swapin needs gfp_mask for loop on tmpfs
[linux-2.6.git] / mm / shmem.c
1 /*
2  * Resizable virtual memory filesystem for Linux.
3  *
4  * Copyright (C) 2000 Linus Torvalds.
5  *               2000 Transmeta Corp.
6  *               2000-2001 Christoph Rohland
7  *               2000-2001 SAP AG
8  *               2002 Red Hat Inc.
9  * Copyright (C) 2002-2005 Hugh Dickins.
10  * Copyright (C) 2002-2005 VERITAS Software Corporation.
11  * Copyright (C) 2004 Andi Kleen, SuSE Labs
12  *
13  * Extended attribute support for tmpfs:
14  * Copyright (c) 2004, Luke Kenneth Casson Leighton <lkcl@lkcl.net>
15  * Copyright (c) 2004 Red Hat, Inc., James Morris <jmorris@redhat.com>
16  *
17  * This file is released under the GPL.
18  */
19
20 /*
21  * This virtual memory filesystem is heavily based on the ramfs. It
22  * extends ramfs by the ability to use swap and honor resource limits
23  * which makes it a completely usable filesystem.
24  */
25
26 #include <linux/module.h>
27 #include <linux/init.h>
28 #include <linux/fs.h>
29 #include <linux/xattr.h>
30 #include <linux/exportfs.h>
31 #include <linux/generic_acl.h>
32 #include <linux/mm.h>
33 #include <linux/mman.h>
34 #include <linux/file.h>
35 #include <linux/swap.h>
36 #include <linux/pagemap.h>
37 #include <linux/string.h>
38 #include <linux/slab.h>
39 #include <linux/backing-dev.h>
40 #include <linux/shmem_fs.h>
41 #include <linux/mount.h>
42 #include <linux/writeback.h>
43 #include <linux/vfs.h>
44 #include <linux/blkdev.h>
45 #include <linux/security.h>
46 #include <linux/swapops.h>
47 #include <linux/mempolicy.h>
48 #include <linux/namei.h>
49 #include <linux/ctype.h>
50 #include <linux/migrate.h>
51 #include <linux/highmem.h>
52
53 #include <asm/uaccess.h>
54 #include <asm/div64.h>
55 #include <asm/pgtable.h>
56
57 /* This magic number is used in glibc for posix shared memory */
58 #define TMPFS_MAGIC     0x01021994
59
60 #define ENTRIES_PER_PAGE (PAGE_CACHE_SIZE/sizeof(unsigned long))
61 #define ENTRIES_PER_PAGEPAGE (ENTRIES_PER_PAGE*ENTRIES_PER_PAGE)
62 #define BLOCKS_PER_PAGE  (PAGE_CACHE_SIZE/512)
63
64 #define SHMEM_MAX_INDEX  (SHMEM_NR_DIRECT + (ENTRIES_PER_PAGEPAGE/2) * (ENTRIES_PER_PAGE+1))
65 #define SHMEM_MAX_BYTES  ((unsigned long long)SHMEM_MAX_INDEX << PAGE_CACHE_SHIFT)
66
67 #define VM_ACCT(size)    (PAGE_CACHE_ALIGN(size) >> PAGE_SHIFT)
68
69 /* info->flags needs VM_flags to handle pagein/truncate races efficiently */
70 #define SHMEM_PAGEIN     VM_READ
71 #define SHMEM_TRUNCATE   VM_WRITE
72
73 /* Definition to limit shmem_truncate's steps between cond_rescheds */
74 #define LATENCY_LIMIT    64
75
76 /* Pretend that each entry is of this size in directory's i_size */
77 #define BOGO_DIRENT_SIZE 20
78
79 /* Flag allocation requirements to shmem_getpage and shmem_swp_alloc */
80 enum sgp_type {
81         SGP_QUICK,      /* don't try more than file page cache lookup */
82         SGP_READ,       /* don't exceed i_size, don't allocate page */
83         SGP_CACHE,      /* don't exceed i_size, may allocate page */
84         SGP_WRITE,      /* may exceed i_size, may allocate page */
85         SGP_FAULT,      /* same as SGP_CACHE, return with page locked */
86 };
87
88 static int shmem_getpage(struct inode *inode, unsigned long idx,
89                          struct page **pagep, enum sgp_type sgp, int *type);
90
91 static inline struct page *shmem_dir_alloc(gfp_t gfp_mask)
92 {
93         /*
94          * The above definition of ENTRIES_PER_PAGE, and the use of
95          * BLOCKS_PER_PAGE on indirect pages, assume PAGE_CACHE_SIZE:
96          * might be reconsidered if it ever diverges from PAGE_SIZE.
97          *
98          * Mobility flags are masked out as swap vectors cannot move
99          */
100         return alloc_pages((gfp_mask & ~GFP_MOVABLE_MASK) | __GFP_ZERO,
101                                 PAGE_CACHE_SHIFT-PAGE_SHIFT);
102 }
103
104 static inline void shmem_dir_free(struct page *page)
105 {
106         __free_pages(page, PAGE_CACHE_SHIFT-PAGE_SHIFT);
107 }
108
109 static struct page **shmem_dir_map(struct page *page)
110 {
111         return (struct page **)kmap_atomic(page, KM_USER0);
112 }
113
114 static inline void shmem_dir_unmap(struct page **dir)
115 {
116         kunmap_atomic(dir, KM_USER0);
117 }
118
119 static swp_entry_t *shmem_swp_map(struct page *page)
120 {
121         return (swp_entry_t *)kmap_atomic(page, KM_USER1);
122 }
123
124 static inline void shmem_swp_balance_unmap(void)
125 {
126         /*
127          * When passing a pointer to an i_direct entry, to code which
128          * also handles indirect entries and so will shmem_swp_unmap,
129          * we must arrange for the preempt count to remain in balance.
130          * What kmap_atomic of a lowmem page does depends on config
131          * and architecture, so pretend to kmap_atomic some lowmem page.
132          */
133         (void) kmap_atomic(ZERO_PAGE(0), KM_USER1);
134 }
135
136 static inline void shmem_swp_unmap(swp_entry_t *entry)
137 {
138         kunmap_atomic(entry, KM_USER1);
139 }
140
141 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
142 {
143         return sb->s_fs_info;
144 }
145
146 /*
147  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
148  * for shared memory and for shared anonymous (/dev/zero) mappings
149  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
150  * consistent with the pre-accounting of private mappings ...
151  */
152 static inline int shmem_acct_size(unsigned long flags, loff_t size)
153 {
154         return (flags & VM_ACCOUNT)?
155                 security_vm_enough_memory(VM_ACCT(size)): 0;
156 }
157
158 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
159 {
160         if (flags & VM_ACCOUNT)
161                 vm_unacct_memory(VM_ACCT(size));
162 }
163
164 /*
165  * ... whereas tmpfs objects are accounted incrementally as
166  * pages are allocated, in order to allow huge sparse files.
167  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
168  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
169  */
170 static inline int shmem_acct_block(unsigned long flags)
171 {
172         return (flags & VM_ACCOUNT)?
173                 0: security_vm_enough_memory(VM_ACCT(PAGE_CACHE_SIZE));
174 }
175
176 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
177 {
178         if (!(flags & VM_ACCOUNT))
179                 vm_unacct_memory(pages * VM_ACCT(PAGE_CACHE_SIZE));
180 }
181
182 static const struct super_operations shmem_ops;
183 static const struct address_space_operations shmem_aops;
184 static const struct file_operations shmem_file_operations;
185 static const struct inode_operations shmem_inode_operations;
186 static const struct inode_operations shmem_dir_inode_operations;
187 static const struct inode_operations shmem_special_inode_operations;
188 static struct vm_operations_struct shmem_vm_ops;
189
190 static struct backing_dev_info shmem_backing_dev_info  __read_mostly = {
191         .ra_pages       = 0,    /* No readahead */
192         .capabilities   = BDI_CAP_NO_ACCT_DIRTY | BDI_CAP_NO_WRITEBACK,
193         .unplug_io_fn   = default_unplug_io_fn,
194 };
195
196 static LIST_HEAD(shmem_swaplist);
197 static DEFINE_SPINLOCK(shmem_swaplist_lock);
198
199 static void shmem_free_blocks(struct inode *inode, long pages)
200 {
201         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
202         if (sbinfo->max_blocks) {
203                 spin_lock(&sbinfo->stat_lock);
204                 sbinfo->free_blocks += pages;
205                 inode->i_blocks -= pages*BLOCKS_PER_PAGE;
206                 spin_unlock(&sbinfo->stat_lock);
207         }
208 }
209
210 /*
211  * shmem_recalc_inode - recalculate the size of an inode
212  *
213  * @inode: inode to recalc
214  *
215  * We have to calculate the free blocks since the mm can drop
216  * undirtied hole pages behind our back.
217  *
218  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
219  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
220  *
221  * It has to be called with the spinlock held.
222  */
223 static void shmem_recalc_inode(struct inode *inode)
224 {
225         struct shmem_inode_info *info = SHMEM_I(inode);
226         long freed;
227
228         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
229         if (freed > 0) {
230                 info->alloced -= freed;
231                 shmem_unacct_blocks(info->flags, freed);
232                 shmem_free_blocks(inode, freed);
233         }
234 }
235
236 /*
237  * shmem_swp_entry - find the swap vector position in the info structure
238  *
239  * @info:  info structure for the inode
240  * @index: index of the page to find
241  * @page:  optional page to add to the structure. Has to be preset to
242  *         all zeros
243  *
244  * If there is no space allocated yet it will return NULL when
245  * page is NULL, else it will use the page for the needed block,
246  * setting it to NULL on return to indicate that it has been used.
247  *
248  * The swap vector is organized the following way:
249  *
250  * There are SHMEM_NR_DIRECT entries directly stored in the
251  * shmem_inode_info structure. So small files do not need an addional
252  * allocation.
253  *
254  * For pages with index > SHMEM_NR_DIRECT there is the pointer
255  * i_indirect which points to a page which holds in the first half
256  * doubly indirect blocks, in the second half triple indirect blocks:
257  *
258  * For an artificial ENTRIES_PER_PAGE = 4 this would lead to the
259  * following layout (for SHMEM_NR_DIRECT == 16):
260  *
261  * i_indirect -> dir --> 16-19
262  *            |      +-> 20-23
263  *            |
264  *            +-->dir2 --> 24-27
265  *            |        +-> 28-31
266  *            |        +-> 32-35
267  *            |        +-> 36-39
268  *            |
269  *            +-->dir3 --> 40-43
270  *                     +-> 44-47
271  *                     +-> 48-51
272  *                     +-> 52-55
273  */
274 static swp_entry_t *shmem_swp_entry(struct shmem_inode_info *info, unsigned long index, struct page **page)
275 {
276         unsigned long offset;
277         struct page **dir;
278         struct page *subdir;
279
280         if (index < SHMEM_NR_DIRECT) {
281                 shmem_swp_balance_unmap();
282                 return info->i_direct+index;
283         }
284         if (!info->i_indirect) {
285                 if (page) {
286                         info->i_indirect = *page;
287                         *page = NULL;
288                 }
289                 return NULL;                    /* need another page */
290         }
291
292         index -= SHMEM_NR_DIRECT;
293         offset = index % ENTRIES_PER_PAGE;
294         index /= ENTRIES_PER_PAGE;
295         dir = shmem_dir_map(info->i_indirect);
296
297         if (index >= ENTRIES_PER_PAGE/2) {
298                 index -= ENTRIES_PER_PAGE/2;
299                 dir += ENTRIES_PER_PAGE/2 + index/ENTRIES_PER_PAGE;
300                 index %= ENTRIES_PER_PAGE;
301                 subdir = *dir;
302                 if (!subdir) {
303                         if (page) {
304                                 *dir = *page;
305                                 *page = NULL;
306                         }
307                         shmem_dir_unmap(dir);
308                         return NULL;            /* need another page */
309                 }
310                 shmem_dir_unmap(dir);
311                 dir = shmem_dir_map(subdir);
312         }
313
314         dir += index;
315         subdir = *dir;
316         if (!subdir) {
317                 if (!page || !(subdir = *page)) {
318                         shmem_dir_unmap(dir);
319                         return NULL;            /* need a page */
320                 }
321                 *dir = subdir;
322                 *page = NULL;
323         }
324         shmem_dir_unmap(dir);
325         return shmem_swp_map(subdir) + offset;
326 }
327
328 static void shmem_swp_set(struct shmem_inode_info *info, swp_entry_t *entry, unsigned long value)
329 {
330         long incdec = value? 1: -1;
331
332         entry->val = value;
333         info->swapped += incdec;
334         if ((unsigned long)(entry - info->i_direct) >= SHMEM_NR_DIRECT) {
335                 struct page *page = kmap_atomic_to_page(entry);
336                 set_page_private(page, page_private(page) + incdec);
337         }
338 }
339
340 /*
341  * shmem_swp_alloc - get the position of the swap entry for the page.
