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