tmpfs: make tmpfs scalable with percpu_counter for used blocks
[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         .unplug_io_fn   = default_unplug_io_fn,
228 };
229
230 static LIST_HEAD(shmem_swaplist);
231 static DEFINE_MUTEX(shmem_swaplist_mutex);
232
233 static void shmem_free_blocks(struct inode *inode, long pages)
234 {
235         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
236         if (sbinfo->max_blocks) {
237                 percpu_counter_add(&sbinfo->used_blocks, -pages);
238                 spin_lock(&inode->i_lock);
239                 inode->i_blocks -= pages*BLOCKS_PER_PAGE;
240                 spin_unlock(&inode->i_lock);
241         }
242 }
243
244 static int shmem_reserve_inode(struct super_block *sb)
245 {
246         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
247         if (sbinfo->max_inodes) {
248                 spin_lock(&sbinfo->stat_lock);
249                 if (!sbinfo->free_inodes) {
250                         spin_unlock(&sbinfo->stat_lock);
251                         return -ENOSPC;
252                 }
253                 sbinfo->free_inodes--;
254                 spin_unlock(&sbinfo->stat_lock);
255         }
256         return 0;
257 }
258
259 static void shmem_free_inode(struct super_block *sb)
260 {
261         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
262         if (sbinfo->max_inodes) {
263                 spin_lock(&sbinfo->stat_lock);
264                 sbinfo->free_inodes++;
265                 spin_unlock(&sbinfo->stat_lock);
266         }
267 }
268
269 /**
270  * shmem_recalc_inode - recalculate the size of an inode
271  * @inode: inode to recalc
272  *
273  * We have to calculate the free blocks since the mm can drop
274  * undirtied hole pages behind our back.
275  *
276  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
277  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
278  *
279  * It has to be called with the spinlock held.
280  */
281 static void shmem_recalc_inode(struct inode *inode)
282 {
283         struct shmem_inode_info *info = SHMEM_I(inode);
284         long freed;
285
286         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
287         if (freed > 0) {
288                 info->alloced -= freed;
289                 shmem_unacct_blocks(info->flags, freed);
290                 shmem_free_blocks(inode, freed);
291         }
292 }
293
294 /**
295  * shmem_swp_entry - find the swap vector position in the info structure
296  * @info:  info structure for the inode
297  * @index: index of the page to find
298  * @page:  optional page to add to the structure. Has to be preset to
299  *         all zeros
300  *
301  * If there is no space allocated yet it will return NULL when
302  * page is NULL, else it will use the page for the needed block,
303  * setting it to NULL on return to indicate that it has been used.
304  *
305  * The swap vector is organized the following way:
306  *
307  * There are SHMEM_NR_DIRECT entries directly stored in the
308  * shmem_inode_info structure. So small files do not need an addional
309  * allocation.
310  *
311  * For pages with index > SHMEM_NR_DIRECT there is the pointer
312  * i_indirect which points to a page which holds in the first half
313  * doubly indirect blocks, in the second half triple indirect blocks:
314  *
315  * For an artificial ENTRIES_PER_PAGE = 4 this would lead to the
316  * following layout (for SHMEM_NR_DIRECT == 16):
317  *
318  * i_indirect -> dir --> 16-19
319  *            |      +-> 20-23
320  *            |
321  *            +-->dir2 --> 24-27
322  *            |        +-> 28-31
323  *            |        +-> 32-35
324  *            |        +-> 36-39
325  *            |
326  *            +-->dir3 --> 40-43
327  *                     +-> 44-47
328  *                     +-> 48-51
329  *                     +-> 52-55
330  */
331 static swp_entry_t *shmem_swp_entry(struct shmem_inode_info *info, unsigned long index, struct page **page)
332 {
333         unsigned long offset;
334         struct page **dir;
335         struct page *subdir;
336
337         if (index < SHMEM_NR_DIRECT) {
338                 shmem_swp_balance_unmap();
339                 return info->i_direct+index;
340         }
341         if (!info->i_indirect) {
342                 if (page) {
343                         info->i_indirect = *page;
344                         *page = NULL;
345                 }
346                 return NULL;                    /* need another page */
347         }
348
349         index -= SHMEM_NR_DIRECT;
350         offset = index % ENTRIES_PER_PAGE;
351         index /= ENTRIES_PER_PAGE;
352         dir = shmem_dir_map(info->i_indirect);
353
354         if (index >= ENTRIES_PER_PAGE/2) {
355                 index -= ENTRIES_PER_PAGE/2;
356                 dir += ENTRIES_PER_PAGE/2 + index/ENTRIES_PER_PAGE;
357                 index %= ENTRIES_PER_PAGE;
358                 subdir = *dir;
359                 if (!subdir) {
360                         if (page) {
361                                 *dir = *page;
362                                 *page = NULL;
363                         }
364                         shmem_dir_unmap(dir);
365                         return NULL;            /* need another page */
366                 }
367                 shmem_dir_unmap(dir);
368                 dir = shmem_dir_map(subdir);
369         }
370
371         dir += index;
372         subdir = *dir;
373         if (!subdir) {
374                 if (!page || !(subdir = *page)) {
375                         shmem_dir_unmap(dir);
376                         return NULL;            /* need a page */
377                 }
378                 *dir = subdir;
379                 *page = NULL;
380         }
381         shmem_dir_unmap(dir);
382         return shmem_swp_map(subdir) + offset;
383 }
384
385 static void shmem_swp_set(struct shmem_inode_info *info, swp_entry_t *entry, unsigned long value)
386 {
387         long incdec = value? 1: -1;
388
389         entry->val = value;
390         info->swapped += incdec;
391         if ((unsigned long)(entry - info->i_direct) >= SHMEM_NR_DIRECT) {
392                 struct page *page = kmap_atomic_to_page(entry);
393                 set_page_private(page, page_private(page) + incdec);
394         }
395 }
396
397 /**
398  * shmem_swp_alloc - get the position of the swap entry for the page.
399  * @info:       info structure for the inode
400  * @index:      index of the page to find
401  * @sgp:        check and recheck i_size? skip allocation?
402  *
403  * If the entry does not exist, allocate it.
404  */
405 static swp_entry_t *shmem_swp_alloc(struct shmem_inode_info *info, unsigned long index, enum sgp_type sgp)
406 {
407         struct inode *inode = &info->vfs_inode;
408         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
409         struct page *page = NULL;
410         swp_entry_t *entry;
411
412         if (sgp != SGP_WRITE &&
413             ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode))
414                 return ERR_PTR(-EINVAL);
415
416         while (!(entry = shmem_swp_entry(info, index, &page))) {
417                 if (sgp == SGP_READ)
418                         return shmem_swp_map(ZERO_PAGE(0));
419                 /*
420                  * Test used_blocks against 1 less max_blocks, since we have 1 data
421                  * page (and perhaps indirect index pages) yet to allocate:
422                  * a waste to allocate index if we cannot allocate data.
423                  */
424                 if (sbinfo->max_blocks) {
425                         if (percpu_counter_compare(&sbinfo->used_blocks, (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         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)
770                                         && newsize != inode->i_size) {
771                 struct page *page = NULL;
772
773                 if (newsize < inode->i_size) {
774                         /*
775                          * If truncating down to a partial page, then
776                          * if that page is already allocated, hold it
777                          * in memory until the truncation is over, so
778                          * truncate_partial_page cannnot miss it were
779                          * it assigned to swap.
780                          */
781                         if (newsize & (PAGE_CACHE_SIZE-1)) {
782                                 (void) shmem_getpage(inode,
783                                         newsize >> PAGE_CACHE_SHIFT,
784                                                 &page, SGP_READ, NULL);
785                                 if (page)
786                                         unlock_page(page);
787                         }
788                         /*
789                          * Reset SHMEM_PAGEIN flag so that shmem_truncate can
790                          * detect if any pages might have been added to cache
791                          * after truncate_inode_pages.  But we needn't bother
792                          * if it's being fully truncated to zero-length: the
793                          * nrpages check is efficient enough in that case.
