tmpfs: fix shmem_swaplist races
[linux-2.6.git] / mm / shmem.c
1 /*
2  * Resizable virtual memory filesystem for Linux.
3  *
4  * Copyright (C) 2000 Linus Torvalds.
5  *               2000 Transmeta Corp.
6  *               2000-2001 Christoph Rohland
7  *               2000-2001 SAP AG
8  *               2002 Red Hat Inc.
9  * Copyright (C) 2002-2005 Hugh Dickins.
10  * Copyright (C) 2002-2005 VERITAS Software Corporation.
11  * Copyright (C) 2004 Andi Kleen, SuSE Labs
12  *
13  * Extended attribute support for tmpfs:
14  * Copyright (c) 2004, Luke Kenneth Casson Leighton <lkcl@lkcl.net>
15  * Copyright (c) 2004 Red Hat, Inc., James Morris <jmorris@redhat.com>
16  *
17  * This file is released under the GPL.
18  */
19
20 /*
21  * This virtual memory filesystem is heavily based on the ramfs. It
22  * extends ramfs by the ability to use swap and honor resource limits
23  * which makes it a completely usable filesystem.
24  */
25
26 #include <linux/module.h>
27 #include <linux/init.h>
28 #include <linux/fs.h>
29 #include <linux/xattr.h>
30 #include <linux/exportfs.h>
31 #include <linux/generic_acl.h>
32 #include <linux/mm.h>
33 #include <linux/mman.h>
34 #include <linux/file.h>
35 #include <linux/swap.h>
36 #include <linux/pagemap.h>
37 #include <linux/string.h>
38 #include <linux/slab.h>
39 #include <linux/backing-dev.h>
40 #include <linux/shmem_fs.h>
41 #include <linux/mount.h>
42 #include <linux/writeback.h>
43 #include <linux/vfs.h>
44 #include <linux/blkdev.h>
45 #include <linux/security.h>
46 #include <linux/swapops.h>
47 #include <linux/mempolicy.h>
48 #include <linux/namei.h>
49 #include <linux/ctype.h>
50 #include <linux/migrate.h>
51 #include <linux/highmem.h>
52
53 #include <asm/uaccess.h>
54 #include <asm/div64.h>
55 #include <asm/pgtable.h>
56
57 /* This magic number is used in glibc for posix shared memory */
58 #define TMPFS_MAGIC     0x01021994
59
60 #define ENTRIES_PER_PAGE (PAGE_CACHE_SIZE/sizeof(unsigned long))
61 #define ENTRIES_PER_PAGEPAGE (ENTRIES_PER_PAGE*ENTRIES_PER_PAGE)
62 #define BLOCKS_PER_PAGE  (PAGE_CACHE_SIZE/512)
63
64 #define SHMEM_MAX_INDEX  (SHMEM_NR_DIRECT + (ENTRIES_PER_PAGEPAGE/2) * (ENTRIES_PER_PAGE+1))
65 #define SHMEM_MAX_BYTES  ((unsigned long long)SHMEM_MAX_INDEX << PAGE_CACHE_SHIFT)
66
67 #define VM_ACCT(size)    (PAGE_CACHE_ALIGN(size) >> PAGE_SHIFT)
68
69 /* info->flags needs VM_flags to handle pagein/truncate races efficiently */
70 #define SHMEM_PAGEIN     VM_READ
71 #define SHMEM_TRUNCATE   VM_WRITE
72
73 /* Definition to limit shmem_truncate's steps between cond_rescheds */
74 #define LATENCY_LIMIT    64
75
76 /* Pretend that each entry is of this size in directory's i_size */
77 #define BOGO_DIRENT_SIZE 20
78
79 /* Flag allocation requirements to shmem_getpage and shmem_swp_alloc */
80 enum sgp_type {
81         SGP_READ,       /* don't exceed i_size, don't allocate page */
82         SGP_CACHE,      /* don't exceed i_size, may allocate page */
83         SGP_DIRTY,      /* like SGP_CACHE, but set new page dirty */
84         SGP_WRITE,      /* may exceed i_size, may allocate page */
85 };
86
87 static int shmem_getpage(struct inode *inode, unsigned long idx,
88                          struct page **pagep, enum sgp_type sgp, int *type);
89
90 static inline struct page *shmem_dir_alloc(gfp_t gfp_mask)
91 {
92         /*
93          * The above definition of ENTRIES_PER_PAGE, and the use of
94          * BLOCKS_PER_PAGE on indirect pages, assume PAGE_CACHE_SIZE:
95          * might be reconsidered if it ever diverges from PAGE_SIZE.
96          *
97          * Mobility flags are masked out as swap vectors cannot move
98          */
99         return alloc_pages((gfp_mask & ~GFP_MOVABLE_MASK) | __GFP_ZERO,
100                                 PAGE_CACHE_SHIFT-PAGE_SHIFT);
101 }
102
103 static inline void shmem_dir_free(struct page *page)
104 {
105         __free_pages(page, PAGE_CACHE_SHIFT-PAGE_SHIFT);
106 }
107
108 static struct page **shmem_dir_map(struct page *page)
109 {
110         return (struct page **)kmap_atomic(page, KM_USER0);
111 }
112
113 static inline void shmem_dir_unmap(struct page **dir)
114 {
115         kunmap_atomic(dir, KM_USER0);
116 }
117
118 static swp_entry_t *shmem_swp_map(struct page *page)
119 {
120         return (swp_entry_t *)kmap_atomic(page, KM_USER1);
121 }
122
123 static inline void shmem_swp_balance_unmap(void)
124 {
125         /*
126          * When passing a pointer to an i_direct entry, to code which
127          * also handles indirect entries and so will shmem_swp_unmap,
128          * we must arrange for the preempt count to remain in balance.
129          * What kmap_atomic of a lowmem page does depends on config
130          * and architecture, so pretend to kmap_atomic some lowmem page.
131          */
132         (void) kmap_atomic(ZERO_PAGE(0), KM_USER1);
133 }
134
135 static inline void shmem_swp_unmap(swp_entry_t *entry)
136 {
137         kunmap_atomic(entry, KM_USER1);
138 }
139
140 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
141 {
142         return sb->s_fs_info;
143 }
144
145 /*
146  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
147  * for shared memory and for shared anonymous (/dev/zero) mappings
148  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
149  * consistent with the pre-accounting of private mappings ...
150  */
151 static inline int shmem_acct_size(unsigned long flags, loff_t size)
152 {
153         return (flags & VM_ACCOUNT)?
154                 security_vm_enough_memory(VM_ACCT(size)): 0;
155 }
156
157 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
158 {
159         if (flags & VM_ACCOUNT)
160                 vm_unacct_memory(VM_ACCT(size));
161 }
162
163 /*
164  * ... whereas tmpfs objects are accounted incrementally as
165  * pages are allocated, in order to allow huge sparse files.
166  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
167  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
168  */
169 static inline int shmem_acct_block(unsigned long flags)
170 {
171         return (flags & VM_ACCOUNT)?
172                 0: security_vm_enough_memory(VM_ACCT(PAGE_CACHE_SIZE));
173 }
174
175 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
176 {
177         if (!(flags & VM_ACCOUNT))
178                 vm_unacct_memory(pages * VM_ACCT(PAGE_CACHE_SIZE));
179 }
180
181 static const struct super_operations shmem_ops;
182 static const struct address_space_operations shmem_aops;
183 static const struct file_operations shmem_file_operations;
184 static const struct inode_operations shmem_inode_operations;
185 static const struct inode_operations shmem_dir_inode_operations;
186 static const struct inode_operations shmem_special_inode_operations;
187 static struct vm_operations_struct shmem_vm_ops;
188
189 static struct backing_dev_info shmem_backing_dev_info  __read_mostly = {
190         .ra_pages       = 0,    /* No readahead */
191         .capabilities   = BDI_CAP_NO_ACCT_DIRTY | BDI_CAP_NO_WRITEBACK,
192         .unplug_io_fn   = default_unplug_io_fn,
193 };
194
195 static LIST_HEAD(shmem_swaplist);
196 static DEFINE_MUTEX(shmem_swaplist_mutex);
197
198 static void shmem_free_blocks(struct inode *inode, long pages)
199 {
200         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
201         if (sbinfo->max_blocks) {
202                 spin_lock(&sbinfo->stat_lock);
203                 sbinfo->free_blocks += pages;
204                 inode->i_blocks -= pages*BLOCKS_PER_PAGE;
205                 spin_unlock(&sbinfo->stat_lock);
206         }
207 }
208
209 static int shmem_reserve_inode(struct super_block *sb)
210 {
211         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
212         if (sbinfo->max_inodes) {
213                 spin_lock(&sbinfo->stat_lock);
214                 if (!sbinfo->free_inodes) {
215                         spin_unlock(&sbinfo->stat_lock);
216                         return -ENOSPC;
217                 }
218                 sbinfo->free_inodes--;
219                 spin_unlock(&sbinfo->stat_lock);
220         }
221         return 0;
222 }
223
224 static void shmem_free_inode(struct super_block *sb)
225 {
226         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
227         if (sbinfo->max_inodes) {
228                 spin_lock(&sbinfo->stat_lock);
229                 sbinfo->free_inodes++;
230                 spin_unlock(&sbinfo->stat_lock);
231         }
232 }
233
234 /*
235  * shmem_recalc_inode - recalculate the size of an inode
236  *
237  * @inode: inode to recalc
238  *
239  * We have to calculate the free blocks since the mm can drop
240  * undirtied hole pages behind our back.
241  *
242  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
243  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
244  *
245  * It has to be called with the spinlock held.
246  */
247 static void shmem_recalc_inode(struct inode *inode)
248 {
249         struct shmem_inode_info *info = SHMEM_I(inode);
250         long freed;
251
252         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
253         if (freed > 0) {
254                 info->alloced -= freed;
255                 shmem_unacct_blocks(info->flags, freed);
256                 shmem_free_blocks(inode, freed);
257         }
258 }
259
260 /*
261  * shmem_swp_entry - find the swap vector position in the info structure
262  *
263  * @info:  info structure for the inode
264  * @index: index of the page to find
265  * @page:  optional page to add to the structure. Has to be preset to
266  *         all zeros
267  *
268  * If there is no space allocated yet it will return NULL when
269  * page is NULL, else it will use the page for the needed block,
270  * setting it to NULL on return to indicate that it has been used.
271  *
272  * The swap vector is organized the following way:
273  *
274  * There are SHMEM_NR_DIRECT entries directly stored in the
275  * shmem_inode_info structure. So small files do not need an addional
276  * allocation.
