memcg: add the pagefault count into memcg stats
[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  * tiny-shmem:
18  * Copyright (c) 2004, 2008 Matt Mackall <mpm@selenic.com>
19  *
20  * This file is released under the GPL.
21  */
22
23 #include <linux/fs.h>
24 #include <linux/init.h>
25 #include <linux/vfs.h>
26 #include <linux/mount.h>
27 #include <linux/pagemap.h>
28 #include <linux/file.h>
29 #include <linux/mm.h>
30 #include <linux/module.h>
31 #include <linux/percpu_counter.h>
32 #include <linux/swap.h>
33
34 static struct vfsmount *shm_mnt;
35
36 #ifdef CONFIG_SHMEM
37 /*
38  * This virtual memory filesystem is heavily based on the ramfs. It
39  * extends ramfs by the ability to use swap and honor resource limits
40  * which makes it a completely usable filesystem.
41  */
42
43 #include <linux/xattr.h>
44 #include <linux/exportfs.h>
45 #include <linux/posix_acl.h>
46 #include <linux/generic_acl.h>
47 #include <linux/mman.h>
48 #include <linux/string.h>
49 #include <linux/slab.h>
50 #include <linux/backing-dev.h>
51 #include <linux/shmem_fs.h>
52 #include <linux/writeback.h>
53 #include <linux/blkdev.h>
54 #include <linux/security.h>
55 #include <linux/swapops.h>
56 #include <linux/mempolicy.h>
57 #include <linux/namei.h>
58 #include <linux/ctype.h>
59 #include <linux/migrate.h>
60 #include <linux/highmem.h>
61 #include <linux/seq_file.h>
62 #include <linux/magic.h>
63
64 #include <asm/uaccess.h>
65 #include <asm/div64.h>
66 #include <asm/pgtable.h>
67
68 /*
69  * The maximum size of a shmem/tmpfs file is limited by the maximum size of
70  * its triple-indirect swap vector - see illustration at shmem_swp_entry().
71  *
72  * With 4kB page size, maximum file size is just over 2TB on a 32-bit kernel,
73  * but one eighth of that on a 64-bit kernel.  With 8kB page size, maximum
74  * file size is just over 4TB on a 64-bit kernel, but 16TB on a 32-bit kernel,
75  * MAX_LFS_FILESIZE being then more restrictive than swap vector layout.
76  *
77  * We use / and * instead of shifts in the definitions below, so that the swap
78  * vector can be tested with small even values (e.g. 20) for ENTRIES_PER_PAGE.
79  */
80 #define ENTRIES_PER_PAGE (PAGE_CACHE_SIZE/sizeof(unsigned long))
81 #define ENTRIES_PER_PAGEPAGE ((unsigned long long)ENTRIES_PER_PAGE*ENTRIES_PER_PAGE)
82
83 #define SHMSWP_MAX_INDEX (SHMEM_NR_DIRECT + (ENTRIES_PER_PAGEPAGE/2) * (ENTRIES_PER_PAGE+1))
84 #define SHMSWP_MAX_BYTES (SHMSWP_MAX_INDEX << PAGE_CACHE_SHIFT)
85
86 #define SHMEM_MAX_BYTES  min_t(unsigned long long, SHMSWP_MAX_BYTES, MAX_LFS_FILESIZE)
87 #define SHMEM_MAX_INDEX  ((unsigned long)((SHMEM_MAX_BYTES+1) >> PAGE_CACHE_SHIFT))
88
89 #define BLOCKS_PER_PAGE  (PAGE_CACHE_SIZE/512)
90 #define VM_ACCT(size)    (PAGE_CACHE_ALIGN(size) >> PAGE_SHIFT)
91
92 /* info->flags needs VM_flags to handle pagein/truncate races efficiently */
93 #define SHMEM_PAGEIN     VM_READ
94 #define SHMEM_TRUNCATE   VM_WRITE
95
96 /* Definition to limit shmem_truncate's steps between cond_rescheds */
97 #define LATENCY_LIMIT    64
98
99 /* Pretend that each entry is of this size in directory's i_size */
100 #define BOGO_DIRENT_SIZE 20
101
102 struct shmem_xattr {
103         struct list_head list;  /* anchored by shmem_inode_info->xattr_list */
104         char *name;             /* xattr name */
105         size_t size;
106         char value[0];
107 };
108
109 /* Flag allocation requirements to shmem_getpage and shmem_swp_alloc */
110 enum sgp_type {
111         SGP_READ,       /* don't exceed i_size, don't allocate page */
112         SGP_CACHE,      /* don't exceed i_size, may allocate page */
113         SGP_DIRTY,      /* like SGP_CACHE, but set new page dirty */
114         SGP_WRITE,      /* may exceed i_size, may allocate page */
115 };
116
117 #ifdef CONFIG_TMPFS
118 static unsigned long shmem_default_max_blocks(void)
119 {
120         return totalram_pages / 2;
121 }
122
123 static unsigned long shmem_default_max_inodes(void)
124 {
125         return min(totalram_pages - totalhigh_pages, totalram_pages / 2);
126 }
127 #endif
128
129 static int shmem_getpage(struct inode *inode, unsigned long idx,
130                          struct page **pagep, enum sgp_type sgp, int *type);
131
132 static inline struct page *shmem_dir_alloc(gfp_t gfp_mask)
133 {
134         /*
135          * The above definition of ENTRIES_PER_PAGE, and the use of
136          * BLOCKS_PER_PAGE on indirect pages, assume PAGE_CACHE_SIZE:
137          * might be reconsidered if it ever diverges from PAGE_SIZE.
138          *
139          * Mobility flags are masked out as swap vectors cannot move
140          */
141         return alloc_pages((gfp_mask & ~GFP_MOVABLE_MASK) | __GFP_ZERO,
142                                 PAGE_CACHE_SHIFT-PAGE_SHIFT);
143 }
144
145 static inline void shmem_dir_free(struct page *page)
146 {
147         __free_pages(page, PAGE_CACHE_SHIFT-PAGE_SHIFT);
148 }
149
150 static struct page **shmem_dir_map(struct page *page)
151 {
152         return (struct page **)kmap_atomic(page, KM_USER0);
153 }
154
155 static inline void shmem_dir_unmap(struct page **dir)
156 {
157         kunmap_atomic(dir, KM_USER0);
158 }
159
160 static swp_entry_t *shmem_swp_map(struct page *page)
161 {
162         return (swp_entry_t *)kmap_atomic(page, KM_USER1);
163 }
164
165 static inline void shmem_swp_balance_unmap(void)
166 {
167         /*
168          * When passing a pointer to an i_direct entry, to code which
169          * also handles indirect entries and so will shmem_swp_unmap,
170          * we must arrange for the preempt count to remain in balance.
171          * What kmap_atomic of a lowmem page does depends on config
172          * and architecture, so pretend to kmap_atomic some lowmem page.
173          */
174         (void) kmap_atomic(ZERO_PAGE(0), KM_USER1);
175 }
176
177 static inline void shmem_swp_unmap(swp_entry_t *entry)
178 {
179         kunmap_atomic(entry, KM_USER1);
180 }
181
182 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
183 {
184         return sb->s_fs_info;
185 }
186
187 /*
188  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
189  * for shared memory and for shared anonymous (/dev/zero) mappings
190  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
191  * consistent with the pre-accounting of private mappings ...
192  */
193 static inline int shmem_acct_size(unsigned long flags, loff_t size)
194 {
195         return (flags & VM_NORESERVE) ?
196                 0 : security_vm_enough_memory_kern(VM_ACCT(size));
197 }
198
199 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
200 {
201         if (!(flags & VM_NORESERVE))
202                 vm_unacct_memory(VM_ACCT(size));
203 }
204
205 /*
206  * ... whereas tmpfs objects are accounted incrementally as
207  * pages are allocated, in order to allow huge sparse files.
208  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
209  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
210  */
211 static inline int shmem_acct_block(unsigned long flags)
212 {
213         return (flags & VM_NORESERVE) ?
214                 security_vm_enough_memory_kern(VM_ACCT(PAGE_CACHE_SIZE)) : 0;
215 }
216
217 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
218 {
219         if (flags & VM_NORESERVE)
220                 vm_unacct_memory(pages * VM_ACCT(PAGE_CACHE_SIZE));
221 }
222
223 static const struct super_operations shmem_ops;
224 static const struct address_space_operations shmem_aops;
225 static const struct file_operations shmem_file_operations;
226 static const struct inode_operations shmem_inode_operations;
227 static const struct inode_operations shmem_dir_inode_operations;
228 static const struct inode_operations shmem_special_inode_operations;
229 static const struct vm_operations_struct shmem_vm_ops;
230
231 static struct backing_dev_info shmem_backing_dev_info  __read_mostly = {
232         .ra_pages       = 0,    /* No readahead */
233         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_SWAP_BACKED,
234 };
235
236 static LIST_HEAD(shmem_swaplist);
237 static DEFINE_MUTEX(shmem_swaplist_mutex);
238
239 static void shmem_free_blocks(struct inode *inode, long pages)
240 {
241         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
242         if (sbinfo->max_blocks) {
243                 percpu_counter_add(&sbinfo->used_blocks, -pages);
244                 spin_lock(&inode->i_lock);
245                 inode->i_blocks -= pages*BLOCKS_PER_PAGE;
246                 spin_unlock(&inode->i_lock);
247         }
248 }
249
250 static int shmem_reserve_inode(struct super_block *sb)
251 {
252         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
253         if (sbinfo->max_inodes) {
254                 spin_lock(&sbinfo->stat_lock);
255                 if (!sbinfo->free_inodes) {
256                         spin_unlock(&sbinfo->stat_lock);
257                         return -ENOSPC;
258                 }
259                 sbinfo->free_inodes--;
260                 spin_unlock(&sbinfo->stat_lock);
261         }
262         return 0;
263 }
264
265 static void shmem_free_inode(struct super_block *sb)
266 {
267         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
268         if (sbinfo->max_inodes) {
269                 spin_lock(&sbinfo->stat_lock);
270                 sbinfo->free_inodes++;
271                 spin_unlock(&sbinfo->stat_lock);
272         }
273 }
274
275 /**
276  * shmem_recalc_inode - recalculate the size of an inode
277  * @inode: inode to recalc
278  *
279  * We have to calculate the free blocks since the mm can drop
280  * undirtied hole pages behind our back.
281  *
282  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
283  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
284  *
285  * It has to be called with the spinlock held.
286  */
287 static void shmem_recalc_inode(struct inode *inode)
288 {
289         struct shmem_inode_info *info = SHMEM_I(inode);
290         long freed;
291
292         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
293         if (freed > 0) {
294                 info->alloced -= freed;
295                 shmem_unacct_blocks(info->flags, freed);
296                 shmem_free_blocks(inode, freed);
297         }
298 }
299
300 /**
301  * shmem_swp_entry - find the swap vector position in the info structure
302  * @info:  info structure for the inode
303  * @index: index of the page to find
304  * @page:  optional page to add to the structure. Has to be preset to
305  *         all zeros
306  *
307  * If there is no space allocated yet it will return NULL when
308  * page is NULL, else it will use the page for the needed block,
309  * setting it to NULL on return to indicate that it has been used.
310  *
311  * The swap vector is organized the following way:
312  *
313  * There are SHMEM_NR_DIRECT entries directly stored in the
314  * shmem_inode_info structure. So small files do not need an addional
315  * allocation.
316  *
317  * For pages with index > SHMEM_NR_DIRECT there is the pointer
318  * i_indirect which points to a page which holds in the first half
319  * doubly indirect blocks, in the second half triple indirect blocks:
320  *
321  * For an artificial ENTRIES_PER_PAGE = 4 this would lead to the
322  * following layout (for SHMEM_NR_DIRECT == 16):
323  *
324  * i_indirect -> dir --> 16-19
325  *            |      +-> 20-23
326  *            |
327  *            +-->dir2 --> 24-27
328  *            |        +-> 28-31
329  *            |        +-> 32-35
330  *            |        +-> 36-39
331  *            |
332  *            +-->dir3 --> 40-43
333  *                     +-> 44-47
334  *                     +-> 48-51
335  *                     +-> 52-55
336  */
337 static swp_entry_t *shmem_swp_entry(struct shmem_inode_info *info, unsigned long index, struct page **page)
338 {
339         unsigned long offset;
340         struct page **dir;
341         struct page *subdir;
342
343         if (index < SHMEM_NR_DIRECT) {
344                 shmem_swp_balance_unmap();
345                 return info->i_direct+index;
346         }
347         if (!info->i_indirect) {
348                 if (page) {
349                         info->i_indirect = *page;
350                         *page = NULL;
351                 }
352                 return NULL;                    /* need another page */
353         }
354
355         index -= SHMEM_NR_DIRECT;
356         offset = index % ENTRIES_PER_PAGE;
357         index /= ENTRIES_PER_PAGE;
358         dir = shmem_dir_map(info->i_indirect);
359
360         if (index >= ENTRIES_PER_PAGE/2) {
361                 index -= ENTRIES_PER_PAGE/2;
362                 dir += ENTRIES_PER_PAGE/2 + index/ENTRIES_PER_PAGE;
363                 index %= ENTRIES_PER_PAGE;
364                 subdir = *dir;
365                 if (!subdir) {
366                         if (page) {
367                                 *dir = *page;
368                                 *page = NULL;
369                         }
370                         shmem_dir_unmap(dir);
371                         return NULL;            /* need another page */
372                 }
373                 shmem_dir_unmap(dir);
374                 dir = shmem_dir_map(subdir);
375         }
376
377         dir += index;
378         subdir = *dir;
379         if (!subdir) {
380                 if (!page || !(subdir = *page)) {
381                         shmem_dir_unmap(dir);
382                         return NULL;            /* need a page */
383                 }
384                 *dir = subdir;
385                 *page = NULL;
386         }
387         shmem_dir_unmap(dir);
388         return shmem_swp_map(subdir) + offset;
389 }
390
391 static void shmem_swp_set(struct shmem_inode_info *info, swp_entry_t *entry, unsigned long value)
392 {
393         long incdec = value? 1: -1;
394
395         entry->val = value;
396         info->swapped += incdec;
397         if ((unsigned long)(entry - info->i_direct) >= SHMEM_NR_DIRECT) {
398                 struct page *page = kmap_atomic_to_page(entry);
399                 set_page_private(page, page_private(page) + incdec);
400         }
401 }
402
403 /**
404  * shmem_swp_alloc - get the position of the swap entry for the page.
