Integrity: IMA file free imbalance
[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 #include <linux/magic.h>
54 #include <linux/ima.h>
55
56 #include <asm/uaccess.h>
57 #include <asm/div64.h>
58 #include <asm/pgtable.h>
59
60 #define ENTRIES_PER_PAGE (PAGE_CACHE_SIZE/sizeof(unsigned long))
61 #define ENTRIES_PER_PAGEPAGE (ENTRIES_PER_PAGE*ENTRIES_PER_PAGE)
62 #define BLOCKS_PER_PAGE  (PAGE_CACHE_SIZE/512)
63
64 #define SHMEM_MAX_INDEX  (SHMEM_NR_DIRECT + (ENTRIES_PER_PAGEPAGE/2) * (ENTRIES_PER_PAGE+1))
65 #define SHMEM_MAX_BYTES  ((unsigned long long)SHMEM_MAX_INDEX << PAGE_CACHE_SHIFT)
66
67 #define VM_ACCT(size)    (PAGE_CACHE_ALIGN(size) >> PAGE_SHIFT)
68
69 /* info->flags needs VM_flags to handle pagein/truncate races efficiently */
70 #define SHMEM_PAGEIN     VM_READ
71 #define SHMEM_TRUNCATE   VM_WRITE
72
73 /* Definition to limit shmem_truncate's steps between cond_rescheds */
74 #define LATENCY_LIMIT    64
75
76 /* Pretend that each entry is of this size in directory's i_size */
77 #define BOGO_DIRENT_SIZE 20
78
79 /* Flag allocation requirements to shmem_getpage and shmem_swp_alloc */
80 enum sgp_type {
81         SGP_READ,       /* don't exceed i_size, don't allocate page */
82         SGP_CACHE,      /* don't exceed i_size, may allocate page */
83         SGP_DIRTY,      /* like SGP_CACHE, but set new page dirty */
84         SGP_WRITE,      /* may exceed i_size, may allocate page */
85 };
86
87 #ifdef CONFIG_TMPFS
88 static unsigned long shmem_default_max_blocks(void)
89 {
90         return totalram_pages / 2;
91 }
92
93 static unsigned long shmem_default_max_inodes(void)
94 {
95         return min(totalram_pages - totalhigh_pages, totalram_pages / 2);
96 }
97 #endif
98
99 static int shmem_getpage(struct inode *inode, unsigned long idx,
100                          struct page **pagep, enum sgp_type sgp, int *type);
101
102 static inline struct page *shmem_dir_alloc(gfp_t gfp_mask)
103 {
104         /*
105          * The above definition of ENTRIES_PER_PAGE, and the use of
106          * BLOCKS_PER_PAGE on indirect pages, assume PAGE_CACHE_SIZE:
107          * might be reconsidered if it ever diverges from PAGE_SIZE.
108          *
109          * Mobility flags are masked out as swap vectors cannot move
110          */
111         return alloc_pages((gfp_mask & ~GFP_MOVABLE_MASK) | __GFP_ZERO,
112                                 PAGE_CACHE_SHIFT-PAGE_SHIFT);
113 }
114
115 static inline void shmem_dir_free(struct page *page)
116 {
117         __free_pages(page, PAGE_CACHE_SHIFT-PAGE_SHIFT);
118 }
119
120 static struct page **shmem_dir_map(struct page *page)
121 {
122         return (struct page **)kmap_atomic(page, KM_USER0);
123 }
124
125 static inline void shmem_dir_unmap(struct page **dir)
126 {
127         kunmap_atomic(dir, KM_USER0);
128 }
129
130 static swp_entry_t *shmem_swp_map(struct page *page)
131 {
132         return (swp_entry_t *)kmap_atomic(page, KM_USER1);
133 }
134
135 static inline void shmem_swp_balance_unmap(void)
136 {
137         /*
138          * When passing a pointer to an i_direct entry, to code which
139          * also handles indirect entries and so will shmem_swp_unmap,
140          * we must arrange for the preempt count to remain in balance.
141          * What kmap_atomic of a lowmem page does depends on config
142          * and architecture, so pretend to kmap_atomic some lowmem page.
143          */
144         (void) kmap_atomic(ZERO_PAGE(0), KM_USER1);
145 }
146
147 static inline void shmem_swp_unmap(swp_entry_t *entry)
148 {
149         kunmap_atomic(entry, KM_USER1);
150 }
151
152 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
153 {
154         return sb->s_fs_info;
155 }
156
157 /*
158  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
159  * for shared memory and for shared anonymous (/dev/zero) mappings
160  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
161  * consistent with the pre-accounting of private mappings ...
162  */
163 static inline int shmem_acct_size(unsigned long flags, loff_t size)
164 {
165         return (flags & VM_ACCOUNT) ?
166                 security_vm_enough_memory_kern(VM_ACCT(size)) : 0;
167 }
168
169 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
170 {
171         if (flags & VM_ACCOUNT)
172                 vm_unacct_memory(VM_ACCT(size));
173 }
174
175 /*
176  * ... whereas tmpfs objects are accounted incrementally as
177  * pages are allocated, in order to allow huge sparse files.
178  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
179  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
180  */
181 static inline int shmem_acct_block(unsigned long flags)
182 {
183         return (flags & VM_ACCOUNT) ?
184                 0 : security_vm_enough_memory_kern(VM_ACCT(PAGE_CACHE_SIZE));
185 }
186
187 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
188 {
189         if (!(flags & VM_ACCOUNT))
190                 vm_unacct_memory(pages * VM_ACCT(PAGE_CACHE_SIZE));
191 }
192
193 static const struct super_operations shmem_ops;
194 static const struct address_space_operations shmem_aops;
195 static const struct file_operations shmem_file_operations;
196 static const struct inode_operations shmem_inode_operations;
197 static const struct inode_operations shmem_dir_inode_operations;
198 static const struct inode_operations shmem_special_inode_operations;
199 static struct vm_operations_struct shmem_vm_ops;
200
201 static struct backing_dev_info shmem_backing_dev_info  __read_mostly = {
202         .ra_pages       = 0,    /* No readahead */
203         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_SWAP_BACKED,
204         .unplug_io_fn   = default_unplug_io_fn,
205 };
206
207 static LIST_HEAD(shmem_swaplist);
208 static DEFINE_MUTEX(shmem_swaplist_mutex);
209
210 static void shmem_free_blocks(struct inode *inode, long pages)
211 {
212         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
213         if (sbinfo->max_blocks) {
214                 spin_lock(&sbinfo->stat_lock);
215                 sbinfo->free_blocks += pages;
216                 inode->i_blocks -= pages*BLOCKS_PER_PAGE;
217                 spin_unlock(&sbinfo->stat_lock);
218         }
219 }
220
221 static int shmem_reserve_inode(struct super_block *sb)
222 {
223         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
224         if (sbinfo->max_inodes) {
225                 spin_lock(&sbinfo->stat_lock);
226                 if (!sbinfo->free_inodes) {
227                         spin_unlock(&sbinfo->stat_lock);
228                         return -ENOSPC;
229                 }
230                 sbinfo->free_inodes--;
231                 spin_unlock(&sbinfo->stat_lock);
232         }
233         return 0;
234 }
235
236 static void shmem_free_inode(struct super_block *sb)
237 {
238         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
239         if (sbinfo->max_inodes) {
240                 spin_lock(&sbinfo->stat_lock);
241                 sbinfo->free_inodes++;
242                 spin_unlock(&sbinfo->stat_lock);
243         }
244 }
245
246 /**
247  * shmem_recalc_inode - recalculate the size of an inode
248  * @inode: inode to recalc
249  *
250  * We have to calculate the free blocks since the mm can drop
251  * undirtied hole pages behind our back.
252  *
253  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
254  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
255  *
256  * It has to be called with the spinlock held.
257  */
258 static void shmem_recalc_inode(struct inode *inode)
259 {
260         struct shmem_inode_info *info = SHMEM_I(inode);
261         long freed;
262
263         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
264         if (freed > 0) {
265                 info->alloced -= freed;
266                 shmem_unacct_blocks(info->flags, freed);
267                 shmem_free_blocks(inode, freed);
268         }
269 }
270
271 /**
272  * shmem_swp_entry - find the swap vector position in the info structure
273  * @info:  info structure for the inode
274  * @index: index of the page to find
275  * @page:  optional page to add to the structure. Has to be preset to
276  *         all zeros
277  *
278  * If there is no space allocated yet it will return NULL when
279  * page is NULL, else it will use the page for the needed block,
280  * setting it to NULL on return to indicate that it has been used.
281  *
282  * The swap vector is organized the following way:
283  *
284  * There are SHMEM_NR_DIRECT entries directly stored in the
285  * shmem_inode_info structure. So small files do not need an addional
286  * allocation.
287  *
288  * For pages with index > SHMEM_NR_DIRECT there is the pointer
289  * i_indirect which points to a page which holds in the first half
290  * doubly indirect blocks, in the second half triple indirect blocks:
291  *
292  * For an artificial ENTRIES_PER_PAGE = 4 this would lead to the
293  * following layout (for SHMEM_NR_DIRECT == 16):
294  *
295  * i_indirect -> dir --> 16-19
296  *            |      +-> 20-23
297  *            |
298  *            +-->dir2 --> 24-27
299  *            |        +-> 28-31
300  *            |        +-> 32-35
301  *            |        +-> 36-39
302  *            |
303  *            +-->dir3 --> 40-43
304  *                     +-> 44-47
305  *                     +-> 48-51
306  *                     +-> 52-55
307  */
308 static swp_entry_t *shmem_swp_entry(struct shmem_inode_info *info, unsigned long index, struct page **page)
309 {
310         unsigned long offset;
311         struct page **dir;
312         struct page *subdir;
313
314         if (index < SHMEM_NR_DIRECT) {
315                 shmem_swp_balance_unmap();
316                 return info->i_direct+index;
317         }
318         if (!info->i_indirect) {
319                 if (page) {
320                         info->i_indirect = *page;
321                         *page = NULL;
322                 }
323                 return NULL;                    /* need another page */
324         }
325
326         index -= SHMEM_NR_DIRECT;
327         offset = index % ENTRIES_PER_PAGE;
328         index /= ENTRIES_PER_PAGE;
329         dir = shmem_dir_map(info->i_indirect);
330
331         if (index >= ENTRIES_PER_PAGE/2) {
332                 index -= ENTRIES_PER_PAGE/2;
333                 dir += ENTRIES_PER_PAGE/2 + index/ENTRIES_PER_PAGE;
334                 index %= ENTRIES_PER_PAGE;
335                 subdir = *dir;
336                 if (!subdir) {
337                         if (page) {
338                                 *dir = *page;
339                                 *page = NULL;
340                         }
341                         shmem_dir_unmap(dir);
342                         return NULL;            /* need another page */
343                 }
344                 shmem_dir_unmap(dir);
345                 dir = shmem_dir_map(subdir);
346         }
347
348         dir += index;
349         subdir = *dir;
350         if (!subdir) {
351                 if (!page || !(subdir = *page)) {
352                         shmem_dir_unmap(dir);
353                         return NULL;            /* need a page */
354                 }
355                 *dir = subdir;
356                 *page = NULL;
357         }
358         shmem_dir_unmap(dir);
359         return shmem_swp_map(subdir) + offset;
360 }
361
362 static void shmem_swp_set(struct shmem_inode_info *info, swp_entry_t *entry, unsigned long value)
363 {
364         long incdec = value? 1: -1;
365
366         entry->val = value;
367         info->swapped += incdec;
368         if ((unsigned long)(entry - info->i_direct) >= SHMEM_NR_DIRECT) {
369                 struct page *page = kmap_atomic_to_page(entry);
370                 set_page_private(page, page_private(page) + incdec);
371         }
372 }
373
374 /**
375  * shmem_swp_alloc - get the position of the swap entry for the page.