342  *                   If it does not exist allocate the entry.
343  *
344  * @info:       info structure for the inode
345  * @index:      index of the page to find
346  * @sgp:        check and recheck i_size? skip allocation?
347  */
348 static swp_entry_t *shmem_swp_alloc(struct shmem_inode_info *info, unsigned long index, enum sgp_type sgp)
349 {
350         struct inode *inode = &info->vfs_inode;
351         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
352         struct page *page = NULL;
353         swp_entry_t *entry;
354
355         if (sgp != SGP_WRITE &&
356             ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode))
357                 return ERR_PTR(-EINVAL);
358
359         while (!(entry = shmem_swp_entry(info, index, &page))) {
360                 if (sgp == SGP_READ)
361                         return shmem_swp_map(ZERO_PAGE(0));
362                 /*
363                  * Test free_blocks against 1 not 0, since we have 1 data
364                  * page (and perhaps indirect index pages) yet to allocate:
365                  * a waste to allocate index if we cannot allocate data.
366                  */
367                 if (sbinfo->max_blocks) {
368                         spin_lock(&sbinfo->stat_lock);
369                         if (sbinfo->free_blocks <= 1) {
370                                 spin_unlock(&sbinfo->stat_lock);
371                                 return ERR_PTR(-ENOSPC);
372                         }
373                         sbinfo->free_blocks--;
374                         inode->i_blocks += BLOCKS_PER_PAGE;
375                         spin_unlock(&sbinfo->stat_lock);
376                 }
377
378                 spin_unlock(&info->lock);
379                 page = shmem_dir_alloc(mapping_gfp_mask(inode->i_mapping));
380                 if (page)
381                         set_page_private(page, 0);
382                 spin_lock(&info->lock);
383
384                 if (!page) {
385                         shmem_free_blocks(inode, 1);
386                         return ERR_PTR(-ENOMEM);
387                 }
388                 if (sgp != SGP_WRITE &&
389                     ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
390                         entry = ERR_PTR(-EINVAL);
391                         break;
392                 }
393                 if (info->next_index <= index)
394                         info->next_index = index + 1;
395         }
396         if (page) {
397                 /* another task gave its page, or truncated the file */
398                 shmem_free_blocks(inode, 1);
399                 shmem_dir_free(page);
400         }
401         if (info->next_index <= index && !IS_ERR(entry))
402                 info->next_index = index + 1;
403         return entry;
404 }
405
406 /*
407  * shmem_free_swp - free some swap entries in a directory
408  *
409  * @dir:        pointer to the directory
410  * @edir:       pointer after last entry of the directory
411  * @punch_lock: pointer to spinlock when needed for the holepunch case
412  */
413 static int shmem_free_swp(swp_entry_t *dir, swp_entry_t *edir,
414                                                 spinlock_t *punch_lock)
415 {
416         spinlock_t *punch_unlock = NULL;
417         swp_entry_t *ptr;
418         int freed = 0;
419
420         for (ptr = dir; ptr < edir; ptr++) {
421                 if (ptr->val) {
422                         if (unlikely(punch_lock)) {
423                                 punch_unlock = punch_lock;
424                                 punch_lock = NULL;
425                                 spin_lock(punch_unlock);
426                                 if (!ptr->val)
427                                         continue;
428                         }
429                         free_swap_and_cache(*ptr);
430                         *ptr = (swp_entry_t){0};
431                         freed++;
432                 }
433         }
434         if (punch_unlock)
435                 spin_unlock(punch_unlock);
436         return freed;
437 }
438
439 static int shmem_map_and_free_swp(struct page *subdir, int offset,
440                 int limit, struct page ***dir, spinlock_t *punch_lock)
441 {
442         swp_entry_t *ptr;
443         int freed = 0;
444
445         ptr = shmem_swp_map(subdir);
446         for (; offset < limit; offset += LATENCY_LIMIT) {
447                 int size = limit - offset;
448                 if (size > LATENCY_LIMIT)
449                         size = LATENCY_LIMIT;
450                 freed += shmem_free_swp(ptr+offset, ptr+offset+size,
451                                                         punch_lock);
452                 if (need_resched()) {
453                         shmem_swp_unmap(ptr);
454                         if (*dir) {
455                                 shmem_dir_unmap(*dir);
456                                 *dir = NULL;
457                         }
458                         cond_resched();
459                         ptr = shmem_swp_map(subdir);
460                 }
461         }
462         shmem_swp_unmap(ptr);
463         return freed;
464 }
465
466 static void shmem_free_pages(struct list_head *next)
467 {
468         struct page *page;
469         int freed = 0;
470
471         do {
472                 page = container_of(next, struct page, lru);
473                 next = next->next;
474                 shmem_dir_free(page);
475                 freed++;
476                 if (freed >= LATENCY_LIMIT) {
477                         cond_resched();
478                         freed = 0;
479                 }
480         } while (next);
481 }
482
483 static void shmem_truncate_range(struct inode *inode, loff_t start, loff_t end)
484 {
485         struct shmem_inode_info *info = SHMEM_I(inode);
486         unsigned long idx;
487         unsigned long size;
488         unsigned long limit;
489         unsigned long stage;
490         unsigned long diroff;
491         struct page **dir;
492         struct page *topdir;
493         struct page *middir;
494         struct page *subdir;
495         swp_entry_t *ptr;
496         LIST_HEAD(pages_to_free);
497         long nr_pages_to_free = 0;
498         long nr_swaps_freed = 0;
499         int offset;
500         int freed;
501         int punch_hole;
502         spinlock_t *needs_lock;
503         spinlock_t *punch_lock;
504         unsigned long upper_limit;
505
506         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
507         idx = (start + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
508         if (idx >= info->next_index)
509                 return;
510
511         spin_lock(&info->lock);
512         info->flags |= SHMEM_TRUNCATE;
513         if (likely(end == (loff_t) -1)) {
514                 limit = info->next_index;
515                 upper_limit = SHMEM_MAX_INDEX;
516                 info->next_index = idx;
517                 needs_lock = NULL;
518                 punch_hole = 0;
519         } else {
520                 if (end + 1 >= inode->i_size) { /* we may free a little more */
521                         limit = (inode->i_size + PAGE_CACHE_SIZE - 1) >>
522                                                         PAGE_CACHE_SHIFT;
523                         upper_limit = SHMEM_MAX_INDEX;
524                 } else {
525                         limit = (end + 1) >> PAGE_CACHE_SHIFT;
526                         upper_limit = limit;
527                 }
528                 needs_lock = &info->lock;
529                 punch_hole = 1;
530         }
531
532         topdir = info->i_indirect;
533         if (topdir && idx <= SHMEM_NR_DIRECT && !punch_hole) {
534                 info->i_indirect = NULL;
535                 nr_pages_to_free++;
536                 list_add(&topdir->lru, &pages_to_free);
537         }
538         spin_unlock(&info->lock);
539
540         if (info->swapped && idx < SHMEM_NR_DIRECT) {
541                 ptr = info->i_direct;
542                 size = limit;
543                 if (size > SHMEM_NR_DIRECT)
544                         size = SHMEM_NR_DIRECT;
545                 nr_swaps_freed = shmem_free_swp(ptr+idx, ptr+size, needs_lock);
546         }
547
548         /*
549          * If there are no indirect blocks or we are punching a hole
550          * below indirect blocks, nothing to be done.
551          */
552         if (!topdir || limit <= SHMEM_NR_DIRECT)
553                 goto done2;
554
555         /*
556          * The truncation case has already dropped info->lock, and we're safe
557          * because i_size and next_index have already been lowered, preventing
558          * access beyond.  But in the punch_hole case, we still need to take
559          * the lock when updating the swap directory, because there might be
560          * racing accesses by shmem_getpage(SGP_CACHE), shmem_unuse_inode or
561          * shmem_writepage.  However, whenever we find we can remove a whole
562          * directory page (not at the misaligned start or end of the range),
563          * we first NULLify its pointer in the level above, and then have no
564          * need to take the lock when updating its contents: needs_lock and
565          * punch_lock (either pointing to info->lock or NULL) manage this.