405  * @info:       info structure for the inode
406  * @index:      index of the page to find
407  * @sgp:        check and recheck i_size? skip allocation?
408  *
409  * If the entry does not exist, allocate it.
410  */
411 static swp_entry_t *shmem_swp_alloc(struct shmem_inode_info *info, unsigned long index, enum sgp_type sgp)
412 {
413         struct inode *inode = &info->vfs_inode;
414         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
415         struct page *page = NULL;
416         swp_entry_t *entry;
417
418         if (sgp != SGP_WRITE &&
419             ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode))
420                 return ERR_PTR(-EINVAL);
421
422         while (!(entry = shmem_swp_entry(info, index, &page))) {
423                 if (sgp == SGP_READ)
424                         return shmem_swp_map(ZERO_PAGE(0));
425                 /*
426                  * Test used_blocks against 1 less max_blocks, since we have 1 data
427                  * page (and perhaps indirect index pages) yet to allocate:
428                  * a waste to allocate index if we cannot allocate data.
429                  */
430                 if (sbinfo->max_blocks) {
431                         if (percpu_counter_compare(&sbinfo->used_blocks,
432                                                 sbinfo->max_blocks - 1) >= 0)
433                                 return ERR_PTR(-ENOSPC);
434                         percpu_counter_inc(&sbinfo->used_blocks);
435                         spin_lock(&inode->i_lock);
436                         inode->i_blocks += BLOCKS_PER_PAGE;
437                         spin_unlock(&inode->i_lock);
438                 }
439
440                 spin_unlock(&info->lock);
441                 page = shmem_dir_alloc(mapping_gfp_mask(inode->i_mapping));
442                 spin_lock(&info->lock);
443
444                 if (!page) {
445                         shmem_free_blocks(inode, 1);
446                         return ERR_PTR(-ENOMEM);
447                 }
448                 if (sgp != SGP_WRITE &&
449                     ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
450                         entry = ERR_PTR(-EINVAL);
451                         break;
452                 }
453                 if (info->next_index <= index)
454                         info->next_index = index + 1;
455         }
456         if (page) {
457                 /* another task gave its page, or truncated the file */
458                 shmem_free_blocks(inode, 1);
459                 shmem_dir_free(page);
460         }
461         if (info->next_index <= index && !IS_ERR(entry))
462                 info->next_index = index + 1;
463         return entry;
464 }
465
466 /**
467  * shmem_free_swp - free some swap entries in a directory
468  * @dir:        pointer to the directory
469  * @edir:       pointer after last entry of the directory
470  * @punch_lock: pointer to spinlock when needed for the holepunch case
471  */
472 static int shmem_free_swp(swp_entry_t *dir, swp_entry_t *edir,
473                                                 spinlock_t *punch_lock)
474 {
475         spinlock_t *punch_unlock = NULL;
476         swp_entry_t *ptr;
477         int freed = 0;
478
479         for (ptr = dir; ptr < edir; ptr++) {
480                 if (ptr->val) {
481                         if (unlikely(punch_lock)) {
482                                 punch_unlock = punch_lock;
483                                 punch_lock = NULL;
484                                 spin_lock(punch_unlock);
485                                 if (!ptr->val)
486                                         continue;
487                         }
488                         free_swap_and_cache(*ptr);
489                         *ptr = (swp_entry_t){0};
490                         freed++;
491                 }
492         }
493         if (punch_unlock)
494                 spin_unlock(punch_unlock);
495         return freed;
496 }
497
498 static int shmem_map_and_free_swp(struct page *subdir, int offset,
499                 int limit, struct page ***dir, spinlock_t *punch_lock)
500 {
501         swp_entry_t *ptr;
502         int freed = 0;
503
504         ptr = shmem_swp_map(subdir);
505         for (; offset < limit; offset += LATENCY_LIMIT) {
506                 int size = limit - offset;
507                 if (size > LATENCY_LIMIT)
508                         size = LATENCY_LIMIT;
509                 freed += shmem_free_swp(ptr+offset, ptr+offset+size,
510                                                         punch_lock);
511                 if (need_resched()) {
512                         shmem_swp_unmap(ptr);
513                         if (*dir) {
514                                 shmem_dir_unmap(*dir);
515                                 *dir = NULL;
516                         }
517                         cond_resched();
518                         ptr = shmem_swp_map(subdir);
519                 }
520         }
521         shmem_swp_unmap(ptr);
522         return freed;
523 }
524
525 static void shmem_free_pages(struct list_head *next)
526 {
527         struct page *page;
528         int freed = 0;
529
530         do {
531                 page = container_of(next, struct page, lru);
532                 next = next->next;
533                 shmem_dir_free(page);
534                 freed++;
535                 if (freed >= LATENCY_LIMIT) {
536                         cond_resched();
537                         freed = 0;
538                 }
539         } while (next);
540 }
541
542 static void shmem_truncate_range(struct inode *inode, loff_t start, loff_t end)
543 {
544         struct shmem_inode_info *info = SHMEM_I(inode);
545         unsigned long idx;
546         unsigned long size;
547         unsigned long limit;
548         unsigned long stage;
549         unsigned long diroff;
550         struct page **dir;
551         struct page *topdir;
552         struct page *middir;
553         struct page *subdir;
554         swp_entry_t *ptr;
555         LIST_HEAD(pages_to_free);
556         long nr_pages_to_free = 0;
557         long nr_swaps_freed = 0;
558         int offset;
559         int freed;
560         int punch_hole;
561         spinlock_t *needs_lock;
562         spinlock_t *punch_lock;
563         unsigned long upper_limit;
564
565         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
566         idx = (start + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
567         if (idx >= info->next_index)
568                 return;
569
570         spin_lock(&info->lock);
571         info->flags |= SHMEM_TRUNCATE;
572         if (likely(end == (loff_t) -1)) {
573                 limit = info->next_index;
574                 upper_limit = SHMEM_MAX_INDEX;
575                 info->next_index = idx;
576                 needs_lock = NULL;
577                 punch_hole = 0;
578         } else {
579                 if (end + 1 >= inode->i_size) { /* we may free a little more */
580                         limit = (inode->i_size + PAGE_CACHE_SIZE - 1) >>
581                                                         PAGE_CACHE_SHIFT;
582                         upper_limit = SHMEM_MAX_INDEX;
583                 } else {
584                         limit = (end + 1) >> PAGE_CACHE_SHIFT;
585                         upper_limit = limit;
586                 }
587                 needs_lock = &info->lock;
588                 punch_hole = 1;
589         }
590
591         topdir = info->i_indirect;
592         if (topdir && idx <= SHMEM_NR_DIRECT && !punch_hole) {
593                 info->i_indirect = NULL;
594                 nr_pages_to_free++;
595                 list_add(&topdir->lru, &pages_to_free);
596         }
597         spin_unlock(&info->lock);
598
599         if (info->swapped && idx < SHMEM_NR_DIRECT) {
600                 ptr = info->i_direct;
601                 size = limit;
602                 if (size > SHMEM_NR_DIRECT)
603                         size = SHMEM_NR_DIRECT;
604                 nr_swaps_freed = shmem_free_swp(ptr+idx, ptr+size, needs_lock);
605         }
606
607         /*
608          * If there are no indirect blocks or we are punching a hole
609          * below indirect blocks, nothing to be done.
610          */
611         if (!topdir || limit <= SHMEM_NR_DIRECT)
612                 goto done2;
613
614         /*
615          * The truncation case has already dropped info->lock, and we're safe
616          * because i_size and next_index have already been lowered, preventing
617          * access beyond.  But in the punch_hole case, we still need to take
618          * the lock when updating the swap directory, because there might be
619          * racing accesses by shmem_getpage(SGP_CACHE), shmem_unuse_inode or
620          * shmem_writepage.  However, whenever we find we can remove a whole
621          * directory page (not at the misaligned start or end of the range),
622          * we first NULLify its pointer in the level above, and then have no
623          * need to take the lock when updating its contents: needs_lock and
624          * punch_lock (either pointing to info->lock or NULL) manage this.
625          */
626
627         upper_limit -= SHMEM_NR_DIRECT;
628         limit -= SHMEM_NR_DIRECT;
629         idx = (idx > SHMEM_NR_DIRECT)? (idx - SHMEM_NR_DIRECT): 0;
630         offset = idx % ENTRIES_PER_PAGE;
631         idx -= offset;
632
633         dir = shmem_dir_map(topdir);
634         stage = ENTRIES_PER_PAGEPAGE/2;
635         if (idx < ENTRIES_PER_PAGEPAGE/2) {
636                 middir = topdir;
637                 diroff = idx/ENTRIES_PER_PAGE;
638         } else {
639                 dir += ENTRIES_PER_PAGE/2;
640                 dir += (idx - ENTRIES_PER_PAGEPAGE/2)/ENTRIES_PER_PAGEPAGE;
641                 while (stage <= idx)
642                         stage += ENTRIES_PER_PAGEPAGE;
643                 middir = *dir;
644                 if (*dir) {
645                         diroff = ((idx - ENTRIES_PER_PAGEPAGE/2) %
646                                 ENTRIES_PER_PAGEPAGE) / ENTRIES_PER_PAGE;
647                         if (!diroff && !offset && upper_limit >= stage) {
648                                 if (needs_lock) {
649                                         spin_lock(needs_lock);
650                                         *dir = NULL;
651                                         spin_unlock(needs_lock);
652                                         needs_lock = NULL;
653                                 } else
654                                         *dir = NULL;
655                                 nr_pages_to_free++;
656                                 list_add(&middir->lru, &pages_to_free);
657                         }
658                         shmem_dir_unmap(dir);
659                         dir = shmem_dir_map(middir);
660                 } else {
661                         diroff = 0;
662                         offset = 0;
663                         idx = stage;
664                 }
665         }
666
667         for (; idx < limit; idx += ENTRIES_PER_PAGE, diroff++) {
668                 if (unlikely(idx == stage)) {
669                         shmem_dir_unmap(dir);
670                         dir = shmem_dir_map(topdir) +
671                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
672                         while (!*dir) {
673                                 dir++;
674                                 idx += ENTRIES_PER_PAGEPAGE;
675                                 if (idx >= limit)
676                                         goto done1;
677                         }
678                         stage = idx + ENTRIES_PER_PAGEPAGE;
679                         middir = *dir;
680                         if (punch_hole)
681                                 needs_lock = &info->lock;
682                         if (upper_limit >= stage) {
683                                 if (needs_lock) {
684                                         spin_lock(needs_lock);
685                                         *dir = NULL;
686                                         spin_unlock(needs_lock);
687                                         needs_lock = NULL;
688                                 } else
689                                         *dir = NULL;
690                                 nr_pages_to_free++;
691                                 list_add(&middir->lru, &pages_to_free);
692                         }
693                         shmem_dir_unmap(dir);
694                         cond_resched();
695                         dir = shmem_dir_map(middir);
696                         diroff = 0;
697                 }
698                 punch_lock = needs_lock;
699                 subdir = dir[diroff];
700                 if (subdir && !offset && upper_limit-idx >= ENTRIES_PER_PAGE) {
701                         if (needs_lock) {
702                                 spin_lock(needs_lock);
703                                 dir[diroff] = NULL;
704                                 spin_unlock(needs_lock);
705                                 punch_lock = NULL;
706                         } else
707                                 dir[diroff] = NULL;
708                         nr_pages_to_free++;
709                         list_add(&subdir->lru, &pages_to_free);
710                 }
711                 if (subdir && page_private(subdir) /* has swap entries */) {
712                         size = limit - idx;
713                         if (size > ENTRIES_PER_PAGE)
714                                 size = ENTRIES_PER_PAGE;
715                         freed = shmem_map_and_free_swp(subdir,
716                                         offset, size, &dir, punch_lock);
717                         if (!dir)
718                                 dir = shmem_dir_map(middir);
719                         nr_swaps_freed += freed;
720                         if (offset || punch_lock) {
721                                 spin_lock(&info->lock);
722                                 set_page_private(subdir,
723                                         page_private(subdir) - freed);
724                                 spin_unlock(&info->lock);
725                         } else
726                                 BUG_ON(page_private(subdir) != freed);
727                 }
728                 offset = 0;
729         }
730 done1:
731         shmem_dir_unmap(dir);
732 done2:
733         if (inode->i_mapping->nrpages && (info->flags & SHMEM_PAGEIN)) {
734                 /*
735                  * Call truncate_inode_pages again: racing shmem_unuse_inode
736                  * may have swizzled a page in from swap since
737                  * truncate_pagecache or generic_delete_inode did it, before we
738                  * lowered next_index.  Also, though shmem_getpage checks
739                  * i_size before adding to cache, no recheck after: so fix the
740                  * narrow window there too.