794                          */
795                         if (newsize) {
796                                 struct shmem_inode_info *info = SHMEM_I(inode);
797                                 spin_lock(&info->lock);
798                                 info->flags &= ~SHMEM_PAGEIN;
799                                 spin_unlock(&info->lock);
800                         }
801                 }
802
803                 error = simple_setsize(inode, newsize);
804                 if (page)
805                         page_cache_release(page);
806                 if (error)
807                         return error;
808                 shmem_truncate_range(inode, newsize, (loff_t)-1);
809         }
810
811         error = inode_change_ok(inode, attr);
812         if (!error)
813                 generic_setattr(inode, attr);
814 #ifdef CONFIG_TMPFS_POSIX_ACL
815         if (!error && (attr->ia_valid & ATTR_MODE))
816                 error = generic_acl_chmod(inode);
817 #endif
818         return error;
819 }
820
821 static void shmem_delete_inode(struct inode *inode)
822 {
823         struct shmem_inode_info *info = SHMEM_I(inode);
824
825         if (inode->i_mapping->a_ops == &shmem_aops) {
826                 truncate_inode_pages(inode->i_mapping, 0);
827                 shmem_unacct_size(info->flags, inode->i_size);
828                 inode->i_size = 0;
829                 shmem_truncate_range(inode, 0, (loff_t)-1);
830                 if (!list_empty(&info->swaplist)) {
831                         mutex_lock(&shmem_swaplist_mutex);
832                         list_del_init(&info->swaplist);
833                         mutex_unlock(&shmem_swaplist_mutex);
834                 }
835         }
836         BUG_ON(inode->i_blocks);
837         shmem_free_inode(inode->i_sb);
838         clear_inode(inode);
839 }
840
841 static inline int shmem_find_swp(swp_entry_t entry, swp_entry_t *dir, swp_entry_t *edir)
842 {
843         swp_entry_t *ptr;
844
845         for (ptr = dir; ptr < edir; ptr++) {
846                 if (ptr->val == entry.val)
847                         return ptr - dir;
848         }
849         return -1;
850 }
851
852 static int shmem_unuse_inode(struct shmem_inode_info *info, swp_entry_t entry, struct page *page)
853 {
854         struct inode *inode;
855         unsigned long idx;
856         unsigned long size;
857         unsigned long limit;
858         unsigned long stage;
859         struct page **dir;
860         struct page *subdir;
861         swp_entry_t *ptr;
862         int offset;
863         int error;
864
865         idx = 0;
866         ptr = info->i_direct;
867         spin_lock(&info->lock);
868         if (!info->swapped) {
869                 list_del_init(&info->swaplist);
870                 goto lost2;
871         }
872         limit = info->next_index;
873         size = limit;
874         if (size > SHMEM_NR_DIRECT)
875                 size = SHMEM_NR_DIRECT;
876         offset = shmem_find_swp(entry, ptr, ptr+size);
877         if (offset >= 0)
878                 goto found;
879         if (!info->i_indirect)
880                 goto lost2;
881
882         dir = shmem_dir_map(info->i_indirect);
883         stage = SHMEM_NR_DIRECT + ENTRIES_PER_PAGEPAGE/2;
884
885         for (idx = SHMEM_NR_DIRECT; idx < limit; idx += ENTRIES_PER_PAGE, dir++) {
886                 if (unlikely(idx == stage)) {
887                         shmem_dir_unmap(dir-1);
888                         if (cond_resched_lock(&info->lock)) {
889                                 /* check it has not been truncated */
890                                 if (limit > info->next_index) {
891                                         limit = info->next_index;
892                                         if (idx >= limit)
893                                                 goto lost2;
894                                 }
895                         }
896                         dir = shmem_dir_map(info->i_indirect) +
897                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
898                         while (!*dir) {
899                                 dir++;
900                                 idx += ENTRIES_PER_PAGEPAGE;
901                                 if (idx >= limit)
902                                         goto lost1;
903                         }
904                         stage = idx + ENTRIES_PER_PAGEPAGE;
905                         subdir = *dir;
906                         shmem_dir_unmap(dir);
907                         dir = shmem_dir_map(subdir);
908                 }
909                 subdir = *dir;
910                 if (subdir && page_private(subdir)) {
911                         ptr = shmem_swp_map(subdir);
912                         size = limit - idx;
913                         if (size > ENTRIES_PER_PAGE)
914                                 size = ENTRIES_PER_PAGE;
915                         offset = shmem_find_swp(entry, ptr, ptr+size);
916                         shmem_swp_unmap(ptr);
917                         if (offset >= 0) {
918                                 shmem_dir_unmap(dir);
919                                 goto found;
920                         }
921                 }
922         }
923 lost1:
924         shmem_dir_unmap(dir-1);
925 lost2:
926         spin_unlock(&info->lock);
927         return 0;
928 found:
929         idx += offset;
930         inode = igrab(&info->vfs_inode);
931         spin_unlock(&info->lock);
932
933         /*
934          * Move _head_ to start search for next from here.
935          * But be careful: shmem_delete_inode checks list_empty without taking
936          * mutex, and there's an instant in list_move_tail when info->swaplist
937          * would appear empty, if it were the only one on shmem_swaplist.  We
938          * could avoid doing it if inode NULL; or use this minor optimization.
939          */
940         if (shmem_swaplist.next != &info->swaplist)
941                 list_move_tail(&shmem_swaplist, &info->swaplist);
942         mutex_unlock(&shmem_swaplist_mutex);
943
944         error = 1;
945         if (!inode)
946                 goto out;
947         /*
948          * Charge page using GFP_KERNEL while we can wait.
949          * Charged back to the user(not to caller) when swap account is used.
950          * add_to_page_cache() will be called with GFP_NOWAIT.
951          */
952         error = mem_cgroup_cache_charge(page, current->mm, GFP_KERNEL);
953         if (error)
954                 goto out;
955         error = radix_tree_preload(GFP_KERNEL);
956         if (error) {
957                 mem_cgroup_uncharge_cache_page(page);
958                 goto out;
959         }
960         error = 1;
961
962         spin_lock(&info->lock);
963         ptr = shmem_swp_entry(info, idx, NULL);
964         if (ptr && ptr->val == entry.val) {
965                 error = add_to_page_cache_locked(page, inode->i_mapping,
966                                                 idx, GFP_NOWAIT);
967                 /* does mem_cgroup_uncharge_cache_page on error */
968         } else  /* we must compensate for our precharge above */
969                 mem_cgroup_uncharge_cache_page(page);
970
971         if (error == -EEXIST) {
972                 struct page *filepage = find_get_page(inode->i_mapping, idx);
973                 error = 1;
974                 if (filepage) {
975                         /*
976                          * There might be a more uptodate page coming down
977                          * from a stacked writepage: forget our swappage if so.
978                          */
979                         if (PageUptodate(filepage))
980                                 error = 0;
981                         page_cache_release(filepage);
982                 }
983         }
984         if (!error) {
985                 delete_from_swap_cache(page);
986                 set_page_dirty(page);
987                 info->flags |= SHMEM_PAGEIN;
988                 shmem_swp_set(info, ptr, 0);
989                 swap_free(entry);
990                 error = 1;      /* not an error, but entry was found */
991         }
992         if (ptr)
993                 shmem_swp_unmap(ptr);
994         spin_unlock(&info->lock);
995         radix_tree_preload_end();
996 out:
997         unlock_page(page);
998         page_cache_release(page);
999         iput(inode);            /* allows for NULL */
1000         return error;
1001 }
1002
1003 /*
1004  * shmem_unuse() search for an eventually swapped out shmem page.
1005  */
1006 int shmem_unuse(swp_entry_t entry, struct page *page)
1007 {
1008         struct list_head *p, *next;
1009         struct shmem_inode_info *info;
1010         int found = 0;
1011
1012         mutex_lock(&shmem_swaplist_mutex);
1013         list_for_each_safe(p, next, &shmem_swaplist) {
1014                 info = list_entry(p, struct shmem_inode_info, swaplist);
1015                 found = shmem_unuse_inode(info, entry, page);
1016                 cond_resched();
1017                 if (found)
1018                         goto out;
1019         }
1020         mutex_unlock(&shmem_swaplist_mutex);
1021         /*
1022          * Can some race bring us here?  We've been holding page lock,
1023          * so I think not; but would rather try again later than BUG()
1024          */
1025         unlock_page(page);
1026         page_cache_release(page);
1027 out:
1028         return (found < 0) ? found : 0;
1029 }
1030
1031 /*
1032  * Move the page from the page cache to the swap cache.
1033  */
1034 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1035 {
1036         struct shmem_inode_info *info;
1037         swp_entry_t *entry, swap;
1038         struct address_space *mapping;
1039         unsigned long index;
1040         struct inode *inode;
1041
1042         BUG_ON(!PageLocked(page));
1043         mapping = page->mapping;
1044         index = page->index;
1045         inode = mapping->host;
1046         info = SHMEM_I(inode);
1047         if (info->flags & VM_LOCKED)
1048                 goto redirty;
1049         if (!total_swap_pages)
1050                 goto redirty;
1051
1052         /*
1053          * shmem_backing_dev_info's capabilities prevent regular writeback or
1054          * sync from ever calling shmem_writepage; but a stacking filesystem
1055          * may use the ->writepage of its underlying filesystem, in which case
1056          * tmpfs should write out to swap only in response to memory pressure,
1057          * and not for the writeback threads or sync.  However, in those cases,
1058          * we do still want to check if there's a redundant swappage to be
1059          * discarded.
1060          */
1061         if (wbc->for_reclaim)
1062                 swap = get_swap_page();
1063         else
1064                 swap.val = 0;
1065
1066         spin_lock(&info->lock);
1067         if (index >= info->next_index) {
1068                 BUG_ON(!(info->flags & SHMEM_TRUNCATE));
1069                 goto unlock;
1070         }
1071         entry = shmem_swp_entry(info, index, NULL);
1072         if (entry->val) {
1073                 /*
1074                  * The more uptodate page coming down from a stacked
1075                  * writepage should replace our old swappage.