566          */
567
568         upper_limit -= SHMEM_NR_DIRECT;
569         limit -= SHMEM_NR_DIRECT;
570         idx = (idx > SHMEM_NR_DIRECT)? (idx - SHMEM_NR_DIRECT): 0;
571         offset = idx % ENTRIES_PER_PAGE;
572         idx -= offset;
573
574         dir = shmem_dir_map(topdir);
575         stage = ENTRIES_PER_PAGEPAGE/2;
576         if (idx < ENTRIES_PER_PAGEPAGE/2) {
577                 middir = topdir;
578                 diroff = idx/ENTRIES_PER_PAGE;
579         } else {
580                 dir += ENTRIES_PER_PAGE/2;
581                 dir += (idx - ENTRIES_PER_PAGEPAGE/2)/ENTRIES_PER_PAGEPAGE;
582                 while (stage <= idx)
583                         stage += ENTRIES_PER_PAGEPAGE;
584                 middir = *dir;
585                 if (*dir) {
586                         diroff = ((idx - ENTRIES_PER_PAGEPAGE/2) %
587                                 ENTRIES_PER_PAGEPAGE) / ENTRIES_PER_PAGE;
588                         if (!diroff && !offset && upper_limit >= stage) {
589                                 if (needs_lock) {
590                                         spin_lock(needs_lock);
591                                         *dir = NULL;
592                                         spin_unlock(needs_lock);
593                                         needs_lock = NULL;
594                                 } else
595                                         *dir = NULL;
596                                 nr_pages_to_free++;
597                                 list_add(&middir->lru, &pages_to_free);
598                         }
599                         shmem_dir_unmap(dir);
600                         dir = shmem_dir_map(middir);
601                 } else {
602                         diroff = 0;
603                         offset = 0;
604                         idx = stage;
605                 }
606         }
607
608         for (; idx < limit; idx += ENTRIES_PER_PAGE, diroff++) {
609                 if (unlikely(idx == stage)) {
610                         shmem_dir_unmap(dir);
611                         dir = shmem_dir_map(topdir) +
612                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
613                         while (!*dir) {
614                                 dir++;
615                                 idx += ENTRIES_PER_PAGEPAGE;
616                                 if (idx >= limit)
617                                         goto done1;
618                         }
619                         stage = idx + ENTRIES_PER_PAGEPAGE;
620                         middir = *dir;
621                         if (punch_hole)
622                                 needs_lock = &info->lock;
623                         if (upper_limit >= stage) {
624                                 if (needs_lock) {
625                                         spin_lock(needs_lock);
626                                         *dir = NULL;
627                                         spin_unlock(needs_lock);
628                                         needs_lock = NULL;
629                                 } else
630                                         *dir = NULL;
631                                 nr_pages_to_free++;
632                                 list_add(&middir->lru, &pages_to_free);
633                         }
634                         shmem_dir_unmap(dir);
635                         cond_resched();
636                         dir = shmem_dir_map(middir);
637                         diroff = 0;
638                 }
639                 punch_lock = needs_lock;
640                 subdir = dir[diroff];
641                 if (subdir && !offset && upper_limit-idx >= ENTRIES_PER_PAGE) {
642                         if (needs_lock) {
643                                 spin_lock(needs_lock);
644                                 dir[diroff] = NULL;
645                                 spin_unlock(needs_lock);
646                                 punch_lock = NULL;
647                         } else
648                                 dir[diroff] = NULL;
649                         nr_pages_to_free++;
650                         list_add(&subdir->lru, &pages_to_free);
651                 }
652                 if (subdir && page_private(subdir) /* has swap entries */) {
653                         size = limit - idx;
654                         if (size > ENTRIES_PER_PAGE)
655                                 size = ENTRIES_PER_PAGE;
656                         freed = shmem_map_and_free_swp(subdir,
657                                         offset, size, &dir, punch_lock);
658                         if (!dir)
659                                 dir = shmem_dir_map(middir);
660                         nr_swaps_freed += freed;
661                         if (offset || punch_lock) {
662                                 spin_lock(&info->lock);
663                                 set_page_private(subdir,
664                                         page_private(subdir) - freed);
665                                 spin_unlock(&info->lock);
666                         } else
667                                 BUG_ON(page_private(subdir) != freed);
668                 }
669                 offset = 0;
670         }
671 done1:
672         shmem_dir_unmap(dir);
673 done2:
674         if (inode->i_mapping->nrpages && (info->flags & SHMEM_PAGEIN)) {
675                 /*
676                  * Call truncate_inode_pages again: racing shmem_unuse_inode
677                  * may have swizzled a page in from swap since vmtruncate or
678                  * generic_delete_inode did it, before we lowered next_index.
679                  * Also, though shmem_getpage checks i_size before adding to
680                  * cache, no recheck after: so fix the narrow window there too.
681                  *
682                  * Recalling truncate_inode_pages_range and unmap_mapping_range
683                  * every time for punch_hole (which never got a chance to clear
684                  * SHMEM_PAGEIN at the start of vmtruncate_range) is expensive,
685                  * yet hardly ever necessary: try to optimize them out later.
686                  */
687                 truncate_inode_pages_range(inode->i_mapping, start, end);
688                 if (punch_hole)
689                         unmap_mapping_range(inode->i_mapping, start,
690                                                         end - start, 1);
691         }
692
693         spin_lock(&info->lock);
694         info->flags &= ~SHMEM_TRUNCATE;
695         info->swapped -= nr_swaps_freed;
696         if (nr_pages_to_free)
697                 shmem_free_blocks(inode, nr_pages_to_free);
698         shmem_recalc_inode(inode);
699         spin_unlock(&info->lock);
700
701         /*
702          * Empty swap vector directory pages to be freed?
703          */
704         if (!list_empty(&pages_to_free)) {
705                 pages_to_free.prev->next = NULL;
706                 shmem_free_pages(pages_to_free.next);
707         }
708 }
709
710 static void shmem_truncate(struct inode *inode)
711 {
712         shmem_truncate_range(inode, inode->i_size, (loff_t)-1);
713 }
714
715 static int shmem_notify_change(struct dentry *dentry, struct iattr *attr)
716 {
717         struct inode *inode = dentry->d_inode;
718         struct page *page = NULL;
719         int error;
720
721         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
722                 if (attr->ia_size < inode->i_size) {
723                         /*
724                          * If truncating down to a partial page, then
725                          * if that page is already allocated, hold it
726                          * in memory until the truncation is over, so
727                          * truncate_partial_page cannnot miss it were
728                          * it assigned to swap.
729                          */
730                         if (attr->ia_size & (PAGE_CACHE_SIZE-1)) {
731                                 (void) shmem_getpage(inode,
732                                         attr->ia_size>>PAGE_CACHE_SHIFT,
733                                                 &page, SGP_READ, NULL);
734                         }
735                         /*
736                          * Reset SHMEM_PAGEIN flag so that shmem_truncate can
737                          * detect if any pages might have been added to cache
738                          * after truncate_inode_pages.  But we needn't bother
739                          * if it's being fully truncated to zero-length: the
740                          * nrpages check is efficient enough in that case.
741                          */
742                         if (attr->ia_size) {
743                                 struct shmem_inode_info *info = SHMEM_I(inode);
744                                 spin_lock(&info->lock);
745                                 info->flags &= ~SHMEM_PAGEIN;
746                                 spin_unlock(&info->lock);
747                         }
748                 }
749         }
750
751         error = inode_change_ok(inode, attr);
752         if (!error)
753                 error = inode_setattr(inode, attr);
754 #ifdef CONFIG_TMPFS_POSIX_ACL
755         if (!error && (attr->ia_valid & ATTR_MODE))
756                 error = generic_acl_chmod(inode, &shmem_acl_ops);
757 #endif
758         if (page)
759                 page_cache_release(page);
760         return error;
761 }
762
763 static void shmem_delete_inode(struct inode *inode)
764 {
765         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
766         struct shmem_inode_info *info = SHMEM_I(inode);
767
768         if (inode->i_op->truncate == shmem_truncate) {
769                 truncate_inode_pages(inode->i_mapping, 0);
770                 shmem_unacct_size(info->flags, inode->i_size);
771                 inode->i_size = 0;
772                 shmem_truncate(inode);
773                 if (!list_empty(&info->swaplist)) {
774                         spin_lock(&shmem_swaplist_lock);
775                         list_del_init(&info->swaplist);
776                         spin_unlock(&shmem_swaplist_lock);
777                 }
778         }
779         BUG_ON(inode->i_blocks);
780         if (sbinfo->max_inodes) {
781                 spin_lock(&sbinfo->stat_lock);
782                 sbinfo->free_inodes++;
783                 spin_unlock(&sbinfo->stat_lock);
784         }
785         clear_inode(inode);
786 }
787
788 static inline int shmem_find_swp(swp_entry_t entry, swp_entry_t *dir, swp_entry_t *edir)
789 {
790         swp_entry_t *ptr;
791
792         for (ptr = dir; ptr < edir; ptr++) {
793                 if (ptr->val == entry.val)
794                         return ptr - dir;
795         }
796         return -1;
797 }
798
799 static int shmem_unuse_inode(struct shmem_inode_info *info, swp_entry_t entry, struct page *page)
800 {
801         struct inode *inode;
802         unsigned long idx;
803         unsigned long size;
804         unsigned long limit;
805         unsigned long stage;
806         struct page **dir;
807         struct page *subdir;
808         swp_entry_t *ptr;
809         int offset;
810
811         idx = 0;
812         ptr = info->i_direct;
813         spin_lock(&info->lock);
814         limit = info->next_index;
815         size = limit;
816         if (size > SHMEM_NR_DIRECT)
817                 size = SHMEM_NR_DIRECT;
818         offset = shmem_find_swp(entry, ptr, ptr+size);
819         if (offset >= 0) {
820                 shmem_swp_balance_unmap();
821                 goto found;
822         }
823         if (!info->i_indirect)
824                 goto lost2;
825
826         dir = shmem_dir_map(info->i_indirect);
827         stage = SHMEM_NR_DIRECT + ENTRIES_PER_PAGEPAGE/2;
828
829         for (idx = SHMEM_NR_DIRECT; idx < limit; idx += ENTRIES_PER_PAGE, dir++) {
830                 if (unlikely(idx == stage)) {
831                         shmem_dir_unmap(dir-1);
832                         dir = shmem_dir_map(info->i_indirect) +
833                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
834                         while (!*dir) {
835                                 dir++;
836                                 idx += ENTRIES_PER_PAGEPAGE;
837                                 if (idx >= limit)
838                                         goto lost1;
839                         }
840                         stage = idx + ENTRIES_PER_PAGEPAGE;
841                         subdir = *dir;
842                         shmem_dir_unmap(dir);
843                         dir = shmem_dir_map(subdir);
844                 }
845                 subdir = *dir;
846                 if (subdir && page_private(subdir)) {
847                         ptr = shmem_swp_map(subdir);
848                         size = limit - idx;
849                         if (size > ENTRIES_PER_PAGE)
850                                 size = ENTRIES_PER_PAGE;
851                         offset = shmem_find_swp(entry, ptr, ptr+size);
852                         if (offset >= 0) {
853                                 shmem_dir_unmap(dir);
854                                 goto found;
855                         }
856                         shmem_swp_unmap(ptr);
857                 }
858         }
859 lost1:
860         shmem_dir_unmap(dir-1);
861 lost2:
862         spin_unlock(&info->lock);
863         return 0;
864 found:
865         idx += offset;
866         inode = &info->vfs_inode;
867         if (move_from_swap_cache(page, idx, inode->i_mapping) == 0) {
868                 info->flags |= SHMEM_PAGEIN;
869                 shmem_swp_set(info, ptr + offset, 0);
870         }
871         shmem_swp_unmap(ptr);
872         spin_unlock(&info->lock);
873         /*
874          * Decrement swap count even when the entry is left behind:
875          * try_to_unuse will skip over mms, then reincrement count.
876          */
877         swap_free(entry);
878         return 1;
879 }
880
881 /*
882  * shmem_unuse() search for an eventually swapped out shmem page.
883  */
884 int shmem_unuse(swp_entry_t entry, struct page *page)
885 {
886         struct list_head *p, *next;
887         struct shmem_inode_info *info;
888         int found = 0;
889
890         spin_lock(&shmem_swaplist_lock);
891         list_for_each_safe(p, next, &shmem_swaplist) {
892                 info = list_entry(p, struct shmem_inode_info, swaplist);
893                 if (!info->swapped)
894                         list_del_init(&info->swaplist);
895                 else if (shmem_unuse_inode(info, entry, page)) {
896                         /* move head to start search for next from here */
897                         list_move_tail(&shmem_swaplist, &info->swaplist);
898                         found = 1;
899                         break;
900                 }
901         }
902         spin_unlock(&shmem_swaplist_lock);
903         return found;
904 }
905
906 /*
907  * Move the page from the page cache to the swap cache.
908  */
909 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
910 {
911         struct shmem_inode_info *info;
912         swp_entry_t *entry, swap;
913         struct address_space *mapping;
914         unsigned long index;
915         struct inode *inode;
916
917         BUG_ON(!PageLocked(page));
918         /*
919          * shmem_backing_dev_info's capabilities prevent regular writeback or
920          * sync from ever calling shmem_writepage; but a stacking filesystem
921          * may use the ->writepage of its underlying filesystem, in which case
922          * we want to do nothing when that underlying filesystem is tmpfs
923          * (writing out to swap is useful as a response to memory pressure, but
924          * of no use to stabilize the data) - just redirty the page, unlock it
925          * and claim success in this case.  AOP_WRITEPAGE_ACTIVATE, and the
926          * page_mapped check below, must be avoided unless we're in reclaim.
