SL*B: drop kmem cache argument from constructor
[linux-2.6.git] / mm / shmem.c
1 /*
2  * Resizable virtual memory filesystem for Linux.
3  *
4  * Copyright (C) 2000 Linus Torvalds.
5  *               2000 Transmeta Corp.
6  *               2000-2001 Christoph Rohland
7  *               2000-2001 SAP AG
8  *               2002 Red Hat Inc.
9  * Copyright (C) 2002-2005 Hugh Dickins.
10  * Copyright (C) 2002-2005 VERITAS Software Corporation.
11  * Copyright (C) 2004 Andi Kleen, SuSE Labs
12  *
13  * Extended attribute support for tmpfs:
14  * Copyright (c) 2004, Luke Kenneth Casson Leighton <lkcl@lkcl.net>
15  * Copyright (c) 2004 Red Hat, Inc., James Morris <jmorris@redhat.com>
16  *
17  * This file is released under the GPL.
18  */
19
20 /*
21  * This virtual memory filesystem is heavily based on the ramfs. It
22  * extends ramfs by the ability to use swap and honor resource limits
23  * which makes it a completely usable filesystem.
24  */
25
26 #include <linux/module.h>
27 #include <linux/init.h>
28 #include <linux/fs.h>
29 #include <linux/xattr.h>
30 #include <linux/exportfs.h>
31 #include <linux/generic_acl.h>
32 #include <linux/mm.h>
33 #include <linux/mman.h>
34 #include <linux/file.h>
35 #include <linux/swap.h>
36 #include <linux/pagemap.h>
37 #include <linux/string.h>
38 #include <linux/slab.h>
39 #include <linux/backing-dev.h>
40 #include <linux/shmem_fs.h>
41 #include <linux/mount.h>
42 #include <linux/writeback.h>
43 #include <linux/vfs.h>
44 #include <linux/blkdev.h>
45 #include <linux/security.h>
46 #include <linux/swapops.h>
47 #include <linux/mempolicy.h>
48 #include <linux/namei.h>
49 #include <linux/ctype.h>
50 #include <linux/migrate.h>
51 #include <linux/highmem.h>
52 #include <linux/seq_file.h>
53
54 #include <asm/uaccess.h>
55 #include <asm/div64.h>
56 #include <asm/pgtable.h>
57
58 /* This magic number is used in glibc for posix shared memory */
59 #define TMPFS_MAGIC     0x01021994
60
61 #define ENTRIES_PER_PAGE (PAGE_CACHE_SIZE/sizeof(unsigned long))
62 #define ENTRIES_PER_PAGEPAGE (ENTRIES_PER_PAGE*ENTRIES_PER_PAGE)
63 #define BLOCKS_PER_PAGE  (PAGE_CACHE_SIZE/512)
64
65 #define SHMEM_MAX_INDEX  (SHMEM_NR_DIRECT + (ENTRIES_PER_PAGEPAGE/2) * (ENTRIES_PER_PAGE+1))
66 #define SHMEM_MAX_BYTES  ((unsigned long long)SHMEM_MAX_INDEX << PAGE_CACHE_SHIFT)
67
68 #define VM_ACCT(size)    (PAGE_CACHE_ALIGN(size) >> PAGE_SHIFT)
69
70 /* info->flags needs VM_flags to handle pagein/truncate races efficiently */
71 #define SHMEM_PAGEIN     VM_READ
72 #define SHMEM_TRUNCATE   VM_WRITE
73
74 /* Definition to limit shmem_truncate's steps between cond_rescheds */
75 #define LATENCY_LIMIT    64
76
77 /* Pretend that each entry is of this size in directory's i_size */
78 #define BOGO_DIRENT_SIZE 20
79
80 /* Flag allocation requirements to shmem_getpage and shmem_swp_alloc */
81 enum sgp_type {
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_DIRTY,      /* like SGP_CACHE, but set new page dirty */
85         SGP_WRITE,      /* may exceed i_size, may allocate page */
86 };
87
88 #ifdef CONFIG_TMPFS
89 static unsigned long shmem_default_max_blocks(void)
90 {
91         return totalram_pages / 2;
92 }
93
94 static unsigned long shmem_default_max_inodes(void)
95 {
96         return min(totalram_pages - totalhigh_pages, totalram_pages / 2);
97 }
98 #endif
99
100 static int shmem_getpage(struct inode *inode, unsigned long idx,
101                          struct page **pagep, enum sgp_type sgp, int *type);
102
103 static inline struct page *shmem_dir_alloc(gfp_t gfp_mask)
104 {
105         /*
106          * The above definition of ENTRIES_PER_PAGE, and the use of
107          * BLOCKS_PER_PAGE on indirect pages, assume PAGE_CACHE_SIZE:
108          * might be reconsidered if it ever diverges from PAGE_SIZE.
109          *
110          * Mobility flags are masked out as swap vectors cannot move
111          */
112         return alloc_pages((gfp_mask & ~GFP_MOVABLE_MASK) | __GFP_ZERO,
113                                 PAGE_CACHE_SHIFT-PAGE_SHIFT);
114 }
115
116 static inline void shmem_dir_free(struct page *page)
117 {
118         __free_pages(page, PAGE_CACHE_SHIFT-PAGE_SHIFT);
119 }
120
121 static struct page **shmem_dir_map(struct page *page)
122 {
123         return (struct page **)kmap_atomic(page, KM_USER0);
124 }
125
126 static inline void shmem_dir_unmap(struct page **dir)
127 {
128         kunmap_atomic(dir, KM_USER0);
129 }
130
131 static swp_entry_t *shmem_swp_map(struct page *page)
132 {
133         return (swp_entry_t *)kmap_atomic(page, KM_USER1);
134 }
135
136 static inline void shmem_swp_balance_unmap(void)
137 {
138         /*
139          * When passing a pointer to an i_direct entry, to code which
140          * also handles indirect entries and so will shmem_swp_unmap,
141          * we must arrange for the preempt count to remain in balance.
142          * What kmap_atomic of a lowmem page does depends on config
143          * and architecture, so pretend to kmap_atomic some lowmem page.
144          */
145         (void) kmap_atomic(ZERO_PAGE(0), KM_USER1);
146 }
147
148 static inline void shmem_swp_unmap(swp_entry_t *entry)
149 {
150         kunmap_atomic(entry, KM_USER1);
151 }
152
153 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
154 {
155         return sb->s_fs_info;
156 }
157
158 /*
159  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
160  * for shared memory and for shared anonymous (/dev/zero) mappings
161  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
162  * consistent with the pre-accounting of private mappings ...
163  */
164 static inline int shmem_acct_size(unsigned long flags, loff_t size)
165 {
166         return (flags & VM_ACCOUNT)?
167                 security_vm_enough_memory(VM_ACCT(size)): 0;
168 }
169
170 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
171 {
172         if (flags & VM_ACCOUNT)
173                 vm_unacct_memory(VM_ACCT(size));
174 }
175
176 /*
177  * ... whereas tmpfs objects are accounted incrementally as
178  * pages are allocated, in order to allow huge sparse files.
179  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
180  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
181  */
182 static inline int shmem_acct_block(unsigned long flags)
183 {
184         return (flags & VM_ACCOUNT)?
185                 0: security_vm_enough_memory(VM_ACCT(PAGE_CACHE_SIZE));
186 }
187
188 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
189 {
190         if (!(flags & VM_ACCOUNT))
191                 vm_unacct_memory(pages * VM_ACCT(PAGE_CACHE_SIZE));
192 }
193
194 static const struct super_operations shmem_ops;
195 static const struct address_space_operations shmem_aops;
196 static const struct file_operations shmem_file_operations;
197 static const struct inode_operations shmem_inode_operations;
198 static const struct inode_operations shmem_dir_inode_operations;
199 static const struct inode_operations shmem_special_inode_operations;
200 static struct vm_operations_struct shmem_vm_ops;
201
202 static struct backing_dev_info shmem_backing_dev_info  __read_mostly = {
203         .ra_pages       = 0,    /* No readahead */
204         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK,
205         .unplug_io_fn   = default_unplug_io_fn,
206 };
207
208 static LIST_HEAD(shmem_swaplist);
209 static DEFINE_MUTEX(shmem_swaplist_mutex);
210
211 static void shmem_free_blocks(struct inode *inode, long pages)
212 {
213         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
214         if (sbinfo->max_blocks) {
215                 spin_lock(&sbinfo->stat_lock);
216                 sbinfo->free_blocks += pages;
217                 inode->i_blocks -= pages*BLOCKS_PER_PAGE;
218                 spin_unlock(&sbinfo->stat_lock);
219         }
220 }
221
222 static int shmem_reserve_inode(struct super_block *sb)
223 {
224         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
225         if (sbinfo->max_inodes) {
226                 spin_lock(&sbinfo->stat_lock);
227                 if (!sbinfo->free_inodes) {
228                         spin_unlock(&sbinfo->stat_lock);
229                         return -ENOSPC;
230                 }
231                 sbinfo->free_inodes--;
232                 spin_unlock(&sbinfo->stat_lock);
233         }
234         return 0;
235 }
236
237 static void shmem_free_inode(struct super_block *sb)
238 {
239         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
240         if (sbinfo->max_inodes) {
241                 spin_lock(&sbinfo->stat_lock);
242                 sbinfo->free_inodes++;
243                 spin_unlock(&sbinfo->stat_lock);
244         }
245 }
246
247 /**
248  * shmem_recalc_inode - recalculate the size of an inode
249  * @inode: inode to recalc
250  *
251  * We have to calculate the free blocks since the mm can drop
252  * undirtied hole pages behind our back.
253  *
254  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
255  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
256  *
257  * It has to be called with the spinlock held.
258  */
259 static void shmem_recalc_inode(struct inode *inode)
260 {
261         struct shmem_inode_info *info = SHMEM_I(inode);
262         long freed;
263
264         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
265         if (freed > 0) {
266                 info->alloced -= freed;
267                 shmem_unacct_blocks(info->flags, freed);
268                 shmem_free_blocks(inode, freed);
269         }
270 }
271
272 /**
273  * shmem_swp_entry - find the swap vector position in the info structure
274  * @info:  info structure for the inode
275  * @index: index of the page to find
276  * @page:  optional page to add to the structure. Has to be preset to
277  *         all zeros
278  *
279  * If there is no space allocated yet it will return NULL when
280  * page is NULL, else it will use the page for the needed block,
281  * setting it to NULL on return to indicate that it has been used.
282  *
283  * The swap vector is organized the following way:
284  *
285  * There are SHMEM_NR_DIRECT entries directly stored in the
286  * shmem_inode_info structure. So small files do not need an addional
287  * allocation.
