new_dentry->d_inode) {
1872                 (void) shmem_unlink(new_dir, new_dentry);
1873                 if (they_are_dirs)
1874                         drop_nlink(old_dir);
1875         } else if (they_are_dirs) {
1876                 drop_nlink(old_dir);
1877                 inc_nlink(new_dir);
1878         }
1879
1880         old_dir->i_size -= BOGO_DIRENT_SIZE;
1881         new_dir->i_size += BOGO_DIRENT_SIZE;
1882         old_dir->i_ctime = old_dir->i_mtime =
1883         new_dir->i_ctime = new_dir->i_mtime =
1884         inode->i_ctime = CURRENT_TIME;
1885         return 0;
1886 }
1887
1888 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
1889 {
1890         int error;
1891         int len;
1892         struct inode *inode;
1893         struct page *page = NULL;
1894         char *kaddr;
1895         struct shmem_inode_info *info;
1896
1897         len = strlen(symname) + 1;
1898         if (len > PAGE_CACHE_SIZE)
1899                 return -ENAMETOOLONG;
1900
1901         inode = shmem_get_inode(dir->i_sb, S_IFLNK|S_IRWXUGO, 0);
1902         if (!inode)
1903                 return -ENOSPC;
1904
1905         error = security_inode_init_security(inode, dir, NULL, NULL,
1906                                              NULL);
1907         if (error) {
1908                 if (error != -EOPNOTSUPP) {
1909                         iput(inode);
1910                         return error;
1911                 }
1912                 error = 0;
1913         }
1914
1915         info = SHMEM_I(inode);
1916         inode->i_size = len-1;
1917         if (len <= (char *)inode - (char *)info) {
1918                 /* do it inline */
1919                 memcpy(info, symname, len);
1920                 inode->i_op = &shmem_symlink_inline_operations;
1921         } else {
1922                 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
1923                 if (error) {
1924                         iput(inode);
1925                         return error;
1926                 }
1927                 inode->i_op = &shmem_symlink_inode_operations;
1928                 kaddr = kmap_atomic(page, KM_USER0);
1929                 memcpy(kaddr, symname, len);
1930                 kunmap_atomic(kaddr, KM_USER0);
1931                 set_page_dirty(page);
1932                 page_cache_release(page);
1933         }
1934         if (dir->i_mode & S_ISGID)
1935                 inode->i_gid = dir->i_gid;
1936         dir->i_size += BOGO_DIRENT_SIZE;
1937         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1938         d_instantiate(dentry, inode);
1939         dget(dentry);
1940         return 0;
1941 }
1942
1943 static void *shmem_follow_link_inline(struct dentry *dentry, struct nameidata *nd)
1944 {
1945         nd_set_link(nd, (char *)SHMEM_I(dentry->d_inode));
1946         return NULL;
1947 }
1948
1949 static void *shmem_follow_link(struct dentry *dentry, struct nameidata *nd)
1950 {
1951         struct page *page = NULL;
1952         int res = shmem_getpage(dentry->d_inode, 0, &page, SGP_READ, NULL);
1953         nd_set_link(nd, res ? ERR_PTR(res) : kmap(page));
1954         return page;
1955 }
1956
1957 static void shmem_put_link(struct dentry *dentry, struct nameidata *nd, void *cookie)
1958 {
1959         if (!IS_ERR(nd_get_link(nd))) {
1960                 struct page *page = cookie;
1961                 kunmap(page);
1962                 mark_page_accessed(page);
1963                 page_cache_release(page);
1964         }
1965 }
1966
1967 static const struct inode_operations shmem_symlink_inline_operations = {
1968         .readlink       = generic_readlink,
1969         .follow_link    = shmem_follow_link_inline,
1970 };
1971
1972 static const struct inode_operations shmem_symlink_inode_operations = {
1973         .truncate       = shmem_truncate,
1974         .readlink       = generic_readlink,
1975         .follow_link    = shmem_follow_link,
1976         .put_link       = shmem_put_link,
1977 };
1978
1979 #ifdef CONFIG_TMPFS_POSIX_ACL
1980 /**
1981  * Superblocks without xattr inode operations will get security.* xattr
1982  * support from the VFS "for free". As soon as we have any other xattrs
1983  * like ACLs, we also need to implement the security.* handlers at
1984  * filesystem level, though.