927          */
928         if (!wbc->for_reclaim) {
929                 set_page_dirty(page);
930                 unlock_page(page);
931                 return 0;
932         }
933         BUG_ON(page_mapped(page));
934
935         mapping = page->mapping;
936         index = page->index;
937         inode = mapping->host;
938         info = SHMEM_I(inode);
939         if (info->flags & VM_LOCKED)
940                 goto redirty;
941         swap = get_swap_page();
942         if (!swap.val)
943                 goto redirty;
944
945         spin_lock(&info->lock);
946         shmem_recalc_inode(inode);
947         if (index >= info->next_index) {
948                 BUG_ON(!(info->flags & SHMEM_TRUNCATE));
949                 goto unlock;
950         }
951         entry = shmem_swp_entry(info, index, NULL);
952         BUG_ON(!entry);
953         BUG_ON(entry->val);
954
955         if (move_to_swap_cache(page, swap) == 0) {
956                 shmem_swp_set(info, entry, swap.val);
957                 shmem_swp_unmap(entry);
958                 spin_unlock(&info->lock);
959                 if (list_empty(&info->swaplist)) {
960                         spin_lock(&shmem_swaplist_lock);
961                         /* move instead of add in case we're racing */
962                         list_move_tail(&info->swaplist, &shmem_swaplist);
963                         spin_unlock(&shmem_swaplist_lock);
964                 }
965                 unlock_page(page);
966                 return 0;
967         }
968
969         shmem_swp_unmap(entry);
970 unlock:
971         spin_unlock(&info->lock);
972         swap_free(swap);
973 redirty:
974         set_page_dirty(page);
975         return AOP_WRITEPAGE_ACTIVATE;  /* Return with the page locked */
976 }
977
978 #ifdef CONFIG_NUMA
979 static inline int shmem_parse_mpol(char *value, int *policy, nodemask_t *policy_nodes)
980 {
981         char *nodelist = strchr(value, ':');
982         int err = 1;
983
984         if (nodelist) {
985                 /* NUL-terminate policy string */
986                 *nodelist++ = '\0';
987                 if (nodelist_parse(nodelist, *policy_nodes))
988                         goto out;
989                 if (!nodes_subset(*policy_nodes, node_states[N_HIGH_MEMORY]))
990                         goto out;
991         }
992         if (!strcmp(value, "default")) {
993                 *policy = MPOL_DEFAULT;
994                 /* Don't allow a nodelist */
995                 if (!nodelist)
996                         err = 0;
997         } else if (!strcmp(value, "prefer")) {
998                 *policy = MPOL_PREFERRED;
999                 /* Insist on a nodelist of one node only */
1000                 if (nodelist) {
1001                         char *rest = nodelist;
1002                         while (isdigit(*rest))
1003                                 rest++;
1004                         if (!*rest)
1005                                 err = 0;
1006                 }
1007         } else if (!strcmp(value, "bind")) {
1008                 *policy = MPOL_BIND;
1009                 /* Insist on a nodelist */
1010                 if (nodelist)
1011                         err = 0;
1012         } else if (!strcmp(value, "interleave")) {
1013                 *policy = MPOL_INTERLEAVE;
1014                 /*
1015                  * Default to online nodes with memory if no nodelist
1016                  */
1017                 if (!nodelist)
1018                         *policy_nodes = node_states[N_HIGH_MEMORY];
1019                 err = 0;
1020         }
1021 out:
1022         /* Restore string for error message */
1023         if (nodelist)
1024                 *--nodelist = ':';
1025         return err;
1026 }
1027
1028 static struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1029                         struct shmem_inode_info *info, unsigned long idx)
1030 {
1031         struct vm_area_struct pvma;
1032         struct page *page;
1033
1034         /* Create a pseudo vma that just contains the policy */
1035         pvma.vm_start = 0;
1036         pvma.vm_pgoff = idx;
1037         pvma.vm_ops = NULL;
1038         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, idx);
1039         page = swapin_readahead(entry, gfp, &pvma, 0);
1040         mpol_free(pvma.vm_policy);
1041         return page;
1042 }
1043
1044 static struct page *shmem_alloc_page(gfp_t gfp,
1045                         struct shmem_inode_info *info, unsigned long idx)
1046 {
1047         struct vm_area_struct pvma;
1048         struct page *page;
1049
1050         /* Create a pseudo vma that just contains the policy */
1051         pvma.vm_start = 0;
1052         pvma.vm_pgoff = idx;
1053         pvma.vm_ops = NULL;
1054         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, idx);
1055         page = alloc_page_vma(gfp, &pvma, 0);
1056         mpol_free(pvma.vm_policy);
1057         return page;
1058 }
1059 #else
1060 static inline int shmem_parse_mpol(char *value, int *policy,
1061                                                 nodemask_t *policy_nodes)
1062 {
1063         return 1;
1064 }
1065
1066 static inline struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1067                         struct shmem_inode_info *info, unsigned long idx)
1068 {
1069         return swapin_readahead(entry, gfp, NULL, 0);
1070 }
1071
1072 static inline struct page *shmem_alloc_page(gfp_t gfp,
1073                         struct shmem_inode_info *info, unsigned long idx)
1074 {
1075         return alloc_page(gfp);
1076 }
1077 #endif
1078
1079 /*
1080  * shmem_getpage - either get the page from swap or allocate a new one
1081  *
1082  * If we allocate a new one we do not mark it dirty. That's up to the
1083  * vm. If we swap it in we mark it dirty since we also free the swap
1084  * entry since a page cannot live in both the swap and page cache
1085  */
1086 static int shmem_getpage(struct inode *inode, unsigned long idx,
1087                         struct page **pagep, enum sgp_type sgp, int *type)
1088 {
1089         struct address_space *mapping = inode->i_mapping;
1090         struct shmem_inode_info *info = SHMEM_I(inode);
1091         struct shmem_sb_info *sbinfo;
1092         struct page *filepage = *pagep;
1093         struct page *swappage;
1094         swp_entry_t *entry;
1095         swp_entry_t swap;
1096         gfp_t gfp;
1097         int error;
1098
1099         if (idx >= SHMEM_MAX_INDEX)
1100                 return -EFBIG;
1101
1102         if (type)
1103                 *type = 0;
1104
1105         /*
1106          * Normally, filepage is NULL on entry, and either found
1107          * uptodate immediately, or allocated and zeroed, or read
1108          * in under swappage, which is then assigned to filepage.
1109          * But shmem_readpage and shmem_write_begin pass in a locked
1110          * filepage, which may be found not uptodate by other callers
1111          * too, and may need to be copied from the swappage read in.
1112          */
1113 repeat:
1114         if (!filepage)
1115                 filepage = find_lock_page(mapping, idx);
1116         if (filepage && PageUptodate(filepage))
1117                 goto done;
1118         error = 0;
1119         if (sgp == SGP_QUICK)
1120                 goto failed;
1121         gfp = mapping_gfp_mask(mapping);
1122
1123         spin_lock(&info->lock);
1124         shmem_recalc_inode(inode);
1125         entry = shmem_swp_alloc(info, idx, sgp);
1126         if (IS_ERR(entry)) {
1127                 spin_unlock(&info->lock);
1128                 error = PTR_ERR(entry);
1129                 goto failed;
1130         }
1131         swap = *entry;
1132
1133         if (swap.val) {
1134                 /* Look it up and read it in.. */
1135                 swappage = lookup_swap_cache(swap);
1136                 if (!swappage) {
1137                         shmem_swp_unmap(entry);
1138                         /* here we actually do the io */
1139                         if (type && !(*type & VM_FAULT_MAJOR)) {
1140                                 __count_vm_event(PGMAJFAULT);
1141                                 *type |= VM_FAULT_MAJOR;
1142                         }
1143                         spin_unlock(&info->lock);
1144                         swappage = shmem_swapin(swap, gfp, info, idx);
1145                         if (!swappage) {
1146                                 spin_lock(&info->lock);
1147                                 entry = shmem_swp_alloc(info, idx, sgp);
1148                                 if (IS_ERR(entry))
1149                                         error = PTR_ERR(entry);
1150                                 else {
1151                                         if (entry->val == swap.val)
1152                                                 error = -ENOMEM;
1153                                         shmem_swp_unmap(entry);
1154                                 }
1155                                 spin_unlock(&info->lock);
1156                                 if (error)
1157                                         goto failed;
1158                                 goto repeat;
1159                         }
1160                         wait_on_page_locked(swappage);
1161                         page_cache_release(swappage);
1162                         goto repeat;
1163                 }
1164
1165                 /* We have to do this with page locked to prevent races */
1166                 if (TestSetPageLocked(swappage)) {
1167                         shmem_swp_unmap(entry);
1168                         spin_unlock(&info->lock);
1169                         wait_on_page_locked(swappage);
1170                         page_cache_release(swappage);
1171                         goto repeat;
1172                 }
1173                 if (PageWriteback(swappage)) {
1174                         shmem_swp_unmap(entry);
1175                         spin_unlock(&info->lock);
1176                         wait_on_page_writeback(swappage);
1177                         unlock_page(swappage);
1178                         page_cache_release(swappage);
1179                         goto repeat;
1180                 }
1181                 if (!PageUptodate(swappage)) {
1182                         shmem_swp_unmap(entry);
1183                         spin_unlock(&info->lock);
1184                         unlock_page(swappage);
1185                         page_cache_release(swappage);
1186                         error = -EIO;
1187                         goto failed;
1188                 }
1189
1190                 if (filepage) {
1191                         shmem_swp_set(info, entry, 0);
1192                         shmem_swp_unmap(entry);
1193                         delete_from_swap_cache(swappage);
1194                         spin_unlock(&info->lock);
1195                         copy_highpage(filepage, swappage);
1196                         unlock_page(swappage);
1197                         page_cache_release(swappage);
1198                         flush_dcache_page(filepage);
1199                         SetPageUptodate(filepage);
1200                         set_page_dirty(filepage);
1201                         swap_free(swap);
1202                 } else if (!(error = move_from_swap_cache(
1203                                 swappage, idx, mapping))) {
1204                         info->flags |= SHMEM_PAGEIN;
1205                         shmem_swp_set(info, entry, 0);
1206                         shmem_swp_unmap(entry);
1207                         spin_unlock(&info->lock);
1208                         filepage = swappage;
1209                         swap_free(swap);
1210                 } else {
1211                         shmem_swp_unmap(entry);
1212                         spin_unlock(&info->lock);
1213                         unlock_page(swappage);
1214                         page_cache_release(swappage);
1215                         if (error == -ENOMEM) {
1216                                 /* let kswapd refresh zone for GFP_ATOMICs */
1217                                 congestion_wait(WRITE, HZ/50);
1218                         }
1219                         goto repeat;
1220                 }
1221         } else if (sgp == SGP_READ && !filepage) {
1222                 shmem_swp_unmap(entry);
1223                 filepage = find_get_page(mapping, idx);
1224                 if (filepage &&
1225                     (!PageUptodate(filepage) || TestSetPageLocked(filepage))) {
1226                         spin_unlock(&info->lock);
1227                         wait_on_page_locked(filepage);
1228                         page_cache_release(filepage);
1229                         filepage = NULL;
1230                         goto repeat;
1231                 }
1232                 spin_unlock(&info->lock);
1233         } else {
1234                 shmem_swp_unmap(entry);
1235                 sbinfo = SHMEM_SB(inode->i_sb);
1236                 if (sbinfo->max_blocks) {
1237                         spin_lock(&sbinfo->stat_lock);
1238                         if (sbinfo->free_blocks == 0 ||
1239                             shmem_acct_block(info->flags)) {
1240                                 spin_unlock(&sbinfo->stat_lock);
1241                                 spin_unlock(&info->lock);