288  *
289  * For pages with index > SHMEM_NR_DIRECT there is the pointer
290  * i_indirect which points to a page which holds in the first half
291  * doubly indirect blocks, in the second half triple indirect blocks:
292  *
293  * For an artificial ENTRIES_PER_PAGE = 4 this would lead to the
294  * following layout (for SHMEM_NR_DIRECT == 16):
295  *
296  * i_indirect -> dir --> 16-19
297  *            |      +-> 20-23
298  *            |
299  *            +-->dir2 --> 24-27
300  *            |        +-> 28-31
301  *            |        +-> 32-35
302  *            |        +-> 36-39
303  *            |
304  *            +-->dir3 --> 40-43
305  *                     +-> 44-47
306  *                     +-> 48-51
307  *                     +-> 52-55
308  */
309 static swp_entry_t *shmem_swp_entry(struct shmem_inode_info *info, unsigned long index, struct page **page)
310 {
311         unsigned long offset;
312         struct page **dir;
313         struct page *subdir;
314
315         if (index < SHMEM_NR_DIRECT) {
316                 shmem_swp_balance_unmap();
317                 return info->i_direct+index;
318         }
319         if (!info->i_indirect) {
320                 if (page) {
321                         info->i_indirect = *page;
322                         *page = NULL;
323                 }
324                 return NULL;                    /* need another page */
325         }
326
327         index -= SHMEM_NR_DIRECT;
328         offset = index % ENTRIES_PER_PAGE;
329         index /= ENTRIES_PER_PAGE;
330         dir = shmem_dir_map(info->i_indirect);
331
332         if (index >= ENTRIES_PER_PAGE/2) {
333                 index -= ENTRIES_PER_PAGE/2;
334                 dir += ENTRIES_PER_PAGE/2 + index/ENTRIES_PER_PAGE;
335                 index %= ENTRIES_PER_PAGE;
336                 subdir = *dir;
337                 if (!subdir) {
338                         if (page) {
339                                 *dir = *page;
340                                 *page = NULL;
341                         }
342                         shmem_dir_unmap(dir);
343                         return NULL;            /* need another page */
344                 }
345                 shmem_dir_unmap(dir);
346                 dir = shmem_dir_map(subdir);
347         }
348
349         dir += index;
350         subdir = *dir;
351         if (!subdir) {
352                 if (!page || !(subdir = *page)) {
353                         shmem_dir_unmap(dir);
354                         return NULL;            /* need a page */
355                 }
356                 *dir = subdir;
357                 *page = NULL;
358         }
359         shmem_dir_unmap(dir);
360         return shmem_swp_map(subdir) + offset;
361 }
362
363 static void shmem_swp_set(struct shmem_inode_info *info, swp_entry_t *entry, unsigned long value)
364 {
365         long incdec = value? 1: -1;
366
367         entry->val = value;
368         info->swapped += incdec;
369         if ((unsigned long)(entry - info->i_direct) >= SHMEM_NR_DIRECT) {
370                 struct page *page = kmap_atomic_to_page(entry);
371                 set_page_private(page, page_private(page) + incdec);
372         }
373 }
374
375 /**
376  * shmem_swp_alloc - get the position of the swap entry for the page.
377  * @info:       info structure for the inode
378  * @index:      index of the page to find
379  * @sgp:        check and recheck i_size? skip allocation?
380  *
381  * If the entry does not exist, allocate it.
382  */
383 static swp_entry_t *shmem_swp_alloc(struct shmem_inode_info *info, unsigned long index, enum sgp_type sgp)
384 {
385         struct inode *inode = &info->vfs_inode;
386         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
387         struct page *page = NULL;
388         swp_entry_t *entry;
389
390         if (sgp != SGP_WRITE &&
391             ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode))
392                 return ERR_PTR(-EINVAL);
393
394         while (!(entry = shmem_swp_entry(info, index, &page))) {
395                 if (sgp == SGP_READ)
396                         return shmem_swp_map(ZERO_PAGE(0));
397                 /*
398                  * Test free_blocks against 1 not 0, since we have 1 data
399                  * page (and perhaps indirect index pages) yet to allocate:
400                  * a waste to allocate index if we cannot allocate data.
401                  */
402                 if (sbinfo->max_blocks) {
403                         spin_lock(&sbinfo->stat_lock);
404                         if (sbinfo->free_blocks <= 1) {
405                                 spin_unlock(&sbinfo->stat_lock);
406                                 return ERR_PTR(-ENOSPC);
407                         }
408                         sbinfo->free_blocks--;
409                         inode->i_blocks += BLOCKS_PER_PAGE;
410                         spin_unlock(&sbinfo->stat_lock);
411                 }
412
413                 spin_unlock(&info->lock);
414                 page = shmem_dir_alloc(mapping_gfp_mask(inode->i_mapping));
415                 if (page)
416                         set_page_private(page, 0);
417                 spin_lock(&info->lock);
418
419                 if (!page) {
420                         shmem_free_blocks(inode, 1);
421                         return ERR_PTR(-ENOMEM);
422                 }
423                 if (sgp != SGP_WRITE &&
424                     ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
425                         entry = ERR_PTR(-EINVAL);
426                         break;
427                 }
428                 if (info->next_index <= index)
429                         info->next_index = index + 1;
430         }
431         if (page) {
432                 /* another task gave its page, or truncated the file */
433                 shmem_free_blocks(inode, 1);
434                 shmem_dir_free(page);
435         }
436         if (info->next_index <= index && !IS_ERR(entry))
437                 info->next_index = index + 1;
438         return entry;
439 }
440
441 /**
442  * shmem_free_swp - free some swap entries in a directory
443  * @dir:        pointer to the directory
444  * @edir:       pointer after last entry of the directory
445  * @punch_lock: pointer to spinlock when needed for the holepunch case
446  */
447 static int shmem_free_swp(swp_entry_t *dir, swp_entry_t *edir,
448                                                 spinlock_t *punch_lock)
449 {
450         spinlock_t *punch_unlock = NULL;
451         swp_entry_t *ptr;
452         int freed = 0;
453
454         for (ptr = dir; ptr < edir; ptr++) {
455                 if (ptr->val) {
456                         if (unlikely(punch_lock)) {
457                                 punch_unlock = punch_lock;
458                                 punch_lock = NULL;
459                                 spin_lock(punch_unlock);
460                                 if (!ptr->val)
461                                         continue;
462                         }
463                         free_swap_and_cache(*ptr);
464                         *ptr = (swp_entry_t){0};
465                         freed++;
466                 }
467         }
468         if (punch_unlock)
469                 spin_unlock(punch_unlock);
470         return freed;
471 }
472
473 static int shmem_map_and_free_swp(struct page *subdir, int offset,
474                 int limit, struct page ***dir, spinlock_t *punch_lock)
475 {
476         swp_entry_t *ptr;
477         int freed = 0;
478
479         ptr = shmem_swp_map(subdir);
480         for (; offset < limit; offset += LATENCY_LIMIT) {
481                 int size = limit - offset;
482                 if (size > LATENCY_LIMIT)
483                         size = LATENCY_LIMIT;
484                 freed += shmem_free_swp(ptr+offset, ptr+offset+size,
485                                                         punch_lock);
486                 if (need_resched()) {
487                         shmem_swp_unmap(ptr);
488                         if (*dir) {
489                                 shmem_dir_unmap(*dir);
490                                 *dir = NULL;
491                         }
492                         cond_resched();
493                         ptr = shmem_swp_map(subdir);
494                 }
495         }
496         shmem_swp_unmap(ptr);
497         return freed;
498 }
499
500 static void shmem_free_pages(struct list_head *next)
501 {
502         struct page *page;
503         int freed = 0;
504
505         do {
506                 page = container_of(next, struct page, lru);
507                 next = next->next;
508                 shmem_dir_free(page);
509                 freed++;
510                 if (freed >= LATENCY_LIMIT) {
511                         cond_resched();
512                         freed = 0;
513                 }
514         } while (next);
515 }
516
517 static void shmem_truncate_range(struct inode *inode, loff_t start, loff_t end)
518 {
519         struct shmem_inode_info *info = SHMEM_I(inode);
520         unsigned long idx;
521         unsigned long size;
522         unsigned long limit;
523         unsigned long stage;
524         unsigned long diroff;
525         struct page **dir;
526         struct page *topdir;
527         struct page *middir;
528         struct page *subdir;
529         swp_entry_t *ptr;
530         LIST_HEAD(pages_to_free);
531         long nr_pages_to_free = 0;
532         long nr_swaps_freed = 0;
533         int offset;
534         int freed;
535         int punch_hole;
536         spinlock_t *needs_lock;
537         spinlock_t *punch_lock;
538         unsigned long upper_limit;
539
540         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
541         idx = (start + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
542         if (idx >= info->next_index)
543                 return;
544
545         spin_lock(&info->lock);
546         info->flags |= SHMEM_TRUNCATE;
547         if (likely(end == (loff_t) -1)) {
548                 limit = info->next_index;
549                 upper_limit = SHMEM_MAX_INDEX;
550                 info->next_index = idx;
551                 needs_lock = NULL;
552                 punch_hole = 0;
553         } else {
554                 if (end + 1 >= inode->i_size) { /* we may free a little more */
555                         limit = (inode->i_size + PAGE_CACHE_SIZE - 1) >>
556                                                         PAGE_CACHE_SHIFT;
557                         upper_limit = SHMEM_MAX_INDEX;
558                 } else {
559                         limit = (end + 1) >> PAGE_CACHE_SHIFT;
560                         upper_limit = limit;
561                 }
562                 needs_lock = &info->lock;
563                 punch_hole = 1;
564         }
565
566         topdir = info->i_indirect;
567         if (topdir && idx <= SHMEM_NR_DIRECT && !punch_hole) {
568                 info->i_indirect = NULL;
569                 nr_pages_to_free++;
570                 list_add(&topdir->lru, &pages_to_free);
571         }
572         spin_unlock(&info->lock);
573
574         if (info->swapped && idx < SHMEM_NR_DIRECT) {
575                 ptr = info->i_direct;
576                 size = limit;
577                 if (size > SHMEM_NR_DIRECT)
578                         size = SHMEM_NR_DIRECT;
579                 nr_swaps_freed = shmem_free_swp(ptr+idx, ptr+size, needs_lock);
580         }
581
582         /*
583          * If there are no indirect blocks or we are punching a hole
584          * below indirect blocks, nothing to be done.
585          */
586         if (!topdir || limit <= SHMEM_NR_DIRECT)
587                 goto done2;
588
589         /*
590          * The truncation case has already dropped info->lock, and we're safe
591          * because i_size and next_index have already been lowered, preventing
592          * access beyond.  But in the punch_hole case, we still need to take
593          * the lock when updating the swap directory, because there might be
594          * racing accesses by shmem_getpage(SGP_CACHE), shmem_unuse_inode or
595          * shmem_writepage.  However, whenever we find we can remove a whole
596          * directory page (not at the misaligned start or end of the range),
597          * we first NULLify its pointer in the level above, and then have no
598          * need to take the lock when updating its contents: needs_lock and
599          * punch_lock (either pointing to info->lock or NULL) manage this.