1985  */
1986
1987 static size_t shmem_xattr_security_list(struct inode *inode, char *list,
1988                                         size_t list_len, const char *name,
1989                                         size_t name_len)
1990 {
1991         return security_inode_listsecurity(inode, list, list_len);
1992 }
1993
1994 static int shmem_xattr_security_get(struct inode *inode, const char *name,
1995                                     void *buffer, size_t size)
1996 {
1997         if (strcmp(name, "") == 0)
1998                 return -EINVAL;
1999         return security_inode_getsecurity(inode, name, buffer, size,
2000                                           -EOPNOTSUPP);
2001 }
2002
2003 static int shmem_xattr_security_set(struct inode *inode, const char *name,
2004                                     const void *value, size_t size, int flags)
2005 {
2006         if (strcmp(name, "") == 0)
2007                 return -EINVAL;
2008         return security_inode_setsecurity(inode, name, value, size, flags);
2009 }
2010
2011 static struct xattr_handler shmem_xattr_security_handler = {
2012         .prefix = XATTR_SECURITY_PREFIX,
2013         .list   = shmem_xattr_security_list,
2014         .get    = shmem_xattr_security_get,
2015         .set    = shmem_xattr_security_set,
2016 };
2017
2018 static struct xattr_handler *shmem_xattr_handlers[] = {
2019         &shmem_xattr_acl_access_handler,
2020         &shmem_xattr_acl_default_handler,
2021         &shmem_xattr_security_handler,
2022         NULL
2023 };
2024 #endif
2025
2026 static struct dentry *shmem_get_parent(struct dentry *child)
2027 {
2028         return ERR_PTR(-ESTALE);
2029 }
2030
2031 static int shmem_match(struct inode *ino, void *vfh)
2032 {
2033         __u32 *fh = vfh;
2034         __u64 inum = fh[2];
2035         inum = (inum << 32) | fh[1];
2036         return ino->i_ino == inum && fh[0] == ino->i_generation;
2037 }
2038
2039 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
2040                 struct fid *fid, int fh_len, int fh_type)
2041 {
2042         struct inode *inode;
2043         struct dentry *dentry = NULL;
2044         u64 inum = fid->raw[2];
2045         inum = (inum << 32) | fid->raw[1];
2046
2047         if (fh_len < 3)
2048                 return NULL;
2049
2050         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
2051                         shmem_match, fid->raw);
2052         if (inode) {
2053                 dentry = d_find_alias(inode);
2054                 iput(inode);
2055         }
2056
2057         return dentry;
2058 }
2059
2060 static int shmem_encode_fh(struct dentry *dentry, __u32 *fh, int *len,
2061                                 int connectable)
2062 {
2063         struct inode *inode = dentry->d_inode;
2064
2065         if (*len < 3)
2066                 return 255;
2067
2068         if (hlist_unhashed(&inode->i_hash)) {
2069                 /* Unfortunately insert_inode_hash is not idempotent,
2070                  * so as we hash inodes here rather than at creation
2071                  * time, we need a lock to ensure we only try
2072                  * to do it once
2073                  */
2074                 static DEFINE_SPINLOCK(lock);
2075                 spin_lock(&lock);
2076                 if (hlist_unhashed(&inode->i_hash))
2077                         __insert_inode_hash(inode,
2078                                             inode->i_ino + inode->i_generation);
2079                 spin_unlock(&lock);
2080         }
2081
2082         fh[0] = inode->i_generation;
2083         fh[1] = inode->i_ino;
2084         fh[2] = ((__u64)inode->i_ino) >> 32;
2085
2086         *len = 3;
2087         return 1;
2088 }
2089
2090 static const struct export_operations shmem_export_ops = {
2091         .get_parent     = shmem_get_parent,
2092         .encode_fh      = shmem_encode_fh,
2093         .fh_to_dentry   = shmem_fh_to_dentry,
2094 };
2095
2096 static int shmem_parse_options(char *options, int *mode, uid_t *uid,
2097         gid_t *gid, unsigned long *blocks, unsigned long *inodes,
2098         int *policy, nodemask_t *policy_nodes)
2099 {
2100         char *this_char, *value, *rest;
2101
2102         while (options != NULL) {
2103                 this_char = options;
2104                 for (;;) {
2105                         /*
2106                          * NUL-terminate this option: unfortunately,
2107                          * mount options form a comma-separated list,
2108                          * but mpol's nodelist may also contain commas.