1242                                 error = -ENOSPC;
1243                                 goto failed;
1244                         }
1245                         sbinfo->free_blocks--;
1246                         inode->i_blocks += BLOCKS_PER_PAGE;
1247                         spin_unlock(&sbinfo->stat_lock);
1248                 } else if (shmem_acct_block(info->flags)) {
1249                         spin_unlock(&info->lock);
1250                         error = -ENOSPC;
1251                         goto failed;
1252                 }
1253
1254                 if (!filepage) {
1255                         spin_unlock(&info->lock);
1256                         filepage = shmem_alloc_page(gfp, info, idx);
1257                         if (!filepage) {
1258                                 shmem_unacct_blocks(info->flags, 1);
1259                                 shmem_free_blocks(inode, 1);
1260                                 error = -ENOMEM;
1261                                 goto failed;
1262                         }
1263
1264                         spin_lock(&info->lock);
1265                         entry = shmem_swp_alloc(info, idx, sgp);
1266                         if (IS_ERR(entry))
1267                                 error = PTR_ERR(entry);
1268                         else {
1269                                 swap = *entry;
1270                                 shmem_swp_unmap(entry);
1271                         }
1272                         if (error || swap.val || 0 != add_to_page_cache_lru(
1273                                         filepage, mapping, idx, GFP_ATOMIC)) {
1274                                 spin_unlock(&info->lock);
1275                                 page_cache_release(filepage);
1276                                 shmem_unacct_blocks(info->flags, 1);
1277                                 shmem_free_blocks(inode, 1);
1278                                 filepage = NULL;
1279                                 if (error)
1280                                         goto failed;
1281                                 goto repeat;
1282                         }
1283                         info->flags |= SHMEM_PAGEIN;
1284                 }
1285
1286                 info->alloced++;
1287                 spin_unlock(&info->lock);
1288                 clear_highpage(filepage);
1289                 flush_dcache_page(filepage);
1290                 SetPageUptodate(filepage);
1291         }
1292 done:
1293         if (*pagep != filepage) {
1294                 *pagep = filepage;
1295                 if (sgp != SGP_FAULT)
1296                         unlock_page(filepage);
1297
1298         }
1299         return 0;
1300
1301 failed:
1302         if (*pagep != filepage) {
1303                 unlock_page(filepage);
1304                 page_cache_release(filepage);
1305         }
1306         return error;
1307 }
1308
1309 static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1310 {
1311         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1312         int error;
1313         int ret;
1314
1315         if (((loff_t)vmf->pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
1316                 return VM_FAULT_SIGBUS;
1317
1318         error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_FAULT, &ret);
1319         if (error)
1320                 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1321
1322         mark_page_accessed(vmf->page);
1323         return ret | VM_FAULT_LOCKED;
1324 }
1325
1326 #ifdef CONFIG_NUMA
1327 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *new)
1328 {
1329         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1330         return mpol_set_shared_policy(&SHMEM_I(i)->policy, vma, new);
1331 }
1332
1333 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
1334                                           unsigned long addr)
1335 {
1336         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1337         unsigned long idx;
1338
1339         idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1340         return mpol_shared_policy_lookup(&SHMEM_I(i)->policy, idx);
1341 }
1342 #endif
1343
1344 int shmem_lock(struct file *file, int lock, struct user_struct *user)
1345 {
1346         struct inode *inode = file->f_path.dentry->d_inode;
1347         struct shmem_inode_info *info = SHMEM_I(inode);
1348         int retval = -ENOMEM;
1349
1350         spin_lock(&info->lock);
1351         if (lock && !(info->flags & VM_LOCKED)) {
1352                 if (!user_shm_lock(inode->i_size, user))
1353                         goto out_nomem;
1354                 info->flags |= VM_LOCKED;
1355         }
1356         if (!lock && (info->flags & VM_LOCKED) && user) {
1357                 user_shm_unlock(inode->i_size, user);
1358                 info->flags &= ~VM_LOCKED;
1359         }
1360         retval = 0;
1361 out_nomem:
1362         spin_unlock(&info->lock);
1363         return retval;
1364 }
1365
1366 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1367 {
1368         file_accessed(file);
1369         vma->vm_ops = &shmem_vm_ops;
1370         vma->vm_flags |= VM_CAN_NONLINEAR;
1371         return 0;
1372 }
1373
1374 static struct inode *
1375 shmem_get_inode(struct super_block *sb, int mode, dev_t dev)
1376 {
1377         struct inode *inode;
1378         struct shmem_inode_info *info;
1379         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1380
1381         if (sbinfo->max_inodes) {
1382                 spin_lock(&sbinfo->stat_lock);
1383                 if (!sbinfo->free_inodes) {
1384                         spin_unlock(&sbinfo->stat_lock);
1385                         return NULL;
1386                 }
1387                 sbinfo->free_inodes--;
1388                 spin_unlock(&sbinfo->stat_lock);
1389         }
1390
1391         inode = new_inode(sb);
1392         if (inode) {
1393                 inode->i_mode = mode;
1394                 inode->i_uid = current->fsuid;
1395                 inode->i_gid = current->fsgid;
1396                 inode->i_blocks = 0;
1397                 inode->i_mapping->a_ops = &shmem_aops;
1398                 inode->i_mapping->backing_dev_info = &shmem_backing_dev_info;
1399                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1400                 inode->i_generation = get_seconds();
1401                 info = SHMEM_I(inode);
1402                 memset(info, 0, (char *)inode - (char *)info);
1403                 spin_lock_init(&info->lock);
1404                 INIT_LIST_HEAD(&info->swaplist);
1405
1406                 switch (mode & S_IFMT) {
1407                 default:
1408                         inode->i_op = &shmem_special_inode_operations;
1409                         init_special_inode(inode, mode, dev);
1410                         break;
1411                 case S_IFREG:
1412                         inode->i_op = &shmem_inode_operations;
1413                         inode->i_fop = &shmem_file_operations;
1414                         mpol_shared_policy_init(&info->policy, sbinfo->policy,
1415                                                         &sbinfo->policy_nodes);
1416                         break;
1417                 case S_IFDIR:
1418                         inc_nlink(inode);
1419                         /* Some things misbehave if size == 0 on a directory */
1420                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
1421                         inode->i_op = &shmem_dir_inode_operations;
1422                         inode->i_fop = &simple_dir_operations;
1423                         break;
1424                 case S_IFLNK:
1425                         /*
1426                          * Must not load anything in the rbtree,
1427                          * mpol_free_shared_policy will not be called.
1428                          */
1429                         mpol_shared_policy_init(&info->policy, MPOL_DEFAULT,
1430                                                 NULL);
1431                         break;
1432                 }
1433         } else if (sbinfo->max_inodes) {
1434                 spin_lock(&sbinfo->stat_lock);
1435                 sbinfo->free_inodes++;
1436                 spin_unlock(&sbinfo->stat_lock);
1437         }
1438         return inode;
1439 }
1440
1441 #ifdef CONFIG_TMPFS
1442 static const struct inode_operations shmem_symlink_inode_operations;
1443 static const struct inode_operations shmem_symlink_inline_operations;
1444
1445 /*
1446  * Normally tmpfs avoids the use of shmem_readpage and shmem_write_begin;
1447  * but providing them allows a tmpfs file to be used for splice, sendfile, and
1448  * below the loop driver, in the generic fashion that many filesystems support.
1449  */
1450 static int shmem_readpage(struct file *file, struct page *page)
1451 {
1452         struct inode *inode = page->mapping->host;
1453         int error = shmem_getpage(inode, page->index, &page, SGP_CACHE, NULL);
1454         unlock_page(page);
1455         return error;
1456 }
1457
1458 static int
1459 shmem_write_begin(struct file *file, struct address_space *mapping,
1460                         loff_t pos, unsigned len, unsigned flags,
1461                         struct page **pagep, void **fsdata)
1462 {
1463         struct inode *inode = mapping->host;
1464         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1465         *pagep = NULL;
1466         return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1467 }
1468
1469 static int
1470 shmem_write_end(struct file *file, struct address_space *mapping,
1471                         loff_t pos, unsigned len, unsigned copied,
1472                         struct page *page, void *fsdata)
1473 {
1474         struct inode *inode = mapping->host;
1475
1476         set_page_dirty(page);
1477         page_cache_release(page);
1478
1479         if (pos+copied > inode->i_size)
1480                 i_size_write(inode, pos+copied);
1481
1482         return copied;
1483 }
1484
1485 static ssize_t
1486 shmem_file_write(struct file *file, const char __user *buf, size_t count, loff_t *ppos)
1487 {
1488         struct inode    *inode = file->f_path.dentry->d_inode;
1489         loff_t          pos;
1490         unsigned long   written;
1491         ssize_t         err;
1492
1493         if ((ssize_t) count < 0)
1494                 return -EINVAL;
1495
1496         if (!access_ok(VERIFY_READ, buf, count))
1497                 return -EFAULT;
1498
1499         mutex_lock(&inode->i_mutex);
1500
1501         pos = *ppos;
1502         written = 0;
1503
1504         err = generic_write_checks(file, &pos, &count, 0);
1505         if (err || !count)
1506                 goto out;
1507
1508         err = remove_suid(file->f_path.dentry);
1509         if (err)
1510                 goto out;
1511
1512         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1513
1514         do {
1515                 struct page *page = NULL;
1516                 unsigned long bytes, index, offset;
1517                 char *kaddr;
1518                 int left;
1519
1520                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
1521                 index = pos >> PAGE_CACHE_SHIFT;
1522                 bytes = PAGE_CACHE_SIZE - offset;
1523                 if (bytes > count)
1524                         bytes = count;
1525
1526                 /*
1527                  * We don't hold page lock across copy from user -
1528                  * what would it guard against? - so no deadlock here.
1529                  * But it still may be a good idea to prefault below.
1530                  */
1531
1532                 err = shmem_getpage(inode, index, &page, SGP_WRITE, NULL);
1533                 if (err)
1534                         break;
1535
1536                 left = bytes;
1537                 if (PageHighMem(page)) {
1538                         volatile unsigned char dummy;
1539                         __get_user(dummy, buf);
1540                         __get_user(dummy, buf + bytes - 1);
1541
1542                         kaddr = kmap_atomic(page, KM_USER0);
1543                         left = __copy_from_user_inatomic(kaddr + offset,
1544                                                         buf, bytes);
1545                         kunmap_atomic(kaddr, KM_USER0);
1546                 }
1547                 if (left) {
1548                         kaddr = kmap(page);
1549                         left = __copy_from_user(kaddr + offset, buf, bytes);
1550                         kunmap(page);
1551                 }
1552
1553                 written += bytes;
1554                 count -= bytes;
1555                 pos += bytes;
1556                 buf += bytes;
1557                 if (pos > inode->i_size)
1558                         i_size_write(inode, pos);
1559
1560                 flush_dcache_page(page);
1561                 set_page_dirty(page);
1562                 mark_page_accessed(page);
1563                 page_cache_release(page);
1564
1565                 if (left) {
1566                         pos -= left;
1567                         written -= left;
1568                         err = -EFAULT;
1569                         break;
1570                 }
1571
1572                 /*
1573                  * Our dirty pages are not counted in nr_dirty,
1574                  * and we do not attempt to balance dirty pages.