600          */
601
602         upper_limit -= SHMEM_NR_DIRECT;
603         limit -= SHMEM_NR_DIRECT;
604         idx = (idx > SHMEM_NR_DIRECT)? (idx - SHMEM_NR_DIRECT): 0;
605         offset = idx % ENTRIES_PER_PAGE;
606         idx -= offset;
607
608         dir = shmem_dir_map(topdir);
609         stage = ENTRIES_PER_PAGEPAGE/2;
610         if (idx < ENTRIES_PER_PAGEPAGE/2) {
611                 middir = topdir;
612                 diroff = idx/ENTRIES_PER_PAGE;
613         } else {
614                 dir += ENTRIES_PER_PAGE/2;
615                 dir += (idx - ENTRIES_PER_PAGEPAGE/2)/ENTRIES_PER_PAGEPAGE;
616                 while (stage <= idx)
617                         stage += ENTRIES_PER_PAGEPAGE;
618                 middir = *dir;
619                 if (*dir) {
620                         diroff = ((idx - ENTRIES_PER_PAGEPAGE/2) %
621                                 ENTRIES_PER_PAGEPAGE) / ENTRIES_PER_PAGE;
622                         if (!diroff && !offset && upper_limit >= stage) {
623                                 if (needs_lock) {
624                                         spin_lock(needs_lock);
625                                         *dir = NULL;
626                                         spin_unlock(needs_lock);
627                                         needs_lock = NULL;
628                                 } else
629                                         *dir = NULL;
630                                 nr_pages_to_free++;
631                                 list_add(&middir->lru, &pages_to_free);
632                         }
633                         shmem_dir_unmap(dir);
634                         dir = shmem_dir_map(middir);
635                 } else {
636                         diroff = 0;
637                         offset = 0;
638                         idx = stage;
639                 }
640         }
641
642         for (; idx < limit; idx += ENTRIES_PER_PAGE, diroff++) {
643                 if (unlikely(idx == stage)) {
644                         shmem_dir_unmap(dir);
645                         dir = shmem_dir_map(topdir) +
646                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
647                         while (!*dir) {
648                                 dir++;
649                                 idx += ENTRIES_PER_PAGEPAGE;
650                                 if (idx >= limit)
651                                         goto done1;
652                         }
653                         stage = idx + ENTRIES_PER_PAGEPAGE;
654                         middir = *dir;
655                         if (punch_hole)
656                                 needs_lock = &info->lock;
657                         if (upper_limit >= stage) {
658                                 if (needs_lock) {
659                                         spin_lock(needs_lock);
660                                         *dir = NULL;
661                                         spin_unlock(needs_lock);
662                                         needs_lock = NULL;
663                                 } else
664                                         *dir = NULL;
665                                 nr_pages_to_free++;
666                                 list_add(&middir->lru, &pages_to_free);
667                         }
668                         shmem_dir_unmap(dir);
669                         cond_resched();
670                         dir = shmem_dir_map(middir);
671                         diroff = 0;
672                 }
673                 punch_lock = needs_lock;
674                 subdir = dir[diroff];
675                 if (subdir && !offset && upper_limit-idx >= ENTRIES_PER_PAGE) {
676                         if (needs_lock) {
677                                 spin_lock(needs_lock);
678                                 dir[diroff] = NULL;
679                                 spin_unlock(needs_lock);
680                                 punch_lock = NULL;
681                         } else
682                                 dir[diroff] = NULL;
683                         nr_pages_to_free++;
684                         list_add(&subdir->lru, &pages_to_free);
685                 }
686                 if (subdir && page_private(subdir) /* has swap entries */) {
687                         size = limit - idx;
688                         if (size > ENTRIES_PER_PAGE)
689                                 size = ENTRIES_PER_PAGE;
690                         freed = shmem_map_and_free_swp(subdir,
691                                         offset, size, &dir, punch_lock);
692                         if (!dir)
693                                 dir = shmem_dir_map(middir);
694                         nr_swaps_freed += freed;
695                         if (offset || punch_lock) {
696                                 spin_lock(&info->lock);
697                                 set_page_private(subdir,
698                                         page_private(subdir) - freed);
699                                 spin_unlock(&info->lock);
700                         } else
701                                 BUG_ON(page_private(subdir) != freed);
702                 }
703                 offset = 0;
704         }
705 done1:
706         shmem_dir_unmap(dir);
707 done2:
708         if (inode->i_mapping->nrpages && (info->flags & SHMEM_PAGEIN)) {
709                 /*
710                  * Call truncate_inode_pages again: racing shmem_unuse_inode
711                  * may have swizzled a page in from swap since vmtruncate or
712                  * generic_delete_inode did it, before we lowered next_index.
713                  * Also, though shmem_getpage checks i_size before adding to
714                  * cache, no recheck after: so fix the narrow window there too.
715                  *
716                  * Recalling truncate_inode_pages_range and unmap_mapping_range
717                  * every time for punch_hole (which never got a chance to clear
718                  * SHMEM_PAGEIN at the start of vmtruncate_range) is expensive,
719                  * yet hardly ever necessary: try to optimize them out later.
720                  */
721                 truncate_inode_pages_range(inode->i_mapping, start, end);
722                 if (punch_hole)
723                         unmap_mapping_range(inode->i_mapping, start,
724                                                         end - start, 1);
725         }
726
727         spin_lock(&info->lock);
728         info->flags &= ~SHMEM_TRUNCATE;
729         info->swapped -= nr_swaps_freed;
730         if (nr_pages_to_free)
731                 shmem_free_blocks(inode, nr_pages_to_free);
732         shmem_recalc_inode(inode);
733         spin_unlock(&info->lock);
734
735         /*
736          * Empty swap vector directory pages to be freed?
737          */
738         if (!list_empty(&pages_to_free)) {
739                 pages_to_free.prev->next = NULL;
740                 shmem_free_pages(pages_to_free.next);
741         }
742 }
743
744 static void shmem_truncate(struct inode *inode)
745 {
746         shmem_truncate_range(inode, inode->i_size, (loff_t)-1);
747 }
748
749 static int shmem_notify_change(struct dentry *dentry, struct iattr *attr)
750 {
751         struct inode *inode = dentry->d_inode;
752         struct page *page = NULL;
753         int error;
754
755         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
756                 if (attr->ia_size < inode->i_size) {
757                         /*
758                          * If truncating down to a partial page, then
759                          * if that page is already allocated, hold it
760                          * in memory until the truncation is over, so
761                          * truncate_partial_page cannnot miss it were
762                          * it assigned to swap.
763                          */
764                         if (attr->ia_size & (PAGE_CACHE_SIZE-1)) {
765                                 (void) shmem_getpage(inode,
766                                         attr->ia_size>>PAGE_CACHE_SHIFT,
767                                                 &page, SGP_READ, NULL);
768                                 if (page)
769                                         unlock_page(page);
770                         }
771                         /*
772                          * Reset SHMEM_PAGEIN flag so that shmem_truncate can
773                          * detect if any pages might have been added to cache
774                          * after truncate_inode_pages.  But we needn't bother
775                          * if it's being fully truncated to zero-length: the
776                          * nrpages check is efficient enough in that case.
777                          */
778                         if (attr->ia_size) {
779                                 struct shmem_inode_info *info = SHMEM_I(inode);
780                                 spin_lock(&info->lock);
781                                 info->flags &= ~SHMEM_PAGEIN;
782                                 spin_unlock(&info->lock);
783                         }
784                 }
785         }
786
787         error = inode_change_ok(inode, attr);
788         if (!error)
789                 error = inode_setattr(inode, attr);
790 #ifdef CONFIG_TMPFS_POSIX_ACL
791         if (!error && (attr->ia_valid & ATTR_MODE))
792                 error = generic_acl_chmod(inode, &shmem_acl_ops);
793 #endif
794         if (page)
795                 page_cache_release(page);
796         return error;
797 }
798
799 static void shmem_delete_inode(struct inode *inode)
800 {
801         struct shmem_inode_info *info = SHMEM_I(inode);
802
803         if (inode->i_op->truncate == shmem_truncate) {
804                 truncate_inode_pages(inode->i_mapping, 0);
805                 shmem_unacct_size(info->flags, inode->i_size);
806                 inode->i_size = 0;
807                 shmem_truncate(inode);
808                 if (!list_empty(&info->swaplist)) {
809                         mutex_lock(&shmem_swaplist_mutex);
810                         list_del_init(&info->swaplist);
811                         mutex_unlock(&shmem_swaplist_mutex);
812                 }
813         }
814         BUG_ON(inode->i_blocks);
815         shmem_free_inode(inode->i_sb);
816         clear_inode(inode);
817 }
818
819 static inline int shmem_find_swp(swp_entry_t entry, swp_entry_t *dir, swp_entry_t *edir)
820 {
821         swp_entry_t *ptr;
822
823         for (ptr = dir; ptr < edir; ptr++) {
824                 if (ptr->val == entry.val)
825                         return ptr - dir;
826         }
827         return -1;
828 }
829
830 static int shmem_unuse_inode(struct shmem_inode_info *info, swp_entry_t entry, struct page *page)
831 {
832         struct inode *inode;
833         unsigned long idx;
834         unsigned long size;
835         unsigned long limit;
836         unsigned long stage;
837         struct page **dir;
838         struct page *subdir;
839         swp_entry_t *ptr;
840         int offset;
841         int error;
842
843         idx = 0;
844         ptr = info->i_direct;
845         spin_lock(&info->lock);
846         if (!info->swapped) {
847                 list_del_init(&info->swaplist);
848                 goto lost2;
849         }
850         limit = info->next_index;
851         size = limit;
852         if (size > SHMEM_NR_DIRECT)
853                 size = SHMEM_NR_DIRECT;
854         offset = shmem_find_swp(entry, ptr, ptr+size);
855         if (offset >= 0)
856                 goto found;
857         if (!info->i_indirect)
858                 goto lost2;
859
860         dir = shmem_dir_map(info->i_indirect);
861         stage = SHMEM_NR_DIRECT + ENTRIES_PER_PAGEPAGE/2;
862
863         for (idx = SHMEM_NR_DIRECT; idx < limit; idx += ENTRIES_PER_PAGE, dir++) {
864                 if (unlikely(idx == stage)) {
865                         shmem_dir_unmap(dir-1);
866                         if (cond_resched_lock(&info->lock)) {
867                                 /* check it has not been truncated */
868                                 if (limit > info->next_index) {
869                                         limit = info->next_index;
870                                         if (idx >= limit)
871                                                 goto lost2;
872                                 }
873                         }
874                         dir = shmem_dir_map(info->i_indirect) +
875                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
876                         while (!*dir) {
877                                 dir++;
878                                 idx += ENTRIES_PER_PAGEPAGE;
879                                 if (idx >= limit)
880                                         goto lost1;
881                         }
882                         stage = idx + ENTRIES_PER_PAGEPAGE;
883                         subdir = *dir;
884                         shmem_dir_unmap(dir);
885                         dir = shmem_dir_map(subdir);
886                 }
887                 subdir = *dir;
888                 if (subdir && page_private(subdir)) {
889                         ptr = shmem_swp_map(subdir);
890                         size = limit - idx;
891                         if (size > ENTRIES_PER_PAGE)
892                                 size = ENTRIES_PER_PAGE;
893                         offset = shmem_find_swp(entry, ptr, ptr+size);
894                         shmem_swp_unmap(ptr);
895                         if (offset >= 0) {
896                                 shmem_dir_unmap(dir);
897                                 goto found;
898                         }
899                 }
900         }
901 lost1:
902         shmem_dir_unmap(dir-1);
903 lost2:
904         spin_unlock(&info->lock);
905         return 0;
906 found:
907         idx += offset;
908         inode = igrab(&info->vfs_inode);
909         spin_unlock(&info->lock);
910
911         /*
912          * Move _head_ to start search for next from here.
913          * But be careful: shmem_delete_inode checks list_empty without taking
914          * mutex, and there's an instant in list_move_tail when info->swaplist
915          * would appear empty, if it were the only one on shmem_swaplist.  We
916          * could avoid doing it if inode NULL; or use this minor optimization.