277  *
278  * For pages with index > SHMEM_NR_DIRECT there is the pointer
279  * i_indirect which points to a page which holds in the first half
280  * doubly indirect blocks, in the second half triple indirect blocks:
281  *
282  * For an artificial ENTRIES_PER_PAGE = 4 this would lead to the
283  * following layout (for SHMEM_NR_DIRECT == 16):
284  *
285  * i_indirect -> dir --> 16-19
286  *            |      +-> 20-23
287  *            |
288  *            +-->dir2 --> 24-27
289  *            |        +-> 28-31
290  *            |        +-> 32-35
291  *            |        +-> 36-39
292  *            |
293  *            +-->dir3 --> 40-43
294  *                     +-> 44-47
295  *                     +-> 48-51
296  *                     +-> 52-55
297  */
298 static swp_entry_t *shmem_swp_entry(struct shmem_inode_info *info, unsigned long index, struct page **page)
299 {
300         unsigned long offset;
301         struct page **dir;
302         struct page *subdir;
303
304         if (index < SHMEM_NR_DIRECT) {
305                 shmem_swp_balance_unmap();
306                 return info->i_direct+index;
307         }
308         if (!info->i_indirect) {
309                 if (page) {
310                         info->i_indirect = *page;
311                         *page = NULL;
312                 }
313                 return NULL;                    /* need another page */
314         }
315
316         index -= SHMEM_NR_DIRECT;
317         offset = index % ENTRIES_PER_PAGE;
318         index /= ENTRIES_PER_PAGE;
319         dir = shmem_dir_map(info->i_indirect);
320
321         if (index >= ENTRIES_PER_PAGE/2) {
322                 index -= ENTRIES_PER_PAGE/2;
323                 dir += ENTRIES_PER_PAGE/2 + index/ENTRIES_PER_PAGE;
324                 index %= ENTRIES_PER_PAGE;
325                 subdir = *dir;
326                 if (!subdir) {
327                         if (page) {
328                                 *dir = *page;
329                                 *page = NULL;
330                         }
331                         shmem_dir_unmap(dir);
332                         return NULL;            /* need another page */
333                 }
334                 shmem_dir_unmap(dir);
335                 dir = shmem_dir_map(subdir);
336         }
337
338         dir += index;
339         subdir = *dir;
340         if (!subdir) {
341                 if (!page || !(subdir = *page)) {
342                         shmem_dir_unmap(dir);
343                         return NULL;            /* need a page */
344                 }
345                 *dir = subdir;
346                 *page = NULL;
347         }
348         shmem_dir_unmap(dir);
349         return shmem_swp_map(subdir) + offset;
350 }
351
352 static void shmem_swp_set(struct shmem_inode_info *info, swp_entry_t *entry, unsigned long value)
353 {
354         long incdec = value? 1: -1;
355
356         entry->val = value;
357         info->swapped += incdec;
358         if ((unsigned long)(entry - info->i_direct) >= SHMEM_NR_DIRECT) {
359                 struct page *page = kmap_atomic_to_page(entry);
360                 set_page_private(page, page_private(page) + incdec);
361         }
362 }
363
364 /*
365  * shmem_swp_alloc - get the position of the swap entry for the page.
366  *                   If it does not exist allocate the entry.
367  *
368  * @info:       info structure for the inode
369  * @index:      index of the page to find
370  * @sgp:        check and recheck i_size? skip allocation?
371  */
372 static swp_entry_t *shmem_swp_alloc(struct shmem_inode_info *info, unsigned long index, enum sgp_type sgp)
373 {
374         struct inode *inode = &info->vfs_inode;
375         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
376         struct page *page = NULL;
377         swp_entry_t *entry;
378
379         if (sgp != SGP_WRITE &&
380             ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode))
381                 return ERR_PTR(-EINVAL);
382
383         while (!(entry = shmem_swp_entry(info, index, &page))) {
384                 if (sgp == SGP_READ)
385                         return shmem_swp_map(ZERO_PAGE(0));
386                 /*
387                  * Test free_blocks against 1 not 0, since we have 1 data
388                  * page (and perhaps indirect index pages) yet to allocate:
389                  * a waste to allocate index if we cannot allocate data.
390                  */
391                 if (sbinfo->max_blocks) {
392                         spin_lock(&sbinfo->stat_lock);
393                         if (sbinfo->free_blocks <= 1) {
394                                 spin_unlock(&sbinfo->stat_lock);
395                                 return ERR_PTR(-ENOSPC);
396                         }
397                         sbinfo->free_blocks--;
398                         inode->i_blocks += BLOCKS_PER_PAGE;
399                         spin_unlock(&sbinfo->stat_lock);
400                 }
401
402                 spin_unlock(&info->lock);
403                 page = shmem_dir_alloc(mapping_gfp_mask(inode->i_mapping));
404                 if (page)
405                         set_page_private(page, 0);
406                 spin_lock(&info->lock);
407
408                 if (!page) {
409                         shmem_free_blocks(inode, 1);
410                         return ERR_PTR(-ENOMEM);
411                 }
412                 if (sgp != SGP_WRITE &&
413                     ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
414                         entry = ERR_PTR(-EINVAL);
415                         break;
416                 }
417                 if (info->next_index <= index)
418                         info->next_index = index + 1;
419         }
420         if (page) {
421                 /* another task gave its page, or truncated the file */
422                 shmem_free_blocks(inode, 1);
423                 shmem_dir_free(page);
424         }
425         if (info->next_index <= index && !IS_ERR(entry))
426                 info->next_index = index + 1;
427         return entry;
428 }
429
430 /*
431  * shmem_free_swp - free some swap entries in a directory
432  *
433  * @dir:        pointer to the directory
434  * @edir:       pointer after last entry of the directory
435  * @punch_lock: pointer to spinlock when needed for the holepunch case
436  */
437 static int shmem_free_swp(swp_entry_t *dir, swp_entry_t *edir,
438                                                 spinlock_t *punch_lock)
439 {
440         spinlock_t *punch_unlock = NULL;
441         swp_entry_t *ptr;
442         int freed = 0;
443
444         for (ptr = dir; ptr < edir; ptr++) {
445                 if (ptr->val) {
446                         if (unlikely(punch_lock)) {
447                                 punch_unlock = punch_lock;
448                                 punch_lock = NULL;
449                                 spin_lock(punch_unlock);
450                                 if (!ptr->val)
451                                         continue;
452                         }
453                         free_swap_and_cache(*ptr);
454                         *ptr = (swp_entry_t){0};
455                         freed++;
456                 }
457         }
458         if (punch_unlock)
459                 spin_unlock(punch_unlock);
460         return freed;
461 }
462
463 static int shmem_map_and_free_swp(struct page *subdir, int offset,
464                 int limit, struct page ***dir, spinlock_t *punch_lock)
465 {
466         swp_entry_t *ptr;
467         int freed = 0;
468
469         ptr = shmem_swp_map(subdir);
470         for (; offset < limit; offset += LATENCY_LIMIT) {
471                 int size = limit - offset;
472                 if (size > LATENCY_LIMIT)
473                         size = LATENCY_LIMIT;
474                 freed += shmem_free_swp(ptr+offset, ptr+offset+size,
475                                                         punch_lock);
476                 if (need_resched()) {
477                         shmem_swp_unmap(ptr);
478                         if (*dir) {
479                                 shmem_dir_unmap(*dir);
480                                 *dir = NULL;
481                         }
482                         cond_resched();
483                         ptr = shmem_swp_map(subdir);
484                 }
485         }
486         shmem_swp_unmap(ptr);
487         return freed;
488 }
489
490 static void shmem_free_pages(struct list_head *next)
491 {
492         struct page *page;
493         int freed = 0;
494
495         do {
496                 page = container_of(next, struct page, lru);
497                 next = next->next;
498                 shmem_dir_free(page);
499                 freed++;
500                 if (freed >= LATENCY_LIMIT) {
501                         cond_resched();
502                         freed = 0;
503                 }
504         } while (next);
505 }
506
507 static void shmem_truncate_range(struct inode *inode, loff_t start, loff_t end)
508 {
509         struct shmem_inode_info *info = SHMEM_I(inode);
510         unsigned long idx;
511         unsigned long size;
512         unsigned long limit;
513         unsigned long stage;
514         unsigned long diroff;
515         struct page **dir;
516         struct page *topdir;
517         struct page *middir;
518         struct page *subdir;
519         swp_entry_t *ptr;
520         LIST_HEAD(pages_to_free);
521         long nr_pages_to_free = 0;
522         long nr_swaps_freed = 0;
523         int offset;
524         int freed;
525         int punch_hole;
526         spinlock_t *needs_lock;
527         spinlock_t *punch_lock;
528         unsigned long upper_limit;
529
530         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
531         idx = (start + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
532         if (idx >= info->next_index)
533                 return;
534
535         spin_lock(&info->lock);
536         info->flags |= SHMEM_TRUNCATE;
537         if (likely(end == (loff_t) -1)) {
538                 limit = info->next_index;
539                 upper_limit = SHMEM_MAX_INDEX;
540                 info->next_index = idx;
541                 needs_lock = NULL;
542                 punch_hole = 0;
543         } else {
544                 if (end + 1 >= inode->i_size) { /* we may free a little more */
545                         limit = (inode->i_size + PAGE_CACHE_SIZE - 1) >>
546                                                         PAGE_CACHE_SHIFT;
547                         upper_limit = SHMEM_MAX_INDEX;
548                 } else {
549                         limit = (end + 1) >> PAGE_CACHE_SHIFT;
550                         upper_limit = limit;
551                 }
552                 needs_lock = &info->lock;
553                 punch_hole = 1;
554         }
555
556         topdir = info->i_indirect;
557         if (topdir && idx <= SHMEM_NR_DIRECT && !punch_hole) {
558                 info->i_indirect = NULL;
559                 nr_pages_to_free++;
560                 list_add(&topdir->lru, &pages_to_free);
561         }
562         spin_unlock(&info->lock);
563
564         if (info->swapped && idx < SHMEM_NR_DIRECT) {
565                 ptr = info->i_direct;
566                 size = limit;
567                 if (size > SHMEM_NR_DIRECT)
568                         size = SHMEM_NR_DIRECT;
569                 nr_swaps_freed = shmem_free_swp(ptr+idx, ptr+size, needs_lock);
570         }
571
572         /*
573          * If there are no indirect blocks or we are punching a hole
574          * below indirect blocks, nothing to be done.
575          */
576         if (!topdir || limit <= SHMEM_NR_DIRECT)
577                 goto done2;
578
579         /*
580          * The truncation case has already dropped info->lock, and we're safe
581          * because i_size and next_index have already been lowered, preventing
582          * access beyond.  But in the punch_hole case, we still need to take
583          * the lock when updating the swap directory, because there might be
584          * racing accesses by shmem_getpage(SGP_CACHE), shmem_unuse_inode or
585          * shmem_writepage.  However, whenever we find we can remove a whole
586          * directory page (not at the misaligned start or end of the range),
587          * we first NULLify its pointer in the level above, and then have no
588          * need to take the lock when updating its contents: needs_lock and
589          * punch_lock (either pointing to info->lock or NULL) manage this.