1575                  */
1576
1577                 cond_resched();
1578         } while (count);
1579
1580         *ppos = pos;
1581         if (written)
1582                 err = written;
1583 out:
1584         mutex_unlock(&inode->i_mutex);
1585         return err;
1586 }
1587
1588 static void do_shmem_file_read(struct file *filp, loff_t *ppos, read_descriptor_t *desc, read_actor_t actor)
1589 {
1590         struct inode *inode = filp->f_path.dentry->d_inode;
1591         struct address_space *mapping = inode->i_mapping;
1592         unsigned long index, offset;
1593
1594         index = *ppos >> PAGE_CACHE_SHIFT;
1595         offset = *ppos & ~PAGE_CACHE_MASK;
1596
1597         for (;;) {
1598                 struct page *page = NULL;
1599                 unsigned long end_index, nr, ret;
1600                 loff_t i_size = i_size_read(inode);
1601
1602                 end_index = i_size >> PAGE_CACHE_SHIFT;
1603                 if (index > end_index)
1604                         break;
1605                 if (index == end_index) {
1606                         nr = i_size & ~PAGE_CACHE_MASK;
1607                         if (nr <= offset)
1608                                 break;
1609                 }
1610
1611                 desc->error = shmem_getpage(inode, index, &page, SGP_READ, NULL);
1612                 if (desc->error) {
1613                         if (desc->error == -EINVAL)
1614                                 desc->error = 0;
1615                         break;
1616                 }
1617
1618                 /*
1619                  * We must evaluate after, since reads (unlike writes)
1620                  * are called without i_mutex protection against truncate
1621                  */
1622                 nr = PAGE_CACHE_SIZE;
1623                 i_size = i_size_read(inode);
1624                 end_index = i_size >> PAGE_CACHE_SHIFT;
1625                 if (index == end_index) {
1626                         nr = i_size & ~PAGE_CACHE_MASK;
1627                         if (nr <= offset) {
1628                                 if (page)
1629                                         page_cache_release(page);
1630                                 break;
1631                         }
1632                 }
1633                 nr -= offset;
1634
1635                 if (page) {
1636                         /*
1637                          * If users can be writing to this page using arbitrary
1638                          * virtual addresses, take care about potential aliasing
1639                          * before reading the page on the kernel side.
1640                          */
1641                         if (mapping_writably_mapped(mapping))
1642                                 flush_dcache_page(page);
1643                         /*
1644                          * Mark the page accessed if we read the beginning.
1645                          */
1646                         if (!offset)
1647                                 mark_page_accessed(page);
1648                 } else {
1649                         page = ZERO_PAGE(0);
1650                         page_cache_get(page);
1651                 }
1652
1653                 /*
1654                  * Ok, we have the page, and it's up-to-date, so
1655                  * now we can copy it to user space...
1656                  *
1657                  * The actor routine returns how many bytes were actually used..
1658                  * NOTE! This may not be the same as how much of a user buffer
1659                  * we filled up (we may be padding etc), so we can only update
1660                  * "pos" here (the actor routine has to update the user buffer
1661                  * pointers and the remaining count).
1662                  */
1663                 ret = actor(desc, page, offset, nr);
1664                 offset += ret;
1665                 index += offset >> PAGE_CACHE_SHIFT;
1666                 offset &= ~PAGE_CACHE_MASK;
1667
1668                 page_cache_release(page);
1669                 if (ret != nr || !desc->count)
1670                         break;
1671
1672                 cond_resched();
1673         }
1674
1675         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1676         file_accessed(filp);
1677 }
1678
1679 static ssize_t shmem_file_read(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
1680 {
1681         read_descriptor_t desc;
1682
1683         if ((ssize_t) count < 0)
1684                 return -EINVAL;
1685         if (!access_ok(VERIFY_WRITE, buf, count))
1686                 return -EFAULT;
1687         if (!count)
1688                 return 0;
1689
1690         desc.written = 0;
1691         desc.count = count;
1692         desc.arg.buf = buf;
1693         desc.error = 0;
1694
1695         do_shmem_file_read(filp, ppos, &desc, file_read_actor);
1696         if (desc.written)
1697                 return desc.written;
1698         return desc.error;
1699 }
1700
1701 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
1702 {
1703         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
1704
1705         buf->f_type = TMPFS_MAGIC;
1706         buf->f_bsize = PAGE_CACHE_SIZE;
1707         buf->f_namelen = NAME_MAX;
1708         spin_lock(&sbinfo->stat_lock);
1709         if (sbinfo->max_blocks) {
1710                 buf->f_blocks = sbinfo->max_blocks;
1711                 buf->f_bavail = buf->f_bfree = sbinfo->free_blocks;
1712         }
1713         if (sbinfo->max_inodes) {
1714                 buf->f_files = sbinfo->max_inodes;
1715                 buf->f_ffree = sbinfo->free_inodes;
1716         }
1717         /* else leave those fields 0 like simple_statfs */
1718         spin_unlock(&sbinfo->stat_lock);
1719         return 0;
1720 }
1721
1722 /*
1723  * File creation. Allocate an inode, and we're done..
1724  */
1725 static int
1726 shmem_mknod(struct inode *dir, struct dentry *dentry, int mode, dev_t dev)
1727 {
1728         struct inode *inode = shmem_get_inode(dir->i_sb, mode, dev);
1729         int error = -ENOSPC;
1730
1731         if (inode) {
1732                 error = security_inode_init_security(inode, dir, NULL, NULL,
1733                                                      NULL);
1734                 if (error) {
1735                         if (error != -EOPNOTSUPP) {
1736                                 iput(inode);
1737                                 return error;
1738                         }
1739                 }
1740                 error = shmem_acl_init(inode, dir);
1741                 if (error) {
1742                         iput(inode);
1743                         return error;
1744                 }
1745                 if (dir->i_mode & S_ISGID) {
1746                         inode->i_gid = dir->i_gid;
1747                         if (S_ISDIR(mode))
1748                                 inode->i_mode |= S_ISGID;
1749                 }
1750                 dir->i_size += BOGO_DIRENT_SIZE;
1751                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1752                 d_instantiate(dentry, inode);
1753                 dget(dentry); /* Extra count - pin the dentry in core */
1754         }
1755         return error;
1756 }
1757
1758 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, int mode)
1759 {
1760         int error;
1761
1762         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
1763                 return error;
1764         inc_nlink(dir);
1765         return 0;
1766 }
1767
1768 static int shmem_create(struct inode *dir, struct dentry *dentry, int mode,
1769                 struct nameidata *nd)
1770 {
1771         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
1772 }
1773
1774 /*
1775  * Link a file..
1776  */
1777 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
1778 {
1779         struct inode *inode = old_dentry->d_inode;
1780         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1781
1782         /*
1783          * No ordinary (disk based) filesystem counts links as inodes;
1784          * but each new link needs a new dentry, pinning lowmem, and
1785          * tmpfs dentries cannot be pruned until they are unlinked.
1786          */
1787         if (sbinfo->max_inodes) {
1788                 spin_lock(&sbinfo->stat_lock);
1789                 if (!sbinfo->free_inodes) {
1790                         spin_unlock(&sbinfo->stat_lock);
1791                         return -ENOSPC;
1792                 }
1793                 sbinfo->free_inodes--;
1794                 spin_unlock(&sbinfo->stat_lock);
1795         }
1796
1797         dir->i_size += BOGO_DIRENT_SIZE;
1798         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1799         inc_nlink(inode);
1800         atomic_inc(&inode->i_count);    /* New dentry reference */
1801         dget(dentry);           /* Extra pinning count for the created dentry */
1802         d_instantiate(dentry, inode);
1803         return 0;
1804 }
1805
1806 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
1807 {
1808         struct inode *inode = dentry->d_inode;
1809
1810         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode)) {
1811                 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1812                 if (sbinfo->max_inodes) {
1813                         spin_lock(&sbinfo->stat_lock);
1814                         sbinfo->free_inodes++;
1815                         spin_unlock(&sbinfo->stat_lock);
1816                 }
1817         }
1818
1819         dir->i_size -= BOGO_DIRENT_SIZE;
1820         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1821         drop_nlink(inode);
1822         dput(dentry);   /* Undo the count from "create" - this does all the work */
1823         return 0;
1824 }
1825
1826 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
1827 {
1828         if (!simple_empty(dentry))
1829                 return -ENOTEMPTY;
1830
1831         drop_nlink(dentry->d_inode);
1832         drop_nlink(dir);
1833         return shmem_unlink(dir, dentry);
1834 }
1835
1836 /*
1837  * The VFS layer already does all the dentry stuff for rename,
1838  * we just have to decrement the usage count for the target if
1839  * it exists so that the VFS layer correctly free's it when it
1840  * gets overwritten.
1841  */
1842 static int shmem_rename(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
1843 {
1844         struct inode *inode = old_dentry->d_inode;
1845         int they_are_dirs = S_ISDIR(inode->i_mode);
1846
1847         if (!simple_empty(new_dentry))
1848                 return -ENOTEMPTY;
1849
1850         if (new_dentry->d_inode) {
1851                 (void) shmem_unlink(new_dir, new_dentry);
1852                 if (they_are_dirs)
1853                         drop_nlink(old_dir);
1854         } else if (they_are_dirs) {
1855                 drop_nlink(old_dir);
1856                 inc_nlink(new_dir);
1857         }
1858
1859         old_dir->i_size -= BOGO_DIRENT_SIZE;
1860         new_dir->i_size += BOGO_DIRENT_SIZE;
1861         old_dir->i_ctime = old_dir->i_mtime =
1862         new_dir->i_ctime = new_dir->i_mtime =
1863         inode->i_ctime = CURRENT_TIME;
1864         return 0;
1865 }
1866
1867 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
1868 {
1869         int error;
1870         int len;
1871         struct inode *inode;
1872         struct page *page = NULL;
1873         char *kaddr;
1874         struct shmem_inode_info *info;
1875
1876         len = strlen(symname) + 1;
1877         if (len > PAGE_CACHE_SIZE)
1878                 return -ENAMETOOLONG;
1879
1880         inode = shmem_get_inode(dir->i_sb, S_IFLNK|S_IRWXUGO, 0);
1881         if (!inode)
1882                 return -ENOSPC;
1883
1884         error = security_inode_init_security(inode, dir, NULL, NULL,
1885                                              NULL);
1886         if (error) {
1887                 if (error != -EOPNOTSUPP) {
1888                         iput(inode);
1889                         return error;
1890                 }
1891                 error = 0;
1892         }
1893
1894         info = SHMEM_I(inode);
1895         inode->i_size = len-1;
1896         if (len <= (char *)inode - (char *)info) {
1897                 /* do it inline */
1898                 memcpy(info, symname, len);
1899                 inode->i_op = &shmem_symlink_inline_operations;
1900         } else {
1901                 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
1902                 if (error) {
1903                         iput(inode);
1904                         return error;
1905                 }
1906                 inode->i_op = &shmem_symlink_inode_operations;
1907                 kaddr = kmap_atomic(page, KM_USER0);
1908                 memcpy(kaddr, symname, len);
1909                 kunmap_atomic(kaddr, KM_USER0);
1910                 set_page_dirty(page);
1911                 page_cache_release(page);
1912         }
1913         if (dir->i_mode & S_ISGID)
1914                 inode->i_gid = dir->i_gid;
1915         dir->i_size += BOGO_DIRENT_SIZE;
1916         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1917         d_instantiate(dentry, inode);
1918         dget(dentry);
1919         return 0;
1920 }
1921
1922 static void *shmem_follow_link_inline(struct dentry *dentry, struct nameidata *nd)
1923 {
1924         nd_set_link(nd, (char *)SHMEM_I(dentry->d_inode));
1925         return NULL;
1926 }
1927
1928 static void *shmem_follow_link(struct dentry *dentry, struct nameidata *nd)
1929 {
1930         struct page *page = NULL;
1931         int res = shmem_getpage(dentry->d_inode, 0, &page, SGP_READ, NULL);
1932         nd_set_link(nd, res ? ERR_PTR(res) : kmap(page));
1933         return page;
1934 }
1935
1936 static void shmem_put_link(struct dentry *dentry, struct nameidata *nd, void *cookie)
1937 {
1938         if (!IS_ERR(nd_get_link(nd))) {
1939                 struct page *page = cookie;
1940                 kunmap(page);
1941                 mark_page_accessed(page);
1942                 page_cache_release(page);
1943         }
1944 }
1945
1946 static const struct inode_operations shmem_symlink_inline_operations = {
1947         .readlink       = generic_readlink,
1948         .follow_link    = shmem_follow_link_inline,
1949 };
1950
1951 static const struct inode_operations shmem_symlink_inode_operations = {
1952         .truncate       = shmem_truncate,
1953         .readlink       = generic_readlink,
1954         .follow_link    = shmem_follow_link,
1955         .put_link       = shmem_put_link,
1956 };
1957
1958 #ifdef CONFIG_TMPFS_POSIX_ACL
1959 /**
1960  * Superblocks without xattr inode operations will get security.* xattr
1961  * support from the VFS "for free". As soon as we have any other xattrs
1962  * like ACLs, we also need to implement the security.* handlers at
1963  * filesystem level, though.