2109                          */
2110                         options = strchr(options, ',');
2111                         if (options == NULL)
2112                                 break;
2113                         options++;
2114                         if (!isdigit(*options)) {
2115                                 options[-1] = '\0';
2116                                 break;
2117                         }
2118                 }
2119                 if (!*this_char)
2120                         continue;
2121                 if ((value = strchr(this_char,'=')) != NULL) {
2122                         *value++ = 0;
2123                 } else {
2124                         printk(KERN_ERR
2125                             "tmpfs: No value for mount option '%s'\n",
2126                             this_char);
2127                         return 1;
2128                 }
2129
2130                 if (!strcmp(this_char,"size")) {
2131                         unsigned long long size;
2132                         size = memparse(value,&rest);
2133                         if (*rest == '%') {
2134                                 size <<= PAGE_SHIFT;
2135                                 size *= totalram_pages;
2136                                 do_div(size, 100);
2137                                 rest++;
2138                         }
2139                         if (*rest)
2140                                 goto bad_val;
2141                         *blocks = size >> PAGE_CACHE_SHIFT;
2142                 } else if (!strcmp(this_char,"nr_blocks")) {
2143                         *blocks = memparse(value,&rest);
2144                         if (*rest)
2145                                 goto bad_val;
2146                 } else if (!strcmp(this_char,"nr_inodes")) {
2147                         *inodes = memparse(value,&rest);
2148                         if (*rest)
2149                                 goto bad_val;
2150                 } else if (!strcmp(this_char,"mode")) {
2151                         if (!mode)
2152                                 continue;
2153                         *mode = simple_strtoul(value,&rest,8);
2154                         if (*rest)
2155                                 goto bad_val;
2156                 } else if (!strcmp(this_char,"uid")) {
2157                         if (!uid)
2158                                 continue;
2159                         *uid = simple_strtoul(value,&rest,0);
2160                         if (*rest)
2161                                 goto bad_val;
2162                 } else if (!strcmp(this_char,"gid")) {
2163                         if (!gid)
2164                                 continue;
2165                         *gid = simple_strtoul(value,&rest,0);
2166                         if (*rest)
2167                                 goto bad_val;
2168                 } else if (!strcmp(this_char,"mpol")) {
2169                         if (shmem_parse_mpol(value,policy,policy_nodes))
2170                                 goto bad_val;
2171                 } else {
2172                         printk(KERN_ERR "tmpfs: Bad mount option %s\n",
2173                                this_char);
2174                         return 1;
2175                 }
2176         }
2177         return 0;
2178
2179 bad_val:
2180         printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
2181                value, this_char);
2182         return 1;
2183
2184 }
2185
2186 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2187 {
2188         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2189         unsigned long max_blocks = sbinfo->max_blocks;
2190         unsigned long max_inodes = sbinfo->max_inodes;
2191         int policy = sbinfo->policy;
2192         nodemask_t policy_nodes = sbinfo->policy_nodes;
2193         unsigned long blocks;
2194         unsigned long inodes;
2195         int error = -EINVAL;
2196
2197         if (shmem_parse_options(data, NULL, NULL, NULL, &max_blocks,
2198                                 &max_inodes, &policy, &policy_nodes))
2199                 return error;
2200
2201         spin_lock(&sbinfo->stat_lock);
2202         blocks = sbinfo->max_blocks - sbinfo->free_blocks;
2203         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2204         if (max_blocks < blocks)
2205                 goto out;
2206         if (max_inodes < inodes)
2207                 goto out;
2208         /*
2209          * Those tests also disallow limited->unlimited while any are in
2210          * use, so i_blocks will always be zero when max_blocks is zero;
2211          * but we must separately disallow unlimited->limited, because
2212          * in that case we have no record of how much is already in use.