741                  *
742                  * Recalling truncate_inode_pages_range and unmap_mapping_range
743                  * every time for punch_hole (which never got a chance to clear
744                  * SHMEM_PAGEIN at the start of vmtruncate_range) is expensive,
745                  * yet hardly ever necessary: try to optimize them out later.
746                  */
747                 truncate_inode_pages_range(inode->i_mapping, start, end);
748                 if (punch_hole)
749                         unmap_mapping_range(inode->i_mapping, start,
750                                                         end - start, 1);
751         }
752
753         spin_lock(&info->lock);
754         info->flags &= ~SHMEM_TRUNCATE;
755         info->swapped -= nr_swaps_freed;
756         if (nr_pages_to_free)
757                 shmem_free_blocks(inode, nr_pages_to_free);
758         shmem_recalc_inode(inode);
759         spin_unlock(&info->lock);
760
761         /*
762          * Empty swap vector directory pages to be freed?
763          */
764         if (!list_empty(&pages_to_free)) {
765                 pages_to_free.prev->next = NULL;
766                 shmem_free_pages(pages_to_free.next);
767         }
768 }
769
770 static int shmem_notify_change(struct dentry *dentry, struct iattr *attr)
771 {
772         struct inode *inode = dentry->d_inode;
773         loff_t newsize = attr->ia_size;
774         int error;
775
776         error = inode_change_ok(inode, attr);
777         if (error)
778                 return error;
779
780         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)
781                                         && newsize != inode->i_size) {
782                 struct page *page = NULL;
783
784                 if (newsize < inode->i_size) {
785                         /*
786                          * If truncating down to a partial page, then
787                          * if that page is already allocated, hold it
788                          * in memory until the truncation is over, so
789                          * truncate_partial_page cannot miss it were
790                          * it assigned to swap.
791                          */
792                         if (newsize & (PAGE_CACHE_SIZE-1)) {
793                                 (void) shmem_getpage(inode,
794                                         newsize >> PAGE_CACHE_SHIFT,
795                                                 &page, SGP_READ, NULL);
796                                 if (page)
797                                         unlock_page(page);
798                         }
799                         /*
800                          * Reset SHMEM_PAGEIN flag so that shmem_truncate can
801                          * detect if any pages might have been added to cache
802                          * after truncate_inode_pages.  But we needn't bother
803                          * if it's being fully truncated to zero-length: the
804                          * nrpages check is efficient enough in that case.
805                          */
806                         if (newsize) {
807                                 struct shmem_inode_info *info = SHMEM_I(inode);
808                                 spin_lock(&info->lock);
809                                 info->flags &= ~SHMEM_PAGEIN;
810                                 spin_unlock(&info->lock);
811                         }
812                 }
813
814                 /* XXX(truncate): truncate_setsize should be called last */
815                 truncate_setsize(inode, newsize);
816                 if (page)
817                         page_cache_release(page);
818                 shmem_truncate_range(inode, newsize, (loff_t)-1);
819         }
820
821         setattr_copy(inode, attr);
822 #ifdef CONFIG_TMPFS_POSIX_ACL
823         if (attr->ia_valid & ATTR_MODE)
824                 error = generic_acl_chmod(inode);
825 #endif
826         return error;
827 }
828
829 static void shmem_evict_inode(struct inode *inode)
830 {
831         struct shmem_inode_info *info = SHMEM_I(inode);
832         struct shmem_xattr *xattr, *nxattr;
833
834         if (inode->i_mapping->a_ops == &shmem_aops) {
835                 truncate_inode_pages(inode->i_mapping, 0);
836                 shmem_unacct_size(info->flags, inode->i_size);
837                 inode->i_size = 0;
838                 shmem_truncate_range(inode, 0, (loff_t)-1);
839                 if (!list_empty(&info->swaplist)) {
840                         mutex_lock(&shmem_swaplist_mutex);
841                         list_del_init(&info->swaplist);
842                         mutex_unlock(&shmem_swaplist_mutex);
843                 }
844         }
845
846         list_for_each_entry_safe(xattr, nxattr, &info->xattr_list, list) {
847                 kfree(xattr->name);
848                 kfree(xattr);
849         }
850         BUG_ON(inode->i_blocks);
851         shmem_free_inode(inode->i_sb);
852         end_writeback(inode);
853 }
854
855 static inline int shmem_find_swp(swp_entry_t entry, swp_entry_t *dir, swp_entry_t *edir)
856 {
857         swp_entry_t *ptr;
858
859         for (ptr = dir; ptr < edir; ptr++) {
860                 if (ptr->val == entry.val)
861                         return ptr - dir;
862         }
863         return -1;
864 }
865
866 static int shmem_unuse_inode(struct shmem_inode_info *info, swp_entry_t entry, struct page *page)
867 {
868         struct address_space *mapping;
869         unsigned long idx;
870         unsigned long size;
871         unsigned long limit;
872         unsigned long stage;
873         struct page **dir;
874         struct page *subdir;
875         swp_entry_t *ptr;
876         int offset;
877         int error;
878
879         idx = 0;
880         ptr = info->i_direct;
881         spin_lock(&info->lock);
882         if (!info->swapped) {
883                 list_del_init(&info->swaplist);
884                 goto lost2;
885         }
886         limit = info->next_index;
887         size = limit;
888         if (size > SHMEM_NR_DIRECT)
889                 size = SHMEM_NR_DIRECT;
890         offset = shmem_find_swp(entry, ptr, ptr+size);
891         if (offset >= 0) {
892                 shmem_swp_balance_unmap();
893                 goto found;
894         }
895         if (!info->i_indirect)
896                 goto lost2;
897
898         dir = shmem_dir_map(info->i_indirect);
899         stage = SHMEM_NR_DIRECT + ENTRIES_PER_PAGEPAGE/2;
900
901         for (idx = SHMEM_NR_DIRECT; idx < limit; idx += ENTRIES_PER_PAGE, dir++) {
902                 if (unlikely(idx == stage)) {
903                         shmem_dir_unmap(dir-1);
904                         if (cond_resched_lock(&info->lock)) {
905                                 /* check it has not been truncated */
906                                 if (limit > info->next_index) {
907                                         limit = info->next_index;
908                                         if (idx >= limit)
909                                                 goto lost2;
910                                 }
911                         }
912                         dir = shmem_dir_map(info->i_indirect) +
913                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
914                         while (!*dir) {
915                                 dir++;
916                                 idx += ENTRIES_PER_PAGEPAGE;
917                                 if (idx >= limit)
918                                         goto lost1;
919                         }
920                         stage = idx + ENTRIES_PER_PAGEPAGE;
921                         subdir = *dir;
922                         shmem_dir_unmap(dir);
923                         dir = shmem_dir_map(subdir);
924                 }
925                 subdir = *dir;
926                 if (subdir && page_private(subdir)) {
927                         ptr = shmem_swp_map(subdir);
928                         size = limit - idx;
929                         if (size > ENTRIES_PER_PAGE)
930                                 size = ENTRIES_PER_PAGE;
931                         offset = shmem_find_swp(entry, ptr, ptr+size);
932                         shmem_swp_unmap(ptr);
933                         if (offset >= 0) {
934                                 shmem_dir_unmap(dir);
935                                 ptr = shmem_swp_map(subdir);
936                                 goto found;
937                         }
938                 }
939         }
940 lost1:
941         shmem_dir_unmap(dir-1);
942 lost2:
943         spin_unlock(&info->lock);
944         return 0;
945 found:
946         idx += offset;
947         ptr += offset;
948
949         /*
950          * Move _head_ to start search for next from here.
951          * But be careful: shmem_evict_inode checks list_empty without taking
952          * mutex, and there's an instant in list_move_tail when info->swaplist
953          * would appear empty, if it were the only one on shmem_swaplist.  We
954          * could avoid doing it if inode NULL; or use this minor optimization.
955          */
956         if (shmem_swaplist.next != &info->swaplist)
957                 list_move_tail(&shmem_swaplist, &info->swaplist);
958
959         /*
960          * We rely on shmem_swaplist_mutex, not only to protect the swaplist,
961          * but also to hold up shmem_evict_inode(): so inode cannot be freed
962          * beneath us (pagelock doesn't help until the page is in pagecache).
963          */
964         mapping = info->vfs_inode.i_mapping;
965         error = add_to_page_cache_locked(page, mapping, idx, GFP_NOWAIT);
966         /* which does mem_cgroup_uncharge_cache_page on error */
967
968         if (error == -EEXIST) {
969                 struct page *filepage = find_get_page(mapping, idx);
970                 error = 1;
971                 if (filepage) {
972                         /*
973                          * There might be a more uptodate page coming down
974                          * from a stacked writepage: forget our swappage if so.
975                          */
976                         if (PageUptodate(filepage))
977                                 error = 0;
978                         page_cache_release(filepage);
979                 }
980         }
981         if (!error) {
982                 delete_from_swap_cache(page);
983                 set_page_dirty(page);
984                 info->flags |= SHMEM_PAGEIN;
985                 shmem_swp_set(info, ptr, 0);
986                 swap_free(entry);
987                 error = 1;      /* not an error, but entry was found */
988         }
989         shmem_swp_unmap(ptr);
990         spin_unlock(&info->lock);
991         return error;
992 }
993
994 /*
995  * shmem_unuse() search for an eventually swapped out shmem page.
996  */
997 int shmem_unuse(swp_entry_t entry, struct page *page)
998 {
999         struct list_head *p, *next;
1000         struct shmem_inode_info *info;
1001         int found = 0;
1002         int error;
1003
1004         /*
1005          * Charge page using GFP_KERNEL while we can wait, before taking
1006          * the shmem_swaplist_mutex which might hold up shmem_writepage().
1007          * Charged back to the user (not to caller) when swap account is used.
1008          * add_to_page_cache() will be called with GFP_NOWAIT.
1009          */
1010         error = mem_cgroup_cache_charge(page, current->mm, GFP_KERNEL);
1011         if (error)
1012                 goto out;
1013         /*
1014          * Try to preload while we can wait, to not make a habit of
1015          * draining atomic reserves; but don't latch on to this cpu,
1016          * it's okay if sometimes we get rescheduled after this.
1017          */
1018         error = radix_tree_preload(GFP_KERNEL);
1019         if (error)
1020                 goto uncharge;
1021         radix_tree_preload_end();
1022
1023         mutex_lock(&shmem_swaplist_mutex);
1024         list_for_each_safe(p, next, &shmem_swaplist) {
1025                 info = list_entry(p, struct shmem_inode_info, swaplist);
1026                 found = shmem_unuse_inode(info, entry, page);
1027                 cond_resched();
1028                 if (found)
1029                         break;
1030         }
1031         mutex_unlock(&shmem_swaplist_mutex);
1032
1033 uncharge:
1034         if (!found)
1035                 mem_cgroup_uncharge_cache_page(page);
1036         if (found < 0)
1037                 error = found;
1038 out:
1039         unlock_page(page);
1040         page_cache_release(page);
1041         return error;
1042 }
1043
1044 /*
1045  * Move the page from the page cache to the swap cache.
1046  */
1047 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1048 {
1049         struct shmem_inode_info *info;
1050         swp_entry_t *entry, swap;
1051         struct address_space *mapping;
1052         unsigned long index;
1053         struct inode *inode;
1054
1055         BUG_ON(!PageLocked(page));
1056         mapping = page->mapping;
1057         index = page->index;
1058         inode = mapping->host;
1059         info = SHMEM_I(inode);
1060         if (info->flags & VM_LOCKED)
1061                 goto redirty;
1062         if (!total_swap_pages)
1063                 goto redirty;
1064
1065         /*
1066          * shmem_backing_dev_info's capabilities prevent regular writeback or
1067          * sync from ever calling shmem_writepage; but a stacking filesystem
1068          * may use the ->writepage of its underlying filesystem, in which case
1069          * tmpfs should write out to swap only in response to memory pressure,
1070          * and not for the writeback threads or sync.  However, in those cases,
1071          * we do still want to check if there's a redundant swappage to be
1072          * discarded.
1073          */
1074         if (wbc->for_reclaim)
1075                 swap = get_swap_page();
1076         else
1077                 swap.val = 0;
1078
1079         /*
1080          * Add inode to shmem_unuse()'s list of swapped-out inodes,
1081          * if it's not already there.  Do it now because we cannot take
1082          * mutex while holding spinlock, and must do so before the page
1083          * is moved to swap cache, when its pagelock no longer protects
1084          * the inode from eviction.  But don't unlock the mutex until
1085          * we've taken the spinlock, because shmem_unuse_inode() will
1086          * prune a !swapped inode from the swaplist under both locks.
1087          */
1088         if (swap.val) {
1089                 mutex_lock(&shmem_swaplist_mutex);
1090                 if (list_empty(&info->swaplist))
1091                         list_add_tail(&info->swaplist, &shmem_swaplist);
1092         }
1093
1094         spin_lock(&info->lock);
1095         if (swap.val)
1096                 mutex_unlock(&shmem_swaplist_mutex);
1097
1098         if (index >= info->next_index) {
1099                 BUG_ON(!(info->flags & SHMEM_TRUNCATE));
1100                 goto unlock;
1101         }
1102         entry = shmem_swp_entry(info, index, NULL);
1103         if (entry->val) {
1104                 /*
1105                  * The more uptodate page coming down from a stacked
1106                  * writepage should replace our old swappage.