917          */
918         if (shmem_swaplist.next != &info->swaplist)
919                 list_move_tail(&shmem_swaplist, &info->swaplist);
920         mutex_unlock(&shmem_swaplist_mutex);
921
922         error = 1;
923         if (!inode)
924                 goto out;
925         /* Precharge page using GFP_KERNEL while we can wait */
926         error = mem_cgroup_cache_charge(page, current->mm, GFP_KERNEL);
927         if (error)
928                 goto out;
929         error = radix_tree_preload(GFP_KERNEL);
930         if (error) {
931                 mem_cgroup_uncharge_cache_page(page);
932                 goto out;
933         }
934         error = 1;
935
936         spin_lock(&info->lock);
937         ptr = shmem_swp_entry(info, idx, NULL);
938         if (ptr && ptr->val == entry.val) {
939                 error = add_to_page_cache_locked(page, inode->i_mapping,
940                                                 idx, GFP_NOWAIT);
941                 /* does mem_cgroup_uncharge_cache_page on error */
942         } else  /* we must compensate for our precharge above */
943                 mem_cgroup_uncharge_cache_page(page);
944
945         if (error == -EEXIST) {
946                 struct page *filepage = find_get_page(inode->i_mapping, idx);
947                 error = 1;
948                 if (filepage) {
949                         /*
950                          * There might be a more uptodate page coming down
951                          * from a stacked writepage: forget our swappage if so.
952                          */
953                         if (PageUptodate(filepage))
954                                 error = 0;
955                         page_cache_release(filepage);
956                 }
957         }
958         if (!error) {
959                 delete_from_swap_cache(page);
960                 set_page_dirty(page);
961                 info->flags |= SHMEM_PAGEIN;
962                 shmem_swp_set(info, ptr, 0);
963                 swap_free(entry);
964                 error = 1;      /* not an error, but entry was found */
965         }
966         if (ptr)
967                 shmem_swp_unmap(ptr);
968         spin_unlock(&info->lock);
969         radix_tree_preload_end();
970 out:
971         unlock_page(page);
972         page_cache_release(page);
973         iput(inode);            /* allows for NULL */
974         return error;
975 }
976
977 /*
978  * shmem_unuse() search for an eventually swapped out shmem page.
979  */
980 int shmem_unuse(swp_entry_t entry, struct page *page)
981 {
982         struct list_head *p, *next;
983         struct shmem_inode_info *info;
984         int found = 0;
985
986         mutex_lock(&shmem_swaplist_mutex);
987         list_for_each_safe(p, next, &shmem_swaplist) {
988                 info = list_entry(p, struct shmem_inode_info, swaplist);
989                 found = shmem_unuse_inode(info, entry, page);
990                 cond_resched();
991                 if (found)
992                         goto out;
993         }
994         mutex_unlock(&shmem_swaplist_mutex);
995 out:    return found;   /* 0 or 1 or -ENOMEM */
996 }
997
998 /*
999  * Move the page from the page cache to the swap cache.
1000  */
1001 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1002 {
1003         struct shmem_inode_info *info;
1004         swp_entry_t *entry, swap;
1005         struct address_space *mapping;
1006         unsigned long index;
1007         struct inode *inode;
1008
1009         BUG_ON(!PageLocked(page));
1010         mapping = page->mapping;
1011         index = page->index;
1012         inode = mapping->host;
1013         info = SHMEM_I(inode);
1014         if (info->flags & VM_LOCKED)
1015                 goto redirty;
1016         if (!total_swap_pages)
1017                 goto redirty;
1018
1019         /*
1020          * shmem_backing_dev_info's capabilities prevent regular writeback or
1021          * sync from ever calling shmem_writepage; but a stacking filesystem
1022          * may use the ->writepage of its underlying filesystem, in which case
1023          * tmpfs should write out to swap only in response to memory pressure,
1024          * and not for pdflush or sync.  However, in those cases, we do still
1025          * want to check if there's a redundant swappage to be discarded.
1026          */
1027         if (wbc->for_reclaim)
1028                 swap = get_swap_page();
1029         else
1030                 swap.val = 0;
1031
1032         spin_lock(&info->lock);
1033         if (index >= info->next_index) {
1034                 BUG_ON(!(info->flags & SHMEM_TRUNCATE));
1035                 goto unlock;
1036         }
1037         entry = shmem_swp_entry(info, index, NULL);
1038         if (entry->val) {
1039                 /*
1040                  * The more uptodate page coming down from a stacked
1041                  * writepage should replace our old swappage.
1042                  */
1043                 free_swap_and_cache(*entry);
1044                 shmem_swp_set(info, entry, 0);
1045         }
1046         shmem_recalc_inode(inode);
1047
1048         if (swap.val && add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
1049                 remove_from_page_cache(page);
1050                 shmem_swp_set(info, entry, swap.val);
1051                 shmem_swp_unmap(entry);
1052                 if (list_empty(&info->swaplist))
1053                         inode = igrab(inode);
1054                 else
1055                         inode = NULL;
1056                 spin_unlock(&info->lock);
1057                 swap_duplicate(swap);
1058                 BUG_ON(page_mapped(page));
1059                 page_cache_release(page);       /* pagecache ref */
1060                 set_page_dirty(page);
1061                 unlock_page(page);
1062                 if (inode) {
1063                         mutex_lock(&shmem_swaplist_mutex);
1064                         /* move instead of add in case we're racing */
1065                         list_move_tail(&info->swaplist, &shmem_swaplist);
1066                         mutex_unlock(&shmem_swaplist_mutex);
1067                         iput(inode);
1068                 }
1069                 return 0;
1070         }
1071
1072         shmem_swp_unmap(entry);
1073 unlock:
1074         spin_unlock(&info->lock);
1075         swap_free(swap);
1076 redirty:
1077         set_page_dirty(page);
1078         if (wbc->for_reclaim)
1079                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
1080         unlock_page(page);
1081         return 0;
1082 }
1083
1084 #ifdef CONFIG_NUMA
1085 #ifdef CONFIG_TMPFS
1086 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1087 {
1088         char buffer[64];
1089
1090         if (!mpol || mpol->mode == MPOL_DEFAULT)
1091                 return;         /* show nothing */
1092
1093         mpol_to_str(buffer, sizeof(buffer), mpol, 1);
1094
1095         seq_printf(seq, ",mpol=%s", buffer);
1096 }
1097
1098 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1099 {
1100         struct mempolicy *mpol = NULL;
1101         if (sbinfo->mpol) {
1102                 spin_lock(&sbinfo->stat_lock);  /* prevent replace/use races */
1103                 mpol = sbinfo->mpol;
1104                 mpol_get(mpol);
1105                 spin_unlock(&sbinfo->stat_lock);
1106         }
1107         return mpol;
1108 }
1109 #endif /* CONFIG_TMPFS */
1110
1111 static struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1112                         struct shmem_inode_info *info, unsigned long idx)
1113 {
1114         struct mempolicy mpol, *spol;
1115         struct vm_area_struct pvma;
1116         struct page *page;
1117
1118         spol = mpol_cond_copy(&mpol,
1119                                 mpol_shared_policy_lookup(&info->policy, idx));
1120
1121         /* Create a pseudo vma that just contains the policy */
1122         pvma.vm_start = 0;
1123         pvma.vm_pgoff = idx;
1124         pvma.vm_ops = NULL;
1125         pvma.vm_policy = spol;
1126         page = swapin_readahead(entry, gfp, &pvma, 0);
1127         return page;
1128 }
1129
1130 static struct page *shmem_alloc_page(gfp_t gfp,
1131                         struct shmem_inode_info *info, unsigned long idx)
1132 {
1133         struct vm_area_struct pvma;
1134
1135         /* Create a pseudo vma that just contains the policy */
1136         pvma.vm_start = 0;
1137         pvma.vm_pgoff = idx;
1138         pvma.vm_ops = NULL;
1139         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, idx);
1140
1141         /*
1142          * alloc_page_vma() will drop the shared policy reference
1143          */
1144         return alloc_page_vma(gfp, &pvma, 0);
1145 }
1146 #else /* !CONFIG_NUMA */
1147 #ifdef CONFIG_TMPFS
1148 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *p)
1149 {
1150 }
1151 #endif /* CONFIG_TMPFS */
1152
1153 static inline struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1154                         struct shmem_inode_info *info, unsigned long idx)
1155 {
1156         return swapin_readahead(entry, gfp, NULL, 0);
1157 }
1158
1159 static inline struct page *shmem_alloc_page(gfp_t gfp,
1160                         struct shmem_inode_info *info, unsigned long idx)
1161 {
1162         return alloc_page(gfp);
1163 }
1164 #endif /* CONFIG_NUMA */
1165
1166 #if !defined(CONFIG_NUMA) || !defined(CONFIG_TMPFS)
1167 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1168 {
1169         return NULL;
1170 }
1171 #endif
1172
1173 /*
1174  * shmem_getpage - either get the page from swap or allocate a new one
1175  *
1176  * If we allocate a new one we do not mark it dirty. That's up to the
1177  * vm. If we swap it in we mark it dirty since we also free the swap
1178  * entry since a page cannot live in both the swap and page cache
1179  */
1180 static int shmem_getpage(struct inode *inode, unsigned long idx,
1181                         struct page **pagep, enum sgp_type sgp, int *type)
1182 {
1183         struct address_space *mapping = inode->i_mapping;
1184         struct shmem_inode_info *info = SHMEM_I(inode);
1185         struct shmem_sb_info *sbinfo;
1186         struct page *filepage = *pagep;
1187         struct page *swappage;
1188         swp_entry_t *entry;
1189         swp_entry_t swap;
1190         gfp_t gfp;
1191         int error;
1192
1193         if (idx >= SHMEM_MAX_INDEX)
1194                 return -EFBIG;
1195
1196         if (type)
1197                 *type = 0;
1198
1199         /*
1200          * Normally, filepage is NULL on entry, and either found
1201          * uptodate immediately, or allocated and zeroed, or read
1202          * in under swappage, which is then assigned to filepage.
1203          * But shmem_readpage (required for splice) passes in a locked
1204          * filepage, which may be found not uptodate by other callers
1205          * too, and may need to be copied from the swappage read in.
1206          */
1207 repeat:
1208         if (!filepage)
1209                 filepage = find_lock_page(mapping, idx);
1210         if (filepage && PageUptodate(filepage))
1211                 goto done;
1212         error = 0;
1213         gfp = mapping_gfp_mask(mapping);
1214         if (!filepage) {
1215                 /*
1216                  * Try to preload while we can wait, to not make a habit of
1217                  * draining atomic reserves; but don't latch on to this cpu.