590          */
591
592         upper_limit -= SHMEM_NR_DIRECT;
593         limit -= SHMEM_NR_DIRECT;
594         idx = (idx > SHMEM_NR_DIRECT)? (idx - SHMEM_NR_DIRECT): 0;
595         offset = idx % ENTRIES_PER_PAGE;
596         idx -= offset;
597
598         dir = shmem_dir_map(topdir);
599         stage = ENTRIES_PER_PAGEPAGE/2;
600         if (idx < ENTRIES_PER_PAGEPAGE/2) {
601                 middir = topdir;
602                 diroff = idx/ENTRIES_PER_PAGE;
603         } else {
604                 dir += ENTRIES_PER_PAGE/2;
605                 dir += (idx - ENTRIES_PER_PAGEPAGE/2)/ENTRIES_PER_PAGEPAGE;
606                 while (stage <= idx)
607                         stage += ENTRIES_PER_PAGEPAGE;
608                 middir = *dir;
609                 if (*dir) {
610                         diroff = ((idx - ENTRIES_PER_PAGEPAGE/2) %
611                                 ENTRIES_PER_PAGEPAGE) / ENTRIES_PER_PAGE;
612                         if (!diroff && !offset && upper_limit >= stage) {
613                                 if (needs_lock) {
614                                         spin_lock(needs_lock);
615                                         *dir = NULL;
616                                         spin_unlock(needs_lock);
617                                         needs_lock = NULL;
618                                 } else
619                                         *dir = NULL;
620                                 nr_pages_to_free++;
621                                 list_add(&middir->lru, &pages_to_free);
622                         }
623                         shmem_dir_unmap(dir);
624                         dir = shmem_dir_map(middir);
625                 } else {
626                         diroff = 0;
627                         offset = 0;
628                         idx = stage;
629                 }
630         }
631
632         for (; idx < limit; idx += ENTRIES_PER_PAGE, diroff++) {
633                 if (unlikely(idx == stage)) {
634                         shmem_dir_unmap(dir);
635                         dir = shmem_dir_map(topdir) +
636                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
637                         while (!*dir) {
638                                 dir++;
639                                 idx += ENTRIES_PER_PAGEPAGE;
640                                 if (idx >= limit)
641                                         goto done1;
642                         }
643                         stage = idx + ENTRIES_PER_PAGEPAGE;
644                         middir = *dir;
645                         if (punch_hole)
646                                 needs_lock = &info->lock;
647                         if (upper_limit >= stage) {
648                                 if (needs_lock) {
649                                         spin_lock(needs_lock);
650                                         *dir = NULL;
651                                         spin_unlock(needs_lock);
652                                         needs_lock = NULL;
653                                 } else
654                                         *dir = NULL;
655                                 nr_pages_to_free++;
656                                 list_add(&middir->lru, &pages_to_free);
657                         }
658                         shmem_dir_unmap(dir);
659                         cond_resched();
660                         dir = shmem_dir_map(middir);
661                         diroff = 0;
662                 }
663                 punch_lock = needs_lock;
664                 subdir = dir[diroff];
665                 if (subdir && !offset && upper_limit-idx >= ENTRIES_PER_PAGE) {
666                         if (needs_lock) {
667                                 spin_lock(needs_lock);
668                                 dir[diroff] = NULL;
669                                 spin_unlock(needs_lock);
670                                 punch_lock = NULL;
671                         } else
672                                 dir[diroff] = NULL;
673                         nr_pages_to_free++;
674                         list_add(&subdir->lru, &pages_to_free);
675                 }
676                 if (subdir && page_private(subdir) /* has swap entries */) {
677                         size = limit - idx;
678                         if (size > ENTRIES_PER_PAGE)
679                                 size = ENTRIES_PER_PAGE;
680                         freed = shmem_map_and_free_swp(subdir,
681                                         offset, size, &dir, punch_lock);
682                         if (!dir)
683                                 dir = shmem_dir_map(middir);
684                         nr_swaps_freed += freed;
685                         if (offset || punch_lock) {
686                                 spin_lock(&info->lock);
687                                 set_page_private(subdir,
688                                         page_private(subdir) - freed);
689                                 spin_unlock(&info->lock);
690                         } else
691                                 BUG_ON(page_private(subdir) != freed);
692                 }
693                 offset = 0;
694         }
695 done1:
696         shmem_dir_unmap(dir);
697 done2:
698         if (inode->i_mapping->nrpages && (info->flags & SHMEM_PAGEIN)) {
699                 /*
700                  * Call truncate_inode_pages again: racing shmem_unuse_inode
701                  * may have swizzled a page in from swap since vmtruncate or
702                  * generic_delete_inode did it, before we lowered next_index.
703                  * Also, though shmem_getpage checks i_size before adding to
704                  * cache, no recheck after: so fix the narrow window there too.
705                  *
706                  * Recalling truncate_inode_pages_range and unmap_mapping_range
707                  * every time for punch_hole (which never got a chance to clear
708                  * SHMEM_PAGEIN at the start of vmtruncate_range) is expensive,
709                  * yet hardly ever necessary: try to optimize them out later.
710                  */
711                 truncate_inode_pages_range(inode->i_mapping, start, end);
712                 if (punch_hole)
713                         unmap_mapping_range(inode->i_mapping, start,
714                                                         end - start, 1);
715         }
716
717         spin_lock(&info->lock);
718         info->flags &= ~SHMEM_TRUNCATE;
719         info->swapped -= nr_swaps_freed;
720         if (nr_pages_to_free)
721                 shmem_free_blocks(inode, nr_pages_to_free);
722         shmem_recalc_inode(inode);
723         spin_unlock(&info->lock);
724
725         /*
726          * Empty swap vector directory pages to be freed?
727          */
728         if (!list_empty(&pages_to_free)) {
729                 pages_to_free.prev->next = NULL;
730                 shmem_free_pages(pages_to_free.next);
731         }
732 }
733
734 static void shmem_truncate(struct inode *inode)
735 {
736         shmem_truncate_range(inode, inode->i_size, (loff_t)-1);
737 }
738
739 static int shmem_notify_change(struct dentry *dentry, struct iattr *attr)
740 {
741         struct inode *inode = dentry->d_inode;
742         struct page *page = NULL;
743         int error;
744
745         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
746                 if (attr->ia_size < inode->i_size) {
747                         /*
748                          * If truncating down to a partial page, then
749                          * if that page is already allocated, hold it
750                          * in memory until the truncation is over, so
751                          * truncate_partial_page cannnot miss it were
752                          * it assigned to swap.
753                          */
754                         if (attr->ia_size & (PAGE_CACHE_SIZE-1)) {
755                                 (void) shmem_getpage(inode,
756                                         attr->ia_size>>PAGE_CACHE_SHIFT,
757                                                 &page, SGP_READ, NULL);
758                                 if (page)
759                                         unlock_page(page);
760                         }
761                         /*
762                          * Reset SHMEM_PAGEIN flag so that shmem_truncate can
763                          * detect if any pages might have been added to cache
764                          * after truncate_inode_pages.  But we needn't bother
765                          * if it's being fully truncated to zero-length: the
766                          * nrpages check is efficient enough in that case.
767                          */
768                         if (attr->ia_size) {
769                                 struct shmem_inode_info *info = SHMEM_I(inode);
770                                 spin_lock(&info->lock);
771                                 info->flags &= ~SHMEM_PAGEIN;
772                                 spin_unlock(&info->lock);
773                         }
774                 }
775         }
776
777         error = inode_change_ok(inode, attr);
778         if (!error)
779                 error = inode_setattr(inode, attr);
780 #ifdef CONFIG_TMPFS_POSIX_ACL
781         if (!error && (attr->ia_valid & ATTR_MODE))
782                 error = generic_acl_chmod(inode, &shmem_acl_ops);
783 #endif
784         if (page)
785                 page_cache_release(page);
786         return error;
787 }
788
789 static void shmem_delete_inode(struct inode *inode)
790 {
791         struct shmem_inode_info *info = SHMEM_I(inode);
792
793         if (inode->i_op->truncate == shmem_truncate) {
794                 truncate_inode_pages(inode->i_mapping, 0);
795                 shmem_unacct_size(info->flags, inode->i_size);
796                 inode->i_size = 0;
797                 shmem_truncate(inode);
798                 if (!list_empty(&info->swaplist)) {
799                         mutex_lock(&shmem_swaplist_mutex);
800                         list_del_init(&info->swaplist);
801                         mutex_unlock(&shmem_swaplist_mutex);
802                 }
803         }
804         BUG_ON(inode->i_blocks);
805         shmem_free_inode(inode->i_sb);
806         clear_inode(inode);
807 }
808
809 static inline int shmem_find_swp(swp_entry_t entry, swp_entry_t *dir, swp_entry_t *edir)
810 {
811         swp_entry_t *ptr;
812
813         for (ptr = dir; ptr < edir; ptr++) {
814                 if (ptr->val == entry.val)
815                         return ptr - dir;
816         }
817         return -1;
818 }
819
820 static int shmem_unuse_inode(struct shmem_inode_info *info, swp_entry_t entry, struct page *page)
821 {
822         struct inode *inode;
823         unsigned long idx;
824         unsigned long size;
825         unsigned long limit;
826         unsigned long stage;
827         struct page **dir;
828         struct page *subdir;
829         swp_entry_t *ptr;
830         int offset;
831         int error;
832
833         idx = 0;
834         ptr = info->i_direct;
835         spin_lock(&info->lock);
836         if (!info->swapped) {
837                 list_del_init(&info->swaplist);
838                 goto lost2;
839         }
840         limit = info->next_index;
841         size = limit;
842         if (size > SHMEM_NR_DIRECT)
843                 size = SHMEM_NR_DIRECT;
844         offset = shmem_find_swp(entry, ptr, ptr+size);
845         if (offset >= 0)
846                 goto found;
847         if (!info->i_indirect)
848                 goto lost2;
849
850         dir = shmem_dir_map(info->i_indirect);
851         stage = SHMEM_NR_DIRECT + ENTRIES_PER_PAGEPAGE/2;
852
853         for (idx = SHMEM_NR_DIRECT; idx < limit; idx += ENTRIES_PER_PAGE, dir++) {
854                 if (unlikely(idx == stage)) {
855                         shmem_dir_unmap(dir-1);
856                         if (cond_resched_lock(&info->lock)) {
857                                 /* check it has not been truncated */
858                                 if (limit > info->next_index) {
859                                         limit = info->next_index;
860                                         if (idx >= limit)
861                                                 goto lost2;
862                                 }
863                         }
864                         dir = shmem_dir_map(info->i_indirect) +
865                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
866                         while (!*dir) {
867                                 dir++;
868                                 idx += ENTRIES_PER_PAGEPAGE;
869                                 if (idx >= limit)
870                                         goto lost1;
871                         }
872                         stage = idx + ENTRIES_PER_PAGEPAGE;
873                         subdir = *dir;
874                         shmem_dir_unmap(dir);
875                         dir = shmem_dir_map(subdir);
876                 }
877                 subdir = *dir;
878                 if (subdir && page_private(subdir)) {
879                         ptr = shmem_swp_map(subdir);
880                         size = limit - idx;
881                         if (size > ENTRIES_PER_PAGE)
882                                 size = ENTRIES_PER_PAGE;
883                         offset = shmem_find_swp(entry, ptr, ptr+size);
884                         shmem_swp_unmap(ptr);
885                         if (offset >= 0) {
886                                 shmem_dir_unmap(dir);
887                                 goto found;
888                         }
889                 }
890         }
891 lost1:
892         shmem_dir_unmap(dir-1);
893 lost2:
894         spin_unlock(&info->lock);
895         return 0;
896 found:
897         idx += offset;
898         inode = igrab(&info->vfs_inode);
899         spin_unlock(&info->lock);
900
901         /*
902          * Move _head_ to start search for next from here.
903          * But be careful: shmem_delete_inode checks list_empty without taking
904          * mutex, and there's an instant in list_move_tail when info->swaplist
905          * would appear empty, if it were the only one on shmem_swaplist.  We
906          * could avoid doing it if inode NULL; or use this minor optimization.
907          */
908         if (shmem_swaplist.next != &info->swaplist)
909                 list_move_tail(&shmem_swaplist, &info->swaplist);
910         mutex_unlock(&shmem_swaplist_mutex);
911
912         error = 1;
913         if (!inode)
914                 goto out;
915         error = radix_tree_preload(GFP_KERNEL);
916         if (error)
917                 goto out;
918         error = 1;
919
920         spin_lock(&info->lock);
921         ptr = shmem_swp_entry(info, idx, NULL);
922         if (ptr && ptr->val == entry.val)
923                 error = add_to_page_cache(page, inode->i_mapping,
924                                                 idx, GFP_NOWAIT);
925         if (error == -EEXIST) {
926                 struct page *filepage = find_get_page(inode->i_mapping, idx);
927                 error = 1;
928                 if (filepage) {
929                         /*
930                          * There might be a more uptodate page coming down
931                          * from a stacked writepage: forget our swappage if so.