1107                  */
1108                 free_swap_and_cache(*entry);
1109                 shmem_swp_set(info, entry, 0);
1110         }
1111         shmem_recalc_inode(inode);
1112
1113         if (swap.val && add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
1114                 delete_from_page_cache(page);
1115                 shmem_swp_set(info, entry, swap.val);
1116                 shmem_swp_unmap(entry);
1117                 spin_unlock(&info->lock);
1118                 swap_shmem_alloc(swap);
1119                 BUG_ON(page_mapped(page));
1120                 swap_writepage(page, wbc);
1121                 return 0;
1122         }
1123
1124         shmem_swp_unmap(entry);
1125 unlock:
1126         spin_unlock(&info->lock);
1127         /*
1128          * add_to_swap_cache() doesn't return -EEXIST, so we can safely
1129          * clear SWAP_HAS_CACHE flag.
1130          */
1131         swapcache_free(swap, NULL);
1132 redirty:
1133         set_page_dirty(page);
1134         if (wbc->for_reclaim)
1135                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
1136         unlock_page(page);
1137         return 0;
1138 }
1139
1140 #ifdef CONFIG_NUMA
1141 #ifdef CONFIG_TMPFS
1142 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1143 {
1144         char buffer[64];
1145
1146         if (!mpol || mpol->mode == MPOL_DEFAULT)
1147                 return;         /* show nothing */
1148
1149         mpol_to_str(buffer, sizeof(buffer), mpol, 1);
1150
1151         seq_printf(seq, ",mpol=%s", buffer);
1152 }
1153
1154 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1155 {
1156         struct mempolicy *mpol = NULL;
1157         if (sbinfo->mpol) {
1158                 spin_lock(&sbinfo->stat_lock);  /* prevent replace/use races */
1159                 mpol = sbinfo->mpol;
1160                 mpol_get(mpol);
1161                 spin_unlock(&sbinfo->stat_lock);
1162         }
1163         return mpol;
1164 }
1165 #endif /* CONFIG_TMPFS */
1166
1167 static struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1168                         struct shmem_inode_info *info, unsigned long idx)
1169 {
1170         struct mempolicy mpol, *spol;
1171         struct vm_area_struct pvma;
1172         struct page *page;
1173
1174         spol = mpol_cond_copy(&mpol,
1175                                 mpol_shared_policy_lookup(&info->policy, idx));
1176
1177         /* Create a pseudo vma that just contains the policy */
1178         pvma.vm_start = 0;
1179         pvma.vm_pgoff = idx;
1180         pvma.vm_ops = NULL;
1181         pvma.vm_policy = spol;
1182         page = swapin_readahead(entry, gfp, &pvma, 0);
1183         return page;
1184 }
1185
1186 static struct page *shmem_alloc_page(gfp_t gfp,
1187                         struct shmem_inode_info *info, unsigned long idx)
1188 {
1189         struct vm_area_struct pvma;
1190
1191         /* Create a pseudo vma that just contains the policy */
1192         pvma.vm_start = 0;
1193         pvma.vm_pgoff = idx;
1194         pvma.vm_ops = NULL;
1195         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, idx);
1196
1197         /*
1198          * alloc_page_vma() will drop the shared policy reference
1199          */
1200         return alloc_page_vma(gfp, &pvma, 0);
1201 }
1202 #else /* !CONFIG_NUMA */
1203 #ifdef CONFIG_TMPFS
1204 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *p)
1205 {
1206 }
1207 #endif /* CONFIG_TMPFS */
1208
1209 static inline struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1210                         struct shmem_inode_info *info, unsigned long idx)
1211 {
1212         return swapin_readahead(entry, gfp, NULL, 0);
1213 }
1214
1215 static inline struct page *shmem_alloc_page(gfp_t gfp,
1216                         struct shmem_inode_info *info, unsigned long idx)
1217 {
1218         return alloc_page(gfp);
1219 }
1220 #endif /* CONFIG_NUMA */
1221
1222 #if !defined(CONFIG_NUMA) || !defined(CONFIG_TMPFS)
1223 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1224 {
1225         return NULL;
1226 }
1227 #endif
1228
1229 /*
1230  * shmem_getpage - either get the page from swap or allocate a new one
1231  *
1232  * If we allocate a new one we do not mark it dirty. That's up to the
1233  * vm. If we swap it in we mark it dirty since we also free the swap
1234  * entry since a page cannot live in both the swap and page cache
1235  */
1236 static int shmem_getpage(struct inode *inode, unsigned long idx,
1237                         struct page **pagep, enum sgp_type sgp, int *type)
1238 {
1239         struct address_space *mapping = inode->i_mapping;
1240         struct shmem_inode_info *info = SHMEM_I(inode);
1241         struct shmem_sb_info *sbinfo;
1242         struct page *filepage = *pagep;
1243         struct page *swappage;
1244         struct page *prealloc_page = NULL;
1245         swp_entry_t *entry;
1246         swp_entry_t swap;
1247         gfp_t gfp;
1248         int error;
1249
1250         if (idx >= SHMEM_MAX_INDEX)
1251                 return -EFBIG;
1252
1253         if (type)
1254                 *type = 0;
1255
1256         /*
1257          * Normally, filepage is NULL on entry, and either found
1258          * uptodate immediately, or allocated and zeroed, or read
1259          * in under swappage, which is then assigned to filepage.
1260          * But shmem_readpage (required for splice) passes in a locked
1261          * filepage, which may be found not uptodate by other callers
1262          * too, and may need to be copied from the swappage read in.
1263          */
1264 repeat:
1265         if (!filepage)
1266                 filepage = find_lock_page(mapping, idx);
1267         if (filepage && PageUptodate(filepage))
1268                 goto done;
1269         gfp = mapping_gfp_mask(mapping);
1270         if (!filepage) {
1271                 /*
1272                  * Try to preload while we can wait, to not make a habit of
1273                  * draining atomic reserves; but don't latch on to this cpu.
1274                  */
1275                 error = radix_tree_preload(gfp & ~__GFP_HIGHMEM);
1276                 if (error)
1277                         goto failed;
1278                 radix_tree_preload_end();
1279                 if (sgp != SGP_READ && !prealloc_page) {
1280                         /* We don't care if this fails */
1281                         prealloc_page = shmem_alloc_page(gfp, info, idx);
1282                         if (prealloc_page) {
1283                                 if (mem_cgroup_cache_charge(prealloc_page,
1284                                                 current->mm, GFP_KERNEL)) {
1285                                         page_cache_release(prealloc_page);
1286                                         prealloc_page = NULL;
1287                                 }
1288                         }
1289                 }
1290         }
1291         error = 0;
1292
1293         spin_lock(&info->lock);
1294         shmem_recalc_inode(inode);
1295         entry = shmem_swp_alloc(info, idx, sgp);
1296         if (IS_ERR(entry)) {
1297                 spin_unlock(&info->lock);
1298                 error = PTR_ERR(entry);
1299                 goto failed;
1300         }
1301         swap = *entry;
1302
1303         if (swap.val) {
1304                 /* Look it up and read it in.. */
1305                 swappage = lookup_swap_cache(swap);
1306                 if (!swappage) {
1307                         shmem_swp_unmap(entry);
1308                         spin_unlock(&info->lock);
1309                         /* here we actually do the io */
1310                         if (type)
1311                                 *type |= VM_FAULT_MAJOR;
1312                         swappage = shmem_swapin(swap, gfp, info, idx);
1313                         if (!swappage) {
1314                                 spin_lock(&info->lock);
1315                                 entry = shmem_swp_alloc(info, idx, sgp);
1316                                 if (IS_ERR(entry))
1317                                         error = PTR_ERR(entry);
1318                                 else {
1319                                         if (entry->val == swap.val)
1320                                                 error = -ENOMEM;
1321                                         shmem_swp_unmap(entry);
1322                                 }
1323                                 spin_unlock(&info->lock);
1324                                 if (error)
1325                                         goto failed;
1326                                 goto repeat;
1327                         }
1328                         wait_on_page_locked(swappage);
1329                         page_cache_release(swappage);
1330                         goto repeat;
1331                 }
1332
1333                 /* We have to do this with page locked to prevent races */
1334                 if (!trylock_page(swappage)) {
1335                         shmem_swp_unmap(entry);
1336                         spin_unlock(&info->lock);
1337                         wait_on_page_locked(swappage);
1338                         page_cache_release(swappage);
1339                         goto repeat;
1340                 }
1341                 if (PageWriteback(swappage)) {
1342                         shmem_swp_unmap(entry);
1343                         spin_unlock(&info->lock);
1344                         wait_on_page_writeback(swappage);
1345                         unlock_page(swappage);
1346                         page_cache_release(swappage);
1347                         goto repeat;
1348                 }
1349                 if (!PageUptodate(swappage)) {
1350                         shmem_swp_unmap(entry);
1351                         spin_unlock(&info->lock);
1352                         unlock_page(swappage);
1353                         page_cache_release(swappage);
1354                         error = -EIO;
1355                         goto failed;
1356                 }
1357
1358                 if (filepage) {
1359                         shmem_swp_set(info, entry, 0);
1360                         shmem_swp_unmap(entry);
1361                         delete_from_swap_cache(swappage);
1362                         spin_unlock(&info->lock);
1363                         copy_highpage(filepage, swappage);
1364                         unlock_page(swappage);
1365                         page_cache_release(swappage);
1366                         flush_dcache_page(filepage);
1367                         SetPageUptodate(filepage);
1368                         set_page_dirty(filepage);
1369                         swap_free(swap);
1370                 } else if (!(error = add_to_page_cache_locked(swappage, mapping,
1371                                         idx, GFP_NOWAIT))) {
1372                         info->flags |= SHMEM_PAGEIN;
1373                         shmem_swp_set(info, entry, 0);
1374                         shmem_swp_unmap(entry);
1375                         delete_from_swap_cache(swappage);
1376                         spin_unlock(&info->lock);
1377                         filepage = swappage;
1378                         set_page_dirty(filepage);
1379                         swap_free(swap);
1380                 } else {
1381                         shmem_swp_unmap(entry);
1382                         spin_unlock(&info->lock);
1383                         if (error == -ENOMEM) {
1384                                 /*
1385                                  * reclaim from proper memory cgroup and
1386                                  * call memcg's OOM if needed.
1387                                  */
1388                                 error = mem_cgroup_shmem_charge_fallback(
1389                                                                 swappage,
1390                                                                 current->mm,
1391                                                                 gfp);
1392                                 if (error) {
1393                                         unlock_page(swappage);
1394                                         page_cache_release(swappage);
1395                                         goto failed;
1396                                 }
1397                         }
1398                         unlock_page(swappage);
1399                         page_cache_release(swappage);
1400                         goto repeat;
1401                 }
1402         } else if (sgp == SGP_READ && !filepage) {
1403                 shmem_swp_unmap(entry);
1404                 filepage = find_get_page(mapping, idx);
1405                 if (filepage &&
1406                     (!PageUptodate(filepage) || !trylock_page(filepage))) {
1407                         spin_unlock(&info->lock);
1408                         wait_on_page_locked(filepage);
1409                         page_cache_release(filepage);
1410                         filepage = NULL;
1411                         goto repeat;
1412                 }
1413                 spin_unlock(&info->lock);
1414         } else {
1415                 shmem_swp_unmap(entry);
1416                 sbinfo = SHMEM_SB(inode->i_sb);
1417                 if (sbinfo->max_blocks) {
1418                         if (percpu_counter_compare(&sbinfo->used_blocks,
1419                                                 sbinfo->max_blocks) >= 0 ||
1420                             shmem_acct_block(info->flags))
1421                                 goto nospace;
1422                         percpu_counter_inc(&sbinfo->used_blocks);
1423                         spin_lock(&inode->i_lock);
1424                         inode->i_blocks += BLOCKS_PER_PAGE;
1425                         spin_unlock(&inode->i_lock);
1426                 } else if (shmem_acct_block(info->flags))
1427                         goto nospace;
1428
1429                 if (!filepage) {
1430                         int ret;
1431
1432                         if (!prealloc_page) {
1433                                 spin_unlock(&info->lock);
1434                                 filepage = shmem_alloc_page(gfp, info, idx);
1435                                 if (!filepage) {
1436                                         shmem_unacct_blocks(info->flags, 1);
1437                                         shmem_free_blocks(inode, 1);
1438                                         error = -ENOMEM;
1439                                         goto failed;
1440                                 }
1441                                 SetPageSwapBacked(filepage);
1442
1443                                 /*
1444                                  * Precharge page while we can wait, compensate
1445                                  * after
1446                                  */
1447                                 error = mem_cgroup_cache_charge(filepage,
1448                                         current->mm, GFP_KERNEL);
1449                                 if (error) {
1450                                         page_cache_release(filepage);
1451                                         shmem_unacct_blocks(info->flags, 1);
1452                                         shmem_free_blocks(inode, 1);
1453                                         filepage = NULL;
1454                                         goto failed;
1455                                 }
1456
1457                                 spin_lock(&info->lock);
1458                         } else {
1459                                 filepage = prealloc_page;
1460                                 prealloc_page = NULL;
1461                                 SetPageSwapBacked(filepage);
1462                         }
1463
1464                         entry = shmem_swp_alloc(info, idx, sgp);
1465                         if (IS_ERR(entry))
1466                                 error = PTR_ERR(entry);
1467                         else {
1468                                 swap = *entry;
1469                                 shmem_swp_unmap(entry);
1470                         }
1471                         ret = error || swap.val;
1472                         if (ret)
1473                                 mem_cgroup_uncharge_cache_page(filepage);
1474                         else
1475                                 ret = add_to_page_cache_lru(filepage, mapping,
1476                                                 idx, GFP_NOWAIT);
1477                         /*
1478                          * At add_to_page_cache_lru() failure, uncharge will
1479                          * be done automatically.