1076                  */
1077                 free_swap_and_cache(*entry);
1078                 shmem_swp_set(info, entry, 0);
1079         }
1080         shmem_recalc_inode(inode);
1081
1082         if (swap.val && add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
1083                 remove_from_page_cache(page);
1084                 shmem_swp_set(info, entry, swap.val);
1085                 shmem_swp_unmap(entry);
1086                 if (list_empty(&info->swaplist))
1087                         inode = igrab(inode);
1088                 else
1089                         inode = NULL;
1090                 spin_unlock(&info->lock);
1091                 swap_shmem_alloc(swap);
1092                 BUG_ON(page_mapped(page));
1093                 page_cache_release(page);       /* pagecache ref */
1094                 swap_writepage(page, wbc);
1095                 if (inode) {
1096                         mutex_lock(&shmem_swaplist_mutex);
1097                         /* move instead of add in case we're racing */
1098                         list_move_tail(&info->swaplist, &shmem_swaplist);
1099                         mutex_unlock(&shmem_swaplist_mutex);
1100                         iput(inode);
1101                 }
1102                 return 0;
1103         }
1104
1105         shmem_swp_unmap(entry);
1106 unlock:
1107         spin_unlock(&info->lock);
1108         /*
1109          * add_to_swap_cache() doesn't return -EEXIST, so we can safely
1110          * clear SWAP_HAS_CACHE flag.
1111          */
1112         swapcache_free(swap, NULL);
1113 redirty:
1114         set_page_dirty(page);
1115         if (wbc->for_reclaim)
1116                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
1117         unlock_page(page);
1118         return 0;
1119 }
1120
1121 #ifdef CONFIG_NUMA
1122 #ifdef CONFIG_TMPFS
1123 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1124 {
1125         char buffer[64];
1126
1127         if (!mpol || mpol->mode == MPOL_DEFAULT)
1128                 return;         /* show nothing */
1129
1130         mpol_to_str(buffer, sizeof(buffer), mpol, 1);
1131
1132         seq_printf(seq, ",mpol=%s", buffer);
1133 }
1134
1135 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1136 {
1137         struct mempolicy *mpol = NULL;
1138         if (sbinfo->mpol) {
1139                 spin_lock(&sbinfo->stat_lock);  /* prevent replace/use races */
1140                 mpol = sbinfo->mpol;
1141                 mpol_get(mpol);
1142                 spin_unlock(&sbinfo->stat_lock);
1143         }
1144         return mpol;
1145 }
1146 #endif /* CONFIG_TMPFS */
1147
1148 static struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1149                         struct shmem_inode_info *info, unsigned long idx)
1150 {
1151         struct mempolicy mpol, *spol;
1152         struct vm_area_struct pvma;
1153         struct page *page;
1154
1155         spol = mpol_cond_copy(&mpol,
1156                                 mpol_shared_policy_lookup(&info->policy, idx));
1157
1158         /* Create a pseudo vma that just contains the policy */
1159         pvma.vm_start = 0;
1160         pvma.vm_pgoff = idx;
1161         pvma.vm_ops = NULL;
1162         pvma.vm_policy = spol;
1163         page = swapin_readahead(entry, gfp, &pvma, 0);
1164         return page;
1165 }
1166
1167 static struct page *shmem_alloc_page(gfp_t gfp,
1168                         struct shmem_inode_info *info, unsigned long idx)
1169 {
1170         struct vm_area_struct pvma;
1171
1172         /* Create a pseudo vma that just contains the policy */
1173         pvma.vm_start = 0;
1174         pvma.vm_pgoff = idx;
1175         pvma.vm_ops = NULL;
1176         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, idx);
1177
1178         /*
1179          * alloc_page_vma() will drop the shared policy reference
1180          */
1181         return alloc_page_vma(gfp, &pvma, 0);
1182 }
1183 #else /* !CONFIG_NUMA */
1184 #ifdef CONFIG_TMPFS
1185 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *p)
1186 {
1187 }
1188 #endif /* CONFIG_TMPFS */
1189
1190 static inline struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1191                         struct shmem_inode_info *info, unsigned long idx)
1192 {
1193         return swapin_readahead(entry, gfp, NULL, 0);
1194 }
1195
1196 static inline struct page *shmem_alloc_page(gfp_t gfp,
1197                         struct shmem_inode_info *info, unsigned long idx)
1198 {
1199         return alloc_page(gfp);
1200 }
1201 #endif /* CONFIG_NUMA */
1202
1203 #if !defined(CONFIG_NUMA) || !defined(CONFIG_TMPFS)
1204 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1205 {
1206         return NULL;
1207 }
1208 #endif
1209
1210 /*
1211  * shmem_getpage - either get the page from swap or allocate a new one
1212  *
1213  * If we allocate a new one we do not mark it dirty. That's up to the
1214  * vm. If we swap it in we mark it dirty since we also free the swap
1215  * entry since a page cannot live in both the swap and page cache
1216  */
1217 static int shmem_getpage(struct inode *inode, unsigned long idx,
1218                         struct page **pagep, enum sgp_type sgp, int *type)
1219 {
1220         struct address_space *mapping = inode->i_mapping;
1221         struct shmem_inode_info *info = SHMEM_I(inode);
1222         struct shmem_sb_info *sbinfo;
1223         struct page *filepage = *pagep;
1224         struct page *swappage;
1225         swp_entry_t *entry;
1226         swp_entry_t swap;
1227         gfp_t gfp;
1228         int error;
1229
1230         if (idx >= SHMEM_MAX_INDEX)
1231                 return -EFBIG;
1232
1233         if (type)
1234                 *type = 0;
1235
1236         /*
1237          * Normally, filepage is NULL on entry, and either found
1238          * uptodate immediately, or allocated and zeroed, or read
1239          * in under swappage, which is then assigned to filepage.
1240          * But shmem_readpage (required for splice) passes in a locked
1241          * filepage, which may be found not uptodate by other callers
1242          * too, and may need to be copied from the swappage read in.
1243          */
1244 repeat:
1245         if (!filepage)
1246                 filepage = find_lock_page(mapping, idx);
1247         if (filepage && PageUptodate(filepage))
1248                 goto done;
1249         error = 0;
1250         gfp = mapping_gfp_mask(mapping);
1251         if (!filepage) {
1252                 /*
1253                  * Try to preload while we can wait, to not make a habit of
1254                  * draining atomic reserves; but don't latch on to this cpu.
1255                  */
1256                 error = radix_tree_preload(gfp & ~__GFP_HIGHMEM);
1257                 if (error)
1258                         goto failed;
1259                 radix_tree_preload_end();
1260         }
1261
1262         spin_lock(&info->lock);
1263         shmem_recalc_inode(inode);
1264         entry = shmem_swp_alloc(info, idx, sgp);
1265         if (IS_ERR(entry)) {
1266                 spin_unlock(&info->lock);
1267                 error = PTR_ERR(entry);
1268                 goto failed;
1269         }
1270         swap = *entry;
1271
1272         if (swap.val) {
1273                 /* Look it up and read it in.. */
1274                 swappage = lookup_swap_cache(swap);
1275                 if (!swappage) {
1276                         shmem_swp_unmap(entry);
1277                         /* here we actually do the io */
1278                         if (type && !(*type & VM_FAULT_MAJOR)) {
1279                                 __count_vm_event(PGMAJFAULT);
1280                                 *type |= VM_FAULT_MAJOR;
1281                         }
1282                         spin_unlock(&info->lock);
1283                         swappage = shmem_swapin(swap, gfp, info, idx);
1284                         if (!swappage) {
1285                                 spin_lock(&info->lock);
1286                                 entry = shmem_swp_alloc(info, idx, sgp);
1287                                 if (IS_ERR(entry))
1288                                         error = PTR_ERR(entry);
1289                                 else {
1290                                         if (entry->val == swap.val)
1291                                                 error = -ENOMEM;
1292                                         shmem_swp_unmap(entry);
1293                                 }
1294                                 spin_unlock(&info->lock);
1295                                 if (error)
1296                                         goto failed;
1297                                 goto repeat;
1298                         }
1299                         wait_on_page_locked(swappage);
1300                         page_cache_release(swappage);
1301                         goto repeat;
1302                 }
1303
1304                 /* We have to do this with page locked to prevent races */
1305                 if (!trylock_page(swappage)) {
1306                         shmem_swp_unmap(entry);
1307                         spin_unlock(&info->lock);
1308                         wait_on_page_locked(swappage);
1309                         page_cache_release(swappage);
1310                         goto repeat;
1311                 }
1312                 if (PageWriteback(swappage)) {
1313                         shmem_swp_unmap(entry);
1314                         spin_unlock(&info->lock);
1315                         wait_on_page_writeback(swappage);
1316                         unlock_page(swappage);
1317                         page_cache_release(swappage);
1318                         goto repeat;
1319                 }
1320                 if (!PageUptodate(swappage)) {
1321                         shmem_swp_unmap(entry);
1322                         spin_unlock(&info->lock);
1323                         unlock_page(swappage);
1324                         page_cache_release(swappage);
1325                         error = -EIO;
1326                         goto failed;
1327                 }
1328
1329                 if (filepage) {
1330                         shmem_swp_set(info, entry, 0);
1331                         shmem_swp_unmap(entry);
1332                         delete_from_swap_cache(swappage);
1333                         spin_unlock(&info->lock);
1334                         copy_highpage(filepage, swappage);
1335                         unlock_page(swappage);
1336                         page_cache_release(swappage);
1337                         flush_dcache_page(filepage);
1338                         SetPageUptodate(filepage);
1339                         set_page_dirty(filepage);
1340                         swap_free(swap);
1341                 } else if (!(error = add_to_page_cache_locked(swappage, mapping,
1342                                         idx, GFP_NOWAIT))) {
1343                         info->flags |= SHMEM_PAGEIN;
1344                         shmem_swp_set(info, entry, 0);
1345                         shmem_swp_unmap(entry);
1346                         delete_from_swap_cache(swappage);
1347                         spin_unlock(&info->lock);
1348                         filepage = swappage;
1349                         set_page_dirty(filepage);
1350                         swap_free(swap);
1351                 } else {
1352                         shmem_swp_unmap(entry);
1353                         spin_unlock(&info->lock);
1354                         if (error == -ENOMEM) {
1355                                 /*
1356                                  * reclaim from proper memory cgroup and
1357                                  * call memcg's OOM if needed.