932                          */
933                         if (PageUptodate(filepage))
934                                 error = 0;
935                         page_cache_release(filepage);
936                 }
937         }
938         if (!error) {
939                 delete_from_swap_cache(page);
940                 set_page_dirty(page);
941                 info->flags |= SHMEM_PAGEIN;
942                 shmem_swp_set(info, ptr, 0);
943                 swap_free(entry);
944                 error = 1;      /* not an error, but entry was found */
945         }
946         if (ptr)
947                 shmem_swp_unmap(ptr);
948         spin_unlock(&info->lock);
949         radix_tree_preload_end();
950 out:
951         unlock_page(page);
952         page_cache_release(page);
953         iput(inode);            /* allows for NULL */
954         return error;
955 }
956
957 /*
958  * shmem_unuse() search for an eventually swapped out shmem page.
959  */
960 int shmem_unuse(swp_entry_t entry, struct page *page)
961 {
962         struct list_head *p, *next;
963         struct shmem_inode_info *info;
964         int found = 0;
965
966         mutex_lock(&shmem_swaplist_mutex);
967         list_for_each_safe(p, next, &shmem_swaplist) {
968                 info = list_entry(p, struct shmem_inode_info, swaplist);
969                 found = shmem_unuse_inode(info, entry, page);
970                 cond_resched();
971                 if (found)
972                         goto out;
973         }
974         mutex_unlock(&shmem_swaplist_mutex);
975 out:    return found;   /* 0 or 1 or -ENOMEM */
976 }
977
978 /*
979  * Move the page from the page cache to the swap cache.
980  */
981 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
982 {
983         struct shmem_inode_info *info;
984         swp_entry_t *entry, swap;
985         struct address_space *mapping;
986         unsigned long index;
987         struct inode *inode;
988
989         BUG_ON(!PageLocked(page));
990         mapping = page->mapping;
991         index = page->index;
992         inode = mapping->host;
993         info = SHMEM_I(inode);
994         if (info->flags & VM_LOCKED)
995                 goto redirty;
996         if (!total_swap_pages)
997                 goto redirty;
998
999         /*
1000          * shmem_backing_dev_info's capabilities prevent regular writeback or
1001          * sync from ever calling shmem_writepage; but a stacking filesystem
1002          * may use the ->writepage of its underlying filesystem, in which case
1003          * tmpfs should write out to swap only in response to memory pressure,
1004          * and not for pdflush or sync.  However, in those cases, we do still
1005          * want to check if there's a redundant swappage to be discarded.
1006          */
1007         if (wbc->for_reclaim)
1008                 swap = get_swap_page();
1009         else
1010                 swap.val = 0;
1011
1012         spin_lock(&info->lock);
1013         if (index >= info->next_index) {
1014                 BUG_ON(!(info->flags & SHMEM_TRUNCATE));
1015                 goto unlock;
1016         }
1017         entry = shmem_swp_entry(info, index, NULL);
1018         if (entry->val) {
1019                 /*
1020                  * The more uptodate page coming down from a stacked
1021                  * writepage should replace our old swappage.
1022                  */
1023                 free_swap_and_cache(*entry);
1024                 shmem_swp_set(info, entry, 0);
1025         }
1026         shmem_recalc_inode(inode);
1027
1028         if (swap.val && add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
1029                 remove_from_page_cache(page);
1030                 shmem_swp_set(info, entry, swap.val);
1031                 shmem_swp_unmap(entry);
1032                 if (list_empty(&info->swaplist))
1033                         inode = igrab(inode);
1034                 else
1035                         inode = NULL;
1036                 spin_unlock(&info->lock);
1037                 swap_duplicate(swap);
1038                 BUG_ON(page_mapped(page));
1039                 page_cache_release(page);       /* pagecache ref */
1040                 set_page_dirty(page);
1041                 unlock_page(page);
1042                 if (inode) {
1043                         mutex_lock(&shmem_swaplist_mutex);
1044                         /* move instead of add in case we're racing */
1045                         list_move_tail(&info->swaplist, &shmem_swaplist);
1046                         mutex_unlock(&shmem_swaplist_mutex);
1047                         iput(inode);
1048                 }
1049                 return 0;
1050         }
1051
1052         shmem_swp_unmap(entry);
1053 unlock:
1054         spin_unlock(&info->lock);
1055         swap_free(swap);
1056 redirty:
1057         set_page_dirty(page);
1058         if (wbc->for_reclaim)
1059                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
1060         unlock_page(page);
1061         return 0;
1062 }
1063
1064 #ifdef CONFIG_NUMA
1065 static inline int shmem_parse_mpol(char *value, int *policy, nodemask_t *policy_nodes)
1066 {
1067         char *nodelist = strchr(value, ':');
1068         int err = 1;
1069
1070         if (nodelist) {
1071                 /* NUL-terminate policy string */
1072                 *nodelist++ = '\0';
1073                 if (nodelist_parse(nodelist, *policy_nodes))
1074                         goto out;
1075                 if (!nodes_subset(*policy_nodes, node_states[N_HIGH_MEMORY]))
1076                         goto out;
1077         }
1078         if (!strcmp(value, "default")) {
1079                 *policy = MPOL_DEFAULT;
1080                 /* Don't allow a nodelist */
1081                 if (!nodelist)
1082                         err = 0;
1083         } else if (!strcmp(value, "prefer")) {
1084                 *policy = MPOL_PREFERRED;
1085                 /* Insist on a nodelist of one node only */
1086                 if (nodelist) {
1087                         char *rest = nodelist;
1088                         while (isdigit(*rest))
1089                                 rest++;
1090                         if (!*rest)
1091                                 err = 0;
1092                 }
1093         } else if (!strcmp(value, "bind")) {
1094                 *policy = MPOL_BIND;
1095                 /* Insist on a nodelist */
1096                 if (nodelist)
1097                         err = 0;
1098         } else if (!strcmp(value, "interleave")) {
1099                 *policy = MPOL_INTERLEAVE;
1100                 /*
1101                  * Default to online nodes with memory if no nodelist
1102                  */
1103                 if (!nodelist)
1104                         *policy_nodes = node_states[N_HIGH_MEMORY];
1105                 err = 0;
1106         }
1107 out:
1108         /* Restore string for error message */
1109         if (nodelist)
1110                 *--nodelist = ':';
1111         return err;
1112 }
1113
1114 static struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1115                         struct shmem_inode_info *info, unsigned long idx)
1116 {
1117         struct vm_area_struct pvma;
1118         struct page *page;
1119
1120         /* Create a pseudo vma that just contains the policy */
1121         pvma.vm_start = 0;
1122         pvma.vm_pgoff = idx;
1123         pvma.vm_ops = NULL;
1124         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, idx);
1125         page = swapin_readahead(entry, gfp, &pvma, 0);
1126         mpol_free(pvma.vm_policy);
1127         return page;
1128 }
1129
1130 static struct page *shmem_alloc_page(gfp_t gfp,
1131                         struct shmem_inode_info *info, unsigned long idx)
1132 {
1133         struct vm_area_struct pvma;
1134         struct page *page;
1135
1136         /* Create a pseudo vma that just contains the policy */
1137         pvma.vm_start = 0;
1138         pvma.vm_pgoff = idx;
1139         pvma.vm_ops = NULL;
1140         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, idx);
1141         page = alloc_page_vma(gfp, &pvma, 0);
1142         mpol_free(pvma.vm_policy);
1143         return page;
1144 }
1145 #else
1146 static inline int shmem_parse_mpol(char *value, int *policy,
1147                                                 nodemask_t *policy_nodes)
1148 {
1149         return 1;
1150 }
1151
1152 static inline struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1153                         struct shmem_inode_info *info, unsigned long idx)
1154 {
1155         return swapin_readahead(entry, gfp, NULL, 0);
1156 }
1157
1158 static inline struct page *shmem_alloc_page(gfp_t gfp,
1159                         struct shmem_inode_info *info, unsigned long idx)
1160 {
1161         return alloc_page(gfp);
1162 }
1163 #endif
1164
1165 /*
1166  * shmem_getpage - either get the page from swap or allocate a new one
1167  *
1168  * If we allocate a new one we do not mark it dirty. That's up to the
1169  * vm. If we swap it in we mark it dirty since we also free the swap
1170  * entry since a page cannot live in both the swap and page cache
1171  */
1172 static int shmem_getpage(struct inode *inode, unsigned long idx,
1173                         struct page **pagep, enum sgp_type sgp, int *type)
1174 {
1175         struct address_space *mapping = inode->i_mapping;
1176         struct shmem_inode_info *info = SHMEM_I(inode);
1177         struct shmem_sb_info *sbinfo;
1178         struct page *filepage = *pagep;
1179         struct page *swappage;
1180         swp_entry_t *entry;
1181         swp_entry_t swap;
1182         gfp_t gfp;
1183         int error;
1184
1185         if (idx >= SHMEM_MAX_INDEX)
1186                 return -EFBIG;
1187
1188         if (type)
1189                 *type = 0;
1190
1191         /*
1192          * Normally, filepage is NULL on entry, and either found
1193          * uptodate immediately, or allocated and zeroed, or read
1194          * in under swappage, which is then assigned to filepage.
1195          * But shmem_readpage (required for splice) passes in a locked
1196          * filepage, which may be found not uptodate by other callers
1197          * too, and may need to be copied from the swappage read in.
1198          */
1199 repeat:
1200         if (!filepage)
1201                 filepage = find_lock_page(mapping, idx);
1202         if (filepage && PageUptodate(filepage))
1203                 goto done;
1204         error = 0;
1205         gfp = mapping_gfp_mask(mapping);
1206         if (!filepage) {
1207                 /*
1208                  * Try to preload while we can wait, to not make a habit of
1209                  * draining atomic reserves; but don't latch on to this cpu.