1218                  */
1219                 error = radix_tree_preload(gfp & ~__GFP_HIGHMEM);
1220                 if (error)
1221                         goto failed;
1222                 radix_tree_preload_end();
1223         }
1224
1225         spin_lock(&info->lock);
1226         shmem_recalc_inode(inode);
1227         entry = shmem_swp_alloc(info, idx, sgp);
1228         if (IS_ERR(entry)) {
1229                 spin_unlock(&info->lock);
1230                 error = PTR_ERR(entry);
1231                 goto failed;
1232         }
1233         swap = *entry;
1234
1235         if (swap.val) {
1236                 /* Look it up and read it in.. */
1237                 swappage = lookup_swap_cache(swap);
1238                 if (!swappage) {
1239                         shmem_swp_unmap(entry);
1240                         /* here we actually do the io */
1241                         if (type && !(*type & VM_FAULT_MAJOR)) {
1242                                 __count_vm_event(PGMAJFAULT);
1243                                 *type |= VM_FAULT_MAJOR;
1244                         }
1245                         spin_unlock(&info->lock);
1246                         swappage = shmem_swapin(swap, gfp, info, idx);
1247                         if (!swappage) {
1248                                 spin_lock(&info->lock);
1249                                 entry = shmem_swp_alloc(info, idx, sgp);
1250                                 if (IS_ERR(entry))
1251                                         error = PTR_ERR(entry);
1252                                 else {
1253                                         if (entry->val == swap.val)
1254                                                 error = -ENOMEM;
1255                                         shmem_swp_unmap(entry);
1256                                 }
1257                                 spin_unlock(&info->lock);
1258                                 if (error)
1259                                         goto failed;
1260                                 goto repeat;
1261                         }
1262                         wait_on_page_locked(swappage);
1263                         page_cache_release(swappage);
1264                         goto repeat;
1265                 }
1266
1267                 /* We have to do this with page locked to prevent races */
1268                 if (TestSetPageLocked(swappage)) {
1269                         shmem_swp_unmap(entry);
1270                         spin_unlock(&info->lock);
1271                         wait_on_page_locked(swappage);
1272                         page_cache_release(swappage);
1273                         goto repeat;
1274                 }
1275                 if (PageWriteback(swappage)) {
1276                         shmem_swp_unmap(entry);
1277                         spin_unlock(&info->lock);
1278                         wait_on_page_writeback(swappage);
1279                         unlock_page(swappage);
1280                         page_cache_release(swappage);
1281                         goto repeat;
1282                 }
1283                 if (!PageUptodate(swappage)) {
1284                         shmem_swp_unmap(entry);
1285                         spin_unlock(&info->lock);
1286                         unlock_page(swappage);
1287                         page_cache_release(swappage);
1288                         error = -EIO;
1289                         goto failed;
1290                 }
1291
1292                 if (filepage) {
1293                         shmem_swp_set(info, entry, 0);
1294                         shmem_swp_unmap(entry);
1295                         delete_from_swap_cache(swappage);
1296                         spin_unlock(&info->lock);
1297                         copy_highpage(filepage, swappage);
1298                         unlock_page(swappage);
1299                         page_cache_release(swappage);
1300                         flush_dcache_page(filepage);
1301                         SetPageUptodate(filepage);
1302                         set_page_dirty(filepage);
1303                         swap_free(swap);
1304                 } else if (!(error = add_to_page_cache_locked(swappage, mapping,
1305                                         idx, GFP_NOWAIT))) {
1306                         info->flags |= SHMEM_PAGEIN;
1307                         shmem_swp_set(info, entry, 0);
1308                         shmem_swp_unmap(entry);
1309                         delete_from_swap_cache(swappage);
1310                         spin_unlock(&info->lock);
1311                         filepage = swappage;
1312                         set_page_dirty(filepage);
1313                         swap_free(swap);
1314                 } else {
1315                         shmem_swp_unmap(entry);
1316                         spin_unlock(&info->lock);
1317                         unlock_page(swappage);
1318                         page_cache_release(swappage);
1319                         if (error == -ENOMEM) {
1320                                 /* allow reclaim from this memory cgroup */
1321                                 error = mem_cgroup_shrink_usage(current->mm,
1322                                                                 gfp);
1323                                 if (error)
1324                                         goto failed;
1325                         }
1326                         goto repeat;
1327                 }
1328         } else if (sgp == SGP_READ && !filepage) {
1329                 shmem_swp_unmap(entry);
1330                 filepage = find_get_page(mapping, idx);
1331                 if (filepage &&
1332                     (!PageUptodate(filepage) || TestSetPageLocked(filepage))) {
1333                         spin_unlock(&info->lock);
1334                         wait_on_page_locked(filepage);
1335                         page_cache_release(filepage);
1336                         filepage = NULL;
1337                         goto repeat;
1338                 }
1339                 spin_unlock(&info->lock);
1340         } else {
1341                 shmem_swp_unmap(entry);
1342                 sbinfo = SHMEM_SB(inode->i_sb);
1343                 if (sbinfo->max_blocks) {
1344                         spin_lock(&sbinfo->stat_lock);
1345                         if (sbinfo->free_blocks == 0 ||
1346                             shmem_acct_block(info->flags)) {
1347                                 spin_unlock(&sbinfo->stat_lock);
1348                                 spin_unlock(&info->lock);
1349                                 error = -ENOSPC;
1350                                 goto failed;
1351                         }
1352                         sbinfo->free_blocks--;
1353                         inode->i_blocks += BLOCKS_PER_PAGE;
1354                         spin_unlock(&sbinfo->stat_lock);
1355                 } else if (shmem_acct_block(info->flags)) {
1356                         spin_unlock(&info->lock);
1357                         error = -ENOSPC;
1358                         goto failed;
1359                 }
1360
1361                 if (!filepage) {
1362                         int ret;
1363
1364                         spin_unlock(&info->lock);
1365                         filepage = shmem_alloc_page(gfp, info, idx);
1366                         if (!filepage) {
1367                                 shmem_unacct_blocks(info->flags, 1);
1368                                 shmem_free_blocks(inode, 1);
1369                                 error = -ENOMEM;
1370                                 goto failed;
1371                         }
1372
1373                         /* Precharge page while we can wait, compensate after */
1374                         error = mem_cgroup_cache_charge(filepage, current->mm,
1375                                                         gfp & ~__GFP_HIGHMEM);
1376                         if (error) {
1377                                 page_cache_release(filepage);
1378                                 shmem_unacct_blocks(info->flags, 1);
1379                                 shmem_free_blocks(inode, 1);
1380                                 filepage = NULL;
1381                                 goto failed;
1382                         }
1383
1384                         spin_lock(&info->lock);
1385                         entry = shmem_swp_alloc(info, idx, sgp);
1386                         if (IS_ERR(entry))
1387                                 error = PTR_ERR(entry);
1388                         else {
1389                                 swap = *entry;
1390                                 shmem_swp_unmap(entry);
1391                         }
1392                         ret = error || swap.val;
1393                         if (ret)
1394                                 mem_cgroup_uncharge_cache_page(filepage);
1395                         else
1396                                 ret = add_to_page_cache_lru(filepage, mapping,
1397                                                 idx, GFP_NOWAIT);
1398                         /*
1399                          * At add_to_page_cache_lru() failure, uncharge will
1400                          * be done automatically.
1401                          */
1402                         if (ret) {
1403                                 spin_unlock(&info->lock);
1404                                 page_cache_release(filepage);
1405                                 shmem_unacct_blocks(info->flags, 1);
1406                                 shmem_free_blocks(inode, 1);
1407                                 filepage = NULL;
1408                                 if (error)
1409                                         goto failed;
1410                                 goto repeat;
1411                         }
1412                         info->flags |= SHMEM_PAGEIN;
1413                 }
1414
1415                 info->alloced++;
1416                 spin_unlock(&info->lock);
1417                 clear_highpage(filepage);
1418                 flush_dcache_page(filepage);
1419                 SetPageUptodate(filepage);
1420                 if (sgp == SGP_DIRTY)
1421                         set_page_dirty(filepage);
1422         }
1423 done:
1424         *pagep = filepage;
1425         return 0;
1426
1427 failed:
1428         if (*pagep != filepage) {
1429                 unlock_page(filepage);
1430                 page_cache_release(filepage);
1431         }
1432         return error;
1433 }
1434
1435 static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1436 {
1437         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1438         int error;
1439         int ret;
1440
1441         if (((loff_t)vmf->pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
1442                 return VM_FAULT_SIGBUS;
1443
1444         error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_CACHE, &ret);
1445         if (error)
1446                 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1447
1448         mark_page_accessed(vmf->page);
1449         return ret | VM_FAULT_LOCKED;
1450 }
1451
1452 #ifdef CONFIG_NUMA
1453 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *new)
1454 {
1455         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1456         return mpol_set_shared_policy(&SHMEM_I(i)->policy, vma, new);
1457 }
1458
1459 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
1460                                           unsigned long addr)
1461 {
1462         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1463         unsigned long idx;
1464
1465         idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1466         return mpol_shared_policy_lookup(&SHMEM_I(i)->policy, idx);
1467 }
1468 #endif
1469
1470 int shmem_lock(struct file *file, int lock, struct user_struct *user)
1471 {
1472         struct inode *inode = file->f_path.dentry->d_inode;
1473         struct shmem_inode_info *info = SHMEM_I(inode);
1474         int retval = -ENOMEM;
1475
1476         spin_lock(&info->lock);
1477         if (lock && !(info->flags & VM_LOCKED)) {
1478                 if (!user_shm_lock(inode->i_size, user))
1479                         goto out_nomem;
1480                 info->flags |= VM_LOCKED;
1481         }
1482         if (!lock && (info->flags & VM_LOCKED) && user) {
1483                 user_shm_unlock(inode->i_size, user);
1484                 info->flags &= ~VM_LOCKED;
1485         }
1486         retval = 0;
1487 out_nomem:
1488         spin_unlock(&info->lock);
1489         return retval;
1490 }
1491
1492 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1493 {
1494         file_accessed(file);
1495         vma->vm_ops = &shmem_vm_ops;
1496         vma->vm_flags |= VM_CAN_NONLINEAR;
1497         return 0;
1498 }
1499
1500 static struct inode *
1501 shmem_get_inode(struct super_block *sb, int mode, dev_t dev)
1502 {
1503         struct inode *inode;
1504         struct shmem_inode_info *info;
1505         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1506
1507         if (shmem_reserve_inode(sb))
1508                 return NULL;
1509
1510         inode = new_inode(sb);
1511         if (inode) {
1512                 inode->i_mode = mode;
1513                 inode->i_uid = current->fsuid;
1514                 inode->i_gid = current->fsgid;
1515                 inode->i_blocks = 0;
1516                 inode->i_mapping->a_ops = &shmem_aops;
1517                 inode->i_mapping->backing_dev_info = &shmem_backing_dev_info;
1518                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1519                 inode->i_generation = get_seconds();
1520                 info = SHMEM_I(inode);
1521                 memset(info, 0, (char *)inode - (char *)info);
1522                 spin_lock_init(&info->lock);
1523                 INIT_LIST_HEAD(&info->swaplist);
1524
1525                 switch (mode & S_IFMT) {
1526                 default:
1527                         inode->i_op = &shmem_special_inode_operations;
1528                         init_special_inode(inode, mode, dev);
1529                         break;
1530                 case S_IFREG:
1531                         inode->i_op = &shmem_inode_operations;
1532                         inode->i_fop = &shmem_file_operations;
1533                         mpol_shared_policy_init(&info->policy,
1534                                                  shmem_get_sbmpol(sbinfo));
1535                         break;
1536                 case S_IFDIR:
1537                         inc_nlink(inode);
1538                         /* Some things misbehave if size == 0 on a directory */
1539                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
1540                         inode->i_op = &shmem_dir_inode_operations;
1541                         inode->i_fop = &simple_dir_operations;
1542                         break;
1543                 case S_IFLNK:
1544                         /*
1545                          * Must not load anything in the rbtree,
1546                          * mpol_free_shared_policy will not be called.