376  * @info:       info structure for the inode
377  * @index:      index of the page to find
378  * @sgp:        check and recheck i_size? skip allocation?
379  *
380  * If the entry does not exist, allocate it.
381  */
382 static swp_entry_t *shmem_swp_alloc(struct shmem_inode_info *info, unsigned long index, enum sgp_type sgp)
383 {
384         struct inode *inode = &info->vfs_inode;
385         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
386         struct page *page = NULL;
387         swp_entry_t *entry;
388
389         if (sgp != SGP_WRITE &&
390             ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode))
391                 return ERR_PTR(-EINVAL);
392
393         while (!(entry = shmem_swp_entry(info, index, &page))) {
394                 if (sgp == SGP_READ)
395                         return shmem_swp_map(ZERO_PAGE(0));
396                 /*
397                  * Test free_blocks against 1 not 0, since we have 1 data
398                  * page (and perhaps indirect index pages) yet to allocate:
399                  * a waste to allocate index if we cannot allocate data.
400                  */
401                 if (sbinfo->max_blocks) {
402                         spin_lock(&sbinfo->stat_lock);
403                         if (sbinfo->free_blocks <= 1) {
404                                 spin_unlock(&sbinfo->stat_lock);
405                                 return ERR_PTR(-ENOSPC);
406                         }
407                         sbinfo->free_blocks--;
408                         inode->i_blocks += BLOCKS_PER_PAGE;
409                         spin_unlock(&sbinfo->stat_lock);
410                 }
411
412                 spin_unlock(&info->lock);
413                 page = shmem_dir_alloc(mapping_gfp_mask(inode->i_mapping));
414                 if (page)
415                         set_page_private(page, 0);
416                 spin_lock(&info->lock);
417
418                 if (!page) {
419                         shmem_free_blocks(inode, 1);
420                         return ERR_PTR(-ENOMEM);
421                 }
422                 if (sgp != SGP_WRITE &&
423                     ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
424                         entry = ERR_PTR(-EINVAL);
425                         break;
426                 }
427                 if (info->next_index <= index)
428                         info->next_index = index + 1;
429         }
430         if (page) {
431                 /* another task gave its page, or truncated the file */
432                 shmem_free_blocks(inode, 1);
433                 shmem_dir_free(page);
434         }
435         if (info->next_index <= index && !IS_ERR(entry))
436                 info->next_index = index + 1;
437         return entry;
438 }
439
440 /**
441  * shmem_free_swp - free some swap entries in a directory
442  * @dir:        pointer to the directory
443  * @edir:       pointer after last entry of the directory
444  * @punch_lock: pointer to spinlock when needed for the holepunch case
445  */
446 static int shmem_free_swp(swp_entry_t *dir, swp_entry_t *edir,
447                                                 spinlock_t *punch_lock)
448 {
449         spinlock_t *punch_unlock = NULL;
450         swp_entry_t *ptr;
451         int freed = 0;
452
453         for (ptr = dir; ptr < edir; ptr++) {
454                 if (ptr->val) {
455                         if (unlikely(punch_lock)) {
456                                 punch_unlock = punch_lock;
457                                 punch_lock = NULL;
458                                 spin_lock(punch_unlock);
459                                 if (!ptr->val)
460                                         continue;
461                         }
462                         free_swap_and_cache(*ptr);
463                         *ptr = (swp_entry_t){0};
464                         freed++;
465                 }
466         }
467         if (punch_unlock)
468                 spin_unlock(punch_unlock);
469         return freed;
470 }
471
472 static int shmem_map_and_free_swp(struct page *subdir, int offset,
473                 int limit, struct page ***dir, spinlock_t *punch_lock)
474 {
475         swp_entry_t *ptr;
476         int freed = 0;
477
478         ptr = shmem_swp_map(subdir);
479         for (; offset < limit; offset += LATENCY_LIMIT) {
480                 int size = limit - offset;
481                 if (size > LATENCY_LIMIT)
482                         size = LATENCY_LIMIT;
483                 freed += shmem_free_swp(ptr+offset, ptr+offset+size,
484                                                         punch_lock);
485                 if (need_resched()) {
486                         shmem_swp_unmap(ptr);
487                         if (*dir) {
488                                 shmem_dir_unmap(*dir);
489                                 *dir = NULL;
490                         }
491                         cond_resched();
492                         ptr = shmem_swp_map(subdir);
493                 }
494         }
495         shmem_swp_unmap(ptr);
496         return freed;
497 }
498
499 static void shmem_free_pages(struct list_head *next)
500 {
501         struct page *page;
502         int freed = 0;
503
504         do {
505                 page = container_of(next, struct page, lru);
506                 next = next->next;
507                 shmem_dir_free(page);
508                 freed++;
509                 if (freed >= LATENCY_LIMIT) {
510                         cond_resched();
511                         freed = 0;
512                 }
513         } while (next);
514 }
515
516 static void shmem_truncate_range(struct inode *inode, loff_t start, loff_t end)
517 {
518         struct shmem_inode_info *info = SHMEM_I(inode);
519         unsigned long idx;
520         unsigned long size;
521         unsigned long limit;
522         unsigned long stage;
523         unsigned long diroff;
524         struct page **dir;
525         struct page *topdir;
526         struct page *middir;
527         struct page *subdir;
528         swp_entry_t *ptr;
529         LIST_HEAD(pages_to_free);
530         long nr_pages_to_free = 0;
531         long nr_swaps_freed = 0;
532         int offset;
533         int freed;
534         int punch_hole;
535         spinlock_t *needs_lock;
536         spinlock_t *punch_lock;
537         unsigned long upper_limit;
538
539         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
540         idx = (start + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
541         if (idx >= info->next_index)
542                 return;
543
544         spin_lock(&info->lock);
545         info->flags |= SHMEM_TRUNCATE;
546         if (likely(end == (loff_t) -1)) {
547                 limit = info->next_index;
548                 upper_limit = SHMEM_MAX_INDEX;
549                 info->next_index = idx;
550                 needs_lock = NULL;
551                 punch_hole = 0;
552         } else {
553                 if (end + 1 >= inode->i_size) { /* we may free a little more */
554                         limit = (inode->i_size + PAGE_CACHE_SIZE - 1) >>
555                                                         PAGE_CACHE_SHIFT;
556                         upper_limit = SHMEM_MAX_INDEX;
557                 } else {
558                         limit = (end + 1) >> PAGE_CACHE_SHIFT;
559                         upper_limit = limit;
560                 }
561                 needs_lock = &info->lock;
562                 punch_hole = 1;
563         }
564
565         topdir = info->i_indirect;
566         if (topdir && idx <= SHMEM_NR_DIRECT && !punch_hole) {
567                 info->i_indirect = NULL;
568                 nr_pages_to_free++;
569                 list_add(&topdir->lru, &pages_to_free);
570         }
571         spin_unlock(&info->lock);
572
573         if (info->swapped && idx < SHMEM_NR_DIRECT) {
574                 ptr = info->i_direct;
575                 size = limit;
576                 if (size > SHMEM_NR_DIRECT)
577                         size = SHMEM_NR_DIRECT;
578                 nr_swaps_freed = shmem_free_swp(ptr+idx, ptr+size, needs_lock);
579         }
580
581         /*
582          * If there are no indirect blocks or we are punching a hole
583          * below indirect blocks, nothing to be done.
584          */
585         if (!topdir || limit <= SHMEM_NR_DIRECT)
586                 goto done2;
587
588         /*
589          * The truncation case has already dropped info->lock, and we're safe
590          * because i_size and next_index have already been lowered, preventing
591          * access beyond.  But in the punch_hole case, we still need to take
592          * the lock when updating the swap directory, because there might be
593          * racing accesses by shmem_getpage(SGP_CACHE), shmem_unuse_inode or
594          * shmem_writepage.  However, whenever we find we can remove a whole
595          * directory page (not at the misaligned start or end of the range),
596          * we first NULLify its pointer in the level above, and then have no
597          * need to take the lock when updating its contents: needs_lock and
598          * punch_lock (either pointing to info->lock or NULL) manage this.
599          */
600
601         upper_limit -= SHMEM_NR_DIRECT;
602         limit -= SHMEM_NR_DIRECT;
603         idx = (idx > SHMEM_NR_DIRECT)? (idx - SHMEM_NR_DIRECT): 0;
604         offset = idx % ENTRIES_PER_PAGE;
605         idx -= offset;
606
607         dir = shmem_dir_map(topdir);
608         stage = ENTRIES_PER_PAGEPAGE/2;
609         if (idx < ENTRIES_PER_PAGEPAGE/2) {
610                 middir = topdir;
611                 diroff = idx/ENTRIES_PER_PAGE;
612         } else {
613                 dir += ENTRIES_PER_PAGE/2;
614                 dir += (idx - ENTRIES_PER_PAGEPAGE/2)/ENTRIES_PER_PAGEPAGE;
615                 while (stage <= idx)
616                         stage += ENTRIES_PER_PAGEPAGE;
617                 middir = *dir;
618                 if (*dir) {
619                         diroff = ((idx - ENTRIES_PER_PAGEPAGE/2) %
620                                 ENTRIES_PER_PAGEPAGE) / ENTRIES_PER_PAGE;
621                         if (!diroff && !offset && upper_limit >= stage) {
622                                 if (needs_lock) {
623                                         spin_lock(needs_lock);
624                                         *dir = NULL;
625                                         spin_unlock(needs_lock);
626                                         needs_lock = NULL;
627                                 } else
628                                         *dir = NULL;
629                                 nr_pages_to_free++;
630                                 list_add(&middir->lru, &pages_to_free);
631                         }
632                         shmem_dir_unmap(dir);
633                         dir = shmem_dir_map(middir);
634                 } else {
635                         diroff = 0;
636                         offset = 0;
637                         idx = stage;
638                 }
639         }
640
641         for (; idx < limit; idx += ENTRIES_PER_PAGE, diroff++) {
642                 if (unlikely(idx == stage)) {
643                         shmem_dir_unmap(dir);
644                         dir = shmem_dir_map(topdir) +
645                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
646                         while (!*dir) {
647                                 dir++;
648                                 idx += ENTRIES_PER_PAGEPAGE;
649                                 if (idx >= limit)
650                                         goto done1;
651                         }
652                         stage = idx + ENTRIES_PER_PAGEPAGE;
653                         middir = *dir;
654                         if (punch_hole)
655                                 needs_lock = &info->lock;
656                         if (upper_limit >= stage) {
657                                 if (needs_lock) {
658                                         spin_lock(needs_lock);
659                                         *dir = NULL;
660                                         spin_unlock(needs_lock);
661                                         needs_lock = NULL;
662                                 } else
663                                         *dir = NULL;
664                                 nr_pages_to_free++;
665                                 list_add(&middir->lru, &pages_to_free);
666                         }
667                         shmem_dir_unmap(dir);
668                         cond_resched();
669                         dir = shmem_dir_map(middir);
670                         diroff = 0;
671                 }
672                 punch_lock = needs_lock;
673                 subdir = dir[diroff];
674                 if (subdir && !offset && upper_limit-idx >= ENTRIES_PER_PAGE) {
675                         if (needs_lock) {
676                                 spin_lock(needs_lock);
677                                 dir[diroff] = NULL;
678                                 spin_unlock(needs_lock);
679                                 punch_lock = NULL;
680                         } else
681                                 dir[diroff] = NULL;
682                         nr_pages_to_free++;
683                         list_add(&subdir->lru, &pages_to_free);
684                 }
685                 if (subdir && page_private(subdir) /* has swap entries */) {
686                         size = limit - idx;
687                         if (size > ENTRIES_PER_PAGE)
688                                 size = ENTRIES_PER_PAGE;
689                         freed = shmem_map_and_free_swp(subdir,
690                                         offset, size, &dir, punch_lock);
691                         if (!dir)
692                                 dir = shmem_dir_map(middir);
693                         nr_swaps_freed += freed;
694                         if (offset || punch_lock) {
695                                 spin_lock(&info->lock);
696                                 set_page_private(subdir,
697                                         page_private(subdir) - freed);
698                                 spin_unlock(&info->lock);
699                         } else
700                                 BUG_ON(page_private(subdir) != freed);
701                 }
702                 offset = 0;
703         }
704 done1:
705         shmem_dir_unmap(dir);
706 done2:
707         if (inode->i_mapping->nrpages && (info->flags & SHMEM_PAGEIN)) {
708                 /*
709                  * Call truncate_inode_pages again: racing shmem_unuse_inode
710                  * may have swizzled a page in from swap since vmtruncate or
711                  * generic_delete_inode did it, before we lowered next_index.