1210                  */
1211                 error = radix_tree_preload(gfp & ~__GFP_HIGHMEM);
1212                 if (error)
1213                         goto failed;
1214                 radix_tree_preload_end();
1215         }
1216
1217         spin_lock(&info->lock);
1218         shmem_recalc_inode(inode);
1219         entry = shmem_swp_alloc(info, idx, sgp);
1220         if (IS_ERR(entry)) {
1221                 spin_unlock(&info->lock);
1222                 error = PTR_ERR(entry);
1223                 goto failed;
1224         }
1225         swap = *entry;
1226
1227         if (swap.val) {
1228                 /* Look it up and read it in.. */
1229                 swappage = lookup_swap_cache(swap);
1230                 if (!swappage) {
1231                         shmem_swp_unmap(entry);
1232                         /* here we actually do the io */
1233                         if (type && !(*type & VM_FAULT_MAJOR)) {
1234                                 __count_vm_event(PGMAJFAULT);
1235                                 *type |= VM_FAULT_MAJOR;
1236                         }
1237                         spin_unlock(&info->lock);
1238                         swappage = shmem_swapin(swap, gfp, info, idx);
1239                         if (!swappage) {
1240                                 spin_lock(&info->lock);
1241                                 entry = shmem_swp_alloc(info, idx, sgp);
1242                                 if (IS_ERR(entry))
1243                                         error = PTR_ERR(entry);
1244                                 else {
1245                                         if (entry->val == swap.val)
1246                                                 error = -ENOMEM;
1247                                         shmem_swp_unmap(entry);
1248                                 }
1249                                 spin_unlock(&info->lock);
1250                                 if (error)
1251                                         goto failed;
1252                                 goto repeat;
1253                         }
1254                         wait_on_page_locked(swappage);
1255                         page_cache_release(swappage);
1256                         goto repeat;
1257                 }
1258
1259                 /* We have to do this with page locked to prevent races */
1260                 if (TestSetPageLocked(swappage)) {
1261                         shmem_swp_unmap(entry);
1262                         spin_unlock(&info->lock);
1263                         wait_on_page_locked(swappage);
1264                         page_cache_release(swappage);
1265                         goto repeat;
1266                 }
1267                 if (PageWriteback(swappage)) {
1268                         shmem_swp_unmap(entry);
1269                         spin_unlock(&info->lock);
1270                         wait_on_page_writeback(swappage);
1271                         unlock_page(swappage);
1272                         page_cache_release(swappage);
1273                         goto repeat;
1274                 }
1275                 if (!PageUptodate(swappage)) {
1276                         shmem_swp_unmap(entry);
1277                         spin_unlock(&info->lock);
1278                         unlock_page(swappage);
1279                         page_cache_release(swappage);
1280                         error = -EIO;
1281                         goto failed;
1282                 }
1283
1284                 if (filepage) {
1285                         shmem_swp_set(info, entry, 0);
1286                         shmem_swp_unmap(entry);
1287                         delete_from_swap_cache(swappage);
1288                         spin_unlock(&info->lock);
1289                         copy_highpage(filepage, swappage);
1290                         unlock_page(swappage);
1291                         page_cache_release(swappage);
1292                         flush_dcache_page(filepage);
1293                         SetPageUptodate(filepage);
1294                         set_page_dirty(filepage);
1295                         swap_free(swap);
1296                 } else if (!(error = add_to_page_cache(
1297                                 swappage, mapping, idx, GFP_NOWAIT))) {
1298                         info->flags |= SHMEM_PAGEIN;
1299                         shmem_swp_set(info, entry, 0);
1300                         shmem_swp_unmap(entry);
1301                         delete_from_swap_cache(swappage);
1302                         spin_unlock(&info->lock);
1303                         filepage = swappage;
1304                         set_page_dirty(filepage);
1305                         swap_free(swap);
1306                 } else {
1307                         shmem_swp_unmap(entry);
1308                         spin_unlock(&info->lock);
1309                         unlock_page(swappage);
1310                         page_cache_release(swappage);
1311                         goto repeat;
1312                 }
1313         } else if (sgp == SGP_READ && !filepage) {
1314                 shmem_swp_unmap(entry);
1315                 filepage = find_get_page(mapping, idx);
1316                 if (filepage &&
1317                     (!PageUptodate(filepage) || TestSetPageLocked(filepage))) {
1318                         spin_unlock(&info->lock);
1319                         wait_on_page_locked(filepage);
1320                         page_cache_release(filepage);
1321                         filepage = NULL;
1322                         goto repeat;
1323                 }
1324                 spin_unlock(&info->lock);
1325         } else {
1326                 shmem_swp_unmap(entry);
1327                 sbinfo = SHMEM_SB(inode->i_sb);
1328                 if (sbinfo->max_blocks) {
1329                         spin_lock(&sbinfo->stat_lock);
1330                         if (sbinfo->free_blocks == 0 ||
1331                             shmem_acct_block(info->flags)) {
1332                                 spin_unlock(&sbinfo->stat_lock);
1333                                 spin_unlock(&info->lock);
1334                                 error = -ENOSPC;
1335                                 goto failed;
1336                         }
1337                         sbinfo->free_blocks--;
1338                         inode->i_blocks += BLOCKS_PER_PAGE;
1339                         spin_unlock(&sbinfo->stat_lock);
1340                 } else if (shmem_acct_block(info->flags)) {
1341                         spin_unlock(&info->lock);
1342                         error = -ENOSPC;
1343                         goto failed;
1344                 }
1345
1346                 if (!filepage) {
1347                         spin_unlock(&info->lock);
1348                         filepage = shmem_alloc_page(gfp, info, idx);
1349                         if (!filepage) {
1350                                 shmem_unacct_blocks(info->flags, 1);
1351                                 shmem_free_blocks(inode, 1);
1352                                 error = -ENOMEM;
1353                                 goto failed;
1354                         }
1355
1356                         spin_lock(&info->lock);
1357                         entry = shmem_swp_alloc(info, idx, sgp);
1358                         if (IS_ERR(entry))
1359                                 error = PTR_ERR(entry);
1360                         else {
1361                                 swap = *entry;
1362                                 shmem_swp_unmap(entry);
1363                         }
1364                         if (error || swap.val || 0 != add_to_page_cache_lru(
1365                                         filepage, mapping, idx, GFP_NOWAIT)) {
1366                                 spin_unlock(&info->lock);
1367                                 page_cache_release(filepage);
1368                                 shmem_unacct_blocks(info->flags, 1);
1369                                 shmem_free_blocks(inode, 1);
1370                                 filepage = NULL;
1371                                 if (error)
1372                                         goto failed;
1373                                 goto repeat;
1374                         }
1375                         info->flags |= SHMEM_PAGEIN;
1376                 }
1377
1378                 info->alloced++;
1379                 spin_unlock(&info->lock);
1380                 clear_highpage(filepage);
1381                 flush_dcache_page(filepage);
1382                 SetPageUptodate(filepage);
1383                 if (sgp == SGP_DIRTY)
1384                         set_page_dirty(filepage);
1385         }
1386 done:
1387         *pagep = filepage;
1388         return 0;
1389
1390 failed:
1391         if (*pagep != filepage) {
1392                 unlock_page(filepage);
1393                 page_cache_release(filepage);
1394         }
1395         return error;
1396 }
1397
1398 static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1399 {
1400         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1401         int error;
1402         int ret;
1403
1404         if (((loff_t)vmf->pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
1405                 return VM_FAULT_SIGBUS;
1406
1407         error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_CACHE, &ret);
1408         if (error)
1409                 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1410
1411         mark_page_accessed(vmf->page);
1412         return ret | VM_FAULT_LOCKED;
1413 }
1414
1415 #ifdef CONFIG_NUMA
1416 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *new)
1417 {
1418         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1419         return mpol_set_shared_policy(&SHMEM_I(i)->policy, vma, new);
1420 }
1421
1422 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
1423                                           unsigned long addr)
1424 {
1425         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1426         unsigned long idx;
1427
1428         idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1429         return mpol_shared_policy_lookup(&SHMEM_I(i)->policy, idx);
1430 }
1431 #endif
1432
1433 int shmem_lock(struct file *file, int lock, struct user_struct *user)
1434 {
1435         struct inode *inode = file->f_path.dentry->d_inode;
1436         struct shmem_inode_info *info = SHMEM_I(inode);
1437         int retval = -ENOMEM;
1438
1439         spin_lock(&info->lock);
1440         if (lock && !(info->flags & VM_LOCKED)) {
1441                 if (!user_shm_lock(inode->i_size, user))
1442                         goto out_nomem;
1443                 info->flags |= VM_LOCKED;
1444         }
1445         if (!lock && (info->flags & VM_LOCKED) && user) {
1446                 user_shm_unlock(inode->i_size, user);
1447                 info->flags &= ~VM_LOCKED;
1448         }
1449         retval = 0;
1450 out_nomem:
1451         spin_unlock(&info->lock);
1452         return retval;
1453 }
1454
1455 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1456 {
1457         file_accessed(file);
1458         vma->vm_ops = &shmem_vm_ops;
1459         vma->vm_flags |= VM_CAN_NONLINEAR;
1460         return 0;
1461 }
1462
1463 static struct inode *
1464 shmem_get_inode(struct super_block *sb, int mode, dev_t dev)
1465 {
1466         struct inode *inode;
1467         struct shmem_inode_info *info;
1468         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1469
1470         if (shmem_reserve_inode(sb))
1471                 return NULL;
1472
1473         inode = new_inode(sb);
1474         if (inode) {
1475                 inode->i_mode = mode;
1476                 inode->i_uid = current->fsuid;
1477                 inode->i_gid = current->fsgid;
1478                 inode->i_blocks = 0;
1479                 inode->i_mapping->a_ops = &shmem_aops;
1480                 inode->i_mapping->backing_dev_info = &shmem_backing_dev_info;
1481                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1482                 inode->i_generation = get_seconds();
1483                 info = SHMEM_I(inode);
1484                 memset(info, 0, (char *)inode - (char *)info);
1485                 spin_lock_init(&info->lock);
1486                 INIT_LIST_HEAD(&info->swaplist);
1487
1488                 switch (mode & S_IFMT) {
1489                 default:
1490                         inode->i_op = &shmem_special_inode_operations;
1491                         init_special_inode(inode, mode, dev);
1492                         break;
1493                 case S_IFREG:
1494                         inode->i_op = &shmem_inode_operations;
1495                         inode->i_fop = &shmem_file_operations;
1496                         mpol_shared_policy_init(&info->policy, sbinfo->policy,
1497                                                         &sbinfo->policy_nodes);
1498                         break;
1499                 case S_IFDIR:
1500                         inc_nlink(inode);
1501                         /* Some things misbehave if size == 0 on a directory */
1502                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
1503                         inode->i_op = &shmem_dir_inode_operations;
1504                         inode->i_fop = &simple_dir_operations;
1505                         break;
1506                 case S_IFLNK:
1507                         /*
1508                          * Must not load anything in the rbtree,
1509                          * mpol_free_shared_policy will not be called.
1510                          */
1511                         mpol_shared_policy_init(&info->policy, MPOL_DEFAULT,
1512                                                 NULL);
1513                         break;
1514                 }
1515         } else
1516                 shmem_free_inode(sb);
1517         return inode;
1518 }
1519
1520 #ifdef CONFIG_TMPFS
1521 static const struct inode_operations shmem_symlink_inode_operations;
1522 static const struct inode_operations shmem_symlink_inline_operations;
1523
1524 /*
1525  * Normally tmpfs avoids the use of shmem_readpage and shmem_write_begin;
1526  * but providing them allows a tmpfs file to be used for splice, sendfile, and
1527  * below the loop driver, in the generic fashion that many filesystems support.
1528  */
1529 static int shmem_readpage(struct file *file, struct page *page)
1530 {
1531         struct inode *inode = page->mapping->host;
1532         int error = shmem_getpage(inode, page->index, &page, SGP_CACHE, NULL);
1533         unlock_page(page);
1534         return error;
1535 }
1536
1537 static int
1538 shmem_write_begin(struct file *file, struct address_space *mapping,
1539                         loff_t pos, unsigned len, unsigned flags,
1540                         struct page **pagep, void **fsdata)
1541 {
1542         struct inode *inode = mapping->host;
1543         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1544         *pagep = NULL;
1545         return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1546 }
1547
1548 static int
1549 shmem_write_end(struct file *file, struct address_space *mapping,
1550                         loff_t pos, unsigned len, unsigned copied,
1551                         struct page *page, void *fsdata)
1552 {
1553         struct inode *inode = mapping->host;
1554
1555         if (pos + copied > inode->i_size)
1556                 i_size_write(inode, pos + copied);
1557
1558         unlock_page(page);
1559         set_page_dirty(page);
1560         page_cache_release(page);
1561
1562         return copied;
1563 }
1564
1565 static void do_shmem_file_read(struct file *filp, loff_t *ppos, read_descriptor_t *desc, read_actor_t actor)
1566 {
1567         struct inode *inode = filp->f_path.dentry->d_inode;
1568         struct address_space *mapping = inode->i_mapping;
1569         unsigned long index, offset;
1570         enum sgp_type sgp = SGP_READ;
1571
1572         /*
1573          * Might this read be for a stacking filesystem?  Then when reading
1574          * holes of a sparse file, we actually need to allocate those pages,
1575          * and even mark them dirty, so it cannot exceed the max_blocks limit.