1358                                  */
1359                                 error = mem_cgroup_shmem_charge_fallback(
1360                                                                 swappage,
1361                                                                 current->mm,
1362                                                                 gfp);
1363                                 if (error) {
1364                                         unlock_page(swappage);
1365                                         page_cache_release(swappage);
1366                                         goto failed;
1367                                 }
1368                         }
1369                         unlock_page(swappage);
1370                         page_cache_release(swappage);
1371                         goto repeat;
1372                 }
1373         } else if (sgp == SGP_READ && !filepage) {
1374                 shmem_swp_unmap(entry);
1375                 filepage = find_get_page(mapping, idx);
1376                 if (filepage &&
1377                     (!PageUptodate(filepage) || !trylock_page(filepage))) {
1378                         spin_unlock(&info->lock);
1379                         wait_on_page_locked(filepage);
1380                         page_cache_release(filepage);
1381                         filepage = NULL;
1382                         goto repeat;
1383                 }
1384                 spin_unlock(&info->lock);
1385         } else {
1386                 shmem_swp_unmap(entry);
1387                 sbinfo = SHMEM_SB(inode->i_sb);
1388                 if (sbinfo->max_blocks) {
1389                         if ((percpu_counter_compare(&sbinfo->used_blocks, sbinfo->max_blocks) > 0) ||
1390                             shmem_acct_block(info->flags)) {
1391                                 spin_unlock(&info->lock);
1392                                 error = -ENOSPC;
1393                                 goto failed;
1394                         }
1395                         percpu_counter_inc(&sbinfo->used_blocks);
1396                         spin_lock(&inode->i_lock);
1397                         inode->i_blocks += BLOCKS_PER_PAGE;
1398                         spin_unlock(&inode->i_lock);
1399                 } else if (shmem_acct_block(info->flags)) {
1400                         spin_unlock(&info->lock);
1401                         error = -ENOSPC;
1402                         goto failed;
1403                 }
1404
1405                 if (!filepage) {
1406                         int ret;
1407
1408                         spin_unlock(&info->lock);
1409                         filepage = shmem_alloc_page(gfp, info, idx);
1410                         if (!filepage) {
1411                                 shmem_unacct_blocks(info->flags, 1);
1412                                 shmem_free_blocks(inode, 1);
1413                                 error = -ENOMEM;
1414                                 goto failed;
1415                         }
1416                         SetPageSwapBacked(filepage);
1417
1418                         /* Precharge page while we can wait, compensate after */
1419                         error = mem_cgroup_cache_charge(filepage, current->mm,
1420                                         GFP_KERNEL);
1421                         if (error) {
1422                                 page_cache_release(filepage);
1423                                 shmem_unacct_blocks(info->flags, 1);
1424                                 shmem_free_blocks(inode, 1);
1425                                 filepage = NULL;
1426                                 goto failed;
1427                         }
1428
1429                         spin_lock(&info->lock);
1430                         entry = shmem_swp_alloc(info, idx, sgp);
1431                         if (IS_ERR(entry))
1432                                 error = PTR_ERR(entry);
1433                         else {
1434                                 swap = *entry;
1435                                 shmem_swp_unmap(entry);
1436                         }
1437                         ret = error || swap.val;
1438                         if (ret)
1439                                 mem_cgroup_uncharge_cache_page(filepage);
1440                         else
1441                                 ret = add_to_page_cache_lru(filepage, mapping,
1442                                                 idx, GFP_NOWAIT);
1443                         /*
1444                          * At add_to_page_cache_lru() failure, uncharge will
1445                          * be done automatically.
1446                          */
1447                         if (ret) {
1448                                 spin_unlock(&info->lock);
1449                                 page_cache_release(filepage);
1450                                 shmem_unacct_blocks(info->flags, 1);
1451                                 shmem_free_blocks(inode, 1);
1452                                 filepage = NULL;
1453                                 if (error)
1454                                         goto failed;
1455                                 goto repeat;
1456                         }
1457                         info->flags |= SHMEM_PAGEIN;
1458                 }
1459
1460                 info->alloced++;
1461                 spin_unlock(&info->lock);
1462                 clear_highpage(filepage);
1463                 flush_dcache_page(filepage);
1464                 SetPageUptodate(filepage);
1465                 if (sgp == SGP_DIRTY)
1466                         set_page_dirty(filepage);
1467         }
1468 done:
1469         *pagep = filepage;
1470         return 0;
1471
1472 failed:
1473         if (*pagep != filepage) {
1474                 unlock_page(filepage);
1475                 page_cache_release(filepage);
1476         }
1477         return error;
1478 }
1479
1480 static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1481 {
1482         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1483         int error;
1484         int ret;
1485
1486         if (((loff_t)vmf->pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
1487                 return VM_FAULT_SIGBUS;
1488
1489         error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_CACHE, &ret);
1490         if (error)
1491                 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1492
1493         return ret | VM_FAULT_LOCKED;
1494 }
1495
1496 #ifdef CONFIG_NUMA
1497 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *new)
1498 {
1499         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1500         return mpol_set_shared_policy(&SHMEM_I(i)->policy, vma, new);
1501 }
1502
1503 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
1504                                           unsigned long addr)
1505 {
1506         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1507         unsigned long idx;
1508
1509         idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1510         return mpol_shared_policy_lookup(&SHMEM_I(i)->policy, idx);
1511 }
1512 #endif
1513
1514 int shmem_lock(struct file *file, int lock, struct user_struct *user)
1515 {
1516         struct inode *inode = file->f_path.dentry->d_inode;
1517         struct shmem_inode_info *info = SHMEM_I(inode);
1518         int retval = -ENOMEM;
1519
1520         spin_lock(&info->lock);
1521         if (lock && !(info->flags & VM_LOCKED)) {
1522                 if (!user_shm_lock(inode->i_size, user))
1523                         goto out_nomem;
1524                 info->flags |= VM_LOCKED;
1525                 mapping_set_unevictable(file->f_mapping);
1526         }
1527         if (!lock && (info->flags & VM_LOCKED) && user) {
1528                 user_shm_unlock(inode->i_size, user);
1529                 info->flags &= ~VM_LOCKED;
1530                 mapping_clear_unevictable(file->f_mapping);
1531                 scan_mapping_unevictable_pages(file->f_mapping);
1532         }
1533         retval = 0;
1534
1535 out_nomem:
1536         spin_unlock(&info->lock);
1537         return retval;
1538 }
1539
1540 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1541 {
1542         file_accessed(file);
1543         vma->vm_ops = &shmem_vm_ops;
1544         vma->vm_flags |= VM_CAN_NONLINEAR;
1545         return 0;
1546 }
1547
1548 static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
1549                                      int mode, dev_t dev, unsigned long flags)
1550 {
1551         struct inode *inode;
1552         struct shmem_inode_info *info;
1553         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1554
1555         if (shmem_reserve_inode(sb))
1556                 return NULL;
1557
1558         inode = new_inode(sb);
1559         if (inode) {
1560                 inode_init_owner(inode, dir, mode);
1561                 inode->i_blocks = 0;
1562                 inode->i_mapping->backing_dev_info = &shmem_backing_dev_info;
1563                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1564                 inode->i_generation = get_seconds();
1565                 info = SHMEM_I(inode);
1566                 memset(info, 0, (char *)inode - (char *)info);
1567                 spin_lock_init(&info->lock);
1568                 info->flags = flags & VM_NORESERVE;
1569                 INIT_LIST_HEAD(&info->swaplist);
1570                 cache_no_acl(inode);
1571
1572                 switch (mode & S_IFMT) {
1573                 default:
1574                         inode->i_op = &shmem_special_inode_operations;
1575                         init_special_inode(inode, mode, dev);
1576                         break;
1577                 case S_IFREG:
1578                         inode->i_mapping->a_ops = &shmem_aops;
1579                         inode->i_op = &shmem_inode_operations;
1580                         inode->i_fop = &shmem_file_operations;
1581                         mpol_shared_policy_init(&info->policy,
1582                                                  shmem_get_sbmpol(sbinfo));
1583                         break;
1584                 case S_IFDIR:
1585                         inc_nlink(inode);
1586                         /* Some things misbehave if size == 0 on a directory */
1587                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
1588                         inode->i_op = &shmem_dir_inode_operations;
1589                         inode->i_fop = &simple_dir_operations;
1590                         break;
1591                 case S_IFLNK:
1592                         /*
1593                          * Must not load anything in the rbtree,
1594                          * mpol_free_shared_policy will not be called.
1595                          */
1596                         mpol_shared_policy_init(&info->policy, NULL);
1597                         break;
1598                 }
1599         } else
1600                 shmem_free_inode(sb);
1601         return inode;
1602 }
1603
1604 #ifdef CONFIG_TMPFS
1605 static const struct inode_operations shmem_symlink_inode_operations;
1606 static const struct inode_operations shmem_symlink_inline_operations;
1607
1608 /*
1609  * Normally tmpfs avoids the use of shmem_readpage and shmem_write_begin;
1610  * but providing them allows a tmpfs file to be used for splice, sendfile, and
1611  * below the loop driver, in the generic fashion that many filesystems support.