712                  * Also, though shmem_getpage checks i_size before adding to
713                  * cache, no recheck after: so fix the narrow window there too.
714                  *
715                  * Recalling truncate_inode_pages_range and unmap_mapping_range
716                  * every time for punch_hole (which never got a chance to clear
717                  * SHMEM_PAGEIN at the start of vmtruncate_range) is expensive,
718                  * yet hardly ever necessary: try to optimize them out later.
719                  */
720                 truncate_inode_pages_range(inode->i_mapping, start, end);
721                 if (punch_hole)
722                         unmap_mapping_range(inode->i_mapping, start,
723                                                         end - start, 1);
724         }
725
726         spin_lock(&info->lock);
727         info->flags &= ~SHMEM_TRUNCATE;
728         info->swapped -= nr_swaps_freed;
729         if (nr_pages_to_free)
730                 shmem_free_blocks(inode, nr_pages_to_free);
731         shmem_recalc_inode(inode);
732         spin_unlock(&info->lock);
733
734         /*
735          * Empty swap vector directory pages to be freed?
736          */
737         if (!list_empty(&pages_to_free)) {
738                 pages_to_free.prev->next = NULL;
739                 shmem_free_pages(pages_to_free.next);
740         }
741 }
742
743 static void shmem_truncate(struct inode *inode)
744 {
745         shmem_truncate_range(inode, inode->i_size, (loff_t)-1);
746 }
747
748 static int shmem_notify_change(struct dentry *dentry, struct iattr *attr)
749 {
750         struct inode *inode = dentry->d_inode;
751         struct page *page = NULL;
752         int error;
753
754         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
755                 if (attr->ia_size < inode->i_size) {
756                         /*
757                          * If truncating down to a partial page, then
758                          * if that page is already allocated, hold it
759                          * in memory until the truncation is over, so
760                          * truncate_partial_page cannnot miss it were
761                          * it assigned to swap.
762                          */
763                         if (attr->ia_size & (PAGE_CACHE_SIZE-1)) {
764                                 (void) shmem_getpage(inode,
765                                         attr->ia_size>>PAGE_CACHE_SHIFT,
766                                                 &page, SGP_READ, NULL);
767                                 if (page)
768                                         unlock_page(page);
769                         }
770                         /*
771                          * Reset SHMEM_PAGEIN flag so that shmem_truncate can
772                          * detect if any pages might have been added to cache
773                          * after truncate_inode_pages.  But we needn't bother
774                          * if it's being fully truncated to zero-length: the
775                          * nrpages check is efficient enough in that case.
776                          */
777                         if (attr->ia_size) {
778                                 struct shmem_inode_info *info = SHMEM_I(inode);
779                                 spin_lock(&info->lock);
780                                 info->flags &= ~SHMEM_PAGEIN;
781                                 spin_unlock(&info->lock);
782                         }
783                 }
784         }
785
786         error = inode_change_ok(inode, attr);
787         if (!error)
788                 error = inode_setattr(inode, attr);
789 #ifdef CONFIG_TMPFS_POSIX_ACL
790         if (!error && (attr->ia_valid & ATTR_MODE))
791                 error = generic_acl_chmod(inode, &shmem_acl_ops);
792 #endif
793         if (page)
794                 page_cache_release(page);
795         return error;
796 }
797
798 static void shmem_delete_inode(struct inode *inode)
799 {
800         struct shmem_inode_info *info = SHMEM_I(inode);
801
802         if (inode->i_op->truncate == shmem_truncate) {
803                 truncate_inode_pages(inode->i_mapping, 0);
804                 shmem_unacct_size(info->flags, inode->i_size);
805                 inode->i_size = 0;
806                 shmem_truncate(inode);
807                 if (!list_empty(&info->swaplist)) {
808                         mutex_lock(&shmem_swaplist_mutex);
809                         list_del_init(&info->swaplist);
810                         mutex_unlock(&shmem_swaplist_mutex);
811                 }
812         }
813         BUG_ON(inode->i_blocks);
814         shmem_free_inode(inode->i_sb);
815         clear_inode(inode);
816 }
817
818 static inline int shmem_find_swp(swp_entry_t entry, swp_entry_t *dir, swp_entry_t *edir)
819 {
820         swp_entry_t *ptr;
821
822         for (ptr = dir; ptr < edir; ptr++) {
823                 if (ptr->val == entry.val)
824                         return ptr - dir;
825         }
826         return -1;
827 }
828
829 static int shmem_unuse_inode(struct shmem_inode_info *info, swp_entry_t entry, struct page *page)
830 {
831         struct inode *inode;
832         unsigned long idx;
833         unsigned long size;
834         unsigned long limit;
835         unsigned long stage;
836         struct page **dir;
837         struct page *subdir;
838         swp_entry_t *ptr;
839         int offset;
840         int error;
841
842         idx = 0;
843         ptr = info->i_direct;
844         spin_lock(&info->lock);
845         if (!info->swapped) {
846                 list_del_init(&info->swaplist);
847                 goto lost2;
848         }
849         limit = info->next_index;
850         size = limit;
851         if (size > SHMEM_NR_DIRECT)
852                 size = SHMEM_NR_DIRECT;
853         offset = shmem_find_swp(entry, ptr, ptr+size);
854         if (offset >= 0)
855                 goto found;
856         if (!info->i_indirect)
857                 goto lost2;
858
859         dir = shmem_dir_map(info->i_indirect);
860         stage = SHMEM_NR_DIRECT + ENTRIES_PER_PAGEPAGE/2;
861
862         for (idx = SHMEM_NR_DIRECT; idx < limit; idx += ENTRIES_PER_PAGE, dir++) {
863                 if (unlikely(idx == stage)) {
864                         shmem_dir_unmap(dir-1);
865                         if (cond_resched_lock(&info->lock)) {
866                                 /* check it has not been truncated */
867                                 if (limit > info->next_index) {
868                                         limit = info->next_index;
869                                         if (idx >= limit)
870                                                 goto lost2;
871                                 }
872                         }
873                         dir = shmem_dir_map(info->i_indirect) +
874                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
875                         while (!*dir) {
876                                 dir++;
877                                 idx += ENTRIES_PER_PAGEPAGE;
878                                 if (idx >= limit)
879                                         goto lost1;
880                         }
881                         stage = idx + ENTRIES_PER_PAGEPAGE;
882                         subdir = *dir;
883                         shmem_dir_unmap(dir);
884                         dir = shmem_dir_map(subdir);
885                 }
886                 subdir = *dir;
887                 if (subdir && page_private(subdir)) {
888                         ptr = shmem_swp_map(subdir);
889                         size = limit - idx;
890                         if (size > ENTRIES_PER_PAGE)
891                                 size = ENTRIES_PER_PAGE;
892                         offset = shmem_find_swp(entry, ptr, ptr+size);
893                         shmem_swp_unmap(ptr);
894                         if (offset >= 0) {
895                                 shmem_dir_unmap(dir);
896                                 goto found;
897                         }
898                 }
899         }
900 lost1:
901         shmem_dir_unmap(dir-1);
902 lost2:
903         spin_unlock(&info->lock);
904         return 0;
905 found:
906         idx += offset;
907         inode = igrab(&info->vfs_inode);
908         spin_unlock(&info->lock);
909
910         /*
911          * Move _head_ to start search for next from here.
912          * But be careful: shmem_delete_inode checks list_empty without taking
913          * mutex, and there's an instant in list_move_tail when info->swaplist
914          * would appear empty, if it were the only one on shmem_swaplist.  We
915          * could avoid doing it if inode NULL; or use this minor optimization.
916          */
917         if (shmem_swaplist.next != &info->swaplist)
918                 list_move_tail(&shmem_swaplist, &info->swaplist);
919         mutex_unlock(&shmem_swaplist_mutex);
920
921         error = 1;
922         if (!inode)
923                 goto out;
924         /* Precharge page using GFP_KERNEL while we can wait */
925         error = mem_cgroup_cache_charge(page, current->mm, GFP_KERNEL);
926         if (error)
927                 goto out;
928         error = radix_tree_preload(GFP_KERNEL);
929         if (error) {
930                 mem_cgroup_uncharge_cache_page(page);
931                 goto out;
932         }
933         error = 1;
934
935         spin_lock(&info->lock);
936         ptr = shmem_swp_entry(info, idx, NULL);
937         if (ptr && ptr->val == entry.val) {
938                 error = add_to_page_cache_locked(page, inode->i_mapping,
939                                                 idx, GFP_NOWAIT);
940                 /* does mem_cgroup_uncharge_cache_page on error */
941         } else  /* we must compensate for our precharge above */
942                 mem_cgroup_uncharge_cache_page(page);
943
944         if (error == -EEXIST) {
945                 struct page *filepage = find_get_page(inode->i_mapping, idx);
946                 error = 1;
947                 if (filepage) {
948                         /*
949                          * There might be a more uptodate page coming down
950                          * from a stacked writepage: forget our swappage if so.