1576          */
1577         if (segment_eq(get_fs(), KERNEL_DS))
1578                 sgp = SGP_DIRTY;
1579
1580         index = *ppos >> PAGE_CACHE_SHIFT;
1581         offset = *ppos & ~PAGE_CACHE_MASK;
1582
1583         for (;;) {
1584                 struct page *page = NULL;
1585                 unsigned long end_index, nr, ret;
1586                 loff_t i_size = i_size_read(inode);
1587
1588                 end_index = i_size >> PAGE_CACHE_SHIFT;
1589                 if (index > end_index)
1590                         break;
1591                 if (index == end_index) {
1592                         nr = i_size & ~PAGE_CACHE_MASK;
1593                         if (nr <= offset)
1594                                 break;
1595                 }
1596
1597                 desc->error = shmem_getpage(inode, index, &page, sgp, NULL);
1598                 if (desc->error) {
1599                         if (desc->error == -EINVAL)
1600                                 desc->error = 0;
1601                         break;
1602                 }
1603                 if (page)
1604                         unlock_page(page);
1605
1606                 /*
1607                  * We must evaluate after, since reads (unlike writes)
1608                  * are called without i_mutex protection against truncate
1609                  */
1610                 nr = PAGE_CACHE_SIZE;
1611                 i_size = i_size_read(inode);
1612                 end_index = i_size >> PAGE_CACHE_SHIFT;
1613                 if (index == end_index) {
1614                         nr = i_size & ~PAGE_CACHE_MASK;
1615                         if (nr <= offset) {
1616                                 if (page)
1617                                         page_cache_release(page);
1618                                 break;
1619                         }
1620                 }
1621                 nr -= offset;
1622
1623                 if (page) {
1624                         /*
1625                          * If users can be writing to this page using arbitrary
1626                          * virtual addresses, take care about potential aliasing
1627                          * before reading the page on the kernel side.
1628                          */
1629                         if (mapping_writably_mapped(mapping))
1630                                 flush_dcache_page(page);
1631                         /*
1632                          * Mark the page accessed if we read the beginning.
1633                          */
1634                         if (!offset)
1635                                 mark_page_accessed(page);
1636                 } else {
1637                         page = ZERO_PAGE(0);
1638                         page_cache_get(page);
1639                 }
1640
1641                 /*
1642                  * Ok, we have the page, and it's up-to-date, so
1643                  * now we can copy it to user space...
1644                  *
1645                  * The actor routine returns how many bytes were actually used..
1646                  * NOTE! This may not be the same as how much of a user buffer
1647                  * we filled up (we may be padding etc), so we can only update
1648                  * "pos" here (the actor routine has to update the user buffer
1649                  * pointers and the remaining count).
1650                  */
1651                 ret = actor(desc, page, offset, nr);
1652                 offset += ret;
1653                 index += offset >> PAGE_CACHE_SHIFT;
1654                 offset &= ~PAGE_CACHE_MASK;
1655
1656                 page_cache_release(page);
1657                 if (ret != nr || !desc->count)
1658                         break;
1659
1660                 cond_resched();
1661         }
1662
1663         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1664         file_accessed(filp);
1665 }
1666
1667 static ssize_t shmem_file_read(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
1668 {
1669         read_descriptor_t desc;
1670
1671         if ((ssize_t) count < 0)
1672                 return -EINVAL;
1673         if (!access_ok(VERIFY_WRITE, buf, count))
1674                 return -EFAULT;
1675         if (!count)
1676                 return 0;
1677
1678         desc.written = 0;
1679         desc.count = count;
1680         desc.arg.buf = buf;
1681         desc.error = 0;
1682
1683         do_shmem_file_read(filp, ppos, &desc, file_read_actor);
1684         if (desc.written)
1685                 return desc.written;
1686         return desc.error;
1687 }
1688
1689 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
1690 {
1691         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
1692
1693         buf->f_type = TMPFS_MAGIC;
1694         buf->f_bsize = PAGE_CACHE_SIZE;
1695         buf->f_namelen = NAME_MAX;
1696         spin_lock(&sbinfo->stat_lock);
1697         if (sbinfo->max_blocks) {
1698                 buf->f_blocks = sbinfo->max_blocks;
1699                 buf->f_bavail = buf->f_bfree = sbinfo->free_blocks;
1700         }
1701         if (sbinfo->max_inodes) {
1702                 buf->f_files = sbinfo->max_inodes;
1703                 buf->f_ffree = sbinfo->free_inodes;
1704         }
1705         /* else leave those fields 0 like simple_statfs */
1706         spin_unlock(&sbinfo->stat_lock);
1707         return 0;
1708 }
1709
1710 /*
1711  * File creation. Allocate an inode, and we're done..
1712  */
1713 static int
1714 shmem_mknod(struct inode *dir, struct dentry *dentry, int mode, dev_t dev)
1715 {
1716         struct inode *inode = shmem_get_inode(dir->i_sb, mode, dev);
1717         int error = -ENOSPC;
1718
1719         if (inode) {
1720                 error = security_inode_init_security(inode, dir, NULL, NULL,
1721                                                      NULL);
1722                 if (error) {
1723                         if (error != -EOPNOTSUPP) {
1724                                 iput(inode);
1725                                 return error;
1726                         }
1727                 }
1728                 error = shmem_acl_init(inode, dir);
1729                 if (error) {
1730                         iput(inode);
1731                         return error;
1732                 }
1733                 if (dir->i_mode & S_ISGID) {
1734                         inode->i_gid = dir->i_gid;
1735                         if (S_ISDIR(mode))
1736                                 inode->i_mode |= S_ISGID;
1737                 }
1738                 dir->i_size += BOGO_DIRENT_SIZE;
1739                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1740                 d_instantiate(dentry, inode);
1741                 dget(dentry); /* Extra count - pin the dentry in core */
1742         }
1743         return error;
1744 }
1745
1746 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, int mode)
1747 {
1748         int error;
1749
1750         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
1751                 return error;
1752         inc_nlink(dir);
1753         return 0;
1754 }
1755
1756 static int shmem_create(struct inode *dir, struct dentry *dentry, int mode,
1757                 struct nameidata *nd)
1758 {
1759         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
1760 }
1761
1762 /*
1763  * Link a file..
1764  */
1765 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
1766 {
1767         struct inode *inode = old_dentry->d_inode;
1768         int ret;
1769
1770         /*
1771          * No ordinary (disk based) filesystem counts links as inodes;
1772          * but each new link needs a new dentry, pinning lowmem, and
1773          * tmpfs dentries cannot be pruned until they are unlinked.
1774          */
1775         ret = shmem_reserve_inode(inode->i_sb);
1776         if (ret)
1777                 goto out;
1778
1779         dir->i_size += BOGO_DIRENT_SIZE;
1780         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1781         inc_nlink(inode);
1782         atomic_inc(&inode->i_count);    /* New dentry reference */
1783         dget(dentry);           /* Extra pinning count for the created dentry */
1784         d_instantiate(dentry, inode);
1785 out:
1786         return ret;
1787 }
1788
1789 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
1790 {
1791         struct inode *inode = dentry->d_inode;
1792
1793         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
1794                 shmem_free_inode(inode->i_sb);
1795
1796         dir->i_size -= BOGO_DIRENT_SIZE;
1797         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1798         drop_nlink(inode);
1799         dput(dentry);   /* Undo the count from "create" - this does all the work */
1800         return 0;
1801 }
1802
1803 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
1804 {
1805         if (!simple_empty(dentry))
1806                 return -ENOTEMPTY;
1807
1808         drop_nlink(dentry->d_inode);
1809         drop_nlink(dir);
1810         return shmem_unlink(dir, dentry);
1811 }
1812
1813 /*
1814  * The VFS layer already does all the dentry stuff for rename,
1815  * we just have to decrement the usage count for the target if
1816  * it exists so that the VFS layer correctly free's it when it
1817  * gets overwritten.
1818  */
1819 static int shmem_rename(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
1820 {
1821         struct inode *inode = old_dentry->d_inode;
1822         int they_are_dirs = S_ISDIR(inode->i_mode);
1823
1824         if (!simple_empty(new_dentry))
1825                 return -ENOTEMPTY;
1826
1827         if (new_dentry->d_inode) {
1828                 (void) shmem_unlink(new_dir, new_dentry);
1829                 if (they_are_dirs)
1830                         drop_nlink(old_dir);
1831         } else if (they_are_dirs) {
1832                 drop_nlink(old_dir);
1833                 inc_nlink(new_dir);
1834         }
1835
1836         old_dir->i_size -= BOGO_DIRENT_SIZE;
1837         new_dir->i_size += BOGO_DIRENT_SIZE;
1838         old_dir->i_ctime = old_dir->i_mtime =
1839         new_dir->i_ctime = new_dir->i_mtime =
1840         inode->i_ctime = CURRENT_TIME;
1841         return 0;
1842 }
1843
1844 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
1845 {
1846         int error;
1847         int len;
1848         struct inode *inode;
1849         struct page *page = NULL;
1850         char *kaddr;
1851         struct shmem_inode_info *info;
1852
1853         len = strlen(symname) + 1;
1854         if (len > PAGE_CACHE_SIZE)
1855                 return -ENAMETOOLONG;
1856
1857         inode = shmem_get_inode(dir->i_sb, S_IFLNK|S_IRWXUGO, 0);
1858         if (!inode)
1859                 return -ENOSPC;
1860
1861         error = security_inode_init_security(inode, dir, NULL, NULL,
1862                                              NULL);
1863         if (error) {
1864                 if (error != -EOPNOTSUPP) {
1865                         iput(inode);
1866                         return error;
1867                 }
1868                 error = 0;
1869         }
1870
1871         info = SHMEM_I(inode);
1872         inode->i_size = len-1;
1873         if (len <= (char *)inode - (char *)info) {
1874                 /* do it inline */
1875                 memcpy(info, symname, len);
1876                 inode->i_op = &shmem_symlink_inline_operations;
1877         } else {
1878                 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
1879                 if (error) {
1880                         iput(inode);
1881                         return error;
1882                 }
1883                 unlock_page(page);
1884                 inode->i_op = &shmem_symlink_inode_operations;
1885                 kaddr = kmap_atomic(page, KM_USER0);
1886                 memcpy(kaddr, symname, len);
1887                 kunmap_atomic(kaddr, KM_USER0);
1888                 set_page_dirty(page);
1889                 page_cache_release(page);
1890         }
1891         if (dir->i_mode & S_ISGID)
1892                 inode->i_gid = dir->i_gid;
1893         dir->i_size += BOGO_DIRENT_SIZE;
1894         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1895         d_instantiate(dentry, inode);
1896         dget(dentry);
1897         return 0;
1898 }
1899
1900 static void *shmem_follow_link_inline(struct dentry *dentry, struct nameidata *nd)
1901 {
1902         nd_set_link(nd, (char *)SHMEM_I(dentry->d_inode));
1903         return NULL;
1904 }
1905
1906 static void *shmem_follow_link(struct dentry *dentry, struct nameidata *nd)
1907 {
1908         struct page *page = NULL;
1909         int res = shmem_getpage(dentry->d_inode, 0, &page, SGP_READ, NULL);
1910         nd_set_link(nd, res ? ERR_PTR(res) : kmap(page));
1911         if (page)
1912                 unlock_page(page);
1913         return page;
1914 }
1915
1916 static void shmem_put_link(struct dentry *dentry, struct nameidata *nd, void *cookie)
1917 {
1918         if (!IS_ERR(nd_get_link(nd))) {
1919                 struct page *page = cookie;
1920                 kunmap(page);
1921                 mark_page_accessed(page);
1922                 page_cache_release(page);
1923         }
1924 }
1925
1926 static const struct inode_operations shmem_symlink_inline_operations = {
1927         .readlink       = generic_readlink,
1928         .follow_link    = shmem_follow_link_inline,
1929 };
1930
1931 static const struct inode_operations shmem_symlink_inode_operations = {
1932         .truncate       = shmem_truncate,
1933         .readlink       = generic_readlink,
1934         .follow_link    = shmem_follow_link,
1935         .put_link       = shmem_put_link,
1936 };
1937
1938 #ifdef CONFIG_TMPFS_POSIX_ACL
1939 /**
1940  * Superblocks without xattr inode operations will get security.* xattr
1941  * support from the VFS "for free". As soon as we have any other xattrs
1942  * like ACLs, we also need to implement the security.* handlers at
1943  * filesystem level, though.