1612  */
1613 static int shmem_readpage(struct file *file, struct page *page)
1614 {
1615         struct inode *inode = page->mapping->host;
1616         int error = shmem_getpage(inode, page->index, &page, SGP_CACHE, NULL);
1617         unlock_page(page);
1618         return error;
1619 }
1620
1621 static int
1622 shmem_write_begin(struct file *file, struct address_space *mapping,
1623                         loff_t pos, unsigned len, unsigned flags,
1624                         struct page **pagep, void **fsdata)
1625 {
1626         struct inode *inode = mapping->host;
1627         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1628         *pagep = NULL;
1629         return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1630 }
1631
1632 static int
1633 shmem_write_end(struct file *file, struct address_space *mapping,
1634                         loff_t pos, unsigned len, unsigned copied,
1635                         struct page *page, void *fsdata)
1636 {
1637         struct inode *inode = mapping->host;
1638
1639         if (pos + copied > inode->i_size)
1640                 i_size_write(inode, pos + copied);
1641
1642         set_page_dirty(page);
1643         unlock_page(page);
1644         page_cache_release(page);
1645
1646         return copied;
1647 }
1648
1649 static void do_shmem_file_read(struct file *filp, loff_t *ppos, read_descriptor_t *desc, read_actor_t actor)
1650 {
1651         struct inode *inode = filp->f_path.dentry->d_inode;
1652         struct address_space *mapping = inode->i_mapping;
1653         unsigned long index, offset;
1654         enum sgp_type sgp = SGP_READ;
1655
1656         /*
1657          * Might this read be for a stacking filesystem?  Then when reading
1658          * holes of a sparse file, we actually need to allocate those pages,
1659          * and even mark them dirty, so it cannot exceed the max_blocks limit.
1660          */
1661         if (segment_eq(get_fs(), KERNEL_DS))
1662                 sgp = SGP_DIRTY;
1663
1664         index = *ppos >> PAGE_CACHE_SHIFT;
1665         offset = *ppos & ~PAGE_CACHE_MASK;
1666
1667         for (;;) {
1668                 struct page *page = NULL;
1669                 unsigned long end_index, nr, ret;
1670                 loff_t i_size = i_size_read(inode);
1671
1672                 end_index = i_size >> PAGE_CACHE_SHIFT;
1673                 if (index > end_index)
1674                         break;
1675                 if (index == end_index) {
1676                         nr = i_size & ~PAGE_CACHE_MASK;
1677                         if (nr <= offset)
1678                                 break;
1679                 }
1680
1681                 desc->error = shmem_getpage(inode, index, &page, sgp, NULL);
1682                 if (desc->error) {
1683                         if (desc->error == -EINVAL)
1684                                 desc->error = 0;
1685                         break;
1686                 }
1687                 if (page)
1688                         unlock_page(page);
1689
1690                 /*
1691                  * We must evaluate after, since reads (unlike writes)
1692                  * are called without i_mutex protection against truncate
1693                  */
1694                 nr = PAGE_CACHE_SIZE;
1695                 i_size = i_size_read(inode);
1696                 end_index = i_size >> PAGE_CACHE_SHIFT;
1697                 if (index == end_index) {
1698                         nr = i_size & ~PAGE_CACHE_MASK;
1699                         if (nr <= offset) {
1700                                 if (page)
1701                                         page_cache_release(page);
1702                                 break;
1703                         }
1704                 }
1705                 nr -= offset;
1706
1707                 if (page) {
1708                         /*
1709                          * If users can be writing to this page using arbitrary
1710                          * virtual addresses, take care about potential aliasing
1711                          * before reading the page on the kernel side.
1712                          */
1713                         if (mapping_writably_mapped(mapping))
1714                                 flush_dcache_page(page);
1715                         /*
1716                          * Mark the page accessed if we read the beginning.
1717                          */
1718                         if (!offset)
1719                                 mark_page_accessed(page);
1720                 } else {
1721                         page = ZERO_PAGE(0);
1722                         page_cache_get(page);
1723                 }
1724
1725                 /*
1726                  * Ok, we have the page, and it's up-to-date, so
1727                  * now we can copy it to user space...
1728                  *
1729                  * The actor routine returns how many bytes were actually used..
1730                  * NOTE! This may not be the same as how much of a user buffer
1731                  * we filled up (we may be padding etc), so we can only update
1732                  * "pos" here (the actor routine has to update the user buffer
1733                  * pointers and the remaining count).
1734                  */
1735                 ret = actor(desc, page, offset, nr);
1736                 offset += ret;
1737                 index += offset >> PAGE_CACHE_SHIFT;
1738                 offset &= ~PAGE_CACHE_MASK;
1739
1740                 page_cache_release(page);
1741                 if (ret != nr || !desc->count)
1742                         break;
1743
1744                 cond_resched();
1745         }
1746
1747         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1748         file_accessed(filp);
1749 }
1750
1751 static ssize_t shmem_file_aio_read(struct kiocb *iocb,
1752                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
1753 {
1754         struct file *filp = iocb->ki_filp;
1755         ssize_t retval;
1756         unsigned long seg;
1757         size_t count;
1758         loff_t *ppos = &iocb->ki_pos;
1759
1760         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1761         if (retval)
1762                 return retval;
1763
1764         for (seg = 0; seg < nr_segs; seg++) {
1765                 read_descriptor_t desc;
1766
1767                 desc.written = 0;
1768                 desc.arg.buf = iov[seg].iov_base;
1769                 desc.count = iov[seg].iov_len;
1770                 if (desc.count == 0)
1771                         continue;
1772                 desc.error = 0;
1773                 do_shmem_file_read(filp, ppos, &desc, file_read_actor);
1774                 retval += desc.written;
1775                 if (desc.error) {
1776                         retval = retval ?: desc.error;
1777                         break;
1778                 }
1779                 if (desc.count > 0)
1780                         break;
1781         }
1782         return retval;
1783 }
1784
1785 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
1786 {
1787         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
1788
1789         buf->f_type = TMPFS_MAGIC;
1790         buf->f_bsize = PAGE_CACHE_SIZE;
1791         buf->f_namelen = NAME_MAX;
1792         if (sbinfo->max_blocks) {
1793                 buf->f_blocks = sbinfo->max_blocks;
1794                 buf->f_bavail = buf->f_bfree =
1795                                 sbinfo->max_blocks - percpu_counter_sum(&sbinfo->used_blocks);
1796         }
1797         if (sbinfo->max_inodes) {
1798                 buf->f_files = sbinfo->max_inodes;
1799                 buf->f_ffree = sbinfo->free_inodes;
1800         }
1801         /* else leave those fields 0 like simple_statfs */
1802         return 0;
1803 }
1804
1805 /*
1806  * File creation. Allocate an inode, and we're done..
1807  */
1808 static int
1809 shmem_mknod(struct inode *dir, struct dentry *dentry, int mode, dev_t dev)
1810 {
1811         struct inode *inode;
1812         int error = -ENOSPC;
1813
1814         inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
1815         if (inode) {
1816                 error = security_inode_init_security(inode, dir, NULL, NULL,
1817                                                      NULL);
1818                 if (error) {
1819                         if (error != -EOPNOTSUPP) {
1820                                 iput(inode);
1821                                 return error;
1822                         }
1823                 }
1824 #ifdef CONFIG_TMPFS_POSIX_ACL
1825                 error = generic_acl_init(inode, dir);
1826                 if (error) {
1827                         iput(inode);
1828                         return error;
1829                 }
1830 #else
1831                 error = 0;
1832 #endif
1833                 dir->i_size += BOGO_DIRENT_SIZE;
1834                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1835                 d_instantiate(dentry, inode);
1836                 dget(dentry); /* Extra count - pin the dentry in core */
1837         }
1838         return error;
1839 }
1840
1841 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, int mode)
1842 {
1843         int error;
1844
1845         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
1846                 return error;
1847         inc_nlink(dir);
1848         return 0;
1849 }
1850
1851 static int shmem_create(struct inode *dir, struct dentry *dentry, int mode,
1852                 struct nameidata *nd)
1853 {
1854         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
1855 }
1856
1857 /*
1858  * Link a file..
1859  */
1860 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
1861 {
1862         struct inode *inode = old_dentry->d_inode;
1863         int ret;
1864
1865         /*
1866          * No ordinary (disk based) filesystem counts links as inodes;
1867          * but each new link needs a new dentry, pinning lowmem, and
1868          * tmpfs dentries cannot be pruned until they are unlinked.