951                          */
952                         if (PageUptodate(filepage))
953                                 error = 0;
954                         page_cache_release(filepage);
955                 }
956         }
957         if (!error) {
958                 delete_from_swap_cache(page);
959                 set_page_dirty(page);
960                 info->flags |= SHMEM_PAGEIN;
961                 shmem_swp_set(info, ptr, 0);
962                 swap_free(entry);
963                 error = 1;      /* not an error, but entry was found */
964         }
965         if (ptr)
966                 shmem_swp_unmap(ptr);
967         spin_unlock(&info->lock);
968         radix_tree_preload_end();
969 out:
970         unlock_page(page);
971         page_cache_release(page);
972         iput(inode);            /* allows for NULL */
973         return error;
974 }
975
976 /*
977  * shmem_unuse() search for an eventually swapped out shmem page.
978  */
979 int shmem_unuse(swp_entry_t entry, struct page *page)
980 {
981         struct list_head *p, *next;
982         struct shmem_inode_info *info;
983         int found = 0;
984
985         mutex_lock(&shmem_swaplist_mutex);
986         list_for_each_safe(p, next, &shmem_swaplist) {
987                 info = list_entry(p, struct shmem_inode_info, swaplist);
988                 found = shmem_unuse_inode(info, entry, page);
989                 cond_resched();
990                 if (found)
991                         goto out;
992         }
993         mutex_unlock(&shmem_swaplist_mutex);
994 out:    return found;   /* 0 or 1 or -ENOMEM */
995 }
996
997 /*
998  * Move the page from the page cache to the swap cache.
999  */
1000 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1001 {
1002         struct shmem_inode_info *info;
1003         swp_entry_t *entry, swap;
1004         struct address_space *mapping;
1005         unsigned long index;
1006         struct inode *inode;
1007
1008         BUG_ON(!PageLocked(page));
1009         mapping = page->mapping;
1010         index = page->index;
1011         inode = mapping->host;
1012         info = SHMEM_I(inode);
1013         if (info->flags & VM_LOCKED)
1014                 goto redirty;
1015         if (!total_swap_pages)
1016                 goto redirty;
1017
1018         /*
1019          * shmem_backing_dev_info's capabilities prevent regular writeback or
1020          * sync from ever calling shmem_writepage; but a stacking filesystem
1021          * may use the ->writepage of its underlying filesystem, in which case
1022          * tmpfs should write out to swap only in response to memory pressure,
1023          * and not for pdflush or sync.  However, in those cases, we do still
1024          * want to check if there's a redundant swappage to be discarded.
1025          */
1026         if (wbc->for_reclaim)
1027                 swap = get_swap_page();
1028         else
1029                 swap.val = 0;
1030
1031         spin_lock(&info->lock);
1032         if (index >= info->next_index) {
1033                 BUG_ON(!(info->flags & SHMEM_TRUNCATE));
1034                 goto unlock;
1035         }
1036         entry = shmem_swp_entry(info, index, NULL);
1037         if (entry->val) {
1038                 /*
1039                  * The more uptodate page coming down from a stacked
1040                  * writepage should replace our old swappage.
1041                  */
1042                 free_swap_and_cache(*entry);
1043                 shmem_swp_set(info, entry, 0);
1044         }
1045         shmem_recalc_inode(inode);
1046
1047         if (swap.val && add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
1048                 remove_from_page_cache(page);
1049                 shmem_swp_set(info, entry, swap.val);
1050                 shmem_swp_unmap(entry);
1051                 if (list_empty(&info->swaplist))
1052                         inode = igrab(inode);
1053                 else
1054                         inode = NULL;
1055                 spin_unlock(&info->lock);
1056                 swap_duplicate(swap);
1057                 BUG_ON(page_mapped(page));
1058                 page_cache_release(page);       /* pagecache ref */
1059                 set_page_dirty(page);
1060                 unlock_page(page);
1061                 if (inode) {
1062                         mutex_lock(&shmem_swaplist_mutex);
1063                         /* move instead of add in case we're racing */
1064                         list_move_tail(&info->swaplist, &shmem_swaplist);
1065                         mutex_unlock(&shmem_swaplist_mutex);
1066                         iput(inode);
1067                 }
1068                 return 0;
1069         }
1070
1071         shmem_swp_unmap(entry);
1072 unlock:
1073         spin_unlock(&info->lock);
1074         swap_free(swap);
1075 redirty:
1076         set_page_dirty(page);
1077         if (wbc->for_reclaim)
1078                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
1079         unlock_page(page);
1080         return 0;
1081 }
1082
1083 #ifdef CONFIG_NUMA
1084 #ifdef CONFIG_TMPFS
1085 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1086 {
1087         char buffer[64];
1088
1089         if (!mpol || mpol->mode == MPOL_DEFAULT)
1090                 return;         /* show nothing */
1091
1092         mpol_to_str(buffer, sizeof(buffer), mpol, 1);
1093
1094         seq_printf(seq, ",mpol=%s", buffer);
1095 }
1096
1097 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1098 {
1099         struct mempolicy *mpol = NULL;
1100         if (sbinfo->mpol) {
1101                 spin_lock(&sbinfo->stat_lock);  /* prevent replace/use races */
1102                 mpol = sbinfo->mpol;
1103                 mpol_get(mpol);
1104                 spin_unlock(&sbinfo->stat_lock);
1105         }
1106         return mpol;
1107 }
1108 #endif /* CONFIG_TMPFS */
1109
1110 static struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1111                         struct shmem_inode_info *info, unsigned long idx)
1112 {
1113         struct mempolicy mpol, *spol;
1114         struct vm_area_struct pvma;
1115         struct page *page;
1116
1117         spol = mpol_cond_copy(&mpol,
1118                                 mpol_shared_policy_lookup(&info->policy, idx));
1119
1120         /* Create a pseudo vma that just contains the policy */
1121         pvma.vm_start = 0;
1122         pvma.vm_pgoff = idx;
1123         pvma.vm_ops = NULL;
1124         pvma.vm_policy = spol;
1125         page = swapin_readahead(entry, gfp, &pvma, 0);
1126         return page;
1127 }
1128
1129 static struct page *shmem_alloc_page(gfp_t gfp,
1130                         struct shmem_inode_info *info, unsigned long idx)
1131 {
1132         struct vm_area_struct pvma;
1133
1134         /* Create a pseudo vma that just contains the policy */
1135         pvma.vm_start = 0;
1136         pvma.vm_pgoff = idx;
1137         pvma.vm_ops = NULL;
1138         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, idx);
1139
1140         /*
1141          * alloc_page_vma() will drop the shared policy reference
1142          */
1143         return alloc_page_vma(gfp, &pvma, 0);
1144 }
1145 #else /* !CONFIG_NUMA */
1146 #ifdef CONFIG_TMPFS
1147 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *p)
1148 {
1149 }
1150 #endif /* CONFIG_TMPFS */
1151
1152 static inline struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1153                         struct shmem_inode_info *info, unsigned long idx)
1154 {
1155         return swapin_readahead(entry, gfp, NULL, 0);
1156 }
1157
1158 static inline struct page *shmem_alloc_page(gfp_t gfp,
1159                         struct shmem_inode_info *info, unsigned long idx)
1160 {
1161         return alloc_page(gfp);
1162 }
1163 #endif /* CONFIG_NUMA */
1164
1165 #if !defined(CONFIG_NUMA) || !defined(CONFIG_TMPFS)
1166 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1167 {
1168         return NULL;
1169 }
1170 #endif
1171
1172 /*
1173  * shmem_getpage - either get the page from swap or allocate a new one
1174  *
1175  * If we allocate a new one we do not mark it dirty. That's up to the
1176  * vm. If we swap it in we mark it dirty since we also free the swap
1177  * entry since a page cannot live in both the swap and page cache
1178  */
1179 static int shmem_getpage(struct inode *inode, unsigned long idx,
1180                         struct page **pagep, enum sgp_type sgp, int *type)
1181 {
1182         struct address_space *mapping = inode->i_mapping;
1183         struct shmem_inode_info *info = SHMEM_I(inode);
1184         struct shmem_sb_info *sbinfo;
1185         struct page *filepage = *pagep;
1186         struct page *swappage;
1187         swp_entry_t *entry;
1188         swp_entry_t swap;
1189         gfp_t gfp;
1190         int error;
1191
1192         if (idx >= SHMEM_MAX_INDEX)
1193                 return -EFBIG;
1194
1195         if (type)
1196                 *type = 0;
1197
1198         /*
1199          * Normally, filepage is NULL on entry, and either found
1200          * uptodate immediately, or allocated and zeroed, or read
1201          * in under swappage, which is then assigned to filepage.
1202          * But shmem_readpage (required for splice) passes in a locked
1203          * filepage, which may be found not uptodate by other callers
1204          * too, and may need to be copied from the swappage read in.
1205          */
1206 repeat:
1207         if (!filepage)
1208                 filepage = find_lock_page(mapping, idx);
1209         if (filepage && PageUptodate(filepage))
1210                 goto done;
1211         error = 0;
1212         gfp = mapping_gfp_mask(mapping);
1213         if (!filepage) {
1214                 /*
1215                  * Try to preload while we can wait, to not make a habit of
1216                  * draining atomic reserves; but don't latch on to this cpu.