1480                          */
1481                         if (ret) {
1482                                 spin_unlock(&info->lock);
1483                                 page_cache_release(filepage);
1484                                 shmem_unacct_blocks(info->flags, 1);
1485                                 shmem_free_blocks(inode, 1);
1486                                 filepage = NULL;
1487                                 if (error)
1488                                         goto failed;
1489                                 goto repeat;
1490                         }
1491                         info->flags |= SHMEM_PAGEIN;
1492                 }
1493
1494                 info->alloced++;
1495                 spin_unlock(&info->lock);
1496                 clear_highpage(filepage);
1497                 flush_dcache_page(filepage);
1498                 SetPageUptodate(filepage);
1499                 if (sgp == SGP_DIRTY)
1500                         set_page_dirty(filepage);
1501         }
1502 done:
1503         *pagep = filepage;
1504         error = 0;
1505         goto out;
1506
1507 nospace:
1508         /*
1509          * Perhaps the page was brought in from swap between find_lock_page
1510          * and taking info->lock?  We allow for that at add_to_page_cache_lru,
1511          * but must also avoid reporting a spurious ENOSPC while working on a
1512          * full tmpfs.  (When filepage has been passed in to shmem_getpage, it
1513          * is already in page cache, which prevents this race from occurring.)
1514          */
1515         if (!filepage) {
1516                 struct page *page = find_get_page(mapping, idx);
1517                 if (page) {
1518                         spin_unlock(&info->lock);
1519                         page_cache_release(page);
1520                         goto repeat;
1521                 }
1522         }
1523         spin_unlock(&info->lock);
1524         error = -ENOSPC;
1525 failed:
1526         if (*pagep != filepage) {
1527                 unlock_page(filepage);
1528                 page_cache_release(filepage);
1529         }
1530 out:
1531         if (prealloc_page) {
1532                 mem_cgroup_uncharge_cache_page(prealloc_page);
1533                 page_cache_release(prealloc_page);
1534         }
1535         return error;
1536 }
1537
1538 static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1539 {
1540         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1541         int error;
1542         int ret;
1543
1544         if (((loff_t)vmf->pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
1545                 return VM_FAULT_SIGBUS;
1546
1547         error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_CACHE, &ret);
1548         if (error)
1549                 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1550         if (ret & VM_FAULT_MAJOR) {
1551                 count_vm_event(PGMAJFAULT);
1552                 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1553         }
1554         return ret | VM_FAULT_LOCKED;
1555 }
1556
1557 #ifdef CONFIG_NUMA
1558 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *new)
1559 {
1560         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1561         return mpol_set_shared_policy(&SHMEM_I(i)->policy, vma, new);
1562 }
1563
1564 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
1565                                           unsigned long addr)
1566 {
1567         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1568         unsigned long idx;
1569
1570         idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1571         return mpol_shared_policy_lookup(&SHMEM_I(i)->policy, idx);
1572 }
1573 #endif
1574
1575 int shmem_lock(struct file *file, int lock, struct user_struct *user)
1576 {
1577         struct inode *inode = file->f_path.dentry->d_inode;
1578         struct shmem_inode_info *info = SHMEM_I(inode);
1579         int retval = -ENOMEM;
1580
1581         spin_lock(&info->lock);
1582         if (lock && !(info->flags & VM_LOCKED)) {
1583                 if (!user_shm_lock(inode->i_size, user))
1584                         goto out_nomem;
1585                 info->flags |= VM_LOCKED;
1586                 mapping_set_unevictable(file->f_mapping);
1587         }
1588         if (!lock && (info->flags & VM_LOCKED) && user) {
1589                 user_shm_unlock(inode->i_size, user);
1590                 info->flags &= ~VM_LOCKED;
1591                 mapping_clear_unevictable(file->f_mapping);
1592                 scan_mapping_unevictable_pages(file->f_mapping);
1593         }
1594         retval = 0;
1595
1596 out_nomem:
1597         spin_unlock(&info->lock);
1598         return retval;
1599 }
1600
1601 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1602 {
1603         file_accessed(file);
1604         vma->vm_ops = &shmem_vm_ops;
1605         vma->vm_flags |= VM_CAN_NONLINEAR;
1606         return 0;
1607 }
1608
1609 static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
1610                                      int mode, dev_t dev, unsigned long flags)
1611 {
1612         struct inode *inode;
1613         struct shmem_inode_info *info;
1614         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1615
1616         if (shmem_reserve_inode(sb))
1617                 return NULL;
1618
1619         inode = new_inode(sb);
1620         if (inode) {
1621                 inode->i_ino = get_next_ino();
1622                 inode_init_owner(inode, dir, mode);
1623                 inode->i_blocks = 0;
1624                 inode->i_mapping->backing_dev_info = &shmem_backing_dev_info;
1625                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1626                 inode->i_generation = get_seconds();
1627                 info = SHMEM_I(inode);
1628                 memset(info, 0, (char *)inode - (char *)info);
1629                 spin_lock_init(&info->lock);
1630                 info->flags = flags & VM_NORESERVE;
1631                 INIT_LIST_HEAD(&info->swaplist);
1632                 INIT_LIST_HEAD(&info->xattr_list);
1633                 cache_no_acl(inode);
1634
1635                 switch (mode & S_IFMT) {
1636                 default:
1637                         inode->i_op = &shmem_special_inode_operations;
1638                         init_special_inode(inode, mode, dev);
1639                         break;
1640                 case S_IFREG:
1641                         inode->i_mapping->a_ops = &shmem_aops;
1642                         inode->i_op = &shmem_inode_operations;
1643                         inode->i_fop = &shmem_file_operations;
1644                         mpol_shared_policy_init(&info->policy,
1645                                                  shmem_get_sbmpol(sbinfo));
1646                         break;
1647                 case S_IFDIR:
1648                         inc_nlink(inode);
1649                         /* Some things misbehave if size == 0 on a directory */
1650                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
1651                         inode->i_op = &shmem_dir_inode_operations;
1652                         inode->i_fop = &simple_dir_operations;
1653                         break;
1654                 case S_IFLNK:
1655                         /*
1656                          * Must not load anything in the rbtree,
1657                          * mpol_free_shared_policy will not be called.
1658                          */
1659                         mpol_shared_policy_init(&info->policy, NULL);
1660                         break;
1661                 }
1662         } else
1663                 shmem_free_inode(sb);
1664         return inode;
1665 }
1666
1667 #ifdef CONFIG_TMPFS
1668 static const struct inode_operations shmem_symlink_inode_operations;
1669 static const struct inode_operations shmem_symlink_inline_operations;
1670
1671 /*
1672  * Normally tmpfs avoids the use of shmem_readpage and shmem_write_begin;
1673  * but providing them allows a tmpfs file to be used for splice, sendfile, and
1674  * below the loop driver, in the generic fashion that many filesystems support.
1675  */
1676 static int shmem_readpage(struct file *file, struct page *page)
1677 {
1678         struct inode *inode = page->mapping->host;
1679         int error = shmem_getpage(inode, page->index, &page, SGP_CACHE, NULL);
1680         unlock_page(page);
1681         return error;
1682 }
1683
1684 static int
1685 shmem_write_begin(struct file *file, struct address_space *mapping,
1686                         loff_t pos, unsigned len, unsigned flags,
1687                         struct page **pagep, void **fsdata)
1688 {
1689         struct inode *inode = mapping->host;
1690         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1691         *pagep = NULL;
1692         return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1693 }
1694
1695 static int
1696 shmem_write_end(struct file *file, struct address_space *mapping,
1697                         loff_t pos, unsigned len, unsigned copied,
1698                         struct page *page, void *fsdata)
1699 {
1700         struct inode *inode = mapping->host;
1701
1702         if (pos + copied > inode->i_size)
1703                 i_size_write(inode, pos + copied);
1704
1705         set_page_dirty(page);
1706         unlock_page(page);
1707         page_cache_release(page);
1708
1709         return copied;
1710 }
1711
1712 static void do_shmem_file_read(struct file *filp, loff_t *ppos, read_descriptor_t *desc, read_actor_t actor)
1713 {
1714         struct inode *inode = filp->f_path.dentry->d_inode;
1715         struct address_space *mapping = inode->i_mapping;
1716         unsigned long index, offset;
1717         enum sgp_type sgp = SGP_READ;
1718
1719         /*
1720          * Might this read be for a stacking filesystem?  Then when reading
1721          * holes of a sparse file, we actually need to allocate those pages,
1722          * and even mark them dirty, so it cannot exceed the max_blocks limit.
1723          */
1724         if (segment_eq(get_fs(), KERNEL_DS))
1725                 sgp = SGP_DIRTY;
1726
1727         index = *ppos >> PAGE_CACHE_SHIFT;
1728         offset = *ppos & ~PAGE_CACHE_MASK;
1729
1730         for (;;) {
1731                 struct page *page = NULL;
1732                 unsigned long end_index, nr, ret;
1733                 loff_t i_size = i_size_read(inode);
1734
1735                 end_index = i_size >> PAGE_CACHE_SHIFT;
1736                 if (index > end_index)
1737                         break;
1738                 if (index == end_index) {
1739                         nr = i_size & ~PAGE_CACHE_MASK;
1740                         if (nr <= offset)
1741                                 break;
1742                 }
1743
1744                 desc->error = shmem_getpage(inode, index, &page, sgp, NULL);
1745                 if (desc->error) {
1746                         if (desc->error == -EINVAL)
1747                                 desc->error = 0;
1748                         break;
1749                 }
1750                 if (page)
1751                         unlock_page(page);
1752
1753                 /*
1754                  * We must evaluate after, since reads (unlike writes)
1755                  * are called without i_mutex protection against truncate
1756                  */
1757                 nr = PAGE_CACHE_SIZE;
1758                 i_size = i_size_read(inode);
1759                 end_index = i_size >> PAGE_CACHE_SHIFT;
1760                 if (index == end_index) {
1761                         nr = i_size & ~PAGE_CACHE_MASK;
1762                         if (nr <= offset) {
1763                                 if (page)
1764                                         page_cache_release(page);
1765                                 break;
1766                         }
1767                 }
1768                 nr -= offset;
1769
1770                 if (page) {
1771                         /*
1772                          * If users can be writing to this page using arbitrary
1773                          * virtual addresses, take care about potential aliasing
1774                          * before reading the page on the kernel side.
1775                          */
1776                         if (mapping_writably_mapped(mapping))
1777                                 flush_dcache_page(page);
1778                         /*
1779                          * Mark the page accessed if we read the beginning.
1780                          */
1781                         if (!offset)
1782                                 mark_page_accessed(page);
1783                 } else {
1784                         page = ZERO_PAGE(0);
1785                         page_cache_get(page);
1786                 }
1787
1788                 /*
1789                  * Ok, we have the page, and it's up-to-date, so
1790                  * now we can copy it to user space...
1791                  *
1792                  * The actor routine returns how many bytes were actually used..
1793                  * NOTE! This may not be the same as how much of a user buffer
1794                  * we filled up (we may be padding etc), so we can only update
1795                  * "pos" here (the actor routine has to update the user buffer
1796                  * pointers and the remaining count).
1797                  */
1798                 ret = actor(desc, page, offset, nr);
1799                 offset += ret;
1800                 index += offset >> PAGE_CACHE_SHIFT;
1801                 offset &= ~PAGE_CACHE_MASK;
1802
1803                 page_cache_release(page);
1804                 if (ret != nr || !desc->count)
1805                         break;
1806
1807                 cond_resched();
1808         }
1809
1810         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1811         file_accessed(filp);
1812 }
1813
1814 static ssize_t shmem_file_aio_read(struct kiocb *iocb,
1815                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
1816 {
1817         struct file *filp = iocb->ki_filp;
1818         ssize_t retval;
1819         unsigned long seg;
1820         size_t count;
1821         loff_t *ppos = &iocb->ki_pos;
1822
1823         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1824         if (retval)
1825                 return retval;
1826
1827         for (seg = 0; seg < nr_segs; seg++) {
1828                 read_descriptor_t desc;
1829
1830                 desc.written = 0;
1831                 desc.arg.buf = iov[seg].iov_base;
1832                 desc.count = iov[seg].iov_len;
1833                 if (desc.count == 0)
1834                         continue;
1835                 desc.error = 0;
1836                 do_shmem_file_read(filp, ppos, &desc, file_read_actor);
1837                 retval += desc.written;
1838                 if (desc.error) {
1839                         retval = retval ?: desc.error;
1840                         break;
1841                 }
1842                 if (desc.count > 0)
1843                         break;
1844         }
1845         return retval;
1846 }
1847
1848 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
1849 {
1850         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
1851
1852         buf->f_type = TMPFS_MAGIC;
1853         buf->f_bsize = PAGE_CACHE_SIZE;
1854         buf->f_namelen = NAME_MAX;
1855         if (sbinfo->max_blocks) {
1856                 buf->f_blocks = sbinfo->max_blocks;
1857                 buf->f_bavail = buf->f_bfree =
1858                                 sbinfo->max_blocks - percpu_counter_sum(&sbinfo->used_blocks);
1859         }
1860         if (sbinfo->max_inodes) {
1861                 buf->f_files = sbinfo->max_inodes;
1862                 buf->f_ffree = sbinfo->free_inodes;
1863         }
1864         /* else leave those fields 0 like simple_statfs */
1865         return 0;
1866 }
1867
1868 /*
1869  * File creation. Allocate an inode, and we're done..