1869          */
1870         ret = shmem_reserve_inode(inode->i_sb);
1871         if (ret)
1872                 goto out;
1873
1874         dir->i_size += BOGO_DIRENT_SIZE;
1875         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1876         inc_nlink(inode);
1877         atomic_inc(&inode->i_count);    /* New dentry reference */
1878         dget(dentry);           /* Extra pinning count for the created dentry */
1879         d_instantiate(dentry, inode);
1880 out:
1881         return ret;
1882 }
1883
1884 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
1885 {
1886         struct inode *inode = dentry->d_inode;
1887
1888         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
1889                 shmem_free_inode(inode->i_sb);
1890
1891         dir->i_size -= BOGO_DIRENT_SIZE;
1892         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1893         drop_nlink(inode);
1894         dput(dentry);   /* Undo the count from "create" - this does all the work */
1895         return 0;
1896 }
1897
1898 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
1899 {
1900         if (!simple_empty(dentry))
1901                 return -ENOTEMPTY;
1902
1903         drop_nlink(dentry->d_inode);
1904         drop_nlink(dir);
1905         return shmem_unlink(dir, dentry);
1906 }
1907
1908 /*
1909  * The VFS layer already does all the dentry stuff for rename,
1910  * we just have to decrement the usage count for the target if
1911  * it exists so that the VFS layer correctly free's it when it
1912  * gets overwritten.
1913  */
1914 static int shmem_rename(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
1915 {
1916         struct inode *inode = old_dentry->d_inode;
1917         int they_are_dirs = S_ISDIR(inode->i_mode);
1918
1919         if (!simple_empty(new_dentry))
1920                 return -ENOTEMPTY;
1921
1922         if (new_dentry->d_inode) {
1923                 (void) shmem_unlink(new_dir, new_dentry);
1924                 if (they_are_dirs)
1925                         drop_nlink(old_dir);
1926         } else if (they_are_dirs) {
1927                 drop_nlink(old_dir);
1928                 inc_nlink(new_dir);
1929         }
1930
1931         old_dir->i_size -= BOGO_DIRENT_SIZE;
1932         new_dir->i_size += BOGO_DIRENT_SIZE;
1933         old_dir->i_ctime = old_dir->i_mtime =
1934         new_dir->i_ctime = new_dir->i_mtime =
1935         inode->i_ctime = CURRENT_TIME;
1936         return 0;
1937 }
1938
1939 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
1940 {
1941         int error;
1942         int len;
1943         struct inode *inode;
1944         struct page *page = NULL;
1945         char *kaddr;
1946         struct shmem_inode_info *info;
1947
1948         len = strlen(symname) + 1;
1949         if (len > PAGE_CACHE_SIZE)
1950                 return -ENAMETOOLONG;
1951
1952         inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0, VM_NORESERVE);
1953         if (!inode)
1954                 return -ENOSPC;
1955
1956         error = security_inode_init_security(inode, dir, NULL, NULL,
1957                                              NULL);
1958         if (error) {
1959                 if (error != -EOPNOTSUPP) {
1960                         iput(inode);
1961                         return error;
1962                 }
1963                 error = 0;
1964         }
1965
1966         info = SHMEM_I(inode);
1967         inode->i_size = len-1;
1968         if (len <= (char *)inode - (char *)info) {
1969                 /* do it inline */
1970                 memcpy(info, symname, len);
1971                 inode->i_op = &shmem_symlink_inline_operations;
1972         } else {
1973                 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
1974                 if (error) {
1975                         iput(inode);
1976                         return error;
1977                 }
1978                 inode->i_mapping->a_ops = &shmem_aops;
1979                 inode->i_op = &shmem_symlink_inode_operations;
1980                 kaddr = kmap_atomic(page, KM_USER0);
1981                 memcpy(kaddr, symname, len);
1982                 kunmap_atomic(kaddr, KM_USER0);
1983                 set_page_dirty(page);
1984                 unlock_page(page);
1985                 page_cache_release(page);
1986         }
1987         dir->i_size += BOGO_DIRENT_SIZE;
1988         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1989         d_instantiate(dentry, inode);
1990         dget(dentry);
1991         return 0;
1992 }
1993
1994 static void *shmem_follow_link_inline(struct dentry *dentry, struct nameidata *nd)
1995 {
1996         nd_set_link(nd, (char *)SHMEM_I(dentry->d_inode));
1997         return NULL;
1998 }
1999
2000 static void *shmem_follow_link(struct dentry *dentry, struct nameidata *nd)
2001 {
2002         struct page *page = NULL;
2003         int res = shmem_getpage(dentry->d_inode, 0, &page, SGP_READ, NULL);
2004         nd_set_link(nd, res ? ERR_PTR(res) : kmap(page));
2005         if (page)
2006                 unlock_page(page);
2007         return page;
2008 }
2009
2010 static void shmem_put_link(struct dentry *dentry, struct nameidata *nd, void *cookie)
2011 {
2012         if (!IS_ERR(nd_get_link(nd))) {
2013                 struct page *page = cookie;
2014                 kunmap(page);
2015                 mark_page_accessed(page);
2016                 page_cache_release(page);
2017         }
2018 }
2019
2020 static const struct inode_operations shmem_symlink_inline_operations = {
2021         .readlink       = generic_readlink,
2022         .follow_link    = shmem_follow_link_inline,
2023 };
2024
2025 static const struct inode_operations shmem_symlink_inode_operations = {
2026         .readlink       = generic_readlink,
2027         .follow_link    = shmem_follow_link,
2028         .put_link       = shmem_put_link,
2029 };
2030
2031 #ifdef CONFIG_TMPFS_POSIX_ACL
2032 /*
2033  * Superblocks without xattr inode operations will get security.* xattr
2034  * support from the VFS "for free". As soon as we have any other xattrs
2035  * like ACLs, we also need to implement the security.* handlers at
2036  * filesystem level, though.
2037  */
2038
2039 static size_t shmem_xattr_security_list(struct dentry *dentry, char *list,
2040                                         size_t list_len, const char *name,
2041                                         size_t name_len, int handler_flags)
2042 {
2043         return security_inode_listsecurity(dentry->d_inode, list, list_len);
2044 }
2045
2046 static int shmem_xattr_security_get(struct dentry *dentry, const char *name,
2047                 void *buffer, size_t size, int handler_flags)
2048 {
2049         if (strcmp(name, "") == 0)
2050                 return -EINVAL;
2051         return xattr_getsecurity(dentry->d_inode, name, buffer, size);
2052 }
2053
2054 static int shmem_xattr_security_set(struct dentry *dentry, const char *name,
2055                 const void *value, size_t size, int flags, int handler_flags)
2056 {
2057         if (strcmp(name, "") == 0)
2058                 return -EINVAL;
2059         return security_inode_setsecurity(dentry->d_inode, name, value,
2060                                           size, flags);
2061 }
2062
2063 static const struct xattr_handler shmem_xattr_security_handler = {
2064         .prefix = XATTR_SECURITY_PREFIX,
2065         .list   = shmem_xattr_security_list,
2066         .get    = shmem_xattr_security_get,
2067         .set    = shmem_xattr_security_set,
2068 };
2069
2070 static const struct xattr_handler *shmem_xattr_handlers[] = {
2071         &generic_acl_access_handler,
2072         &generic_acl_default_handler,
2073         &shmem_xattr_security_handler,
2074         NULL
2075 };
2076 #endif
2077
2078 static struct dentry *shmem_get_parent(struct dentry *child)
2079 {
2080         return ERR_PTR(-ESTALE);
2081 }
2082
2083 static int shmem_match(struct inode *ino, void *vfh)
2084 {
2085         __u32 *fh = vfh;
2086         __u64 inum = fh[2];
2087         inum = (inum << 32) | fh[1];
2088         return ino->i_ino == inum && fh[0] == ino->i_generation;
2089 }
2090
2091 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
2092                 struct fid *fid, int fh_len, int fh_type)
2093 {
2094         struct inode *inode;
2095         struct dentry *dentry = NULL;
2096         u64 inum = fid->raw[2];
2097         inum = (inum << 32) | fid->raw[1];
2098
2099         if (fh_len < 3)
2100                 return NULL;
2101
2102         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
2103                         shmem_match, fid->raw);
2104         if (inode) {
2105                 dentry = d_find_alias(inode);
2106                 iput(inode);
2107         }
2108
2109         return dentry;
2110 }
2111
2112 static int shmem_encode_fh(struct dentry *dentry, __u32 *fh, int *len,
2113                                 int connectable)
2114 {
2115         struct inode *inode = dentry->d_inode;
2116
2117         if (*len < 3)
2118                 return 255;
2119
2120         if (hlist_unhashed(&inode->i_hash)) {
2121                 /* Unfortunately insert_inode_hash is not idempotent,
2122                  * so as we hash inodes here rather than at creation
2123                  * time, we need a lock to ensure we only try
2124                  * to do it once
2125                  */
2126                 static DEFINE_SPINLOCK(lock);
2127                 spin_lock(&lock);
2128                 if (hlist_unhashed(&inode->i_hash))
2129                         __insert_inode_hash(inode,
2130                                             inode->i_ino + inode->i_generation);
2131                 spin_unlock(&lock);
2132         }
2133
2134         fh[0] = inode->i_generation;
2135         fh[1] = inode->i_ino;
2136         fh[2] = ((__u64)inode->i_ino) >> 32;
2137
2138         *len = 3;
2139         return 1;
2140 }
2141
2142 static const struct export_operations shmem_export_ops = {
2143         .get_parent     = shmem_get_parent,
2144         .encode_fh      = shmem_encode_fh,
2145         .fh_to_dentry   = shmem_fh_to_dentry,
2146 };
2147
2148 static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
2149                                bool remount)
2150 {
2151         char *this_char, *value, *rest;
2152
2153         while (options != NULL) {
2154                 this_char = options;
2155                 for (;;) {
2156                         /*
2157                          * NUL-terminate this option: unfortunately,
2158                          * mount options form a comma-separated list,
2159                          * but mpol's nodelist may also contain commas.