1964  */
1965
1966 static size_t shmem_xattr_security_list(struct inode *inode, char *list,
1967                                         size_t list_len, const char *name,
1968                                         size_t name_len)
1969 {
1970         return security_inode_listsecurity(inode, list, list_len);
1971 }
1972
1973 static int shmem_xattr_security_get(struct inode *inode, const char *name,
1974                                     void *buffer, size_t size)
1975 {
1976         if (strcmp(name, "") == 0)
1977                 return -EINVAL;
1978         return security_inode_getsecurity(inode, name, buffer, size,
1979                                           -EOPNOTSUPP);
1980 }
1981
1982 static int shmem_xattr_security_set(struct inode *inode, const char *name,
1983                                     const void *value, size_t size, int flags)
1984 {
1985         if (strcmp(name, "") == 0)
1986                 return -EINVAL;
1987         return security_inode_setsecurity(inode, name, value, size, flags);
1988 }
1989
1990 static struct xattr_handler shmem_xattr_security_handler = {
1991         .prefix = XATTR_SECURITY_PREFIX,
1992         .list   = shmem_xattr_security_list,
1993         .get    = shmem_xattr_security_get,
1994         .set    = shmem_xattr_security_set,
1995 };
1996
1997 static struct xattr_handler *shmem_xattr_handlers[] = {
1998         &shmem_xattr_acl_access_handler,
1999         &shmem_xattr_acl_default_handler,
2000         &shmem_xattr_security_handler,
2001         NULL
2002 };
2003 #endif
2004
2005 static struct dentry *shmem_get_parent(struct dentry *child)
2006 {
2007         return ERR_PTR(-ESTALE);
2008 }
2009
2010 static int shmem_match(struct inode *ino, void *vfh)
2011 {
2012         __u32 *fh = vfh;
2013         __u64 inum = fh[2];
2014         inum = (inum << 32) | fh[1];
2015         return ino->i_ino == inum && fh[0] == ino->i_generation;
2016 }
2017
2018 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
2019                 struct fid *fid, int fh_len, int fh_type)
2020 {
2021         struct inode *inode;
2022         struct dentry *dentry = NULL;
2023         u64 inum = fid->raw[2];
2024         inum = (inum << 32) | fid->raw[1];
2025
2026         if (fh_len < 3)
2027                 return NULL;
2028
2029         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
2030                         shmem_match, fid->raw);
2031         if (inode) {
2032                 dentry = d_find_alias(inode);
2033                 iput(inode);
2034         }
2035
2036         return dentry;
2037 }
2038
2039 static int shmem_encode_fh(struct dentry *dentry, __u32 *fh, int *len,
2040                                 int connectable)
2041 {
2042         struct inode *inode = dentry->d_inode;
2043
2044         if (*len < 3)
2045                 return 255;
2046
2047         if (hlist_unhashed(&inode->i_hash)) {
2048                 /* Unfortunately insert_inode_hash is not idempotent,
2049                  * so as we hash inodes here rather than at creation
2050                  * time, we need a lock to ensure we only try
2051                  * to do it once
2052                  */
2053                 static DEFINE_SPINLOCK(lock);
2054                 spin_lock(&lock);
2055                 if (hlist_unhashed(&inode->i_hash))
2056                         __insert_inode_hash(inode,
2057                                             inode->i_ino + inode->i_generation);
2058                 spin_unlock(&lock);
2059         }
2060
2061         fh[0] = inode->i_generation;
2062         fh[1] = inode->i_ino;
2063         fh[2] = ((__u64)inode->i_ino) >> 32;
2064
2065         *len = 3;
2066         return 1;
2067 }
2068
2069 static const struct export_operations shmem_export_ops = {
2070         .get_parent     = shmem_get_parent,
2071         .encode_fh      = shmem_encode_fh,
2072         .fh_to_dentry   = shmem_fh_to_dentry,
2073 };
2074
2075 static int shmem_parse_options(char *options, int *mode, uid_t *uid,
2076         gid_t *gid, unsigned long *blocks, unsigned long *inodes,
2077         int *policy, nodemask_t *policy_nodes)
2078 {
2079         char *this_char, *value, *rest;
2080
2081         while (options != NULL) {
2082                 this_char = options;
2083                 for (;;) {
2084                         /*
2085                          * NUL-terminate this option: unfortunately,
2086                          * mount options form a comma-separated list,
2087                          * but mpol's nodelist may also contain commas.
2088                          */
2089                         options = strchr(options, ',');
2090                         if (options == NULL)
2091                                 break;
2092                         options++;
2093                         if (!isdigit(*options)) {
2094                                 options[-1] = '\0';
2095                                 break;
2096                         }
2097                 }
2098                 if (!*this_char)
2099                         continue;
2100                 if ((value = strchr(this_char,'=')) != NULL) {
2101                         *value++ = 0;
2102                 } else {
2103                         printk(KERN_ERR
2104                             "tmpfs: No value for mount option '%s'\n",
2105                             this_char);
2106                         return 1;
2107                 }
2108
2109                 if (!strcmp(this_char,"size")) {
2110                         unsigned long long size;
2111                         size = memparse(value,&rest);
2112                         if (*rest == '%') {
2113                                 size <<= PAGE_SHIFT;
2114                                 size *= totalram_pages;
2115                                 do_div(size, 100);
2116                                 rest++;
2117                         }
2118                         if (*rest)
2119                                 goto bad_val;
2120                         *blocks = size >> PAGE_CACHE_SHIFT;
2121                 } else if (!strcmp(this_char,"nr_blocks")) {
2122                         *blocks = memparse(value,&rest);
2123                         if (*rest)
2124                                 goto bad_val;
2125                 } else if (!strcmp(this_char,"nr_inodes")) {
2126                         *inodes = memparse(value,&rest);
2127                         if (*rest)
2128                                 goto bad_val;
2129                 } else if (!strcmp(this_char,"mode")) {
2130                         if (!mode)
2131                                 continue;
2132                         *mode = simple_strtoul(value,&rest,8);
2133                         if (*rest)
2134                                 goto bad_val;
2135                 } else if (!strcmp(this_char,"uid")) {
2136                         if (!uid)
2137                                 continue;
2138                         *uid = simple_strtoul(value,&rest,0);
2139                         if (*rest)
2140                                 goto bad_val;
2141                 } else if (!strcmp(this_char,"gid")) {
2142                         if (!gid)
2143                                 continue;
2144                         *gid = simple_strtoul(value,&rest,0);
2145                         if (*rest)
2146                                 goto bad_val;
2147                 } else if (!strcmp(this_char,"mpol")) {
2148                         if (shmem_parse_mpol(value,policy,policy_nodes))
2149                                 goto bad_val;
2150                 } else {
2151                         printk(KERN_ERR "tmpfs: Bad mount option %s\n",
2152                                this_char);
2153                         return 1;
2154                 }
2155         }
2156         return 0;
2157
2158 bad_val:
2159         printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
2160                value, this_char);
2161         return 1;
2162
2163 }
2164
2165 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2166 {
2167         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2168         unsigned long max_blocks = sbinfo->max_blocks;
2169         unsigned long max_inodes = sbinfo->max_inodes;
2170         int policy = sbinfo->policy;
2171         nodemask_t policy_nodes = sbinfo->policy_nodes;
2172         unsigned long blocks;
2173         unsigned long inodes;
2174         int error = -EINVAL;
2175
2176         if (shmem_parse_options(data, NULL, NULL, NULL, &max_blocks,
2177                                 &max_inodes, &policy, &policy_nodes))
2178                 return error;
2179
2180         spin_lock(&sbinfo->stat_lock);
2181         blocks = sbinfo->max_blocks - sbinfo->free_blocks;
2182         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2183         if (max_blocks < blocks)
2184                 goto out;
2185         if (max_inodes < inodes)
2186                 goto out;
2187         /*
2188          * Those tests also disallow limited->unlimited while any are in
2189          * use, so i_blocks will always be zero when max_blocks is zero;
2190          * but we must separately disallow unlimited->limited, because
2191          * in that case we have no record of how much is already in use.
2192          */
2193         if (max_blocks && !sbinfo->max_blocks)
2194                 goto out;
2195         if (max_inodes && !sbinfo->max_inodes)
2196                 goto out;
2197
2198         error = 0;
2199         sbinfo->max_blocks  = max_blocks;
2200         sbinfo->free_blocks = max_blocks - blocks;
2201         sbinfo->max_inodes  = max_inodes;
2202         sbinfo->free_inodes = max_inodes - inodes;
2203         sbinfo->policy = policy;
2204         sbinfo->policy_nodes = policy_nodes;
2205 out:
2206         spin_unlock(&sbinfo->stat_lock);
2207         return error;
2208 }
2209 #endif
2210
2211 static void shmem_put_super(struct super_block *sb)
2212 {
2213         kfree(sb->s_fs_info);
2214         sb->s_fs_info = NULL;
2215 }
2216
2217 static int shmem_fill_super(struct super_block *sb,
2218                             void *data, int silent)
2219 {
2220         struct inode *inode;
2221         struct dentry *root;
2222         int mode   = S_IRWXUGO | S_ISVTX;
2223         uid_t uid = current->fsuid;
2224         gid_t gid = current->fsgid;
2225         int err = -ENOMEM;
2226         struct shmem_sb_info *sbinfo;
2227         unsigned long blocks = 0;
2228         unsigned long inodes = 0;
2229         int policy = MPOL_DEFAULT;
2230         nodemask_t policy_nodes = node_states[N_HIGH_MEMORY];
2231
2232 #ifdef CONFIG_TMPFS
2233         /*
2234          * Per default we only allow half of the physical ram per
2235          * tmpfs instance, limiting inodes to one per page of lowmem;
2236          * but the internal instance is left unlimited.