1217                  */
1218                 error = radix_tree_preload(gfp & ~__GFP_HIGHMEM);
1219                 if (error)
1220                         goto failed;
1221                 radix_tree_preload_end();
1222         }
1223
1224         spin_lock(&info->lock);
1225         shmem_recalc_inode(inode);
1226         entry = shmem_swp_alloc(info, idx, sgp);
1227         if (IS_ERR(entry)) {
1228                 spin_unlock(&info->lock);
1229                 error = PTR_ERR(entry);
1230                 goto failed;
1231         }
1232         swap = *entry;
1233
1234         if (swap.val) {
1235                 /* Look it up and read it in.. */
1236                 swappage = lookup_swap_cache(swap);
1237                 if (!swappage) {
1238                         shmem_swp_unmap(entry);
1239                         /* here we actually do the io */
1240                         if (type && !(*type & VM_FAULT_MAJOR)) {
1241                                 __count_vm_event(PGMAJFAULT);
1242                                 *type |= VM_FAULT_MAJOR;
1243                         }
1244                         spin_unlock(&info->lock);
1245                         swappage = shmem_swapin(swap, gfp, info, idx);
1246                         if (!swappage) {
1247                                 spin_lock(&info->lock);
1248                                 entry = shmem_swp_alloc(info, idx, sgp);
1249                                 if (IS_ERR(entry))
1250                                         error = PTR_ERR(entry);
1251                                 else {
1252                                         if (entry->val == swap.val)
1253                                                 error = -ENOMEM;
1254                                         shmem_swp_unmap(entry);
1255                                 }
1256                                 spin_unlock(&info->lock);
1257                                 if (error)
1258                                         goto failed;
1259                                 goto repeat;
1260                         }
1261                         wait_on_page_locked(swappage);
1262                         page_cache_release(swappage);
1263                         goto repeat;
1264                 }
1265
1266                 /* We have to do this with page locked to prevent races */
1267                 if (!trylock_page(swappage)) {
1268                         shmem_swp_unmap(entry);
1269                         spin_unlock(&info->lock);
1270                         wait_on_page_locked(swappage);
1271                         page_cache_release(swappage);
1272                         goto repeat;
1273                 }
1274                 if (PageWriteback(swappage)) {
1275                         shmem_swp_unmap(entry);
1276                         spin_unlock(&info->lock);
1277                         wait_on_page_writeback(swappage);
1278                         unlock_page(swappage);
1279                         page_cache_release(swappage);
1280                         goto repeat;
1281                 }
1282                 if (!PageUptodate(swappage)) {
1283                         shmem_swp_unmap(entry);
1284                         spin_unlock(&info->lock);
1285                         unlock_page(swappage);
1286                         page_cache_release(swappage);
1287                         error = -EIO;
1288                         goto failed;
1289                 }
1290
1291                 if (filepage) {
1292                         shmem_swp_set(info, entry, 0);
1293                         shmem_swp_unmap(entry);
1294                         delete_from_swap_cache(swappage);
1295                         spin_unlock(&info->lock);
1296                         copy_highpage(filepage, swappage);
1297                         unlock_page(swappage);
1298                         page_cache_release(swappage);
1299                         flush_dcache_page(filepage);
1300                         SetPageUptodate(filepage);
1301                         set_page_dirty(filepage);
1302                         swap_free(swap);
1303                 } else if (!(error = add_to_page_cache_locked(swappage, mapping,
1304                                         idx, GFP_NOWAIT))) {
1305                         info->flags |= SHMEM_PAGEIN;
1306                         shmem_swp_set(info, entry, 0);
1307                         shmem_swp_unmap(entry);
1308                         delete_from_swap_cache(swappage);
1309                         spin_unlock(&info->lock);
1310                         filepage = swappage;
1311                         set_page_dirty(filepage);
1312                         swap_free(swap);
1313                 } else {
1314                         shmem_swp_unmap(entry);
1315                         spin_unlock(&info->lock);
1316                         unlock_page(swappage);
1317                         page_cache_release(swappage);
1318                         if (error == -ENOMEM) {
1319                                 /* allow reclaim from this memory cgroup */
1320                                 error = mem_cgroup_shrink_usage(current->mm,
1321                                                                 gfp);
1322                                 if (error)
1323                                         goto failed;
1324                         }
1325                         goto repeat;
1326                 }
1327         } else if (sgp == SGP_READ && !filepage) {
1328                 shmem_swp_unmap(entry);
1329                 filepage = find_get_page(mapping, idx);
1330                 if (filepage &&
1331                     (!PageUptodate(filepage) || !trylock_page(filepage))) {
1332                         spin_unlock(&info->lock);
1333                         wait_on_page_locked(filepage);
1334                         page_cache_release(filepage);
1335                         filepage = NULL;
1336                         goto repeat;
1337                 }
1338                 spin_unlock(&info->lock);
1339         } else {
1340                 shmem_swp_unmap(entry);
1341                 sbinfo = SHMEM_SB(inode->i_sb);
1342                 if (sbinfo->max_blocks) {
1343                         spin_lock(&sbinfo->stat_lock);
1344                         if (sbinfo->free_blocks == 0 ||
1345                             shmem_acct_block(info->flags)) {
1346                                 spin_unlock(&sbinfo->stat_lock);
1347                                 spin_unlock(&info->lock);
1348                                 error = -ENOSPC;
1349                                 goto failed;
1350                         }
1351                         sbinfo->free_blocks--;
1352                         inode->i_blocks += BLOCKS_PER_PAGE;
1353                         spin_unlock(&sbinfo->stat_lock);
1354                 } else if (shmem_acct_block(info->flags)) {
1355                         spin_unlock(&info->lock);
1356                         error = -ENOSPC;
1357                         goto failed;
1358                 }
1359
1360                 if (!filepage) {
1361                         int ret;
1362
1363                         spin_unlock(&info->lock);
1364                         filepage = shmem_alloc_page(gfp, info, idx);
1365                         if (!filepage) {
1366                                 shmem_unacct_blocks(info->flags, 1);
1367                                 shmem_free_blocks(inode, 1);
1368                                 error = -ENOMEM;
1369                                 goto failed;
1370                         }
1371                         SetPageSwapBacked(filepage);
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                 mapping_set_unevictable(file->f_mapping);
1482         }
1483         if (!lock && (info->flags & VM_LOCKED) && user) {
1484                 user_shm_unlock(inode->i_size, user);
1485                 info->flags &= ~VM_LOCKED;
1486                 mapping_clear_unevictable(file->f_mapping);
1487                 scan_mapping_unevictable_pages(file->f_mapping);
1488         }
1489         retval = 0;
1490
1491 out_nomem:
1492         spin_unlock(&info->lock);
1493         return retval;
1494 }
1495
1496 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1497 {
1498         file_accessed(file);
1499         vma->vm_ops = &shmem_vm_ops;
1500         vma->vm_flags |= VM_CAN_NONLINEAR;
1501         return 0;
1502 }
1503
1504 static struct inode *
1505 shmem_get_inode(struct super_block *sb, int mode, dev_t dev)
1506 {
1507         struct inode *inode;
1508         struct shmem_inode_info *info;
1509         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1510
1511         if (shmem_reserve_inode(sb))
1512                 return NULL;
1513
1514         inode = new_inode(sb);
1515         if (inode) {
1516                 inode->i_mode = mode;
1517                 inode->i_uid = current_fsuid();
1518                 inode->i_gid = current_fsgid();
1519                 inode->i_blocks = 0;
1520                 inode->i_mapping->backing_dev_info = &shmem_backing_dev_info;
1521                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1522                 inode->i_generation = get_seconds();
1523                 info = SHMEM_I(inode);
1524                 memset(info, 0, (char *)inode - (char *)info);
1525                 spin_lock_init(&info->lock);
1526                 INIT_LIST_HEAD(&info->swaplist);
1527
1528                 switch (mode & S_IFMT) {
1529                 default:
1530                         inode->i_op = &shmem_special_inode_operations;
1531                         init_special_inode(inode, mode, dev);
1532                         break;
1533                 case S_IFREG:
1534                         inode->i_mapping->a_ops = &shmem_aops;
1535                         inode->i_op = &shmem_inode_operations;
1536                         inode->i_fop = &shmem_file_operations;
1537                         mpol_shared_policy_init(&info->policy,
1538                                                  shmem_get_sbmpol(sbinfo));
1539                         break;
1540                 case S_IFDIR:
1541                         inc_nlink(inode);
1542                         /* Some things misbehave if size == 0 on a directory */
1543                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
1544                         inode->i_op = &shmem_dir_inode_operations;
1545                         inode->i_fop = &simple_dir_operations;
1546                         break;
1547                 case S_IFLNK:
1548                         /*
1549                          * Must not load anything in the rbtree,
1550                          * mpol_free_shared_policy will not be called.
1551                          */
1552                         mpol_shared_policy_init(&info->policy, NULL);
1553                         break;
1554                 }
1555         } else
1556                 shmem_free_inode(sb);
1557         return inode;
1558 }
1559
1560 #ifdef CONFIG_TMPFS
1561 static const struct inode_operations shmem_symlink_inode_operations;
1562 static const struct inode_operations shmem_symlink_inline_operations;
1563
1564 /*
1565  * Normally tmpfs avoids the use of shmem_readpage and shmem_write_begin;
1566  * but providing them allows a tmpfs file to be used for splice, sendfile, and
1567  * below the loop driver, in the generic fashion that many filesystems support.
1568  */
1569 static int shmem_readpage(struct file *file, struct page *page)
1570 {
1571         struct inode *inode = page->mapping->host;
1572         int error = shmem_getpage(inode, page->index, &page, SGP_CACHE, NULL);
1573         unlock_page(page);
1574         return error;
1575 }
1576
1577 static int
1578 shmem_write_begin(struct file *file, struct address_space *mapping,
1579                         loff_t pos, unsigned len, unsigned flags,
1580                         struct page **pagep, void **fsdata)
1581 {
1582         struct inode *inode = mapping->host;
1583         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1584         *pagep = NULL;
1585         return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1586 }
1587
1588 static int
1589 shmem_write_end(struct file *file, struct address_space *mapping,
1590                         loff_t pos, unsigned len, unsigned copied,
1591                         struct page *page, void *fsdata)
1592 {
1593         struct inode *inode = mapping->host;
1594
1595         if (pos + copied > inode->i_size)
1596                 i_size_write(inode, pos + copied);
1597
1598         unlock_page(page);
1599         set_page_dirty(page);
1600         page_cache_release(page);
1601
1602         return copied;
1603 }
1604
1605 static void do_shmem_file_read(struct file *filp, loff_t *ppos, read_descriptor_t *desc, read_actor_t actor)
1606 {
1607         struct inode *inode = filp->f_path.dentry->d_inode;
1608         struct address_space *mapping = inode->i_mapping;
1609         unsigned long index, offset;
1610         enum sgp_type sgp = SGP_READ;
1611
1612         /*
1613          * Might this read be for a stacking filesystem?  Then when reading
1614          * holes of a sparse file, we actually need to allocate those pages,
1615          * and even mark them dirty, so it cannot exceed the max_blocks limit.
1616          */
1617         if (segment_eq(get_fs(), KERNEL_DS))
1618                 sgp = SGP_DIRTY;
1619
1620         index = *ppos >> PAGE_CACHE_SHIFT;
1621         offset = *ppos & ~PAGE_CACHE_MASK;
1622
1623         for (;;) {
1624                 struct page *page = NULL;
1625                 unsigned long end_index, nr, ret;
1626                 loff_t i_size = i_size_read(inode);
1627
1628                 end_index = i_size >> PAGE_CACHE_SHIFT;
1629                 if (index > end_index)
1630                         break;
1631                 if (index == end_index) {
1632                         nr = i_size & ~PAGE_CACHE_MASK;
1633                         if (nr <= offset)
1634                                 break;
1635                 }
1636
1637                 desc->error = shmem_getpage(inode, index, &page, sgp, NULL);
1638                 if (desc->error) {
1639                         if (desc->error == -EINVAL)
1640                                 desc->error = 0;
1641                         break;
1642                 }
1643                 if (page)
1644                         unlock_page(page);
1645
1646                 /*
1647                  * We must evaluate after, since reads (unlike writes)
1648                  * are called without i_mutex protection against truncate
1649                  */
1650                 nr = PAGE_CACHE_SIZE;
1651                 i_size = i_size_read(inode);
1652                 end_index = i_size >> PAGE_CACHE_SHIFT;
1653                 if (index == end_index) {
1654                         nr = i_size & ~PAGE_CACHE_MASK;
1655                         if (nr <= offset) {
1656                                 if (page)
1657                                         page_cache_release(page);
1658                                 break;
1659                         }
1660                 }
1661                 nr -= offset;
1662
1663                 if (page) {
1664                         /*
1665                          * If users can be writing to this page using arbitrary
1666                          * virtual addresses, take care about potential aliasing
1667                          * before reading the page on the kernel side.