1547                          */
1548                         mpol_shared_policy_init(&info->policy, NULL);
1549                         break;
1550                 }
1551         } else
1552                 shmem_free_inode(sb);
1553         return inode;
1554 }
1555
1556 #ifdef CONFIG_TMPFS
1557 static const struct inode_operations shmem_symlink_inode_operations;
1558 static const struct inode_operations shmem_symlink_inline_operations;
1559
1560 /*
1561  * Normally tmpfs avoids the use of shmem_readpage and shmem_write_begin;
1562  * but providing them allows a tmpfs file to be used for splice, sendfile, and
1563  * below the loop driver, in the generic fashion that many filesystems support.
1564  */
1565 static int shmem_readpage(struct file *file, struct page *page)
1566 {
1567         struct inode *inode = page->mapping->host;
1568         int error = shmem_getpage(inode, page->index, &page, SGP_CACHE, NULL);
1569         unlock_page(page);
1570         return error;
1571 }
1572
1573 static int
1574 shmem_write_begin(struct file *file, struct address_space *mapping,
1575                         loff_t pos, unsigned len, unsigned flags,
1576                         struct page **pagep, void **fsdata)
1577 {
1578         struct inode *inode = mapping->host;
1579         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1580         *pagep = NULL;
1581         return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1582 }
1583
1584 static int
1585 shmem_write_end(struct file *file, struct address_space *mapping,
1586                         loff_t pos, unsigned len, unsigned copied,
1587                         struct page *page, void *fsdata)
1588 {
1589         struct inode *inode = mapping->host;
1590
1591         if (pos + copied > inode->i_size)
1592                 i_size_write(inode, pos + copied);
1593
1594         unlock_page(page);
1595         set_page_dirty(page);
1596         page_cache_release(page);
1597
1598         return copied;
1599 }
1600
1601 static void do_shmem_file_read(struct file *filp, loff_t *ppos, read_descriptor_t *desc, read_actor_t actor)
1602 {
1603         struct inode *inode = filp->f_path.dentry->d_inode;
1604         struct address_space *mapping = inode->i_mapping;
1605         unsigned long index, offset;
1606         enum sgp_type sgp = SGP_READ;
1607
1608         /*
1609          * Might this read be for a stacking filesystem?  Then when reading
1610          * holes of a sparse file, we actually need to allocate those pages,
1611          * and even mark them dirty, so it cannot exceed the max_blocks limit.
1612          */
1613         if (segment_eq(get_fs(), KERNEL_DS))
1614                 sgp = SGP_DIRTY;
1615
1616         index = *ppos >> PAGE_CACHE_SHIFT;
1617         offset = *ppos & ~PAGE_CACHE_MASK;
1618
1619         for (;;) {
1620                 struct page *page = NULL;
1621                 unsigned long end_index, nr, ret;
1622                 loff_t i_size = i_size_read(inode);
1623
1624                 end_index = i_size >> PAGE_CACHE_SHIFT;
1625                 if (index > end_index)
1626                         break;
1627                 if (index == end_index) {
1628                         nr = i_size & ~PAGE_CACHE_MASK;
1629                         if (nr <= offset)
1630                                 break;
1631                 }
1632
1633                 desc->error = shmem_getpage(inode, index, &page, sgp, NULL);
1634                 if (desc->error) {
1635                         if (desc->error == -EINVAL)
1636                                 desc->error = 0;
1637                         break;
1638                 }
1639                 if (page)
1640                         unlock_page(page);
1641
1642                 /*
1643                  * We must evaluate after, since reads (unlike writes)
1644                  * are called without i_mutex protection against truncate
1645                  */
1646                 nr = PAGE_CACHE_SIZE;
1647                 i_size = i_size_read(inode);
1648                 end_index = i_size >> PAGE_CACHE_SHIFT;
1649                 if (index == end_index) {
1650                         nr = i_size & ~PAGE_CACHE_MASK;
1651                         if (nr <= offset) {
1652                                 if (page)
1653                                         page_cache_release(page);
1654                                 break;
1655                         }
1656                 }
1657                 nr -= offset;
1658
1659                 if (page) {
1660                         /*
1661                          * If users can be writing to this page using arbitrary
1662                          * virtual addresses, take care about potential aliasing
1663                          * before reading the page on the kernel side.
1664                          */
1665                         if (mapping_writably_mapped(mapping))
1666                                 flush_dcache_page(page);
1667                         /*
1668                          * Mark the page accessed if we read the beginning.
1669                          */
1670                         if (!offset)
1671                                 mark_page_accessed(page);
1672                 } else {
1673                         page = ZERO_PAGE(0);
1674                         page_cache_get(page);
1675                 }
1676
1677                 /*
1678                  * Ok, we have the page, and it's up-to-date, so
1679                  * now we can copy it to user space...
1680                  *
1681                  * The actor routine returns how many bytes were actually used..
1682                  * NOTE! This may not be the same as how much of a user buffer
1683                  * we filled up (we may be padding etc), so we can only update
1684                  * "pos" here (the actor routine has to update the user buffer
1685                  * pointers and the remaining count).
1686                  */
1687                 ret = actor(desc, page, offset, nr);
1688                 offset += ret;
1689                 index += offset >> PAGE_CACHE_SHIFT;
1690                 offset &= ~PAGE_CACHE_MASK;
1691
1692                 page_cache_release(page);
1693                 if (ret != nr || !desc->count)
1694                         break;
1695
1696                 cond_resched();
1697         }
1698
1699         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1700         file_accessed(filp);
1701 }
1702
1703 static ssize_t shmem_file_aio_read(struct kiocb *iocb,
1704                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
1705 {
1706         struct file *filp = iocb->ki_filp;
1707         ssize_t retval;
1708         unsigned long seg;
1709         size_t count;
1710         loff_t *ppos = &iocb->ki_pos;
1711
1712         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1713         if (retval)
1714                 return retval;
1715
1716         for (seg = 0; seg < nr_segs; seg++) {
1717                 read_descriptor_t desc;
1718
1719                 desc.written = 0;
1720                 desc.arg.buf = iov[seg].iov_base;
1721                 desc.count = iov[seg].iov_len;
1722                 if (desc.count == 0)
1723                         continue;
1724                 desc.error = 0;
1725                 do_shmem_file_read(filp, ppos, &desc, file_read_actor);
1726                 retval += desc.written;
1727                 if (desc.error) {
1728                         retval = retval ?: desc.error;
1729                         break;
1730                 }
1731                 if (desc.count > 0)
1732                         break;
1733         }
1734         return retval;
1735 }
1736
1737 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
1738 {
1739         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
1740
1741         buf->f_type = TMPFS_MAGIC;
1742         buf->f_bsize = PAGE_CACHE_SIZE;
1743         buf->f_namelen = NAME_MAX;
1744         spin_lock(&sbinfo->stat_lock);
1745         if (sbinfo->max_blocks) {
1746                 buf->f_blocks = sbinfo->max_blocks;
1747                 buf->f_bavail = buf->f_bfree = sbinfo->free_blocks;
1748         }
1749         if (sbinfo->max_inodes) {
1750                 buf->f_files = sbinfo->max_inodes;
1751                 buf->f_ffree = sbinfo->free_inodes;
1752         }
1753         /* else leave those fields 0 like simple_statfs */
1754         spin_unlock(&sbinfo->stat_lock);
1755         return 0;
1756 }
1757
1758 /*
1759  * File creation. Allocate an inode, and we're done..
1760  */
1761 static int
1762 shmem_mknod(struct inode *dir, struct dentry *dentry, int mode, dev_t dev)
1763 {
1764         struct inode *inode = shmem_get_inode(dir->i_sb, mode, dev);
1765         int error = -ENOSPC;
1766
1767         if (inode) {
1768                 error = security_inode_init_security(inode, dir, NULL, NULL,
1769                                                      NULL);
1770                 if (error) {
1771                         if (error != -EOPNOTSUPP) {
1772                                 iput(inode);
1773                                 return error;
1774                         }
1775                 }
1776                 error = shmem_acl_init(inode, dir);
1777                 if (error) {
1778                         iput(inode);
1779                         return error;
1780                 }
1781                 if (dir->i_mode & S_ISGID) {
1782                         inode->i_gid = dir->i_gid;
1783                         if (S_ISDIR(mode))
1784                                 inode->i_mode |= S_ISGID;
1785                 }
1786                 dir->i_size += BOGO_DIRENT_SIZE;
1787                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1788                 d_instantiate(dentry, inode);
1789                 dget(dentry); /* Extra count - pin the dentry in core */
1790         }
1791         return error;
1792 }
1793
1794 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, int mode)
1795 {
1796         int error;
1797
1798         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
1799                 return error;
1800         inc_nlink(dir);
1801         return 0;
1802 }
1803
1804 static int shmem_create(struct inode *dir, struct dentry *dentry, int mode,
1805                 struct nameidata *nd)
1806 {
1807         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
1808 }
1809
1810 /*
1811  * Link a file..
1812  */
1813 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
1814 {
1815         struct inode *inode = old_dentry->d_inode;
1816         int ret;
1817
1818         /*
1819          * No ordinary (disk based) filesystem counts links as inodes;
1820          * but each new link needs a new dentry, pinning lowmem, and
1821          * tmpfs dentries cannot be pruned until they are unlinked.
1822          */
1823         ret = shmem_reserve_inode(inode->i_sb);
1824         if (ret)
1825                 goto out;
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 out:
1834         return ret;
1835 }
1836
1837 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
1838 {
1839         struct inode *inode = dentry->d_inode;
1840
1841         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
1842                 shmem_free_inode(inode->i_sb);
1843
1844         dir->i_size -= BOGO_DIRENT_SIZE;
1845         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1846         drop_nlink(inode);
1847         dput(dentry);   /* Undo the count from "create" - this does all the work */
1848         return 0;
1849 }
1850
1851 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
1852 {
1853         if (!simple_empty(dentry))
1854                 return -ENOTEMPTY;
1855
1856         drop_nlink(dentry->d_inode);
1857         drop_nlink(dir);
1858         return shmem_unlink(dir, dentry);
1859 }
1860
1861 /*
1862  * The VFS layer already does all the dentry stuff for rename,
1863  * we just have to decrement the usage count for the target if
1864  * it exists so that the VFS layer correctly free's it when it
1865  * gets overwritten.