1944  */
1945
1946 static size_t shmem_xattr_security_list(struct inode *inode, char *list,
1947                                         size_t list_len, const char *name,
1948                                         size_t name_len)
1949 {
1950         return security_inode_listsecurity(inode, list, list_len);
1951 }
1952
1953 static int shmem_xattr_security_get(struct inode *inode, const char *name,
1954                                     void *buffer, size_t size)
1955 {
1956         if (strcmp(name, "") == 0)
1957                 return -EINVAL;
1958         return security_inode_getsecurity(inode, name, buffer, size,
1959                                           -EOPNOTSUPP);
1960 }
1961
1962 static int shmem_xattr_security_set(struct inode *inode, const char *name,
1963                                     const void *value, size_t size, int flags)
1964 {
1965         if (strcmp(name, "") == 0)
1966                 return -EINVAL;
1967         return security_inode_setsecurity(inode, name, value, size, flags);
1968 }
1969
1970 static struct xattr_handler shmem_xattr_security_handler = {
1971         .prefix = XATTR_SECURITY_PREFIX,
1972         .list   = shmem_xattr_security_list,
1973         .get    = shmem_xattr_security_get,
1974         .set    = shmem_xattr_security_set,
1975 };
1976
1977 static struct xattr_handler *shmem_xattr_handlers[] = {
1978         &shmem_xattr_acl_access_handler,
1979         &shmem_xattr_acl_default_handler,
1980         &shmem_xattr_security_handler,
1981         NULL
1982 };
1983 #endif
1984
1985 static struct dentry *shmem_get_parent(struct dentry *child)
1986 {
1987         return ERR_PTR(-ESTALE);
1988 }
1989
1990 static int shmem_match(struct inode *ino, void *vfh)
1991 {
1992         __u32 *fh = vfh;
1993         __u64 inum = fh[2];
1994         inum = (inum << 32) | fh[1];
1995         return ino->i_ino == inum && fh[0] == ino->i_generation;
1996 }
1997
1998 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
1999                 struct fid *fid, int fh_len, int fh_type)
2000 {
2001         struct inode *inode;
2002         struct dentry *dentry = NULL;
2003         u64 inum = fid->raw[2];
2004         inum = (inum << 32) | fid->raw[1];
2005
2006         if (fh_len < 3)
2007                 return NULL;
2008
2009         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
2010                         shmem_match, fid->raw);
2011         if (inode) {
2012                 dentry = d_find_alias(inode);
2013                 iput(inode);
2014         }
2015
2016         return dentry;
2017 }
2018
2019 static int shmem_encode_fh(struct dentry *dentry, __u32 *fh, int *len,
2020                                 int connectable)
2021 {
2022         struct inode *inode = dentry->d_inode;
2023
2024         if (*len < 3)
2025                 return 255;
2026
2027         if (hlist_unhashed(&inode->i_hash)) {
2028                 /* Unfortunately insert_inode_hash is not idempotent,
2029                  * so as we hash inodes here rather than at creation
2030                  * time, we need a lock to ensure we only try
2031                  * to do it once
2032                  */
2033                 static DEFINE_SPINLOCK(lock);
2034                 spin_lock(&lock);
2035                 if (hlist_unhashed(&inode->i_hash))
2036                         __insert_inode_hash(inode,
2037                                             inode->i_ino + inode->i_generation);
2038                 spin_unlock(&lock);
2039         }
2040
2041         fh[0] = inode->i_generation;
2042         fh[1] = inode->i_ino;
2043         fh[2] = ((__u64)inode->i_ino) >> 32;
2044
2045         *len = 3;
2046         return 1;
2047 }
2048
2049 static const struct export_operations shmem_export_ops = {
2050         .get_parent     = shmem_get_parent,
2051         .encode_fh      = shmem_encode_fh,
2052         .fh_to_dentry   = shmem_fh_to_dentry,
2053 };
2054
2055 static int shmem_parse_options(char *options, int *mode, uid_t *uid,
2056         gid_t *gid, unsigned long *blocks, unsigned long *inodes,
2057         int *policy, nodemask_t *policy_nodes)
2058 {
2059         char *this_char, *value, *rest;
2060
2061         while (options != NULL) {
2062                 this_char = options;
2063                 for (;;) {
2064                         /*
2065                          * NUL-terminate this option: unfortunately,
2066                          * mount options form a comma-separated list,
2067                          * but mpol's nodelist may also contain commas.
2068                          */
2069                         options = strchr(options, ',');
2070                         if (options == NULL)
2071                                 break;
2072                         options++;
2073                         if (!isdigit(*options)) {
2074                                 options[-1] = '\0';
2075                                 break;
2076                         }
2077                 }
2078                 if (!*this_char)
2079                         continue;
2080                 if ((value = strchr(this_char,'=')) != NULL) {
2081                         *value++ = 0;
2082                 } else {
2083                         printk(KERN_ERR
2084                             "tmpfs: No value for mount option '%s'\n",
2085                             this_char);
2086                         return 1;
2087                 }
2088
2089                 if (!strcmp(this_char,"size")) {
2090                         unsigned long long size;
2091                         size = memparse(value,&rest);
2092                         if (*rest == '%') {
2093                                 size <<= PAGE_SHIFT;
2094                                 size *= totalram_pages;
2095                                 do_div(size, 100);
2096                                 rest++;
2097                         }
2098                         if (*rest)
2099                                 goto bad_val;
2100                         *blocks = DIV_ROUND_UP(size, PAGE_CACHE_SIZE);
2101                 } else if (!strcmp(this_char,"nr_blocks")) {
2102                         *blocks = memparse(value,&rest);
2103                         if (*rest)
2104                                 goto bad_val;
2105                 } else if (!strcmp(this_char,"nr_inodes")) {
2106                         *inodes = memparse(value,&rest);
2107                         if (*rest)
2108                                 goto bad_val;
2109                 } else if (!strcmp(this_char,"mode")) {
2110                         if (!mode)
2111                                 continue;
2112                         *mode = simple_strtoul(value,&rest,8);
2113                         if (*rest)
2114                                 goto bad_val;
2115                 } else if (!strcmp(this_char,"uid")) {
2116                         if (!uid)
2117                                 continue;
2118                         *uid = simple_strtoul(value,&rest,0);
2119                         if (*rest)
2120                                 goto bad_val;
2121                 } else if (!strcmp(this_char,"gid")) {
2122                         if (!gid)
2123                                 continue;
2124                         *gid = simple_strtoul(value,&rest,0);
2125                         if (*rest)
2126                                 goto bad_val;
2127                 } else if (!strcmp(this_char,"mpol")) {
2128                         if (shmem_parse_mpol(value,policy,policy_nodes))
2129                                 goto bad_val;
2130                 } else {
2131                         printk(KERN_ERR "tmpfs: Bad mount option %s\n",
2132                                this_char);
2133                         return 1;
2134                 }
2135         }
2136         return 0;
2137
2138 bad_val:
2139         printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
2140                value, this_char);
2141         return 1;
2142
2143 }
2144
2145 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2146 {
2147         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2148         unsigned long max_blocks = sbinfo->max_blocks;
2149         unsigned long max_inodes = sbinfo->max_inodes;
2150         int policy = sbinfo->policy;
2151         nodemask_t policy_nodes = sbinfo->policy_nodes;
2152         unsigned long blocks;
2153         unsigned long inodes;
2154         int error = -EINVAL;
2155
2156         if (shmem_parse_options(data, NULL, NULL, NULL, &max_blocks,
2157                                 &max_inodes, &policy, &policy_nodes))
2158                 return error;
2159
2160         spin_lock(&sbinfo->stat_lock);
2161         blocks = sbinfo->max_blocks - sbinfo->free_blocks;
2162         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2163         if (max_blocks < blocks)
2164                 goto out;
2165         if (max_inodes < inodes)
2166                 goto out;
2167         /*
2168          * Those tests also disallow limited->unlimited while any are in
2169          * use, so i_blocks will always be zero when max_blocks is zero;
2170          * but we must separately disallow unlimited->limited, because
2171          * in that case we have no record of how much is already in use.
2172          */
2173         if (max_blocks && !sbinfo->max_blocks)
2174                 goto out;
2175         if (max_inodes && !sbinfo->max_inodes)
2176                 goto out;
2177
2178         error = 0;
2179         sbinfo->max_blocks  = max_blocks;
2180         sbinfo->free_blocks = max_blocks - blocks;
2181         sbinfo->max_inodes  = max_inodes;
2182         sbinfo->free_inodes = max_inodes - inodes;
2183         sbinfo->policy = policy;
2184         sbinfo->policy_nodes = policy_nodes;
2185 out:
2186         spin_unlock(&sbinfo->stat_lock);
2187         return error;
2188 }
2189 #endif
2190
2191 static void shmem_put_super(struct super_block *sb)
2192 {
2193         kfree(sb->s_fs_info);
2194         sb->s_fs_info = NULL;
2195 }
2196
2197 static int shmem_fill_super(struct super_block *sb,
2198                             void *data, int silent)
2199 {
2200         struct inode *inode;
2201         struct dentry *root;
2202         int mode   = S_IRWXUGO | S_ISVTX;
2203         uid_t uid = current->fsuid;
2204         gid_t gid = current->fsgid;
2205         int err = -ENOMEM;
2206         struct shmem_sb_info *sbinfo;
2207         unsigned long blocks = 0;
2208         unsigned long inodes = 0;
2209         int policy = MPOL_DEFAULT;
2210         nodemask_t policy_nodes = node_states[N_HIGH_MEMORY];
2211
2212 #ifdef CONFIG_TMPFS
2213         /*
2214          * Per default we only allow half of the physical ram per
2215          * tmpfs instance, limiting inodes to one per page of lowmem;
2216          * but the internal instance is left unlimited.