1870  */
1871 static int
1872 shmem_mknod(struct inode *dir, struct dentry *dentry, int mode, dev_t dev)
1873 {
1874         struct inode *inode;
1875         int error = -ENOSPC;
1876
1877         inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
1878         if (inode) {
1879                 error = security_inode_init_security(inode, dir,
1880                                                      &dentry->d_name, NULL,
1881                                                      NULL, NULL);
1882                 if (error) {
1883                         if (error != -EOPNOTSUPP) {
1884                                 iput(inode);
1885                                 return error;
1886                         }
1887                 }
1888 #ifdef CONFIG_TMPFS_POSIX_ACL
1889                 error = generic_acl_init(inode, dir);
1890                 if (error) {
1891                         iput(inode);
1892                         return error;
1893                 }
1894 #else
1895                 error = 0;
1896 #endif
1897                 dir->i_size += BOGO_DIRENT_SIZE;
1898                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1899                 d_instantiate(dentry, inode);
1900                 dget(dentry); /* Extra count - pin the dentry in core */
1901         }
1902         return error;
1903 }
1904
1905 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, int mode)
1906 {
1907         int error;
1908
1909         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
1910                 return error;
1911         inc_nlink(dir);
1912         return 0;
1913 }
1914
1915 static int shmem_create(struct inode *dir, struct dentry *dentry, int mode,
1916                 struct nameidata *nd)
1917 {
1918         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
1919 }
1920
1921 /*
1922  * Link a file..
1923  */
1924 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
1925 {
1926         struct inode *inode = old_dentry->d_inode;
1927         int ret;
1928
1929         /*
1930          * No ordinary (disk based) filesystem counts links as inodes;
1931          * but each new link needs a new dentry, pinning lowmem, and
1932          * tmpfs dentries cannot be pruned until they are unlinked.
1933          */
1934         ret = shmem_reserve_inode(inode->i_sb);
1935         if (ret)
1936                 goto out;
1937
1938         dir->i_size += BOGO_DIRENT_SIZE;
1939         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1940         inc_nlink(inode);
1941         ihold(inode);   /* New dentry reference */
1942         dget(dentry);           /* Extra pinning count for the created dentry */
1943         d_instantiate(dentry, inode);
1944 out:
1945         return ret;
1946 }
1947
1948 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
1949 {
1950         struct inode *inode = dentry->d_inode;
1951
1952         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
1953                 shmem_free_inode(inode->i_sb);
1954
1955         dir->i_size -= BOGO_DIRENT_SIZE;
1956         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1957         drop_nlink(inode);
1958         dput(dentry);   /* Undo the count from "create" - this does all the work */
1959         return 0;
1960 }
1961
1962 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
1963 {
1964         if (!simple_empty(dentry))
1965                 return -ENOTEMPTY;
1966
1967         drop_nlink(dentry->d_inode);
1968         drop_nlink(dir);
1969         return shmem_unlink(dir, dentry);
1970 }
1971
1972 /*
1973  * The VFS layer already does all the dentry stuff for rename,
1974  * we just have to decrement the usage count for the target if
1975  * it exists so that the VFS layer correctly free's it when it
1976  * gets overwritten.
1977  */
1978 static int shmem_rename(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
1979 {
1980         struct inode *inode = old_dentry->d_inode;
1981         int they_are_dirs = S_ISDIR(inode->i_mode);
1982
1983         if (!simple_empty(new_dentry))
1984                 return -ENOTEMPTY;
1985
1986         if (new_dentry->d_inode) {
1987                 (void) shmem_unlink(new_dir, new_dentry);
1988                 if (they_are_dirs)
1989                         drop_nlink(old_dir);
1990         } else if (they_are_dirs) {
1991                 drop_nlink(old_dir);
1992                 inc_nlink(new_dir);
1993         }
1994
1995         old_dir->i_size -= BOGO_DIRENT_SIZE;
1996         new_dir->i_size += BOGO_DIRENT_SIZE;
1997         old_dir->i_ctime = old_dir->i_mtime =
1998         new_dir->i_ctime = new_dir->i_mtime =
1999         inode->i_ctime = CURRENT_TIME;
2000         return 0;
2001 }
2002
2003 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
2004 {
2005         int error;
2006         int len;
2007         struct inode *inode;
2008         struct page *page = NULL;
2009         char *kaddr;
2010         struct shmem_inode_info *info;
2011
2012         len = strlen(symname) + 1;
2013         if (len > PAGE_CACHE_SIZE)
2014                 return -ENAMETOOLONG;
2015
2016         inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0, VM_NORESERVE);
2017         if (!inode)
2018                 return -ENOSPC;
2019
2020         error = security_inode_init_security(inode, dir, &dentry->d_name, NULL,
2021                                              NULL, NULL);
2022         if (error) {
2023                 if (error != -EOPNOTSUPP) {
2024                         iput(inode);
2025                         return error;
2026                 }
2027                 error = 0;
2028         }
2029
2030         info = SHMEM_I(inode);
2031         inode->i_size = len-1;
2032         if (len <= SHMEM_SYMLINK_INLINE_LEN) {
2033                 /* do it inline */
2034                 memcpy(info->inline_symlink, symname, len);
2035                 inode->i_op = &shmem_symlink_inline_operations;
2036         } else {
2037                 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
2038                 if (error) {
2039                         iput(inode);
2040                         return error;
2041                 }
2042                 inode->i_mapping->a_ops = &shmem_aops;
2043                 inode->i_op = &shmem_symlink_inode_operations;
2044                 kaddr = kmap_atomic(page, KM_USER0);
2045                 memcpy(kaddr, symname, len);
2046                 kunmap_atomic(kaddr, KM_USER0);
2047                 set_page_dirty(page);
2048                 unlock_page(page);
2049                 page_cache_release(page);
2050         }
2051         dir->i_size += BOGO_DIRENT_SIZE;
2052         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2053         d_instantiate(dentry, inode);
2054         dget(dentry);
2055         return 0;
2056 }
2057
2058 static void *shmem_follow_link_inline(struct dentry *dentry, struct nameidata *nd)
2059 {
2060         nd_set_link(nd, SHMEM_I(dentry->d_inode)->inline_symlink);
2061         return NULL;
2062 }
2063
2064 static void *shmem_follow_link(struct dentry *dentry, struct nameidata *nd)
2065 {
2066         struct page *page = NULL;
2067         int res = shmem_getpage(dentry->d_inode, 0, &page, SGP_READ, NULL);
2068         nd_set_link(nd, res ? ERR_PTR(res) : kmap(page));
2069         if (page)
2070                 unlock_page(page);
2071         return page;
2072 }
2073
2074 static void shmem_put_link(struct dentry *dentry, struct nameidata *nd, void *cookie)
2075 {
2076         if (!IS_ERR(nd_get_link(nd))) {
2077                 struct page *page = cookie;
2078                 kunmap(page);
2079                 mark_page_accessed(page);
2080                 page_cache_release(page);
2081         }
2082 }
2083
2084 #ifdef CONFIG_TMPFS_XATTR
2085 /*
2086  * Superblocks without xattr inode operations may get some security.* xattr
2087  * support from the LSM "for free". As soon as we have any other xattrs
2088  * like ACLs, we also need to implement the security.* handlers at
2089  * filesystem level, though.
2090  */
2091
2092 static int shmem_xattr_get(struct dentry *dentry, const char *name,
2093                            void *buffer, size_t size)
2094 {
2095         struct shmem_inode_info *info;
2096         struct shmem_xattr *xattr;
2097         int ret = -ENODATA;
2098
2099         info = SHMEM_I(dentry->d_inode);
2100
2101         spin_lock(&info->lock);
2102         list_for_each_entry(xattr, &info->xattr_list, list) {
2103                 if (strcmp(name, xattr->name))
2104                         continue;
2105
2106                 ret = xattr->size;
2107                 if (buffer) {
2108                         if (size < xattr->size)
2109                                 ret = -ERANGE;
2110                         else
2111                                 memcpy(buffer, xattr->value, xattr->size);
2112                 }
2113                 break;
2114         }
2115         spin_unlock(&info->lock);
2116         return ret;
2117 }
2118
2119 static int shmem_xattr_set(struct dentry *dentry, const char *name,
2120                            const void *value, size_t size, int flags)
2121 {
2122         struct inode *inode = dentry->d_inode;
2123         struct shmem_inode_info *info = SHMEM_I(inode);
2124         struct shmem_xattr *xattr;
2125         struct shmem_xattr *new_xattr = NULL;
2126         size_t len;
2127         int err = 0;
2128
2129         /* value == NULL means remove */
2130         if (value) {
2131                 /* wrap around? */
2132                 len = sizeof(*new_xattr) + size;
2133                 if (len <= sizeof(*new_xattr))
2134                         return -ENOMEM;
2135
2136                 new_xattr = kmalloc(len, GFP_KERNEL);
2137                 if (!new_xattr)
2138                         return -ENOMEM;
2139
2140                 new_xattr->name = kstrdup(name, GFP_KERNEL);
2141                 if (!new_xattr->name) {
2142                         kfree(new_xattr);
2143                         return -ENOMEM;
2144                 }
2145
2146                 new_xattr->size = size;
2147                 memcpy(new_xattr->value, value, size);
2148         }
2149
2150         spin_lock(&info->lock);
2151         list_for_each_entry(xattr, &info->xattr_list, list) {
2152                 if (!strcmp(name, xattr->name)) {
2153                         if (flags & XATTR_CREATE) {
2154                                 xattr = new_xattr;
2155                                 err = -EEXIST;
2156                         } else if (new_xattr) {
2157                                 list_replace(&xattr->list, &new_xattr->list);
2158                         } else {
2159                                 list_del(&xattr->list);
2160                         }
2161                         goto out;
2162                 }
2163         }
2164         if (flags & XATTR_REPLACE) {
2165                 xattr = new_xattr;
2166                 err = -ENODATA;
2167         } else {
2168                 list_add(&new_xattr->list, &info->xattr_list);
2169                 xattr = NULL;
2170         }
2171 out:
2172         spin_unlock(&info->lock);
2173         if (xattr)
2174                 kfree(xattr->name);
2175         kfree(xattr);
2176         return err;
2177 }
2178
2179
2180 static const struct xattr_handler *shmem_xattr_handlers[] = {
2181 #ifdef CONFIG_TMPFS_POSIX_ACL
2182         &generic_acl_access_handler,
2183         &generic_acl_default_handler,
2184 #endif
2185         NULL
2186 };
2187
2188 static int shmem_xattr_validate(const char *name)
2189 {
2190         struct { const char *prefix; size_t len; } arr[] = {
2191                 { XATTR_SECURITY_PREFIX, XATTR_SECURITY_PREFIX_LEN },
2192                 { XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN }
2193         };
2194         int i;
2195
2196         for (i = 0; i < ARRAY_SIZE(arr); i++) {
2197                 size_t preflen = arr[i].len;
2198                 if (strncmp(name, arr[i].prefix, preflen) == 0) {
2199                         if (!name[preflen])
2200                                 return -EINVAL;
2201                         return 0;
2202                 }
2203         }
2204         return -EOPNOTSUPP;
2205 }
2206
2207 static ssize_t shmem_getxattr(struct dentry *dentry, const char *name,
2208                               void *buffer, size_t size)
2209 {
2210         int err;
2211
2212         /*
2213          * If this is a request for a synthetic attribute in the system.*
2214          * namespace use the generic infrastructure to resolve a handler
2215          * for it via sb->s_xattr.
2216          */
2217         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2218                 return generic_getxattr(dentry, name, buffer, size);
2219
2220         err = shmem_xattr_validate(name);
2221         if (err)
2222                 return err;
2223
2224         return shmem_xattr_get(dentry, name, buffer, size);
2225 }
2226
2227 static int shmem_setxattr(struct dentry *dentry, const char *name,
2228                           const void *value, size_t size, int flags)
2229 {
2230         int err;
2231
2232         /*
2233          * If this is a request for a synthetic attribute in the system.*
2234          * namespace use the generic infrastructure to resolve a handler
2235          * for it via sb->s_xattr.
2236          */
2237         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2238                 return generic_setxattr(dentry, name, value, size, flags);
2239
2240         err = shmem_xattr_validate(name);
2241         if (err)
2242                 return err;
2243
2244         if (size == 0)
2245                 value = "";  /* empty EA, do not remove */
2246
2247         return shmem_xattr_set(dentry, name, value, size, flags);
2248
2249 }
2250
2251 static int shmem_removexattr(struct dentry *dentry, const char *name)
2252 {
2253         int err;
2254
2255         /*
2256          * If this is a request for a synthetic attribute in the system.*
2257          * namespace use the generic infrastructure to resolve a handler
2258          * for it via sb->s_xattr.