1866  */
1867 static int shmem_rename(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
1868 {
1869         struct inode *inode = old_dentry->d_inode;
1870         int they_are_dirs = S_ISDIR(inode->i_mode);
1871
1872         if (!simple_empty(new_dentry))
1873                 return -ENOTEMPTY;
1874
1875         if (new_dentry->d_inode) {
1876                 (void) shmem_unlink(new_dir, new_dentry);
1877                 if (they_are_dirs)
1878                         drop_nlink(old_dir);
1879         } else if (they_are_dirs) {
1880                 drop_nlink(old_dir);
1881                 inc_nlink(new_dir);
1882         }
1883
1884         old_dir->i_size -= BOGO_DIRENT_SIZE;
1885         new_dir->i_size += BOGO_DIRENT_SIZE;
1886         old_dir->i_ctime = old_dir->i_mtime =
1887         new_dir->i_ctime = new_dir->i_mtime =
1888         inode->i_ctime = CURRENT_TIME;
1889         return 0;
1890 }
1891
1892 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
1893 {
1894         int error;
1895         int len;
1896         struct inode *inode;
1897         struct page *page = NULL;
1898         char *kaddr;
1899         struct shmem_inode_info *info;
1900
1901         len = strlen(symname) + 1;
1902         if (len > PAGE_CACHE_SIZE)
1903                 return -ENAMETOOLONG;
1904
1905         inode = shmem_get_inode(dir->i_sb, S_IFLNK|S_IRWXUGO, 0);
1906         if (!inode)
1907                 return -ENOSPC;
1908
1909         error = security_inode_init_security(inode, dir, NULL, NULL,
1910                                              NULL);
1911         if (error) {
1912                 if (error != -EOPNOTSUPP) {
1913                         iput(inode);
1914                         return error;
1915                 }
1916                 error = 0;
1917         }
1918
1919         info = SHMEM_I(inode);
1920         inode->i_size = len-1;
1921         if (len <= (char *)inode - (char *)info) {
1922                 /* do it inline */
1923                 memcpy(info, symname, len);
1924                 inode->i_op = &shmem_symlink_inline_operations;
1925         } else {
1926                 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
1927                 if (error) {
1928                         iput(inode);
1929                         return error;
1930                 }
1931                 unlock_page(page);
1932                 inode->i_op = &shmem_symlink_inode_operations;
1933                 kaddr = kmap_atomic(page, KM_USER0);
1934                 memcpy(kaddr, symname, len);
1935                 kunmap_atomic(kaddr, KM_USER0);
1936                 set_page_dirty(page);
1937                 page_cache_release(page);
1938         }
1939         if (dir->i_mode & S_ISGID)
1940                 inode->i_gid = dir->i_gid;
1941         dir->i_size += BOGO_DIRENT_SIZE;
1942         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1943         d_instantiate(dentry, inode);
1944         dget(dentry);
1945         return 0;
1946 }
1947
1948 static void *shmem_follow_link_inline(struct dentry *dentry, struct nameidata *nd)
1949 {
1950         nd_set_link(nd, (char *)SHMEM_I(dentry->d_inode));
1951         return NULL;
1952 }
1953
1954 static void *shmem_follow_link(struct dentry *dentry, struct nameidata *nd)
1955 {
1956         struct page *page = NULL;
1957         int res = shmem_getpage(dentry->d_inode, 0, &page, SGP_READ, NULL);
1958         nd_set_link(nd, res ? ERR_PTR(res) : kmap(page));
1959         if (page)
1960                 unlock_page(page);
1961         return page;
1962 }
1963
1964 static void shmem_put_link(struct dentry *dentry, struct nameidata *nd, void *cookie)
1965 {
1966         if (!IS_ERR(nd_get_link(nd))) {
1967                 struct page *page = cookie;
1968                 kunmap(page);
1969                 mark_page_accessed(page);
1970                 page_cache_release(page);
1971         }
1972 }
1973
1974 static const struct inode_operations shmem_symlink_inline_operations = {
1975         .readlink       = generic_readlink,
1976         .follow_link    = shmem_follow_link_inline,
1977 };
1978
1979 static const struct inode_operations shmem_symlink_inode_operations = {
1980         .truncate       = shmem_truncate,
1981         .readlink       = generic_readlink,
1982         .follow_link    = shmem_follow_link,
1983         .put_link       = shmem_put_link,
1984 };
1985
1986 #ifdef CONFIG_TMPFS_POSIX_ACL
1987 /*
1988  * Superblocks without xattr inode operations will get security.* xattr
1989  * support from the VFS "for free". As soon as we have any other xattrs
1990  * like ACLs, we also need to implement the security.* handlers at
1991  * filesystem level, though.
1992  */
1993
1994 static size_t shmem_xattr_security_list(struct inode *inode, char *list,
1995                                         size_t list_len, const char *name,
1996                                         size_t name_len)
1997 {
1998         return security_inode_listsecurity(inode, list, list_len);
1999 }
2000
2001 static int shmem_xattr_security_get(struct inode *inode, const char *name,
2002                                     void *buffer, size_t size)
2003 {
2004         if (strcmp(name, "") == 0)
2005                 return -EINVAL;
2006         return xattr_getsecurity(inode, name, buffer, size);
2007 }
2008
2009 static int shmem_xattr_security_set(struct inode *inode, const char *name,
2010                                     const void *value, size_t size, int flags)
2011 {
2012         if (strcmp(name, "") == 0)
2013                 return -EINVAL;
2014         return security_inode_setsecurity(inode, name, value, size, flags);
2015 }
2016
2017 static struct xattr_handler shmem_xattr_security_handler = {
2018         .prefix = XATTR_SECURITY_PREFIX,
2019         .list   = shmem_xattr_security_list,
2020         .get    = shmem_xattr_security_get,
2021         .set    = shmem_xattr_security_set,
2022 };
2023
2024 static struct xattr_handler *shmem_xattr_handlers[] = {
2025         &shmem_xattr_acl_access_handler,
2026         &shmem_xattr_acl_default_handler,
2027         &shmem_xattr_security_handler,
2028         NULL
2029 };
2030 #endif
2031
2032 static struct dentry *shmem_get_parent(struct dentry *child)
2033 {
2034         return ERR_PTR(-ESTALE);
2035 }
2036
2037 static int shmem_match(struct inode *ino, void *vfh)
2038 {
2039         __u32 *fh = vfh;
2040         __u64 inum = fh[2];
2041         inum = (inum << 32) | fh[1];
2042         return ino->i_ino == inum && fh[0] == ino->i_generation;
2043 }
2044
2045 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
2046                 struct fid *fid, int fh_len, int fh_type)
2047 {
2048         struct inode *inode;
2049         struct dentry *dentry = NULL;
2050         u64 inum = fid->raw[2];
2051         inum = (inum << 32) | fid->raw[1];
2052
2053         if (fh_len < 3)
2054                 return NULL;
2055
2056         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
2057                         shmem_match, fid->raw);
2058         if (inode) {
2059                 dentry = d_find_alias(inode);
2060                 iput(inode);
2061         }
2062
2063         return dentry;
2064 }
2065
2066 static int shmem_encode_fh(struct dentry *dentry, __u32 *fh, int *len,
2067                                 int connectable)
2068 {
2069         struct inode *inode = dentry->d_inode;
2070
2071         if (*len < 3)
2072                 return 255;
2073
2074         if (hlist_unhashed(&inode->i_hash)) {
2075                 /* Unfortunately insert_inode_hash is not idempotent,
2076                  * so as we hash inodes here rather than at creation
2077                  * time, we need a lock to ensure we only try
2078                  * to do it once
2079                  */
2080                 static DEFINE_SPINLOCK(lock);
2081                 spin_lock(&lock);
2082                 if (hlist_unhashed(&inode->i_hash))
2083                         __insert_inode_hash(inode,
2084                                             inode->i_ino + inode->i_generation);
2085                 spin_unlock(&lock);
2086         }
2087
2088         fh[0] = inode->i_generation;
2089         fh[1] = inode->i_ino;
2090         fh[2] = ((__u64)inode->i_ino) >> 32;
2091
2092         *len = 3;
2093         return 1;
2094 }
2095
2096 static const struct export_operations shmem_export_ops = {
2097         .get_parent     = shmem_get_parent,
2098         .encode_fh      = shmem_encode_fh,
2099         .fh_to_dentry   = shmem_fh_to_dentry,
2100 };
2101
2102 static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
2103                                bool remount)
2104 {
2105         char *this_char, *value, *rest;
2106
2107         while (options != NULL) {
2108                 this_char = options;
2109                 for (;;) {
2110                         /*
2111                          * NUL-terminate this option: unfortunately,
2112                          * mount options form a comma-separated list,
2113                          * but mpol's nodelist may also contain commas.
2114                          */
2115                         options = strchr(options, ',');
2116                         if (options == NULL)
2117                                 break;
2118                         options++;
2119                         if (!isdigit(*options)) {
2120                                 options[-1] = '\0';
2121                                 break;
2122                         }
2123                 }
2124                 if (!*this_char)
2125                         continue;
2126                 if ((value = strchr(this_char,'=')) != NULL) {
2127                         *value++ = 0;
2128                 } else {
2129                         printk(KERN_ERR
2130                             "tmpfs: No value for mount option '%s'\n",
2131                             this_char);
2132                         return 1;
2133                 }
2134
2135                 if (!strcmp(this_char,"size")) {
2136                         unsigned long long size;
2137                         size = memparse(value,&rest);
2138                         if (*rest == '%') {
2139                                 size <<= PAGE_SHIFT;
2140                                 size *= totalram_pages;
2141                                 do_div(size, 100);
2142                                 rest++;
2143                         }
2144                         if (*rest)
2145                                 goto bad_val;
2146                         sbinfo->max_blocks =
2147                                 DIV_ROUND_UP(size, PAGE_CACHE_SIZE);
2148                 } else if (!strcmp(this_char,"nr_blocks")) {
2149                         sbinfo->max_blocks = memparse(value, &rest);
2150                         if (*rest)
2151                                 goto bad_val;
2152                 } else if (!strcmp(this_char,"nr_inodes")) {
2153                         sbinfo->max_inodes = memparse(value, &rest);
2154                         if (*rest)
2155                                 goto bad_val;
2156                 } else if (!strcmp(this_char,"mode")) {
2157                         if (remount)
2158                                 continue;
2159                         sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
2160                         if (*rest)
2161                                 goto bad_val;
2162                 } else if (!strcmp(this_char,"uid")) {
2163                         if (remount)
2164                                 continue;
2165                         sbinfo->uid = simple_strtoul(value, &rest, 0);
2166                         if (*rest)
2167                                 goto bad_val;
2168                 } else if (!strcmp(this_char,"gid")) {
2169                         if (remount)
2170                                 continue;
2171                         sbinfo->gid = simple_strtoul(value, &rest, 0);
2172                         if (*rest)
2173                                 goto bad_val;
2174                 } else if (!strcmp(this_char,"mpol")) {
2175                         if (mpol_parse_str(value, &sbinfo->mpol, 1))
2176                                 goto bad_val;
2177                 } else {
2178                         printk(KERN_ERR "tmpfs: Bad mount option %s\n",
2179                                this_char);
2180                         return 1;
2181                 }
2182         }
2183         return 0;
2184
2185 bad_val:
2186         printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
2187                value, this_char);
2188         return 1;
2189
2190 }
2191
2192 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2193 {
2194         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2195         struct shmem_sb_info config = *sbinfo;
2196         unsigned long blocks;
2197         unsigned long inodes;
2198         int error = -EINVAL;
2199
2200         if (shmem_parse_options(data, &config, true))
2201                 return error;
2202
2203         spin_lock(&sbinfo->stat_lock);
2204         blocks = sbinfo->max_blocks - sbinfo->free_blocks;
2205         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2206         if (config.max_blocks < blocks)
2207                 goto out;
2208         if (config.max_inodes < inodes)
2209                 goto out;
2210         /*
2211          * Those tests also disallow limited->unlimited while any are in
2212          * use, so i_blocks will always be zero when max_blocks is zero;
2213          * but we must separately disallow unlimited->limited, because
2214          * in that case we have no record of how much is already in use.