2213          */
2214         if (max_blocks && !sbinfo->max_blocks)
2215                 goto out;
2216         if (max_inodes && !sbinfo->max_inodes)
2217                 goto out;
2218
2219         error = 0;
2220         sbinfo->max_blocks  = max_blocks;
2221         sbinfo->free_blocks = max_blocks - blocks;
2222         sbinfo->max_inodes  = max_inodes;
2223         sbinfo->free_inodes = max_inodes - inodes;
2224         sbinfo->policy = policy;
2225         sbinfo->policy_nodes = policy_nodes;
2226 out:
2227         spin_unlock(&sbinfo->stat_lock);
2228         return error;
2229 }
2230 #endif
2231
2232 static void shmem_put_super(struct super_block *sb)
2233 {
2234         kfree(sb->s_fs_info);
2235         sb->s_fs_info = NULL;
2236 }
2237
2238 static int shmem_fill_super(struct super_block *sb,
2239                             void *data, int silent)
2240 {
2241         struct inode *inode;
2242         struct dentry *root;
2243         int mode   = S_IRWXUGO | S_ISVTX;
2244         uid_t uid = current->fsuid;
2245         gid_t gid = current->fsgid;
2246         int err = -ENOMEM;
2247         struct shmem_sb_info *sbinfo;
2248         unsigned long blocks = 0;
2249         unsigned long inodes = 0;
2250         int policy = MPOL_DEFAULT;
2251         nodemask_t policy_nodes = node_states[N_HIGH_MEMORY];
2252
2253 #ifdef CONFIG_TMPFS
2254         /*
2255          * Per default we only allow half of the physical ram per
2256          * tmpfs instance, limiting inodes to one per page of lowmem;
2257          * but the internal instance is left unlimited.
2258          */
2259         if (!(sb->s_flags & MS_NOUSER)) {
2260                 blocks = totalram_pages / 2;
2261                 inodes = totalram_pages - totalhigh_pages;
2262                 if (inodes > blocks)
2263                         inodes = blocks;
2264                 if (shmem_parse_options(data, &mode, &uid, &gid, &blocks,
2265                                         &inodes, &policy, &policy_nodes))
2266                         return -EINVAL;
2267         }
2268         sb->s_export_op = &shmem_export_ops;
2269 #else
2270         sb->s_flags |= MS_NOUSER;
2271 #endif
2272
2273         /* Round up to L1_CACHE_BYTES to resist false sharing */
2274         sbinfo = kmalloc(max((int)sizeof(struct shmem_sb_info),
2275                                 L1_CACHE_BYTES), GFP_KERNEL);
2276         if (!sbinfo)
2277                 return -ENOMEM;
2278
2279         spin_lock_init(&sbinfo->stat_lock);
2280         sbinfo->max_blocks = blocks;
2281         sbinfo->free_blocks = blocks;
2282         sbinfo->max_inodes = inodes;
2283         sbinfo->free_inodes = inodes;
2284         sbinfo->policy = policy;
2285         sbinfo->policy_nodes = policy_nodes;
2286
2287         sb->s_fs_info = sbinfo;
2288         sb->s_maxbytes = SHMEM_MAX_BYTES;
2289         sb->s_blocksize = PAGE_CACHE_SIZE;
2290         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
2291         sb->s_magic = TMPFS_MAGIC;
2292         sb->s_op = &shmem_ops;
2293         sb->s_time_gran = 1;
2294 #ifdef CONFIG_TMPFS_POSIX_ACL
2295         sb->s_xattr = shmem_xattr_handlers;
2296         sb->s_flags |= MS_POSIXACL;
2297 #endif
2298
2299         inode = shmem_get_inode(sb, S_IFDIR | mode, 0);
2300         if (!inode)
2301                 goto failed;
2302         inode->i_uid = uid;
2303         inode->i_gid = gid;
2304         root = d_alloc_root(inode);
2305         if (!root)
2306                 goto failed_iput;
2307         sb->s_root = root;
2308         return 0;
2309
2310 failed_iput:
2311         iput(inode);
2312 failed:
2313         shmem_put_super(sb);
2314         return err;
2315 }
2316
2317 static struct kmem_cache *shmem_inode_cachep;
2318
2319 static struct inode *shmem_alloc_inode(struct super_block *sb)
2320 {
2321         struct shmem_inode_info *p;
2322         p = (struct shmem_inode_info *)kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
2323         if (!p)
2324                 return NULL;
2325         return &p->vfs_inode;
2326 }
2327
2328 static void shmem_destroy_inode(struct inode *inode)
2329 {
2330         if ((inode->i_mode & S_IFMT) == S_IFREG) {
2331                 /* only struct inode is valid if it's an inline symlink */