1668                          */
1669                         if (mapping_writably_mapped(mapping))
1670                                 flush_dcache_page(page);
1671                         /*
1672                          * Mark the page accessed if we read the beginning.
1673                          */
1674                         if (!offset)
1675                                 mark_page_accessed(page);
1676                 } else {
1677                         page = ZERO_PAGE(0);
1678                         page_cache_get(page);
1679                 }
1680
1681                 /*
1682                  * Ok, we have the page, and it's up-to-date, so
1683                  * now we can copy it to user space...
1684                  *
1685                  * The actor routine returns how many bytes were actually used..
1686                  * NOTE! This may not be the same as how much of a user buffer
1687                  * we filled up (we may be padding etc), so we can only update
1688                  * "pos" here (the actor routine has to update the user buffer
1689                  * pointers and the remaining count).
1690                  */
1691                 ret = actor(desc, page, offset, nr);
1692                 offset += ret;
1693                 index += offset >> PAGE_CACHE_SHIFT;
1694                 offset &= ~PAGE_CACHE_MASK;
1695
1696                 page_cache_release(page);
1697                 if (ret != nr || !desc->count)
1698                         break;
1699
1700                 cond_resched();
1701         }
1702
1703         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1704         file_accessed(filp);
1705 }
1706
1707 static ssize_t shmem_file_aio_read(struct kiocb *iocb,
1708                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
1709 {
1710         struct file *filp = iocb->ki_filp;
1711         ssize_t retval;
1712         unsigned long seg;
1713         size_t count;
1714         loff_t *ppos = &iocb->ki_pos;
1715
1716         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1717         if (retval)
1718                 return retval;
1719
1720         for (seg = 0; seg < nr_segs; seg++) {
1721                 read_descriptor_t desc;
1722
1723                 desc.written = 0;
1724                 desc.arg.buf = iov[seg].iov_base;
1725                 desc.count = iov[seg].iov_len;
1726                 if (desc.count == 0)
1727                         continue;
1728                 desc.error = 0;
1729                 do_shmem_file_read(filp, ppos, &desc, file_read_actor);
1730                 retval += desc.written;
1731                 if (desc.error) {
1732                         retval = retval ?: desc.error;
1733                         break;
1734                 }
1735                 if (desc.count > 0)
1736                         break;
1737         }
1738         return retval;
1739 }
1740
1741 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
1742 {
1743         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
1744
1745         buf->f_type = TMPFS_MAGIC;
1746         buf->f_bsize = PAGE_CACHE_SIZE;
1747         buf->f_namelen = NAME_MAX;
1748         spin_lock(&sbinfo->stat_lock);
1749         if (sbinfo->max_blocks) {
1750                 buf->f_blocks = sbinfo->max_blocks;
1751                 buf->f_bavail = buf->f_bfree = sbinfo->free_blocks;
1752         }
1753         if (sbinfo->max_inodes) {
1754                 buf->f_files = sbinfo->max_inodes;
1755                 buf->f_ffree = sbinfo->free_inodes;
1756         }
1757         /* else leave those fields 0 like simple_statfs */
1758         spin_unlock(&sbinfo->stat_lock);
1759         return 0;
1760 }
1761
1762 /*
1763  * File creation. Allocate an inode, and we're done..
1764  */
1765 static int
1766 shmem_mknod(struct inode *dir, struct dentry *dentry, int mode, dev_t dev)
1767 {
1768         struct inode *inode = shmem_get_inode(dir->i_sb, mode, dev);
1769         int error = -ENOSPC;
1770
1771         if (inode) {
1772                 error = security_inode_init_security(inode, dir, NULL, NULL,
1773                                                      NULL);
1774                 if (error) {
1775                         if (error != -EOPNOTSUPP) {
1776                                 iput(inode);
1777                                 return error;
1778                         }
1779                 }
1780                 error = shmem_acl_init(inode, dir);
1781                 if (error) {
1782                         iput(inode);
1783                         return error;
1784                 }
1785                 if (dir->i_mode & S_ISGID) {
1786                         inode->i_gid = dir->i_gid;
1787                         if (S_ISDIR(mode))
1788                                 inode->i_mode |= S_ISGID;
1789                 }
1790                 dir->i_size += BOGO_DIRENT_SIZE;
1791                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1792                 d_instantiate(dentry, inode);
1793                 dget(dentry); /* Extra count - pin the dentry in core */
1794         }
1795         return error;
1796 }
1797
1798 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, int mode)
1799 {
1800         int error;
1801
1802         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
1803                 return error;
1804         inc_nlink(dir);
1805         return 0;
1806 }
1807
1808 static int shmem_create(struct inode *dir, struct dentry *dentry, int mode,
1809                 struct nameidata *nd)
1810 {
1811         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
1812 }
1813
1814 /*
1815  * Link a file..
1816  */
1817 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
1818 {
1819         struct inode *inode = old_dentry->d_inode;
1820         int ret;
1821
1822         /*
1823          * No ordinary (disk based) filesystem counts links as inodes;
1824          * but each new link needs a new dentry, pinning lowmem, and
1825          * tmpfs dentries cannot be pruned until they are unlinked.
1826          */
1827         ret = shmem_reserve_inode(inode->i_sb);
1828         if (ret)
1829                 goto out;
1830
1831         dir->i_size += BOGO_DIRENT_SIZE;
1832         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1833         inc_nlink(inode);
1834         atomic_inc(&inode->i_count);    /* New dentry reference */
1835         dget(dentry);           /* Extra pinning count for the created dentry */
1836         d_instantiate(dentry, inode);
1837 out:
1838         return ret;
1839 }
1840
1841 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
1842 {
1843         struct inode *inode = dentry->d_inode;
1844
1845         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
1846                 shmem_free_inode(inode->i_sb);
1847
1848         dir->i_size -= BOGO_DIRENT_SIZE;
1849         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1850         drop_nlink(inode);
1851         dput(dentry);   /* Undo the count from "create" - this does all the work */
1852         return 0;
1853 }
1854
1855 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
1856 {
1857         if (!simple_empty(dentry))
1858                 return -ENOTEMPTY;
1859
1860         drop_nlink(dentry->d_inode);
1861         drop_nlink(dir);
1862         return shmem_unlink(dir, dentry);
1863 }
1864
1865 /*
1866  * The VFS layer already does all the dentry stuff for rename,
1867  * we just have to decrement the usage count for the target if
1868  * it exists so that the VFS layer correctly free's it when it
1869  * gets overwritten.
1870  */
1871 static int shmem_rename(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
1872 {
1873         struct inode *inode = old_dentry->d_inode;
1874         int they_are_dirs = S_ISDIR(inode->i_mode);
1875
1876         if (!simple_empty(new_dentry))
1877                 return -ENOTEMPTY;
1878
1879         if (new_dentry->d_inode) {
1880                 (void) shmem_unlink(new_dir, new_dentry);
1881                 if (they_are_dirs)
1882                         drop_nlink(old_dir);
1883         } else if (they_are_dirs) {
1884                 drop_nlink(old_dir);
1885                 inc_nlink(new_dir);
1886         }
1887
1888         old_dir->i_size -= BOGO_DIRENT_SIZE;
1889         new_dir->i_size += BOGO_DIRENT_SIZE;
1890         old_dir->i_ctime = old_dir->i_mtime =
1891         new_dir->i_ctime = new_dir->i_mtime =
1892         inode->i_ctime = CURRENT_TIME;
1893         return 0;
1894 }
1895
1896 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
1897 {
1898         int error;
1899         int len;
1900         struct inode *inode;
1901         struct page *page = NULL;
1902         char *kaddr;
1903         struct shmem_inode_info *info;
1904
1905         len = strlen(symname) + 1;
1906         if (len > PAGE_CACHE_SIZE)
1907                 return -ENAMETOOLONG;
1908
1909         inode = shmem_get_inode(dir->i_sb, S_IFLNK|S_IRWXUGO, 0);
1910         if (!inode)
1911                 return -ENOSPC;
1912
1913         error = security_inode_init_security(inode, dir, NULL, NULL,
1914                                              NULL);
1915         if (error) {
1916                 if (error != -EOPNOTSUPP) {
1917                         iput(inode);
1918                         return error;
1919                 }
1920                 error = 0;
1921         }
1922
1923         info = SHMEM_I(inode);
1924         inode->i_size = len-1;
1925         if (len <= (char *)inode - (char *)info) {
1926                 /* do it inline */
1927                 memcpy(info, symname, len);
1928                 inode->i_op = &shmem_symlink_inline_operations;
1929         } else {
1930                 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
1931                 if (error) {
1932                         iput(inode);
1933                         return error;
1934                 }
1935                 unlock_page(page);
1936                 inode->i_mapping->a_ops = &shmem_aops;
1937                 inode->i_op = &shmem_symlink_inode_operations;
1938                 kaddr = kmap_atomic(page, KM_USER0);
1939                 memcpy(kaddr, symname, len);
1940                 kunmap_atomic(kaddr, KM_USER0);
1941                 set_page_dirty(page);
1942                 page_cache_release(page);
1943         }
1944         if (dir->i_mode & S_ISGID)
1945                 inode->i_gid = dir->i_gid;
1946         dir->i_size += BOGO_DIRENT_SIZE;
1947         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1948         d_instantiate(dentry, inode);
1949         dget(dentry);
1950         return 0;
1951 }
1952
1953 static void *shmem_follow_link_inline(struct dentry *dentry, struct nameidata *nd)
1954 {
1955         nd_set_link(nd, (char *)SHMEM_I(dentry->d_inode));
1956         return NULL;
1957 }
1958
1959 static void *shmem_follow_link(struct dentry *dentry, struct nameidata *nd)
1960 {
1961         struct page *page = NULL;
1962         int res = shmem_getpage(dentry->d_inode, 0, &page, SGP_READ, NULL);
1963         nd_set_link(nd, res ? ERR_PTR(res) : kmap(page));
1964         if (page)
1965                 unlock_page(page);
1966         return page;
1967 }
1968
1969 static void shmem_put_link(struct dentry *dentry, struct nameidata *nd, void *cookie)
1970 {
1971         if (!IS_ERR(nd_get_link(nd))) {
1972                 struct page *page = cookie;
1973                 kunmap(page);
1974                 mark_page_accessed(page);
1975                 page_cache_release(page);
1976         }
1977 }
1978
1979 static const struct inode_operations shmem_symlink_inline_operations = {
1980         .readlink       = generic_readlink,
1981         .follow_link    = shmem_follow_link_inline,
1982 };
1983
1984 static const struct inode_operations shmem_symlink_inode_operations = {
1985         .truncate       = shmem_truncate,
1986         .readlink       = generic_readlink,
1987         .follow_link    = shmem_follow_link,
1988         .put_link       = shmem_put_link,
1989 };
1990
1991 #ifdef CONFIG_TMPFS_POSIX_ACL
1992 /*
1993  * Superblocks without xattr inode operations will get security.* xattr
1994  * support from the VFS "for free". As soon as we have any other xattrs
1995  * like ACLs, we also need to implement the security.* handlers at
1996  * filesystem level, though.