2217          */
2218         if (!(sb->s_flags & MS_NOUSER)) {
2219                 blocks = totalram_pages / 2;
2220                 inodes = totalram_pages - totalhigh_pages;
2221                 if (inodes > blocks)
2222                         inodes = blocks;
2223                 if (shmem_parse_options(data, &mode, &uid, &gid, &blocks,
2224                                         &inodes, &policy, &policy_nodes))
2225                         return -EINVAL;
2226         }
2227         sb->s_export_op = &shmem_export_ops;
2228 #else
2229         sb->s_flags |= MS_NOUSER;
2230 #endif
2231
2232         /* Round up to L1_CACHE_BYTES to resist false sharing */
2233         sbinfo = kmalloc(max((int)sizeof(struct shmem_sb_info),
2234                                 L1_CACHE_BYTES), GFP_KERNEL);
2235         if (!sbinfo)
2236                 return -ENOMEM;
2237
2238         spin_lock_init(&sbinfo->stat_lock);
2239         sbinfo->max_blocks = blocks;
2240         sbinfo->free_blocks = blocks;
2241         sbinfo->max_inodes = inodes;
2242         sbinfo->free_inodes = inodes;
2243         sbinfo->policy = policy;
2244         sbinfo->policy_nodes = policy_nodes;
2245
2246         sb->s_fs_info = sbinfo;
2247         sb->s_maxbytes = SHMEM_MAX_BYTES;
2248         sb->s_blocksize = PAGE_CACHE_SIZE;
2249         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
2250         sb->s_magic = TMPFS_MAGIC;
2251         sb->s_op = &shmem_ops;
2252         sb->s_time_gran = 1;
2253 #ifdef CONFIG_TMPFS_POSIX_ACL
2254         sb->s_xattr = shmem_xattr_handlers;
2255         sb->s_flags |= MS_POSIXACL;
2256 #endif
2257
2258         inode = shmem_get_inode(sb, S_IFDIR | mode, 0);
2259         if (!inode)
2260                 goto failed;
2261         inode->i_uid = uid;
2262         inode->i_gid = gid;
2263         root = d_alloc_root(inode);
2264         if (!root)
2265                 goto failed_iput;
2266         sb->s_root = root;
2267         return 0;
2268
2269 failed_iput:
2270         iput(inode);
2271 failed:
2272         shmem_put_super(sb);
2273         return err;
2274 }
2275
2276 static struct kmem_cache *shmem_inode_cachep;
2277
2278 static struct inode *shmem_alloc_inode(struct super_block *sb)
2279 {
2280         struct shmem_inode_info *p;
2281         p = (struct shmem_inode_info *)kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
2282         if (!p)
2283                 return NULL;
2284         return &p->vfs_inode;
2285 }
2286
2287 static void shmem_destroy_inode(struct inode *inode)
2288 {
2289         if ((inode->i_mode & S_IFMT) == S_IFREG) {
2290                 /* only struct inode is valid if it's an inline symlink */
2291                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
2292         }
2293         shmem_acl_destroy_inode(inode);
2294         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
2295 }
2296
2297 static void init_once(struct kmem_cache *cachep, void *foo)
2298 {
2299         struct shmem_inode_info *p = (struct shmem_inode_info *) foo;
2300
2301         inode_init_once(&p->vfs_inode);
2302 #ifdef CONFIG_TMPFS_POSIX_ACL
2303         p->i_acl = NULL;
2304         p->i_default_acl = NULL;
2305 #endif
2306 }
2307
2308 static int init_inodecache(void)
2309 {
2310         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
2311                                 sizeof(struct shmem_inode_info),
2312                                 0, SLAB_PANIC, init_once);
2313         return 0;
2314 }
2315
2316 static void destroy_inodecache(void)
2317 {
2318         kmem_cache_destroy(shmem_inode_cachep);
2319 }
2320
2321 static const struct address_space_operations shmem_aops = {
2322         .writepage      = shmem_writepage,
2323         .set_page_dirty = __set_page_dirty_no_writeback,
2324 #ifdef CONFIG_TMPFS
2325         .readpage       = shmem_readpage,
2326         .write_begin    = shmem_write_begin,
2327         .write_end      = shmem_write_end,
2328 #endif
2329         .migratepage    = migrate_page,
2330 };
2331
2332 static const struct file_operations shmem_file_operations = {
2333         .mmap           = shmem_mmap,
2334 #ifdef CONFIG_TMPFS
2335         .llseek         = generic_file_llseek,
2336         .read           = shmem_file_read,
2337         .write          = do_sync_write,
2338         .aio_write      = generic_file_aio_write,
2339         .fsync          = simple_sync_file,
2340         .splice_read    = generic_file_splice_read,
2341         .splice_write   = generic_file_splice_write,
2342 #endif
2343 };
2344
2345 static const struct inode_operations shmem_inode_operations = {
2346         .truncate       = shmem_truncate,
2347         .setattr        = shmem_notify_change,
2348         .truncate_range = shmem_truncate_range,
2349 #ifdef CONFIG_TMPFS_POSIX_ACL
2350         .setxattr       = generic_setxattr,
2351         .getxattr       = generic_getxattr,
2352         .listxattr      = generic_listxattr,
2353         .removexattr    = generic_removexattr,
2354         .permission     = shmem_permission,
2355 #endif
2356
2357 };
2358
2359 static const struct inode_operations shmem_dir_inode_operations = {
2360 #ifdef CONFIG_TMPFS
2361         .create         = shmem_create,
2362         .lookup         = simple_lookup,
2363         .link           = shmem_link,
2364         .unlink         = shmem_unlink,
2365         .symlink        = shmem_symlink,
2366         .mkdir          = shmem_mkdir,
2367         .rmdir          = shmem_rmdir,
2368         .mknod          = shmem_mknod,
2369         .rename         = shmem_rename,
2370 #endif
2371 #ifdef CONFIG_TMPFS_POSIX_ACL
2372         .setattr        = shmem_notify_change,
2373         .setxattr       = generic_setxattr,
2374         .getxattr       = generic_getxattr,
2375         .listxattr      = generic_listxattr,
2376         .removexattr    = generic_removexattr,
2377         .permission     = shmem_permission,
2378 #endif
2379 };
2380
2381 static const struct inode_operations shmem_special_inode_operations = {
2382 #ifdef CONFIG_TMPFS_POSIX_ACL
2383         .setattr        = shmem_notify_change,
2384         .setxattr       = generic_setxattr,
2385         .getxattr       = generic_getxattr,
2386         .listxattr      = generic_listxattr,
2387         .removexattr    = generic_removexattr,
2388         .permission     = shmem_permission,
2389 #endif
2390 };
2391
2392 static const struct super_operations shmem_ops = {
2393         .alloc_inode    = shmem_alloc_inode,
2394         .destroy_inode  = shmem_destroy_inode,
2395 #ifdef CONFIG_TMPFS
2396         .statfs         = shmem_statfs,
2397         .remount_fs     = shmem_remount_fs,
2398 #endif
2399         .delete_inode   = shmem_delete_inode,
2400         .drop_inode     = generic_delete_inode,
2401         .put_super      = shmem_put_super,
2402 };
2403
2404 static struct vm_operations_struct shmem_vm_ops = {
2405         .fault          = shmem_fault,
2406 #ifdef CONFIG_NUMA
2407         .set_policy     = shmem_set_policy,
2408         .get_policy     = shmem_get_policy,
2409 #endif
2410 };
2411
2412
2413 static int shmem_get_sb(struct file_system_type *fs_type,
2414         int flags, const char *dev_name, void *data, struct vfsmount *mnt)
2415 {
2416         return get_sb_nodev(fs_type, flags, data, shmem_fill_super, mnt);
2417 }
2418
2419 static struct file_system_type tmpfs_fs_type = {
2420         .owner          = THIS_MODULE,
2421         .name           = "tmpfs",
2422         .get_sb         = shmem_get_sb,
2423         .kill_sb        = kill_litter_super,
2424 };
2425 static struct vfsmount *shm_mnt;
2426
2427 static int __init init_tmpfs(void)
2428 {
2429         int error;
2430
2431         error = bdi_init(&shmem_backing_dev_info);
2432         if (error)
2433                 goto out4;
2434
2435         error = init_inodecache();
2436         if (error)
2437                 goto out3;
2438
2439         error = register_filesystem(&tmpfs_fs_type);
2440         if (error) {
2441                 printk(KERN_ERR "Could not register tmpfs\n");
2442                 goto out2;
2443         }
2444
2445         shm_mnt = vfs_kern_mount(&tmpfs_fs_type, MS_NOUSER,
2446                                 tmpfs_fs_type.name, NULL);
2447         if (IS_ERR(shm_mnt)) {
2448                 error = PTR_ERR(shm_mnt);
2449                 printk(KERN_ERR "Could not kern_mount tmpfs\n");
2450                 goto out1;
2451         }
2452         return 0;
2453
2454 out1:
2455         unregister_filesystem(&tmpfs_fs_type);
2456 out2:
2457         destroy_inodecache();
2458 out3:
2459         bdi_destroy(&shmem_backing_dev_info);
2460 out4:
2461         shm_mnt = ERR_PTR(error);
2462         return error;
2463 }
2464 module_init(init_tmpfs)
2465
2466 /*
2467  * shmem_file_setup - get an unlinked file living in tmpfs
2468  *
2469  * @name: name for dentry (to be seen in /proc/<pid>/maps
2470  * @size: size to be set for the file
2471  *
2472  */
2473 struct file *shmem_file_setup(char *name, loff_t size, unsigned long flags)
2474 {
2475         int error;
2476         struct file *file;
2477         struct inode *inode;
2478         struct dentry *dentry, *root;
2479         struct qstr this;
2480
2481         if (IS_ERR(shm_mnt))
2482                 return (void *)shm_mnt;
2483
2484         if (size < 0 || size > SHMEM_MAX_BYTES)
2485                 return ERR_PTR(-EINVAL);
2486
2487         if (shmem_acct_size(flags, size))
2488                 return ERR_PTR(-ENOMEM);
2489
2490         error = -ENOMEM;
2491         this.name = name;
2492         this.len = strlen(name);
2493         this.hash = 0; /* will go */
2494         root = shm_mnt->mnt_root;
2495         dentry = d_alloc(root, &this);
2496         if (!dentry)
2497                 goto put_memory;
2498
2499         error = -ENFILE;
2500         file = get_empty_filp();
2501         if (!file)
2502                 goto put_dentry;
2503
2504         error = -ENOSPC;
2505         inode = shmem_get_inode(root->d_sb, S_IFREG | S_IRWXUGO, 0);
2506         if (!inode)
2507                 goto close_file;
2508
2509         SHMEM_I(inode)->flags = flags & VM_ACCOUNT;
2510         d_instantiate(dentry, inode);
2511         inode->i_size = size;
2512         inode->i_nlink = 0;     /* It is unlinked */
2513         init_file(file, shm_mnt, dentry, FMODE_WRITE | FMODE_READ,
2514                         &shmem_file_operations);
2515         return file;
2516
2517 close_file:
2518         put_filp(file);
2519 put_dentry:
2520         dput(dentry);
2521 put_memory:
2522         shmem_unacct_size(flags, size);
2523         return ERR_PTR(error);
2524 }
2525
2526 /*
2527  * shmem_zero_setup - setup a shared anonymous mapping
2528  *
2529  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
2530  */
2531 int shmem_zero_setup(struct vm_area_struct *vma)
2532 {
2533         struct file *file;
2534         loff_t size = vma->vm_end - vma->vm_start;
2535
2536         file = shmem_file_setup("dev/zero", size, vma->vm_flags);
2537         if (IS_ERR(file))
2538                 return PTR_ERR(file);
2539
2540         if (vma->vm_file)
2541                 fput(vma->vm_file);
2542         vma->vm_file = file;
2543         vma->vm_ops = &shmem_vm_ops;
2544         return 0;
2545 }