2332                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
2333         }
2334         shmem_acl_destroy_inode(inode);
2335         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
2336 }
2337
2338 static void init_once(struct kmem_cache *cachep, void *foo)
2339 {
2340         struct shmem_inode_info *p = (struct shmem_inode_info *) foo;
2341
2342         inode_init_once(&p->vfs_inode);
2343 #ifdef CONFIG_TMPFS_POSIX_ACL
2344         p->i_acl = NULL;
2345         p->i_default_acl = NULL;
2346 #endif
2347 }
2348
2349 static int init_inodecache(void)
2350 {
2351         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
2352                                 sizeof(struct shmem_inode_info),
2353                                 0, SLAB_PANIC, init_once);
2354         return 0;
2355 }
2356
2357 static void destroy_inodecache(void)
2358 {
2359         kmem_cache_destroy(shmem_inode_cachep);
2360 }
2361
2362 static const struct address_space_operations shmem_aops = {
2363         .writepage      = shmem_writepage,
2364         .set_page_dirty = __set_page_dirty_no_writeback,
2365 #ifdef CONFIG_TMPFS
2366         .readpage       = shmem_readpage,
2367         .write_begin    = shmem_write_begin,
2368         .write_end      = shmem_write_end,
2369 #endif
2370         .migratepage    = migrate_page,
2371 };
2372
2373 static const struct file_operations shmem_file_operations = {
2374         .mmap           = shmem_mmap,
2375 #ifdef CONFIG_TMPFS
2376         .llseek         = generic_file_llseek,
2377         .read           = shmem_file_read,
2378         .write          = shmem_file_write,
2379         .fsync          = simple_sync_file,
2380         .splice_read    = generic_file_splice_read,
2381         .splice_write   = generic_file_splice_write,
2382 #endif
2383 };
2384
2385 static const struct inode_operations shmem_inode_operations = {
2386         .truncate       = shmem_truncate,
2387         .setattr        = shmem_notify_change,
2388         .truncate_range = shmem_truncate_range,
2389 #ifdef CONFIG_TMPFS_POSIX_ACL
2390         .setxattr       = generic_setxattr,
2391         .getxattr       = generic_getxattr,
2392         .listxattr      = generic_listxattr,
2393         .removexattr    = generic_removexattr,
2394         .permission     = shmem_permission,
2395 #endif
2396
2397 };
2398
2399 static const struct inode_operations shmem_dir_inode_operations = {
2400 #ifdef CONFIG_TMPFS
2401         .create         = shmem_create,
2402         .lookup         = simple_lookup,
2403         .link           = shmem_link,
2404         .unlink         = shmem_unlink,
2405         .symlink        = shmem_symlink,
2406         .mkdir          = shmem_mkdir,
2407         .rmdir          = shmem_rmdir,
2408         .mknod          = shmem_mknod,
2409         .rename         = shmem_rename,
2410 #endif
2411 #ifdef CONFIG_TMPFS_POSIX_ACL
2412         .setattr        = shmem_notify_change,
2413         .setxattr       = generic_setxattr,
2414         .getxattr       = generic_getxattr,
2415         .listxattr      = generic_listxattr,
2416         .removexattr    = generic_removexattr,
2417         .permission     = shmem_permission,
2418 #endif
2419 };
2420
2421 static const struct inode_operations shmem_special_inode_operations = {
2422 #ifdef CONFIG_TMPFS_POSIX_ACL
2423         .setattr        = shmem_notify_change,
2424         .setxattr       = generic_setxattr,
2425         .getxattr       = generic_getxattr,
2426         .listxattr      = generic_listxattr,
2427         .removexattr    = generic_removexattr,
2428         .permission     = shmem_permission,
2429 #endif
2430 };
2431
2432 static const struct super_operations shmem_ops = {
2433         .alloc_inode    = shmem_alloc_inode,
2434         .destroy_inode  = shmem_destroy_inode,
2435 #ifdef CONFIG_TMPFS
2436         .statfs         = shmem_statfs,
2437         .remount_fs     = shmem_remount_fs,
2438 #endif
2439         .delete_inode   = shmem_delete_inode,
2440         .drop_inode     = generic_delete_inode,
2441         .put_super      = shmem_put_super,
2442 };
2443
2444 static struct vm_operations_struct shmem_vm_ops = {
2445         .fault          = shmem_fault,
2446 #ifdef CONFIG_NUMA
2447         .set_policy     = shmem_set_policy,