2160                          */
2161                         options = strchr(options, ',');
2162                         if (options == NULL)
2163                                 break;
2164                         options++;
2165                         if (!isdigit(*options)) {
2166                                 options[-1] = '\0';
2167                                 break;
2168                         }
2169                 }
2170                 if (!*this_char)
2171                         continue;
2172                 if ((value = strchr(this_char,'=')) != NULL) {
2173                         *value++ = 0;
2174                 } else {
2175                         printk(KERN_ERR
2176                             "tmpfs: No value for mount option '%s'\n",
2177                             this_char);
2178                         return 1;
2179                 }
2180
2181                 if (!strcmp(this_char,"size")) {
2182                         unsigned long long size;
2183                         size = memparse(value,&rest);
2184                         if (*rest == '%') {
2185                                 size <<= PAGE_SHIFT;
2186                                 size *= totalram_pages;
2187                                 do_div(size, 100);
2188                                 rest++;
2189                         }
2190                         if (*rest)
2191                                 goto bad_val;
2192                         sbinfo->max_blocks =
2193                                 DIV_ROUND_UP(size, PAGE_CACHE_SIZE);
2194                 } else if (!strcmp(this_char,"nr_blocks")) {
2195                         sbinfo->max_blocks = memparse(value, &rest);
2196                         if (*rest)
2197                                 goto bad_val;
2198                 } else if (!strcmp(this_char,"nr_inodes")) {
2199                         sbinfo->max_inodes = memparse(value, &rest);
2200                         if (*rest)
2201                                 goto bad_val;
2202                 } else if (!strcmp(this_char,"mode")) {
2203                         if (remount)
2204                                 continue;
2205                         sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
2206                         if (*rest)
2207                                 goto bad_val;
2208                 } else if (!strcmp(this_char,"uid")) {
2209                         if (remount)
2210                                 continue;
2211                         sbinfo->uid = simple_strtoul(value, &rest, 0);
2212                         if (*rest)
2213                                 goto bad_val;
2214                 } else if (!strcmp(this_char,"gid")) {
2215                         if (remount)
2216                                 continue;
2217                         sbinfo->gid = simple_strtoul(value, &rest, 0);
2218                         if (*rest)
2219                                 goto bad_val;
2220                 } else if (!strcmp(this_char,"mpol")) {
2221                         if (mpol_parse_str(value, &sbinfo->mpol, 1))
2222                                 goto bad_val;
2223                 } else {
2224                         printk(KERN_ERR "tmpfs: Bad mount option %s\n",
2225                                this_char);
2226                         return 1;
2227                 }
2228         }
2229         return 0;
2230
2231 bad_val:
2232         printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
2233                value, this_char);
2234         return 1;
2235
2236 }
2237
2238 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2239 {
2240         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2241         struct shmem_sb_info config = *sbinfo;
2242         unsigned long inodes;
2243         int error = -EINVAL;
2244
2245         if (shmem_parse_options(data, &config, true))
2246                 return error;
2247
2248         spin_lock(&sbinfo->stat_lock);
2249         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2250         if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
2251                 goto out;
2252         if (config.max_inodes < inodes)
2253                 goto out;
2254         /*
2255          * Those tests also disallow limited->unlimited while any are in
2256          * use, so i_blocks will always be zero when max_blocks is zero;
2257          * but we must separately disallow unlimited->limited, because
2258          * in that case we have no record of how much is already in use.
2259          */
2260         if (config.max_blocks && !sbinfo->max_blocks)
2261                 goto out;
2262         if (config.max_inodes && !sbinfo->max_inodes)
2263                 goto out;
2264
2265         error = 0;
2266         sbinfo->max_blocks  = config.max_blocks;
2267         sbinfo->max_inodes  = config.max_inodes;
2268         sbinfo->free_inodes = config.max_inodes - inodes;
2269
2270         mpol_put(sbinfo->mpol);
2271         sbinfo->mpol        = config.mpol;      /* transfers initial ref */
2272 out:
2273         spin_unlock(&sbinfo->stat_lock);
2274         return error;
2275 }
2276
2277 static int shmem_show_options(struct seq_file *seq, struct vfsmount *vfs)
2278 {
2279         struct shmem_sb_info *sbinfo = SHMEM_SB(vfs->mnt_sb);
2280
2281         if (sbinfo->max_blocks != shmem_default_max_blocks())
2282                 seq_printf(seq, ",size=%luk",
2283                         sbinfo->max_blocks << (PAGE_CACHE_SHIFT - 10));
2284         if (sbinfo->max_inodes != shmem_default_max_inodes())
2285                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
2286         if (sbinfo->mode != (S_IRWXUGO | S_ISVTX))
2287                 seq_printf(seq, ",mode=%03o", sbinfo->mode);
2288         if (sbinfo->uid != 0)
2289                 seq_printf(seq, ",uid=%u", sbinfo->uid);
2290         if (sbinfo->gid != 0)
2291                 seq_printf(seq, ",gid=%u", sbinfo->gid);
2292         shmem_show_mpol(seq, sbinfo->mpol);
2293         return 0;
2294 }
2295 #endif /* CONFIG_TMPFS */
2296
2297 static void shmem_put_super(struct super_block *sb)
2298 {
2299         kfree(sb->s_fs_info);
2300         sb->s_fs_info = NULL;
2301 }
2302
2303 int shmem_fill_super(struct super_block *sb, void *data, int silent)
2304 {
2305         struct inode *inode;
2306         struct dentry *root;
2307         struct shmem_sb_info *sbinfo;
2308         int err = -ENOMEM;
2309
2310         /* Round up to L1_CACHE_BYTES to resist false sharing */
2311         sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
2312                                 L1_CACHE_BYTES), GFP_KERNEL);
2313         if (!sbinfo)
2314                 return -ENOMEM;
2315
2316         sbinfo->mode = S_IRWXUGO | S_ISVTX;
2317         sbinfo->uid = current_fsuid();
2318         sbinfo->gid = current_fsgid();
2319         sb->s_fs_info = sbinfo;
2320
2321 #ifdef CONFIG_TMPFS
2322         /*
2323          * Per default we only allow half of the physical ram per
2324          * tmpfs instance, limiting inodes to one per page of lowmem;
2325          * but the internal instance is left unlimited.
2326          */
2327         if (!(sb->s_flags & MS_NOUSER)) {
2328                 sbinfo->max_blocks = shmem_default_max_blocks();
2329                 sbinfo->max_inodes = shmem_default_max_inodes();
2330                 if (shmem_parse_options(data, sbinfo, false)) {
2331                         err = -EINVAL;
2332                         goto failed;
2333                 }
2334         }
2335         sb->s_export_op = &shmem_export_ops;
2336 #else
2337         sb->s_flags |= MS_NOUSER;
2338 #endif
2339
2340         spin_lock_init(&sbinfo->stat_lock);
2341         percpu_counter_init(&sbinfo->used_blocks, 0);
2342         sbinfo->free_inodes = sbinfo->max_inodes;
2343
2344         sb->s_maxbytes = SHMEM_MAX_BYTES;
2345         sb->s_blocksize = PAGE_CACHE_SIZE;
2346         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
2347         sb->s_magic = TMPFS_MAGIC;
2348         sb->s_op = &shmem_ops;
2349         sb->s_time_gran = 1;
2350 #ifdef CONFIG_TMPFS_POSIX_ACL
2351         sb->s_xattr = shmem_xattr_handlers;
2352         sb->s_flags |= MS_POSIXACL;
2353 #endif
2354
2355         inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
2356         if (!inode)
2357                 goto failed;
2358         inode->i_uid = sbinfo->uid;
2359         inode->i_gid = sbinfo->gid;
2360         root = d_alloc_root(inode);
2361         if (!root)
2362                 goto failed_iput;
2363         sb->s_root = root;
2364         return 0;
2365
2366 failed_iput:
2367         iput(inode);
2368 failed:
2369         shmem_put_super(sb);
2370         return err;
2371 }
2372
2373 static struct kmem_cache *shmem_inode_cachep;
2374
2375 static struct inode *shmem_alloc_inode(struct super_block *sb)
2376 {
2377         struct shmem_inode_info *p;
2378         p = (struct shmem_inode_info *)kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
2379         if (!p)
2380                 return NULL;
2381         return &p->vfs_inode;
2382 }
2383
2384 static void shmem_destroy_inode(struct inode *inode)
2385 {
2386         if ((inode->i_mode & S_IFMT) == S_IFREG) {
2387                 /* only struct inode is valid if it's an inline symlink */
2388                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
2389         }
2390         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
2391 }
2392
2393 static void init_once(void *foo)