2237          */
2238         if (!(sb->s_flags & MS_NOUSER)) {
2239                 blocks = totalram_pages / 2;
2240                 inodes = totalram_pages - totalhigh_pages;
2241                 if (inodes > blocks)
2242                         inodes = blocks;
2243                 if (shmem_parse_options(data, &mode, &uid, &gid, &blocks,
2244                                         &inodes, &policy, &policy_nodes))
2245                         return -EINVAL;
2246         }
2247         sb->s_export_op = &shmem_export_ops;
2248 #else
2249         sb->s_flags |= MS_NOUSER;
2250 #endif
2251
2252         /* Round up to L1_CACHE_BYTES to resist false sharing */
2253         sbinfo = kmalloc(max((int)sizeof(struct shmem_sb_info),
2254                                 L1_CACHE_BYTES), GFP_KERNEL);
2255         if (!sbinfo)
2256                 return -ENOMEM;
2257
2258         spin_lock_init(&sbinfo->stat_lock);
2259         sbinfo->max_blocks = blocks;
2260         sbinfo->free_blocks = blocks;
2261         sbinfo->max_inodes = inodes;
2262         sbinfo->free_inodes = inodes;
2263         sbinfo->policy = policy;
2264         sbinfo->policy_nodes = policy_nodes;
2265
2266         sb->s_fs_info = sbinfo;
2267         sb->s_maxbytes = SHMEM_MAX_BYTES;
2268         sb->s_blocksize = PAGE_CACHE_SIZE;
2269         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
2270         sb->s_magic = TMPFS_MAGIC;
2271         sb->s_op = &shmem_ops;
2272         sb->s_time_gran = 1;
2273 #ifdef CONFIG_TMPFS_POSIX_ACL
2274         sb->s_xattr = shmem_xattr_handlers;
2275         sb->s_flags |= MS_POSIXACL;
2276 #endif
2277
2278         inode = shmem_get_inode(sb, S_IFDIR | mode, 0);
2279         if (!inode)
2280                 goto failed;
2281         inode->i_uid = uid;
2282         inode->i_gid = gid;
2283         root = d_alloc_root(inode);
2284         if (!root)
2285                 goto failed_iput;
2286         sb->s_root = root;
2287         return 0;
2288
2289 failed_iput:
2290         iput(inode);
2291 failed:
2292         shmem_put_super(sb);
2293         return err;
2294 }
2295
2296 static struct kmem_cache *shmem_inode_cachep;
2297
2298 static struct inode *shmem_alloc_inode(struct super_block *sb)
2299 {
2300         struct shmem_inode_info *p;
2301         p = (struct shmem_inode_info *)kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
2302         if (!p)
2303                 return NULL;
2304         return &p->vfs_inode;
2305 }
2306
2307 static void shmem_destroy_inode(struct inode *inode)
2308 {
2309         if ((inode->i_mode & S_IFMT) == S_IFREG) {
2310                 /* only struct inode is valid if it's an inline symlink */
2311                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
2312         }
2313         shmem_acl_destroy_inode(inode);
2314         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
2315 }
2316
2317 static void init_once(struct kmem_cache *cachep, void *foo)
2318 {
2319         struct shmem_inode_info *p = (struct shmem_inode_info *) foo;
2320
2321         inode_init_once(&p->vfs_inode);
2322 #ifdef CONFIG_TMPFS_POSIX_ACL
2323         p->i_acl = NULL;
2324         p->i_default_acl = NULL;
2325 #endif
2326 }
2327
2328 static int init_inodecache(void)
2329 {
2330         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
2331                                 sizeof(struct shmem_inode_info),
2332                                 0, SLAB_PANIC, init_once);
2333         return 0;
2334 }
2335
2336 static void destroy_inodecache(void)
2337 {
2338         kmem_cache_destroy(shmem_inode_cachep);
2339 }
2340
2341 static const struct address_space_operations shmem_aops = {
2342         .writepage      = shmem_writepage,
2343         .set_page_dirty = __set_page_dirty_no_writeback,
2344 #ifdef CONFIG_TMPFS
2345         .readpage       = shmem_readpage,
2346         .write_begin    = shmem_write_begin,
2347         .write_end      = shmem_write_end,
2348 #endif
2349         .migratepage    = migrate_page,
2350 };
2351
2352 static const struct file_operations shmem_file_operations = {
2353         .mmap           = shmem_mmap,
2354 #ifdef CONFIG_TMPFS
2355         .llseek         = generic_file_llseek,
2356         .read           = shmem_file_read,
2357         .write          = shmem_file_write,
2358         .fsync          = simple_sync_file,
2359         .splice_read    = generic_file_splice_read,
2360         .splice_write   = generic_file_splice_write,
2361 #endif
2362 };
2363
2364 static const struct inode_operations shmem_inode_operations = {
2365         .truncate       = shmem_truncate,
2366         .setattr        = shmem_notify_change,
2367         .truncate_range = shmem_truncate_range,
2368 #ifdef CONFIG_TMPFS_POSIX_ACL
2369         .setxattr       = generic_setxattr,
2370         .getxattr       = generic_getxattr,
2371         .listxattr      = generic_listxattr,
2372         .removexattr    = generic_removexattr,
2373         .permission     = shmem_permission,
2374 #endif
2375
2376 };
2377
2378 static const struct inode_operations shmem_dir_inode_operations = {
2379 #ifdef CONFIG_TMPFS
2380         .create         = shmem_create,
2381         .lookup         = simple_lookup,
2382         .link           = shmem_link,
2383         .unlink         = shmem_unlink,
2384         .symlink        = shmem_symlink,
2385         .mkdir          = shmem_mkdir,
2386         .rmdir          = shmem_rmdir,
2387         .mknod          = shmem_mknod,
2388         .rename         = shmem_rename,
2389 #endif
2390 #ifdef CONFIG_TMPFS_POSIX_ACL
2391         .setattr        = shmem_notify_change,
2392         .setxattr       = generic_setxattr,
2393         .getxattr       = generic_getxattr,
2394         .listxattr      = generic_listxattr,
2395         .removexattr    = generic_removexattr,
2396         .permission     = shmem_permission,
2397 #endif
2398 };
2399
2400 static const struct inode_operations shmem_special_inode_operations = {
2401 #ifdef CONFIG_TMPFS_POSIX_ACL
2402         .setattr        = shmem_notify_change,
2403         .setxattr       = generic_setxattr,
2404         .getxattr       = generic_getxattr,
2405         .listxattr      = generic_listxattr,
2406         .removexattr    = generic_removexattr,
2407         .permission     = shmem_permission,
2408 #endif
2409 };
2410
2411 static const struct super_operations shmem_ops = {
2412         .alloc_inode    = shmem_alloc_inode,
2413         .destroy_inode  = shmem_destroy_inode,
2414 #ifdef CONFIG_TMPFS
2415         .statfs         = shmem_statfs,
2416         .remount_fs     = shmem_remount_fs,
2417 #endif
2418         .delete_inode   = shmem_delete_inode,
2419         .drop_inode     = generic_delete_inode,
2420         .put_super      = shmem_put_super,
2421 };
2422
2423 static struct vm_operations_struct shmem_vm_ops = {
2424         .fault          = shmem_fault,
2425 #ifdef CONFIG_NUMA
2426         .set_policy     = shmem_set_policy,
2427         .get_policy     = shmem_get_policy,
2428 #endif
2429 };
2430
2431
2432 static int shmem_get_sb(struct file_system_type *fs_type,
2433         int flags, const char *dev_name, void *data, struct vfsmount *mnt)
2434 {
2435         return get_sb_nodev(fs_type, flags, data, shmem_fill_super, mnt);
2436 }
2437
2438 static struct file_system_type tmpfs_fs_type = {
2439         .owner          = THIS_MODULE,
2440         .name           = "tmpfs",
2441         .get_sb         = shmem_get_sb,
2442         .kill_sb        = kill_litter_super,
2443 };
2444 static struct vfsmount *shm_mnt;
2445
2446 static int __init init_tmpfs(void)
2447 {
2448         int error;
2449
2450         error = bdi_init(&shmem_backing_dev_info);
2451         if (error)
2452                 goto out4;
2453
2454         error = init_inodecache();
2455         if (error)
2456                 goto out3;
2457
2458         error = register_filesystem(&tmpfs_fs_type);
2459         if (error) {
2460                 printk(KERN_ERR "Could not register tmpfs\n");
2461                 goto out2;
2462         }
2463
2464         shm_mnt = vfs_kern_mount(&tmpfs_fs_type, MS_NOUSER,
2465                                 tmpfs_fs_type.name, NULL);
2466         if (IS_ERR(shm_mnt)) {
2467                 error = PTR_ERR(shm_mnt);
2468                 printk(KERN_ERR "Could not kern_mount tmpfs\n");
2469                 goto out1;
2470         }
2471         return 0;
2472
2473 out1:
2474         unregister_filesystem(&tmpfs_fs_type);
2475 out2:
2476         destroy_inodecache();
2477 out3:
2478         bdi_destroy(&shmem_backing_dev_info);
2479 out4:
2480         shm_mnt = ERR_PTR(error);
2481         return error;
2482 }
2483 module_init(init_tmpfs)
2484
2485 /*
2486  * shmem_file_setup - get an unlinked file living in tmpfs
2487  *
2488  * @name: name for dentry (to be seen in /proc/<pid>/maps
2489  * @size: size to be set for the file
2490  *
2491  */
2492 struct file *shmem_file_setup(char *name, loff_t size, unsigned long flags)
2493 {
2494         int error;
2495         struct file *file;
2496         struct inode *inode;
2497         struct dentry *dentry, *root;
2498         struct qstr this;
2499
2500         if (IS_ERR(shm_mnt))
2501                 return (void *)shm_mnt;
2502
2503         if (size < 0 || size > SHMEM_MAX_BYTES)
2504                 return ERR_PTR(-EINVAL);
2505
2506         if (shmem_acct_size(flags, size))
2507                 return ERR_PTR(-ENOMEM);
2508
2509         error = -ENOMEM;
2510         this.name = name;
2511         this.len = strlen(name);
2512         this.hash = 0; /* will go */
2513         root = shm_mnt->mnt_root;
2514         dentry = d_alloc(root, &this);
2515         if (!dentry)
2516                 goto put_memory;
2517
2518         error = -ENFILE;
2519         file = get_empty_filp();
2520         if (!file)
2521                 goto put_dentry;
2522
2523         error = -ENOSPC;
2524         inode = shmem_get_inode(root->d_sb, S_IFREG | S_IRWXUGO, 0);
2525         if (!inode)
2526                 goto close_file;
2527
2528         SHMEM_I(inode)->flags = flags & VM_ACCOUNT;
2529         d_instantiate(dentry, inode);
2530         inode->i_size = size;
2531         inode->i_nlink = 0;     /* It is unlinked */
2532         init_file(file, shm_mnt, dentry, FMODE_WRITE | FMODE_READ,
2533                         &shmem_file_operations);
2534         return file;
2535
2536 close_file:
2537         put_filp(file);
2538 put_dentry:
2539         dput(dentry);
2540 put_memory:
2541         shmem_unacct_size(flags, size);
2542         return ERR_PTR(error);
2543 }
2544
2545 /*
2546  * shmem_zero_setup - setup a shared anonymous mapping
2547  *
2548  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
2549  */
2550 int shmem_zero_setup(struct vm_area_struct *vma)
2551 {
2552         struct file *file;
2553         loff_t size = vma->vm_end - vma->vm_start;
2554
2555         file = shmem_file_setup("dev/zero", size, vma->vm_flags);
2556         if (IS_ERR(file))
2557                 return PTR_ERR(file);
2558
2559         if (vma->vm_file)
2560                 fput(vma->vm_file);
2561         vma->vm_file = file;
2562         vma->vm_ops = &shmem_vm_ops;
2563         return 0;
2564 }