2448         .get_policy     = shmem_get_policy,
2449 #endif
2450 };
2451
2452
2453 static int shmem_get_sb(struct file_system_type *fs_type,
2454         int flags, const char *dev_name, void *data, struct vfsmount *mnt)
2455 {
2456         return get_sb_nodev(fs_type, flags, data, shmem_fill_super, mnt);
2457 }
2458
2459 static struct file_system_type tmpfs_fs_type = {
2460         .owner          = THIS_MODULE,
2461         .name           = "tmpfs",
2462         .get_sb         = shmem_get_sb,
2463         .kill_sb        = kill_litter_super,
2464 };
2465 static struct vfsmount *shm_mnt;
2466
2467 static int __init init_tmpfs(void)
2468 {
2469         int error;
2470
2471         error = bdi_init(&shmem_backing_dev_info);
2472         if (error)
2473                 goto out4;
2474
2475         error = init_inodecache();
2476         if (error)
2477                 goto out3;
2478
2479         error = register_filesystem(&tmpfs_fs_type);
2480         if (error) {
2481                 printk(KERN_ERR "Could not register tmpfs\n");
2482                 goto out2;
2483         }
2484
2485         shm_mnt = vfs_kern_mount(&tmpfs_fs_type, MS_NOUSER,
2486                                 tmpfs_fs_type.name, NULL);
2487         if (IS_ERR(shm_mnt)) {
2488                 error = PTR_ERR(shm_mnt);
2489                 printk(KERN_ERR "Could not kern_mount tmpfs\n");
2490                 goto out1;
2491         }
2492         return 0;
2493
2494 out1:
2495         unregister_filesystem(&tmpfs_fs_type);
2496 out2:
2497         destroy_inodecache();
2498 out3:
2499         bdi_destroy(&shmem_backing_dev_info);
2500 out4:
2501         shm_mnt = ERR_PTR(error);
2502         return error;
2503 }
2504 module_init(init_tmpfs)
2505
2506 /*
2507  * shmem_file_setup - get an unlinked file living in tmpfs
2508  *
2509  * @name: name for dentry (to be seen in /proc/<pid>/maps
2510  * @size: size to be set for the file
2511  *
2512  */
2513 struct file *shmem_file_setup(char *name, loff_t size, unsigned long flags)
2514 {
2515         int error;
2516         struct file *file;
2517         struct inode *inode;
2518         struct dentry *dentry, *root;
2519         struct qstr this;
2520
2521         if (IS_ERR(shm_mnt))
2522                 return (void *)shm_mnt;
2523
2524         if (size < 0 || size > SHMEM_MAX_BYTES)
2525                 return ERR_PTR(-EINVAL);
2526
2527         if (shmem_acct_size(flags, size))
2528                 return ERR_PTR(-ENOMEM);
2529
2530         error = -ENOMEM;
2531         this.name = name;
2532         this.len = strlen(name);
2533         this.hash = 0; /* will go */
2534         root = shm_mnt->mnt_root;
2535         dentry = d_alloc(root, &this);
2536         if (!dentry)
2537                 goto put_memory;
2538
2539         error = -ENFILE;
2540         file = get_empty_filp();
2541         if (!file)
2542                 goto put_dentry;
2543
2544         error = -ENOSPC;
2545         inode = shmem_get_inode(root->d_sb, S_IFREG | S_IRWXUGO, 0);
2546         if (!inode)
2547                 goto close_file;
2548
2549         SHMEM_I(inode)->flags = flags & VM_ACCOUNT;
2550         d_instantiate(dentry, inode);
2551         inode->i_size = size;
2552         inode->i_nlink = 0;     /* It is unlinked */
2553         init_file(file, shm_mnt, dentry, FMODE_WRITE | FMODE_READ,
2554                         &shmem_file_operations);
2555         return file;
2556
2557 close_file:
2558         put_filp(file);
2559 put_dentry:
2560         dput(dentry);
2561 put_memory:
2562         shmem_unacct_size(flags, size);
2563         return ERR_PTR(error);
2564 }
2565
2566 /*
2567  * shmem_zero_setup - setup a shared anonymous mapping
2568  *
2569  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
2570  */
2571 int shmem_zero_setup(struct vm_area_struct *vma)
2572 {
2573         struct file *file;
2574         loff_t size = vma->vm_end - vma->vm_start;
2575
2576         file = shmem_file_setup("dev/zero", size, vma->vm_flags);
2577         if (IS_ERR(file))
2578                 return PTR_ERR(file);
2579
2580         if (vma->vm_file)
2581                 fput(vma->vm_file);
2582         vma->vm_file = file;
2583         vma->vm_ops = &shmem_vm_ops;
2584         return 0;
2585 }