1997  */
1998
1999 static size_t shmem_xattr_security_list(struct inode *inode, char *list,
2000                                         size_t list_len, const char *name,
2001                                         size_t name_len)
2002 {
2003         return security_inode_listsecurity(inode, list, list_len);
2004 }
2005
2006 static int shmem_xattr_security_get(struct inode *inode, const char *name,
2007                                     void *buffer, size_t size)
2008 {
2009         if (strcmp(name, "") == 0)
2010                 return -EINVAL;
2011         return xattr_getsecurity(inode, name, buffer, size);
2012 }
2013
2014 static int shmem_xattr_security_set(struct inode *inode, const char *name,
2015                                     const void *value, size_t size, int flags)
2016 {
2017         if (strcmp(name, "") == 0)
2018                 return -EINVAL;
2019         return security_inode_setsecurity(inode, name, value, size, flags);
2020 }
2021
2022 static struct xattr_handler shmem_xattr_security_handler = {
2023         .prefix = XATTR_SECURITY_PREFIX,
2024         .list   = shmem_xattr_security_list,
2025         .get    = shmem_xattr_security_get,
2026         .set    = shmem_xattr_security_set,
2027 };
2028
2029 static struct xattr_handler *shmem_xattr_handlers[] = {
2030         &shmem_xattr_acl_access_handler,
2031         &shmem_xattr_acl_default_handler,
2032         &shmem_xattr_security_handler,
2033         NULL
2034 };
2035 #endif
2036
2037 static struct dentry *shmem_get_parent(struct dentry *child)
2038 {
2039         return ERR_PTR(-ESTALE);
2040 }
2041
2042 static int shmem_match(struct inode *ino, void *vfh)
2043 {
2044         __u32 *fh = vfh;
2045         __u64 inum = fh[2];
2046         inum = (inum << 32) | fh[1];
2047         return ino->i_ino == inum && fh[0] == ino->i_generation;
2048 }
2049
2050 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
2051                 struct fid *fid, int fh_len, int fh_type)
2052 {
2053         struct inode *inode;
2054         struct dentry *dentry = NULL;
2055         u64 inum = fid->raw[2];
2056         inum = (inum << 32) | fid->raw[1];
2057
2058         if (fh_len < 3)
2059                 return NULL;
2060
2061         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
2062                         shmem_match, fid->raw);
2063         if (inode) {
2064                 dentry = d_find_alias(inode);
2065                 iput(inode);
2066         }
2067
2068         return dentry;
2069 }
2070
2071 static int shmem_encode_fh(struct dentry *dentry, __u32 *fh, int *len,
2072                                 int connectable)
2073 {
2074         struct inode *inode = dentry->d_inode;
2075
2076         if (*len < 3)
2077                 return 255;
2078
2079         if (hlist_unhashed(&inode->i_hash)) {
2080                 /* Unfortunately insert_inode_hash is not idempotent,
2081                  * so as we hash inodes here rather than at creation
2082                  * time, we need a lock to ensure we only try
2083                  * to do it once
2084                  */
2085                 static DEFINE_SPINLOCK(lock);
2086                 spin_lock(&lock);
2087                 if (hlist_unhashed(&inode->i_hash))
2088                         __insert_inode_hash(inode,
2089                                             inode->i_ino + inode->i_generation);
2090                 spin_unlock(&lock);
2091         }
2092
2093         fh[0] = inode->i_generation;
2094         fh[1] = inode->i_ino;
2095         fh[2] = ((__u64)inode->i_ino) >> 32;
2096
2097         *len = 3;
2098         return 1;
2099 }
2100
2101 static const struct export_operations shmem_export_ops = {
2102         .get_parent     = shmem_get_parent,
2103         .encode_fh      = shmem_encode_fh,
2104         .fh_to_dentry   = shmem_fh_to_dentry,
2105 };
2106
2107 static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
2108                                bool remount)
2109 {
2110         char *this_char, *value, *rest;
2111
2112         while (options != NULL) {
2113                 this_char = options;
2114                 for (;;) {
2115                         /*
2116                          * NUL-terminate this option: unfortunately,
2117                          * mount options form a comma-separated list,
2118                          * but mpol's nodelist may also contain commas.
2119                          */
2120                         options = strchr(options, ',');
2121                         if (options == NULL)
2122                                 break;
2123                         options++;
2124                         if (!isdigit(*options)) {
2125                                 options[-1] = '\0';
2126                                 break;
2127                         }
2128                 }
2129                 if (!*this_char)
2130                         continue;
2131                 if ((value = strchr(this_char,'=')) != NULL) {
2132                         *value++ = 0;
2133                 } else {
2134                         printk(KERN_ERR
2135                             "tmpfs: No value for mount option '%s'\n",
2136                             this_char);
2137                         return 1;
2138                 }
2139
2140                 if (!strcmp(this_char,"size")) {
2141                         unsigned long long size;
2142                         size = memparse(value,&rest);
2143                         if (*rest == '%') {
2144                                 size <<= PAGE_SHIFT;
2145                                 size *= totalram_pages;
2146                                 do_div(size, 100);
2147                                 rest++;
2148                         }
2149                         if (*rest)
2150                                 goto bad_val;
2151                         sbinfo->max_blocks =
2152                                 DIV_ROUND_UP(size, PAGE_CACHE_SIZE);
2153                 } else if (!strcmp(this_char,"nr_blocks")) {
2154                         sbinfo->max_blocks = memparse(value, &rest);
2155                         if (*rest)
2156                                 goto bad_val;
2157                 } else if (!strcmp(this_char,"nr_inodes")) {
2158                         sbinfo->max_inodes = memparse(value, &rest);
2159                         if (*rest)
2160                                 goto bad_val;
2161                 } else if (!strcmp(this_char,"mode")) {
2162                         if (remount)
2163                                 continue;
2164                         sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
2165                         if (*rest)
2166                                 goto bad_val;
2167                 } else if (!strcmp(this_char,"uid")) {
2168                         if (remount)
2169                                 continue;
2170                         sbinfo->uid = simple_strtoul(value, &rest, 0);
2171                         if (*rest)
2172                                 goto bad_val;
2173                 } else if (!strcmp(this_char,"gid")) {
2174                         if (remount)
2175                                 continue;
2176                         sbinfo->gid = simple_strtoul(value, &rest, 0);
2177                         if (*rest)
2178                                 goto bad_val;
2179                 } else if (!strcmp(this_char,"mpol")) {
2180                         if (mpol_parse_str(value, &sbinfo->mpol, 1))
2181                                 goto bad_val;
2182                 } else {
2183                         printk(KERN_ERR "tmpfs: Bad mount option %s\n",
2184                                this_char);
2185                         return 1;
2186                 }
2187         }
2188         return 0;
2189
2190 bad_val:
2191         printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
2192                value, this_char);
2193         return 1;
2194
2195 }
2196
2197 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2198 {
2199         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2200         struct shmem_sb_info config = *sbinfo;
2201         unsigned long blocks;
2202         unsigned long inodes;
2203         int error = -EINVAL;
2204
2205         if (shmem_parse_options(data, &config, true))
2206                 return error;
2207
2208         spin_lock(&sbinfo->stat_lock);
2209         blocks = sbinfo->max_blocks - sbinfo->free_blocks;
2210         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2211         if (config.max_blocks < blocks)
2212                 goto out;
2213         if (config.max_inodes < inodes)
2214                 goto out;
2215         /*
2216          * Those tests also disallow limited->unlimited while any are in
2217          * use, so i_blocks will always be zero when max_blocks is zero;
2218          * but we must separately disallow unlimited->limited, because
2219          * in that case we have no record of how much is already in use.
2220          */
2221         if (config.max_blocks && !sbinfo->max_blocks)
2222                 goto out;
2223         if (config.max_inodes && !sbinfo->max_inodes)
2224                 goto out;
2225
2226         error = 0;
2227         sbinfo->max_blocks  = config.max_blocks;
2228         sbinfo->free_blocks = config.max_blocks - blocks;
2229         sbinfo->max_inodes  = config.max_inodes;
2230         sbinfo->free_inodes = config.max_inodes - inodes;
2231
2232         mpol_put(sbinfo->mpol);
2233         sbinfo->mpol        = config.mpol;      /* transfers initial ref */
2234 out:
2235         spin_unlock(&sbinfo->stat_lock);
2236         return error;
2237 }
2238
2239 static int shmem_show_options(struct seq_file *seq, struct vfsmount *vfs)
2240 {
2241         struct shmem_sb_info *sbinfo = SHMEM_SB(vfs->mnt_sb);
2242
2243         if (sbinfo->max_blocks != shmem_default_max_blocks())
2244                 seq_printf(seq, ",size=%luk",
2245                         sbinfo->max_blocks << (PAGE_CACHE_SHIFT - 10));
2246         if (sbinfo->max_inodes != shmem_default_max_inodes())
2247                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
2248         if (sbinfo->mode != (S_IRWXUGO | S_ISVTX))
2249                 seq_printf(seq, ",mode=%03o", sbinfo->mode);
2250         if (sbinfo->uid != 0)
2251                 seq_printf(seq, ",uid=%u", sbinfo->uid);
2252         if (sbinfo->gid != 0)
2253                 seq_printf(seq, ",gid=%u", sbinfo->gid);
2254         shmem_show_mpol(seq, sbinfo->mpol);
2255         return 0;
2256 }
2257 #endif /* CONFIG_TMPFS */
2258
2259 static void shmem_put_super(struct super_block *sb)
2260 {
2261         kfree(sb->s_fs_info);
2262         sb->s_fs_info = NULL;
2263 }
2264
2265 static int shmem_fill_super(struct super_block *sb,
2266                             void *data, int silent)
2267 {
2268         struct inode *inode;
2269         struct dentry *root;
2270         struct shmem_sb_info *sbinfo;
2271         int err = -ENOMEM;
2272
2273         /* Round up to L1_CACHE_BYTES to resist false sharing */
2274         sbinfo = kmalloc(max((int)sizeof(struct shmem_sb_info),
2275                                 L1_CACHE_BYTES), GFP_KERNEL);
2276         if (!sbinfo)
2277                 return -ENOMEM;
2278
2279         sbinfo->max_blocks = 0;
2280         sbinfo->max_inodes = 0;
2281         sbinfo->mode = S_IRWXUGO | S_ISVTX;
2282         sbinfo->uid = current_fsuid();
2283         sbinfo->gid = current_fsgid();
2284         sbinfo->mpol = NULL;
2285         sb->s_fs_info = sbinfo;
2286
2287 #ifdef CONFIG_TMPFS
2288         /*
2289          * Per default we only allow half of the physical ram per
2290          * tmpfs instance, limiting inodes to one per page of lowmem;
2291          * but the internal instance is left unlimited.
