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