2259          */
2260         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2261                 return generic_removexattr(dentry, name);
2262
2263         err = shmem_xattr_validate(name);
2264         if (err)
2265                 return err;
2266
2267         return shmem_xattr_set(dentry, name, NULL, 0, XATTR_REPLACE);
2268 }
2269
2270 static bool xattr_is_trusted(const char *name)
2271 {
2272         return !strncmp(name, XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN);
2273 }
2274
2275 static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
2276 {
2277         bool trusted = capable(CAP_SYS_ADMIN);
2278         struct shmem_xattr *xattr;
2279         struct shmem_inode_info *info;
2280         size_t used = 0;
2281
2282         info = SHMEM_I(dentry->d_inode);
2283
2284         spin_lock(&info->lock);
2285         list_for_each_entry(xattr, &info->xattr_list, list) {
2286                 size_t len;
2287
2288                 /* skip "trusted." attributes for unprivileged callers */
2289                 if (!trusted && xattr_is_trusted(xattr->name))
2290                         continue;
2291
2292                 len = strlen(xattr->name) + 1;
2293                 used += len;
2294                 if (buffer) {
2295                         if (size < used) {
2296                                 used = -ERANGE;
2297                                 break;
2298                         }
2299                         memcpy(buffer, xattr->name, len);
2300                         buffer += len;
2301                 }
2302         }
2303         spin_unlock(&info->lock);
2304
2305         return used;
2306 }
2307 #endif /* CONFIG_TMPFS_XATTR */
2308
2309 static const struct inode_operations shmem_symlink_inline_operations = {
2310         .readlink       = generic_readlink,
2311         .follow_link    = shmem_follow_link_inline,
2312 #ifdef CONFIG_TMPFS_XATTR
2313         .setxattr       = shmem_setxattr,
2314         .getxattr       = shmem_getxattr,
2315         .listxattr      = shmem_listxattr,
2316         .removexattr    = shmem_removexattr,
2317 #endif
2318 };
2319
2320 static const struct inode_operations shmem_symlink_inode_operations = {
2321         .readlink       = generic_readlink,
2322         .follow_link    = shmem_follow_link,
2323         .put_link       = shmem_put_link,
2324 #ifdef CONFIG_TMPFS_XATTR
2325         .setxattr       = shmem_setxattr,
2326         .getxattr       = shmem_getxattr,
2327         .listxattr      = shmem_listxattr,
2328         .removexattr    = shmem_removexattr,
2329 #endif
2330 };
2331
2332 static struct dentry *shmem_get_parent(struct dentry *child)
2333 {
2334         return ERR_PTR(-ESTALE);
2335 }
2336
2337 static int shmem_match(struct inode *ino, void *vfh)
2338 {
2339         __u32 *fh = vfh;
2340         __u64 inum = fh[2];
2341         inum = (inum << 32) | fh[1];
2342         return ino->i_ino == inum && fh[0] == ino->i_generation;
2343 }
2344
2345 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
2346                 struct fid *fid, int fh_len, int fh_type)
2347 {
2348         struct inode *inode;
2349         struct dentry *dentry = NULL;
2350         u64 inum = fid->raw[2];
2351         inum = (inum << 32) | fid->raw[1];
2352
2353         if (fh_len < 3)
2354                 return NULL;
2355
2356         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
2357                         shmem_match, fid->raw);
2358         if (inode) {
2359                 dentry = d_find_alias(inode);
2360                 iput(inode);
2361         }
2362
2363         return dentry;
2364 }
2365
2366 static int shmem_encode_fh(struct dentry *dentry, __u32 *fh, int *len,
2367                                 int connectable)
2368 {
2369         struct inode *inode = dentry->d_inode;
2370
2371         if (*len < 3) {
2372                 *len = 3;
2373                 return 255;
2374         }
2375
2376         if (inode_unhashed(inode)) {
2377                 /* Unfortunately insert_inode_hash is not idempotent,
2378                  * so as we hash inodes here rather than at creation
2379                  * time, we need a lock to ensure we only try
2380                  * to do it once
2381                  */
2382                 static DEFINE_SPINLOCK(lock);
2383                 spin_lock(&lock);
2384                 if (inode_unhashed(inode))
2385                         __insert_inode_hash(inode,
2386                                             inode->i_ino + inode->i_generation);
2387                 spin_unlock(&lock);
2388         }
2389
2390         fh[0] = inode->i_generation;
2391         fh[1] = inode->i_ino;
2392         fh[2] = ((__u64)inode->i_ino) >> 32;
2393
2394         *len = 3;
2395         return 1;
2396 }
2397
2398 static const struct export_operations shmem_export_ops = {
2399         .get_parent     = shmem_get_parent,
2400         .encode_fh      = shmem_encode_fh,
2401         .fh_to_dentry   = shmem_fh_to_dentry,
2402 };
2403
2404 static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
2405                                bool remount)
2406 {
2407         char *this_char, *value, *rest;
2408
2409         while (options != NULL) {
2410                 this_char = options;
2411                 for (;;) {
2412                         /*
2413                          * NUL-terminate this option: unfortunately,
2414                          * mount options form a comma-separated list,
2415                          * but mpol's nodelist may also contain commas.
2416                          */
2417                         options = strchr(options, ',');
2418                         if (options == NULL)
2419                                 break;
2420                         options++;
2421                         if (!isdigit(*options)) {
2422                                 options[-1] = '\0';
2423                                 break;
2424                         }
2425                 }
2426                 if (!*this_char)
2427                         continue;
2428                 if ((value = strchr(this_char,'=')) != NULL) {
2429                         *value++ = 0;
2430                 } else {
2431                         printk(KERN_ERR
2432                             "tmpfs: No value for mount option '%s'\n",
2433                             this_char);
2434                         return 1;
2435                 }
2436
2437                 if (!strcmp(this_char,"size")) {
2438                         unsigned long long size;
2439                         size = memparse(value,&rest);
2440                         if (*rest == '%') {
2441                                 size <<= PAGE_SHIFT;
2442                                 size *= totalram_pages;
2443                                 do_div(size, 100);
2444                                 rest++;
2445                         }
2446                         if (*rest)
2447                                 goto bad_val;
2448                         sbinfo->max_blocks =
2449                                 DIV_ROUND_UP(size, PAGE_CACHE_SIZE);
2450                 } else if (!strcmp(this_char,"nr_blocks")) {
2451                         sbinfo->max_blocks = memparse(value, &rest);
2452                         if (*rest)
2453                                 goto bad_val;
2454                 } else if (!strcmp(this_char,"nr_inodes")) {
2455                         sbinfo->max_inodes = memparse(value, &rest);
2456                         if (*rest)
2457                                 goto bad_val;
2458                 } else if (!strcmp(this_char,"mode")) {
2459                         if (remount)
2460                                 continue;
2461                         sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
2462                         if (*rest)
2463                                 goto bad_val;
2464                 } else if (!strcmp(this_char,"uid")) {
2465                         if (remount)
2466                                 continue;
2467                         sbinfo->uid = simple_strtoul(value, &rest, 0);
2468                         if (*rest)
2469                                 goto bad_val;
2470                 } else if (!strcmp(this_char,"gid")) {
2471                         if (remount)
2472                                 continue;
2473                         sbinfo->gid = simple_strtoul(value, &rest, 0);
2474                         if (*rest)
2475                                 goto bad_val;
2476                 } else if (!strcmp(this_char,"mpol")) {
2477                         if (mpol_parse_str(value, &sbinfo->mpol, 1))
2478                                 goto bad_val;
2479                 } else {
2480                         printk(KERN_ERR "tmpfs: Bad mount option %s\n",
2481                                this_char);
2482                         return 1;
2483                 }
2484         }
2485         return 0;
2486
2487 bad_val:
2488         printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
2489                value, this_char);
2490         return 1;
2491
2492 }
2493
2494 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2495 {
2496         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2497         struct shmem_sb_info config = *sbinfo;
2498         unsigned long inodes;
2499         int error = -EINVAL;
2500
2501         if (shmem_parse_options(data, &config, true))
2502                 return error;
2503
2504         spin_lock(&sbinfo->stat_lock);
2505         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2506         if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
2507                 goto out;
2508         if (config.max_inodes < inodes)
2509                 goto out;
2510         /*
2511          * Those tests also disallow limited->unlimited while any are in
2512          * use, so i_blocks will always be zero when max_blocks is zero;
2513          * but we must separately disallow unlimited->limited, because
2514          * in that case we have no record of how much is already in use.
2515          */
2516         if (config.max_blocks && !sbinfo->max_blocks)
2517                 goto out;
2518         if (config.max_inodes && !sbinfo->max_inodes)
2519                 goto out;
2520
2521         error = 0;
2522         sbinfo->max_blocks  = config.max_blocks;
2523         sbinfo->max_inodes  = config.max_inodes;
2524         sbinfo->free_inodes = config.max_inodes - inodes;
2525
2526         mpol_put(sbinfo->mpol);
2527         sbinfo->mpol        = config.mpol;      /* transfers initial ref */
2528 out:
2529         spin_unlock(&sbinfo->stat_lock);
2530         return error;
2531 }
2532
2533 static int shmem_show_options(struct seq_file *seq, struct vfsmount *vfs)
2534 {
2535         struct shmem_sb_info *sbinfo = SHMEM_SB(vfs->mnt_sb);
2536
2537         if (sbinfo->max_blocks != shmem_default_max_blocks())
2538                 seq_printf(seq, ",size=%luk",
2539                         sbinfo->max_blocks << (PAGE_CACHE_SHIFT - 10));
2540         if (sbinfo->max_inodes != shmem_default_max_inodes())
2541                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
2542         if (sbinfo->mode != (S_IRWXUGO | S_ISVTX))
2543                 seq_printf(seq, ",mode=%03o", sbinfo->mode);
2544         if (sbinfo->uid != 0)
2545                 seq_printf(seq, ",uid=%u", sbinfo->uid);
2546         if (sbinfo->gid != 0)
2547                 seq_printf(seq, ",gid=%u", sbinfo->gid);
2548         shmem_show_mpol(seq, sbinfo->mpol);
2549         return 0;
2550 }
2551 #endif /* CONFIG_TMPFS */
2552
2553 static void shmem_put_super(struct super_block *sb)
2554 {
2555         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2556
2557         percpu_counter_destroy(&sbinfo->used_blocks);
2558         kfree(sbinfo);
2559         sb->s_fs_info = NULL;
2560 }
2561
2562 int shmem_fill_super(struct super_block *sb, void *data, int silent)
2563 {
2564         struct inode *inode;
2565         struct dentry *root;
2566         struct shmem_sb_info *sbinfo;
2567         int err = -ENOMEM;
2568
2569         /* Round up to L1_CACHE_BYTES to resist false sharing */
2570         sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
2571                                 L1_CACHE_BYTES), GFP_KERNEL);
2572         if (!sbinfo)
2573                 return -ENOMEM;
2574
2575         sbinfo->mode = S_IRWXUGO | S_ISVTX;
2576         sbinfo->uid = current_fsuid();
2577         sbinfo->gid = current_fsgid();
2578         sb->s_fs_info = sbinfo;
2579
2580 #ifdef CONFIG_TMPFS
2581         /*
2582          * Per default we only allow half of the physical ram per
2583          * tmpfs instance, limiting inodes to one per page of lowmem;
2584          * but the internal instance is left unlimited.