2292          */
2293         if (!(sb->s_flags & MS_NOUSER)) {
2294                 sbinfo->max_blocks = shmem_default_max_blocks();
2295                 sbinfo->max_inodes = shmem_default_max_inodes();
2296                 if (shmem_parse_options(data, sbinfo, false)) {
2297                         err = -EINVAL;
2298                         goto failed;
2299                 }
2300         }
2301         sb->s_export_op = &shmem_export_ops;
2302 #else
2303         sb->s_flags |= MS_NOUSER;
2304 #endif
2305
2306         spin_lock_init(&sbinfo->stat_lock);
2307         sbinfo->free_blocks = sbinfo->max_blocks;
2308         sbinfo->free_inodes = sbinfo->max_inodes;
2309
2310         sb->s_maxbytes = SHMEM_MAX_BYTES;
2311         sb->s_blocksize = PAGE_CACHE_SIZE;
2312         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
2313         sb->s_magic = TMPFS_MAGIC;
2314         sb->s_op = &shmem_ops;
2315         sb->s_time_gran = 1;
2316 #ifdef CONFIG_TMPFS_POSIX_ACL
2317         sb->s_xattr = shmem_xattr_handlers;
2318         sb->s_flags |= MS_POSIXACL;
2319 #endif
2320
2321         inode = shmem_get_inode(sb, S_IFDIR | sbinfo->mode, 0);
2322         if (!inode)
2323                 goto failed;
2324         inode->i_uid = sbinfo->uid;
2325         inode->i_gid = sbinfo->gid;
2326         root = d_alloc_root(inode);
2327         if (!root)
2328                 goto failed_iput;
2329         sb->s_root = root;
2330         return 0;
2331
2332 failed_iput:
2333         iput(inode);
2334 failed:
2335         shmem_put_super(sb);
2336         return err;
2337 }
2338
2339 static struct kmem_cache *shmem_inode_cachep;
2340
2341 static struct inode *shmem_alloc_inode(struct super_block *sb)
2342 {
2343         struct shmem_inode_info *p;
2344         p = (struct shmem_inode_info *)kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
2345         if (!p)
2346                 return NULL;
2347         return &p->vfs_inode;
2348 }
2349
2350 static void shmem_destroy_inode(struct inode *inode)
2351 {
2352         if ((inode->i_mode & S_IFMT) == S_IFREG) {
2353                 /* only struct inode is valid if it's an inline symlink */
2354                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
2355         }
2356         shmem_acl_destroy_inode(inode);
2357         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
2358 }
2359
2360 static void init_once(void *foo)
2361 {
2362         struct shmem_inode_info *p = (struct shmem_inode_info *) foo;
2363
2364         inode_init_once(&p->vfs_inode);
2365 #ifdef CONFIG_TMPFS_POSIX_ACL
2366         p->i_acl = NULL;
2367         p->i_default_acl = NULL;
2368 #endif
2369 }
2370
2371 static int init_inodecache(void)
2372 {
2373         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
2374                                 sizeof(struct shmem_inode_info),
2375                                 0, SLAB_PANIC, init_once);
2376         return 0;
2377 }
2378
2379 static void destroy_inodecache(void)
2380 {
2381         kmem_cache_destroy(shmem_inode_cachep);
2382 }
2383
2384 static const struct address_space_operations shmem_aops = {
2385         .writepage      = shmem_writepage,
2386         .set_page_dirty = __set_page_dirty_no_writeback,
2387 #ifdef CONFIG_TMPFS
2388         .readpage       = shmem_readpage,
2389         .write_begin    = shmem_write_begin,
2390         .write_end      = shmem_write_end,
2391 #endif
2392         .migratepage    = migrate_page,
2393 };
2394
2395 static const struct file_operations shmem_file_operations = {
2396         .mmap           = shmem_mmap,
2397 #ifdef CONFIG_TMPFS
2398         .llseek         = generic_file_llseek,
2399         .read           = do_sync_read,
2400         .write          = do_sync_write,
2401         .aio_read       = shmem_file_aio_read,
2402         .aio_write      = generic_file_aio_write,
2403         .fsync          = simple_sync_file,
2404         .splice_read    = generic_file_splice_read,
2405         .splice_write   = generic_file_splice_write,
2406 #endif
2407 };
2408
2409 static const struct inode_operations shmem_inode_operations = {
2410         .truncate       = shmem_truncate,
2411         .setattr        = shmem_notify_change,
2412         .truncate_range = shmem_truncate_range,
2413 #ifdef CONFIG_TMPFS_POSIX_ACL
2414         .setxattr       = generic_setxattr,
2415         .getxattr       = generic_getxattr,
2416         .listxattr      = generic_listxattr,
2417         .removexattr    = generic_removexattr,
2418         .permission     = shmem_permission,
2419 #endif
2420
2421 };
2422
2423 static const struct inode_operations shmem_dir_inode_operations = {
2424 #ifdef CONFIG_TMPFS
2425         .create         = shmem_create,
2426         .lookup         = simple_lookup,
2427         .link           = shmem_link,
2428         .unlink         = shmem_unlink,
2429         .symlink        = shmem_symlink,
2430         .mkdir          = shmem_mkdir,
2431         .rmdir          = shmem_rmdir,
2432         .mknod          = shmem_mknod,
2433         .rename         = shmem_rename,
2434 #endif
2435 #ifdef CONFIG_TMPFS_POSIX_ACL
2436         .setattr        = shmem_notify_change,
2437         .setxattr       = generic_setxattr,
2438         .getxattr       = generic_getxattr,
2439         .listxattr      = generic_listxattr,
2440         .removexattr    = generic_removexattr,
2441         .permission     = shmem_permission,
2442 #endif
2443 };
2444
2445 static const struct inode_operations shmem_special_inode_operations = {
2446 #ifdef CONFIG_TMPFS_POSIX_ACL
2447         .setattr        = shmem_notify_change,
2448         .setxattr       = generic_setxattr,
2449         .getxattr       = generic_getxattr,
2450         .listxattr      = generic_listxattr,
2451         .removexattr    = generic_removexattr,
2452         .permission     = shmem_permission,
2453 #endif
2454 };
2455
2456 static const struct super_operations shmem_ops = {
2457         .alloc_inode    = shmem_alloc_inode,
2458         .destroy_inode  = shmem_destroy_inode,
2459 #ifdef CONFIG_TMPFS
2460         .statfs         = shmem_statfs,
2461         .remount_fs     = shmem_remount_fs,
2462         .show_options   = shmem_show_options,
2463 #endif
2464         .delete_inode   = shmem_delete_inode,
2465         .drop_inode     = generic_delete_inode,
2466         .put_super      = shmem_put_super,
2467 };
2468
2469 static struct vm_operations_struct shmem_vm_ops = {
2470         .fault          = shmem_fault,
2471 #ifdef CONFIG_NUMA
2472         .set_policy     = shmem_set_policy,
2473         .get_policy     = shmem_get_policy,
2474 #endif
2475 };
2476
2477
2478 static int shmem_get_sb(struct file_system_type *fs_type,
2479         int flags, const char *dev_name, void *data, struct vfsmount *mnt)
2480 {
2481         return get_sb_nodev(fs_type, flags, data, shmem_fill_super, mnt);
2482 }
2483
2484 static struct file_system_type tmpfs_fs_type = {
2485         .owner          = THIS_MODULE,
2486         .name           = "tmpfs",
2487         .get_sb         = shmem_get_sb,
2488         .kill_sb        = kill_litter_super,
2489 };
2490 static struct vfsmount *shm_mnt;
2491
2492 static int __init init_tmpfs(void)
2493 {
2494         int error;
2495
2496         error = bdi_init(&shmem_backing_dev_info);
2497         if (error)
2498                 goto out4;
2499
2500         error = init_inodecache();
2501         if (error)
2502                 goto out3;
2503
2504         error = register_filesystem(&tmpfs_fs_type);
2505         if (error) {
2506                 printk(KERN_ERR "Could not register tmpfs\n");
2507                 goto out2;
2508         }
2509
2510         shm_mnt = vfs_kern_mount(&tmpfs_fs_type, MS_NOUSER,
2511                                 tmpfs_fs_type.name, NULL);
2512         if (IS_ERR(shm_mnt)) {
2513                 error = PTR_ERR(shm_mnt);
2514                 printk(KERN_ERR "Could not kern_mount tmpfs\n");
2515                 goto out1;
2516         }
2517         return 0;
2518
2519 out1:
2520         unregister_filesystem(&tmpfs_fs_type);
2521 out2:
2522         destroy_inodecache();
2523 out3:
2524         bdi_destroy(&shmem_backing_dev_info);
2525 out4:
2526         shm_mnt = ERR_PTR(error);
2527         return error;
2528 }
2529 module_init(init_tmpfs)
2530
2531 /**
2532  * shmem_file_setup - get an unlinked file living in tmpfs
2533  * @name: name for dentry (to be seen in /proc/<pid>/maps
2534  * @size: size to be set for the file
2535  * @flags: vm_flags
2536  */
2537 struct file *shmem_file_setup(char *name, loff_t size, unsigned long flags)
2538 {
2539         int error;
2540         struct file *file;
2541         struct inode *inode;
2542         struct dentry *dentry, *root;
2543         struct qstr this;
2544
2545         if (IS_ERR(shm_mnt))
2546                 return (void *)shm_mnt;
2547
2548         if (size < 0 || size > SHMEM_MAX_BYTES)
2549                 return ERR_PTR(-EINVAL);
2550
2551         if (shmem_acct_size(flags, size))
2552                 return ERR_PTR(-ENOMEM);
2553
2554         error = -ENOMEM;
2555         this.name = name;
2556         this.len = strlen(name);
2557         this.hash = 0; /* will go */
2558         root = shm_mnt->mnt_root;
2559         dentry = d_alloc(root, &this);
2560         if (!dentry)
2561                 goto put_memory;
2562
2563         error = -ENFILE;
2564         file = get_empty_filp();
2565         if (!file)
2566                 goto put_dentry;
2567
2568         error = -ENOSPC;
2569         inode = shmem_get_inode(root->d_sb, S_IFREG | S_IRWXUGO, 0);
2570         if (!inode)
2571                 goto close_file;
2572
2573         SHMEM_I(inode)->flags = flags & VM_ACCOUNT;
2574         d_instantiate(dentry, inode);
2575         inode->i_size = size;
2576         inode->i_nlink = 0;     /* It is unlinked */
2577         init_file(file, shm_mnt, dentry, FMODE_WRITE | FMODE_READ,
2578                         &shmem_file_operations);
2579         return file;
2580
2581 close_file:
2582         put_filp(file);
2583 put_dentry:
2584         dput(dentry);
2585 put_memory:
2586         shmem_unacct_size(flags, size);
2587         return ERR_PTR(error);
2588 }
2589 EXPORT_SYMBOL_GPL(shmem_file_setup);
2590
2591 /**
2592  * shmem_zero_setup - setup a shared anonymous mapping
2593  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
2594  */
2595 int shmem_zero_setup(struct vm_area_struct *vma)
2596 {
2597         struct file *file;
2598         loff_t size = vma->vm_end - vma->vm_start;
2599
2600         file = shmem_file_setup("dev/zero", size, vma->vm_flags);
2601         if (IS_ERR(file))
2602                 return PTR_ERR(file);
2603
2604         ima_shm_check(file);
2605         if (vma->vm_file)
2606                 fput(vma->vm_file);
2607         vma->vm_file = file;
2608         vma->vm_ops = &shmem_vm_ops;
2609         return 0;
2610 }