2215          */
2216         if (config.max_blocks && !sbinfo->max_blocks)
2217                 goto out;
2218         if (config.max_inodes && !sbinfo->max_inodes)
2219                 goto out;
2220
2221         error = 0;
2222         sbinfo->max_blocks  = config.max_blocks;
2223         sbinfo->free_blocks = config.max_blocks - blocks;
2224         sbinfo->max_inodes  = config.max_inodes;
2225         sbinfo->free_inodes = config.max_inodes - inodes;
2226
2227         mpol_put(sbinfo->mpol);
2228         sbinfo->mpol        = config.mpol;      /* transfers initial ref */
2229 out:
2230         spin_unlock(&sbinfo->stat_lock);
2231         return error;
2232 }
2233
2234 static int shmem_show_options(struct seq_file *seq, struct vfsmount *vfs)
2235 {
2236         struct shmem_sb_info *sbinfo = SHMEM_SB(vfs->mnt_sb);
2237
2238         if (sbinfo->max_blocks != shmem_default_max_blocks())
2239                 seq_printf(seq, ",size=%luk",
2240                         sbinfo->max_blocks << (PAGE_CACHE_SHIFT - 10));
2241         if (sbinfo->max_inodes != shmem_default_max_inodes())
2242                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
2243         if (sbinfo->mode != (S_IRWXUGO | S_ISVTX))
2244                 seq_printf(seq, ",mode=%03o", sbinfo->mode);
2245         if (sbinfo->uid != 0)
2246                 seq_printf(seq, ",uid=%u", sbinfo->uid);
2247         if (sbinfo->gid != 0)
2248                 seq_printf(seq, ",gid=%u", sbinfo->gid);
2249         shmem_show_mpol(seq, sbinfo->mpol);
2250         return 0;
2251 }
2252 #endif /* CONFIG_TMPFS */
2253
2254 static void shmem_put_super(struct super_block *sb)
2255 {
2256         kfree(sb->s_fs_info);
2257         sb->s_fs_info = NULL;
2258 }
2259
2260 static int shmem_fill_super(struct super_block *sb,
2261                             void *data, int silent)
2262 {
2263         struct inode *inode;
2264         struct dentry *root;
2265         struct shmem_sb_info *sbinfo;
2266         int err = -ENOMEM;
2267
2268         /* Round up to L1_CACHE_BYTES to resist false sharing */
2269         sbinfo = kmalloc(max((int)sizeof(struct shmem_sb_info),
2270                                 L1_CACHE_BYTES), GFP_KERNEL);
2271         if (!sbinfo)
2272                 return -ENOMEM;
2273
2274         sbinfo->max_blocks = 0;
2275         sbinfo->max_inodes = 0;
2276         sbinfo->mode = S_IRWXUGO | S_ISVTX;
2277         sbinfo->uid = current->fsuid;
2278         sbinfo->gid = current->fsgid;
2279         sbinfo->mpol = NULL;
2280         sb->s_fs_info = sbinfo;
2281
2282 #ifdef CONFIG_TMPFS
2283         /*
2284          * Per default we only allow half of the physical ram per
2285          * tmpfs instance, limiting inodes to one per page of lowmem;
2286          * but the internal instance is left unlimited.
2287          */
2288         if (!(sb->s_flags & MS_NOUSER)) {
2289                 sbinfo->max_blocks = shmem_default_max_blocks();
2290                 sbinfo->max_inodes = shmem_default_max_inodes();
2291                 if (shmem_parse_options(data, sbinfo, false)) {
2292                         err = -EINVAL;
2293                         goto failed;
2294                 }
2295         }
2296         sb->s_export_op = &shmem_export_ops;
2297 #else
2298         sb->s_flags |= MS_NOUSER;
2299 #endif
2300
2301         spin_lock_init(&sbinfo->stat_lock);
2302         sbinfo->free_blocks = sbinfo->max_blocks;
2303         sbinfo->free_inodes = sbinfo->max_inodes;
2304
2305         sb->s_maxbytes = SHMEM_MAX_BYTES;
2306         sb->s_blocksize = PAGE_CACHE_SIZE;
2307         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
2308         sb->s_magic = TMPFS_MAGIC;
2309         sb->s_op = &shmem_ops;
2310         sb->s_time_gran = 1;
2311 #ifdef CONFIG_TMPFS_POSIX_ACL
2312         sb->s_xattr = shmem_xattr_handlers;
2313         sb->s_flags |= MS_POSIXACL;
2314 #endif
2315
2316         inode = shmem_get_inode(sb, S_IFDIR | sbinfo->mode, 0);
2317         if (!inode)
2318                 goto failed;
2319         inode->i_uid = sbinfo->uid;
2320         inode->i_gid = sbinfo->gid;
2321         root = d_alloc_root(inode);
2322         if (!root)
2323                 goto failed_iput;
2324         sb->s_root = root;
2325         return 0;
2326
2327 failed_iput:
2328         iput(inode);
2329 failed:
2330         shmem_put_super(sb);
2331         return err;
2332 }
2333
2334 static struct kmem_cache *shmem_inode_cachep;
2335
2336 static struct inode *shmem_alloc_inode(struct super_block *sb)
2337 {
2338         struct shmem_inode_info *p;
2339         p = (struct shmem_inode_info *)kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
2340         if (!p)
2341                 return NULL;
2342         return &p->vfs_inode;
2343 }
2344
2345 static void shmem_destroy_inode(struct inode *inode)
2346 {
2347         if ((inode->i_mode & S_IFMT) == S_IFREG) {
2348                 /* only struct inode is valid if it's an inline symlink */
2349                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
2350         }
2351         shmem_acl_destroy_inode(inode);
2352         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
2353 }
2354
2355 static void init_once(void *foo)
2356 {
2357         struct shmem_inode_info *p = (struct shmem_inode_info *) foo;
2358
2359         inode_init_once(&p->vfs_inode);
2360 #ifdef CONFIG_TMPFS_POSIX_ACL
2361         p->i_acl = NULL;
2362         p->i_default_acl = NULL;
2363 #endif
2364 }
2365
2366 static int init_inodecache(void)
2367 {
2368         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
2369                                 sizeof(struct shmem_inode_info),
2370                                 0, SLAB_PANIC, init_once);
2371         return 0;
2372 }
2373
2374 static void destroy_inodecache(void)
2375 {
2376         kmem_cache_destroy(shmem_inode_cachep);
2377 }
2378
2379 static const struct address_space_operations shmem_aops = {
2380         .writepage      = shmem_writepage,
2381         .set_page_dirty = __set_page_dirty_no_writeback,
2382 #ifdef CONFIG_TMPFS
2383         .readpage       = shmem_readpage,
2384         .write_begin    = shmem_write_begin,
2385         .write_end      = shmem_write_end,
2386 #endif
2387         .migratepage    = migrate_page,
2388 };
2389
2390 static const struct file_operations shmem_file_operations = {
2391         .mmap           = shmem_mmap,
2392 #ifdef CONFIG_TMPFS
2393         .llseek         = generic_file_llseek,
2394         .read           = do_sync_read,
2395         .write          = do_sync_write,
2396         .aio_read       = shmem_file_aio_read,
2397         .aio_write      = generic_file_aio_write,
2398         .fsync          = simple_sync_file,
2399         .splice_read    = generic_file_splice_read,
2400         .splice_write   = generic_file_splice_write,
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         .show_options   = shmem_show_options,
2458 #endif
2459         .delete_inode   = shmem_delete_inode,
2460         .drop_inode     = generic_delete_inode,
2461         .put_super      = shmem_put_super,
2462 };
2463
2464 static struct vm_operations_struct shmem_vm_ops = {
2465         .fault          = shmem_fault,
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 = bdi_init(&shmem_backing_dev_info);
2492         if (error)
2493                 goto out4;
2494
2495         error = init_inodecache();
2496         if (error)
2497                 goto out3;
2498
2499         error = register_filesystem(&tmpfs_fs_type);
2500         if (error) {
2501                 printk(KERN_ERR "Could not register tmpfs\n");
2502                 goto out2;
2503         }
2504
2505         shm_mnt = vfs_kern_mount(&tmpfs_fs_type, MS_NOUSER,
2506                                 tmpfs_fs_type.name, NULL);
2507         if (IS_ERR(shm_mnt)) {
2508                 error = PTR_ERR(shm_mnt);
2509                 printk(KERN_ERR "Could not kern_mount tmpfs\n");
2510                 goto out1;
2511         }
2512         return 0;
2513
2514 out1:
2515         unregister_filesystem(&tmpfs_fs_type);
2516 out2:
2517         destroy_inodecache();
2518 out3:
2519         bdi_destroy(&shmem_backing_dev_info);
2520 out4:
2521         shm_mnt = ERR_PTR(error);
2522         return error;
2523 }
2524 module_init(init_tmpfs)
2525
2526 /**
2527  * shmem_file_setup - get an unlinked file living in tmpfs
2528  * @name: name for dentry (to be seen in /proc/<pid>/maps
2529  * @size: size to be set for the file
2530  * @flags: vm_flags
2531  */
2532 struct file *shmem_file_setup(char *name, loff_t size, unsigned long flags)
2533 {
2534         int error;
2535         struct file *file;
2536         struct inode *inode;
2537         struct dentry *dentry, *root;
2538         struct qstr this;
2539
2540         if (IS_ERR(shm_mnt))
2541                 return (void *)shm_mnt;
2542
2543         if (size < 0 || size > SHMEM_MAX_BYTES)
2544                 return ERR_PTR(-EINVAL);
2545
2546         if (shmem_acct_size(flags, size))
2547                 return ERR_PTR(-ENOMEM);
2548
2549         error = -ENOMEM;
2550         this.name = name;
2551         this.len = strlen(name);
2552         this.hash = 0; /* will go */
2553         root = shm_mnt->mnt_root;
2554         dentry = d_alloc(root, &this);
2555         if (!dentry)
2556                 goto put_memory;
2557
2558         error = -ENFILE;
2559         file = get_empty_filp();
2560         if (!file)
2561                 goto put_dentry;
2562
2563         error = -ENOSPC;
2564         inode = shmem_get_inode(root->d_sb, S_IFREG | S_IRWXUGO, 0);
2565         if (!inode)
2566                 goto close_file;
2567
2568         SHMEM_I(inode)->flags = flags & VM_ACCOUNT;
2569         d_instantiate(dentry, inode);
2570         inode->i_size = size;
2571         inode->i_nlink = 0;     /* It is unlinked */
2572         init_file(file, shm_mnt, dentry, FMODE_WRITE | FMODE_READ,
2573                         &shmem_file_operations);
2574         return file;
2575
2576 close_file:
2577         put_filp(file);
2578 put_dentry:
2579         dput(dentry);
2580 put_memory:
2581         shmem_unacct_size(flags, size);
2582         return ERR_PTR(error);
2583 }
2584
2585 /**
2586  * shmem_zero_setup - setup a shared anonymous mapping
2587  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
2588  */
2589 int shmem_zero_setup(struct vm_area_struct *vma)
2590 {
2591         struct file *file;
2592         loff_t size = vma->vm_end - vma->vm_start;
2593
2594         file = shmem_file_setup("dev/zero", size, vma->vm_flags);
2595         if (IS_ERR(file))
2596                 return PTR_ERR(file);
2597
2598         if (vma->vm_file)
2599                 fput(vma->vm_file);
2600         vma->vm_file = file;
2601         vma->vm_ops = &shmem_vm_ops;
2602         return 0;
2603 }