2585          */
2586         if (!(sb->s_flags & MS_NOUSER)) {
2587                 sbinfo->max_blocks = shmem_default_max_blocks();
2588                 sbinfo->max_inodes = shmem_default_max_inodes();
2589                 if (shmem_parse_options(data, sbinfo, false)) {
2590                         err = -EINVAL;
2591                         goto failed;
2592                 }
2593         }
2594         sb->s_export_op = &shmem_export_ops;
2595 #else
2596         sb->s_flags |= MS_NOUSER;
2597 #endif
2598
2599         spin_lock_init(&sbinfo->stat_lock);
2600         if (percpu_counter_init(&sbinfo->used_blocks, 0))
2601                 goto failed;
2602         sbinfo->free_inodes = sbinfo->max_inodes;
2603
2604         sb->s_maxbytes = SHMEM_MAX_BYTES;
2605         sb->s_blocksize = PAGE_CACHE_SIZE;
2606         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
2607         sb->s_magic = TMPFS_MAGIC;
2608         sb->s_op = &shmem_ops;
2609         sb->s_time_gran = 1;
2610 #ifdef CONFIG_TMPFS_XATTR
2611         sb->s_xattr = shmem_xattr_handlers;
2612 #endif
2613 #ifdef CONFIG_TMPFS_POSIX_ACL
2614         sb->s_flags |= MS_POSIXACL;
2615 #endif
2616
2617         inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
2618         if (!inode)
2619                 goto failed;
2620         inode->i_uid = sbinfo->uid;
2621         inode->i_gid = sbinfo->gid;
2622         root = d_alloc_root(inode);
2623         if (!root)
2624                 goto failed_iput;
2625         sb->s_root = root;
2626         return 0;
2627
2628 failed_iput:
2629         iput(inode);
2630 failed:
2631         shmem_put_super(sb);
2632         return err;
2633 }
2634
2635 static struct kmem_cache *shmem_inode_cachep;
2636
2637 static struct inode *shmem_alloc_inode(struct super_block *sb)
2638 {
2639         struct shmem_inode_info *p;
2640         p = (struct shmem_inode_info *)kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
2641         if (!p)
2642                 return NULL;
2643         return &p->vfs_inode;
2644 }
2645
2646 static void shmem_i_callback(struct rcu_head *head)
2647 {
2648         struct inode *inode = container_of(head, struct inode, i_rcu);
2649         INIT_LIST_HEAD(&inode->i_dentry);
2650         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
2651 }
2652
2653 static void shmem_destroy_inode(struct inode *inode)
2654 {
2655         if ((inode->i_mode & S_IFMT) == S_IFREG) {
2656                 /* only struct inode is valid if it's an inline symlink */
2657                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
2658         }
2659         call_rcu(&inode->i_rcu, shmem_i_callback);
2660 }
2661
2662 static void init_once(void *foo)
2663 {
2664         struct shmem_inode_info *p = (struct shmem_inode_info *) foo;
2665
2666         inode_init_once(&p->vfs_inode);
2667 }
2668
2669 static int init_inodecache(void)
2670 {
2671         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
2672                                 sizeof(struct shmem_inode_info),
2673                                 0, SLAB_PANIC, init_once);
2674         return 0;
2675 }
2676
2677 static void destroy_inodecache(void)
2678 {
2679         kmem_cache_destroy(shmem_inode_cachep);
2680 }
2681
2682 static const struct address_space_operations shmem_aops = {
2683         .writepage      = shmem_writepage,
2684         .set_page_dirty = __set_page_dirty_no_writeback,
2685 #ifdef CONFIG_TMPFS
2686         .readpage       = shmem_readpage,
2687         .write_begin    = shmem_write_begin,
2688         .write_end      = shmem_write_end,
2689 #endif
2690         .migratepage    = migrate_page,
2691         .error_remove_page = generic_error_remove_page,
2692 };
2693
2694 static const struct file_operations shmem_file_operations = {
2695         .mmap           = shmem_mmap,
2696 #ifdef CONFIG_TMPFS
2697         .llseek         = generic_file_llseek,
2698         .read           = do_sync_read,
2699         .write          = do_sync_write,
2700         .aio_read       = shmem_file_aio_read,
2701         .aio_write      = generic_file_aio_write,
2702         .fsync          = noop_fsync,
2703         .splice_read    = generic_file_splice_read,
2704         .splice_write   = generic_file_splice_write,
2705 #endif
2706 };
2707
2708 static const struct inode_operations shmem_inode_operations = {
2709         .setattr        = shmem_notify_change,
2710         .truncate_range = shmem_truncate_range,
2711 #ifdef CONFIG_TMPFS_XATTR
2712         .setxattr       = shmem_setxattr,
2713         .getxattr       = shmem_getxattr,
2714         .listxattr      = shmem_listxattr,
2715         .removexattr    = shmem_removexattr,
2716 #endif
2717 #ifdef CONFIG_TMPFS_POSIX_ACL
2718         .check_acl      = generic_check_acl,
2719 #endif
2720
2721 };
2722
2723 static const struct inode_operations shmem_dir_inode_operations = {
2724 #ifdef CONFIG_TMPFS
2725         .create         = shmem_create,
2726         .lookup         = simple_lookup,
2727         .link           = shmem_link,
2728         .unlink         = shmem_unlink,
2729         .symlink        = shmem_symlink,
2730         .mkdir          = shmem_mkdir,
2731         .rmdir          = shmem_rmdir,
2732         .mknod          = shmem_mknod,
2733         .rename         = shmem_rename,
2734 #endif
2735 #ifdef CONFIG_TMPFS_XATTR
2736         .setxattr       = shmem_setxattr,
2737         .getxattr       = shmem_getxattr,
2738         .listxattr      = shmem_listxattr,
2739         .removexattr    = shmem_removexattr,
2740 #endif
2741 #ifdef CONFIG_TMPFS_POSIX_ACL
2742         .setattr        = shmem_notify_change,
2743         .check_acl      = generic_check_acl,
2744 #endif
2745 };
2746
2747 static const struct inode_operations shmem_special_inode_operations = {
2748 #ifdef CONFIG_TMPFS_XATTR
2749         .setxattr       = shmem_setxattr,
2750         .getxattr       = shmem_getxattr,
2751         .listxattr      = shmem_listxattr,
2752         .removexattr    = shmem_removexattr,
2753 #endif
2754 #ifdef CONFIG_TMPFS_POSIX_ACL
2755         .setattr        = shmem_notify_change,
2756         .check_acl      = generic_check_acl,
2757 #endif
2758 };
2759
2760 static const struct super_operations shmem_ops = {
2761         .alloc_inode    = shmem_alloc_inode,
2762         .destroy_inode  = shmem_destroy_inode,
2763 #ifdef CONFIG_TMPFS
2764         .statfs         = shmem_statfs,
2765         .remount_fs     = shmem_remount_fs,
2766         .show_options   = shmem_show_options,
2767 #endif
2768         .evict_inode    = shmem_evict_inode,
2769         .drop_inode     = generic_delete_inode,
2770         .put_super      = shmem_put_super,
2771 };
2772
2773 static const struct vm_operations_struct shmem_vm_ops = {
2774         .fault          = shmem_fault,
2775 #ifdef CONFIG_NUMA
2776         .set_policy     = shmem_set_policy,
2777         .get_policy     = shmem_get_policy,
2778 #endif
2779 };
2780
2781
2782 static struct dentry *shmem_mount(struct file_system_type *fs_type,
2783         int flags, const char *dev_name, void *data)
2784 {
2785         return mount_nodev(fs_type, flags, data, shmem_fill_super);
2786 }
2787
2788 static struct file_system_type tmpfs_fs_type = {
2789         .owner          = THIS_MODULE,
2790         .name           = "tmpfs",
2791         .mount          = shmem_mount,
2792         .kill_sb        = kill_litter_super,
2793 };
2794
2795 int __init init_tmpfs(void)
2796 {
2797         int error;
2798
2799         error = bdi_init(&shmem_backing_dev_info);
2800         if (error)
2801                 goto out4;
2802
2803         error = init_inodecache();
2804         if (error)
2805                 goto out3;
2806
2807         error = register_filesystem(&tmpfs_fs_type);
2808         if (error) {
2809                 printk(KERN_ERR "Could not register tmpfs\n");
2810                 goto out2;
2811         }
2812
2813         shm_mnt = vfs_kern_mount(&tmpfs_fs_type, MS_NOUSER,
2814                                 tmpfs_fs_type.name, NULL);
2815         if (IS_ERR(shm_mnt)) {
2816                 error = PTR_ERR(shm_mnt);
2817                 printk(KERN_ERR "Could not kern_mount tmpfs\n");
2818                 goto out1;
2819         }
2820         return 0;
2821
2822 out1:
2823         unregister_filesystem(&tmpfs_fs_type);
2824 out2:
2825         destroy_inodecache();
2826 out3:
2827         bdi_destroy(&shmem_backing_dev_info);
2828 out4:
2829         shm_mnt = ERR_PTR(error);
2830         return error;
2831 }
2832
2833 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2834 /**
2835  * mem_cgroup_get_shmem_target - find a page or entry assigned to the shmem file
2836  * @inode: the inode to be searched
2837  * @pgoff: the offset to be searched
2838  * @pagep: the pointer for the found page to be stored
2839  * @ent: the pointer for the found swap entry to be stored
2840  *
2841  * If a page is found, refcount of it is incremented. Callers should handle
2842  * these refcount.
2843  */
2844 void mem_cgroup_get_shmem_target(struct inode *inode, pgoff_t pgoff,
2845                                         struct page **pagep, swp_entry_t *ent)
2846 {
2847         swp_entry_t entry = { .val = 0 }, *ptr;
2848         struct page *page = NULL;
2849         struct shmem_inode_info *info = SHMEM_I(inode);
2850
2851         if ((pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
2852                 goto out;
2853
2854         spin_lock(&info->lock);
2855         ptr = shmem_swp_entry(info, pgoff, NULL);
2856 #ifdef CONFIG_SWAP
2857         if (ptr && ptr->val) {
2858                 entry.val = ptr->val;
2859                 page = find_get_page(&swapper_space, entry.val);
2860         } else
2861 #endif
2862                 page = find_get_page(inode->i_mapping, pgoff);
2863         if (ptr)
2864                 shmem_swp_unmap(ptr);
2865         spin_unlock(&info->lock);
2866 out:
2867         *pagep = page;
2868         *ent = entry;
2869 }
2870 #endif
2871
2872 #else /* !CONFIG_SHMEM */
2873
2874 /*
2875  * tiny-shmem: simple shmemfs and tmpfs using ramfs code
2876  *
2877  * This is intended for small system where the benefits of the full
2878  * shmem code (swap-backed and resource-limited) are outweighed by
2879  * their complexity. On systems without swap this code should be
2880  * effectively equivalent, but much lighter weight.
2881  */
2882
2883 #include <linux/ramfs.h>
2884
2885 static struct file_system_type tmpfs_fs_type = {
2886         .name           = "tmpfs",
2887         .mount          = ramfs_mount,
2888         .kill_sb        = kill_litter_super,
2889 };
2890
2891 int __init init_tmpfs(void)
2892 {
2893         BUG_ON(register_filesystem(&tmpfs_fs_type) != 0);
2894
2895         shm_mnt = kern_mount(&tmpfs_fs_type);
2896         BUG_ON(IS_ERR(shm_mnt));
2897
2898         return 0;
2899 }
2900
2901 int shmem_unuse(swp_entry_t entry, struct page *page)
2902 {
2903         return 0;
2904 }
2905
2906 int shmem_lock(struct file *file, int lock, struct user_struct *user)
2907 {
2908         return 0;
2909 }
2910
2911 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2912 /**
2913  * mem_cgroup_get_shmem_target - find a page or entry assigned to the shmem file
2914  * @inode: the inode to be searched
2915  * @pgoff: the offset to be searched
2916  * @pagep: the pointer for the found page to be stored
2917  * @ent: the pointer for the found swap entry to be stored
2918  *
2919  * If a page is found, refcount of it is incremented. Callers should handle
2920  * these refcount.
2921  */
2922 void mem_cgroup_get_shmem_target(struct inode *inode, pgoff_t pgoff,
2923                                         struct page **pagep, swp_entry_t *ent)
2924 {
2925         struct page *page = NULL;
2926
2927         if ((pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
2928                 goto out;
2929         page = find_get_page(inode->i_mapping, pgoff);
2930 out:
2931         *pagep = page;
2932         *ent = (swp_entry_t){ .val = 0 };
2933 }
2934 #endif
2935
2936 #define shmem_vm_ops                            generic_file_vm_ops
2937 #define shmem_file_operations                   ramfs_file_operations
2938 #define shmem_get_inode(sb, dir, mode, dev, flags)      ramfs_get_inode(sb, dir, mode, dev)
2939 #define shmem_acct_size(flags, size)            0
2940 #define shmem_unacct_size(flags, size)          do {} while (0)
2941 #define SHMEM_MAX_BYTES                         MAX_LFS_FILESIZE
2942
2943 #endif /* CONFIG_SHMEM */
2944
2945 /* common code */
2946
2947 /**
2948  * shmem_file_setup - get an unlinked file living in tmpfs
2949  * @name: name for dentry (to be seen in /proc/<pid>/maps
2950  * @size: size to be set for the file
2951  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
2952  */
2953 struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
2954 {
2955         int error;
2956         struct file *file;
2957         struct inode *inode;
2958         struct path path;
2959         struct dentry *root;
2960         struct qstr this;
2961
2962         if (IS_ERR(shm_mnt))
2963                 return (void *)shm_mnt;
2964
2965         if (size < 0 || size > SHMEM_MAX_BYTES)
2966                 return ERR_PTR(-EINVAL);
2967
2968         if (shmem_acct_size(flags, size))
2969                 return ERR_PTR(-ENOMEM);
2970
2971         error = -ENOMEM;
2972         this.name = name;
2973         this.len = strlen(name);
2974         this.hash = 0; /* will go */
2975         root = shm_mnt->mnt_root;
2976         path.dentry = d_alloc(root, &this);
2977         if (!path.dentry)
2978                 goto put_memory;
2979         path.mnt = mntget(shm_mnt);
2980
2981         error = -ENOSPC;
2982         inode = shmem_get_inode(root->d_sb, NULL, S_IFREG | S_IRWXUGO, 0, flags);
2983         if (!inode)
2984                 goto put_dentry;
2985
2986         d_instantiate(path.dentry, inode);
2987         inode->i_size = size;
2988         inode->i_nlink = 0;     /* It is unlinked */
2989 #ifndef CONFIG_MMU
2990         error = ramfs_nommu_expand_for_mapping(inode, size);
2991         if (error)
2992                 goto put_dentry;
2993 #endif
2994
2995         error = -ENFILE;
2996         file = alloc_file(&path, FMODE_WRITE | FMODE_READ,
2997                   &shmem_file_operations);
2998         if (!file)
2999                 goto put_dentry;
3000
3001         return file;
3002
3003 put_dentry:
3004         path_put(&path);
3005 put_memory:
3006         shmem_unacct_size(flags, size);
3007         return ERR_PTR(error);
3008 }
3009 EXPORT_SYMBOL_GPL(shmem_file_setup);
3010
3011 /**
3012  * shmem_zero_setup - setup a shared anonymous mapping
3013  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
3014  */
3015 int shmem_zero_setup(struct vm_area_struct *vma)
3016 {
3017         struct file *file;
3018         loff_t size = vma->vm_end - vma->vm_start;
3019
3020         file = shmem_file_setup("dev/zero", size, vma->vm_flags);
3021         if (IS_ERR(file))
3022                 return PTR_ERR(file);
3023
3024         if (vma->vm_file)
3025                 fput(vma->vm_file);
3026         vma->vm_file = file;
3027         vma->vm_ops = &shmem_vm_ops;
3028         vma->vm_flags |= VM_CAN_NONLINEAR;
3029         return 0;
3030 }