2394 {
2395         struct shmem_inode_info *p = (struct shmem_inode_info *) foo;
2396
2397         inode_init_once(&p->vfs_inode);
2398 }
2399
2400 static int init_inodecache(void)
2401 {
2402         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
2403                                 sizeof(struct shmem_inode_info),
2404                                 0, SLAB_PANIC, init_once);
2405         return 0;
2406 }
2407
2408 static void destroy_inodecache(void)
2409 {
2410         kmem_cache_destroy(shmem_inode_cachep);
2411 }
2412
2413 static const struct address_space_operations shmem_aops = {
2414         .writepage      = shmem_writepage,
2415         .set_page_dirty = __set_page_dirty_no_writeback,
2416 #ifdef CONFIG_TMPFS
2417         .readpage       = shmem_readpage,
2418         .write_begin    = shmem_write_begin,
2419         .write_end      = shmem_write_end,
2420 #endif
2421         .migratepage    = migrate_page,
2422         .error_remove_page = generic_error_remove_page,
2423 };
2424
2425 static const struct file_operations shmem_file_operations = {
2426         .mmap           = shmem_mmap,
2427 #ifdef CONFIG_TMPFS
2428         .llseek         = generic_file_llseek,
2429         .read           = do_sync_read,
2430         .write          = do_sync_write,
2431         .aio_read       = shmem_file_aio_read,
2432         .aio_write      = generic_file_aio_write,
2433         .fsync          = noop_fsync,
2434         .splice_read    = generic_file_splice_read,
2435         .splice_write   = generic_file_splice_write,
2436 #endif
2437 };
2438
2439 static const struct inode_operations shmem_inode_operations = {
2440         .setattr        = shmem_notify_change,
2441         .truncate_range = shmem_truncate_range,
2442 #ifdef CONFIG_TMPFS_POSIX_ACL
2443         .setxattr       = generic_setxattr,
2444         .getxattr       = generic_getxattr,
2445         .listxattr      = generic_listxattr,
2446         .removexattr    = generic_removexattr,
2447         .check_acl      = generic_check_acl,
2448 #endif
2449
2450 };
2451
2452 static const struct inode_operations shmem_dir_inode_operations = {
2453 #ifdef CONFIG_TMPFS
2454         .create         = shmem_create,
2455         .lookup         = simple_lookup,
2456         .link           = shmem_link,
2457         .unlink         = shmem_unlink,
2458         .symlink        = shmem_symlink,
2459         .mkdir          = shmem_mkdir,
2460         .rmdir          = shmem_rmdir,
2461         .mknod          = shmem_mknod,
2462         .rename         = shmem_rename,
2463 #endif
2464 #ifdef CONFIG_TMPFS_POSIX_ACL
2465         .setattr        = shmem_notify_change,
2466         .setxattr       = generic_setxattr,
2467         .getxattr       = generic_getxattr,
2468         .listxattr      = generic_listxattr,
2469         .removexattr    = generic_removexattr,
2470         .check_acl      = generic_check_acl,
2471 #endif
2472 };
2473
2474 static const struct inode_operations shmem_special_inode_operations = {
2475 #ifdef CONFIG_TMPFS_POSIX_ACL
2476         .setattr        = shmem_notify_change,
2477         .setxattr       = generic_setxattr,
2478         .getxattr       = generic_getxattr,
2479         .listxattr      = generic_listxattr,
2480         .removexattr    = generic_removexattr,
2481         .check_acl      = generic_check_acl,
2482 #endif
2483 };
2484
2485 static const struct super_operations shmem_ops = {
2486         .alloc_inode    = shmem_alloc_inode,
2487         .destroy_inode  = shmem_destroy_inode,
2488 #ifdef CONFIG_TMPFS
2489         .statfs         = shmem_statfs,
2490         .remount_fs     = shmem_remount_fs,
2491         .show_options   = shmem_show_options,
2492 #endif
2493         .delete_inode   = shmem_delete_inode,
2494         .drop_inode     = generic_delete_inode,
2495         .put_super      = shmem_put_super,
2496 };
2497
2498 static const struct vm_operations_struct shmem_vm_ops = {
2499         .fault          = shmem_fault,
2500 #ifdef CONFIG_NUMA
2501         .set_policy     = shmem_set_policy,
2502         .get_policy     = shmem_get_policy,
2503 #endif
2504 };
2505
2506
2507 static int shmem_get_sb(struct file_system_type *fs_type,
2508         int flags, const char *dev_name, void *data, struct vfsmount *mnt)
2509 {
2510         return get_sb_nodev(fs_type, flags, data, shmem_fill_super, mnt);
2511 }
2512
2513 static struct file_system_type tmpfs_fs_type = {
2514         .owner          = THIS_MODULE,
2515         .name           = "tmpfs",
2516         .get_sb         = shmem_get_sb,
2517         .kill_sb        = kill_litter_super,
2518 };
2519
2520 int __init init_tmpfs(void)
2521 {
2522         int error;
2523
2524         error = bdi_init(&shmem_backing_dev_info);
2525         if (error)
2526                 goto out4;
2527
2528         error = init_inodecache();
2529         if (error)
2530                 goto out3;
2531
2532         error = register_filesystem(&tmpfs_fs_type);
2533         if (error) {
2534                 printk(KERN_ERR "Could not register tmpfs\n");
2535                 goto out2;
2536         }
2537
2538         shm_mnt = vfs_kern_mount(&tmpfs_fs_type, MS_NOUSER,
2539                                 tmpfs_fs_type.name, NULL);
2540         if (IS_ERR(shm_mnt)) {
2541                 error = PTR_ERR(shm_mnt);
2542                 printk(KERN_ERR "Could not kern_mount tmpfs\n");
2543                 goto out1;
2544         }
2545         return 0;
2546
2547 out1:
2548         unregister_filesystem(&tmpfs_fs_type);
2549 out2:
2550         destroy_inodecache();
2551 out3:
2552         bdi_destroy(&shmem_backing_dev_info);
2553 out4:
2554         shm_mnt = ERR_PTR(error);
2555         return error;
2556 }
2557
2558 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2559 /**
2560  * mem_cgroup_get_shmem_target - find a page or entry assigned to the shmem file
2561  * @inode: the inode to be searched
2562  * @pgoff: the offset to be searched
2563  * @pagep: the pointer for the found page to be stored
2564  * @ent: the pointer for the found swap entry to be stored
2565  *
2566  * If a page is found, refcount of it is incremented. Callers should handle
2567  * these refcount.
2568  */
2569 void mem_cgroup_get_shmem_target(struct inode *inode, pgoff_t pgoff,
2570                                         struct page **pagep, swp_entry_t *ent)
2571 {
2572         swp_entry_t entry = { .val = 0 }, *ptr;
2573         struct page *page = NULL;
2574         struct shmem_inode_info *info = SHMEM_I(inode);
2575
2576         if ((pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
2577                 goto out;
2578
2579         spin_lock(&info->lock);
2580         ptr = shmem_swp_entry(info, pgoff, NULL);
2581 #ifdef CONFIG_SWAP
2582         if (ptr && ptr->val) {
2583                 entry.val = ptr->val;
2584                 page = find_get_page(&swapper_space, entry.val);
2585         } else
2586 #endif
2587                 page = find_get_page(inode->i_mapping, pgoff);
2588         if (ptr)
2589                 shmem_swp_unmap(ptr);
2590         spin_unlock(&info->lock);
2591 out:
2592         *pagep = page;
2593         *ent = entry;
2594 }
2595 #endif
2596
2597 #else /* !CONFIG_SHMEM */
2598
2599 /*
2600  * tiny-shmem: simple shmemfs and tmpfs using ramfs code
2601  *
2602  * This is intended for small system where the benefits of the full
2603  * shmem code (swap-backed and resource-limited) are outweighed by
2604  * their complexity. On systems without swap this code should be
2605  * effectively equivalent, but much lighter weight.
2606  */
2607
2608 #include <linux/ramfs.h>
2609
2610 static struct file_system_type tmpfs_fs_type = {
2611         .name           = "tmpfs",
2612         .get_sb         = ramfs_get_sb,
2613         .kill_sb        = kill_litter_super,
2614 };
2615
2616 int __init init_tmpfs(void)
2617 {
2618         BUG_ON(register_filesystem(&tmpfs_fs_type) != 0);
2619
2620         shm_mnt = kern_mount(&tmpfs_fs_type);
2621         BUG_ON(IS_ERR(shm_mnt));
2622
2623         return 0;
2624 }
2625
2626 int shmem_unuse(swp_entry_t entry, struct page *page)
2627 {
2628         return 0;
2629 }
2630
2631 int shmem_lock(struct file *file, int lock, struct user_struct *user)
2632 {
2633         return 0;
2634 }
2635
2636 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2637 /**
2638  * mem_cgroup_get_shmem_target - find a page or entry assigned to the shmem file
2639  * @inode: the inode to be searched
2640  * @pgoff: the offset to be searched
2641  * @pagep: the pointer for the found page to be stored
2642  * @ent: the pointer for the found swap entry to be stored
2643  *
2644  * If a page is found, refcount of it is incremented. Callers should handle
2645  * these refcount.
2646  */
2647 void mem_cgroup_get_shmem_target(struct inode *inode, pgoff_t pgoff,
2648                                         struct page **pagep, swp_entry_t *ent)
2649 {
2650         struct page *page = NULL;
2651
2652         if ((pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
2653                 goto out;
2654         page = find_get_page(inode->i_mapping, pgoff);
2655 out:
2656         *pagep = page;
2657         *ent = (swp_entry_t){ .val = 0 };
2658 }
2659 #endif
2660
2661 #define shmem_vm_ops                            generic_file_vm_ops
2662 #define shmem_file_operations                   ramfs_file_operations
2663 #define shmem_get_inode(sb, dir, mode, dev, flags)      ramfs_get_inode(sb, dir, mode, dev)
2664 #define shmem_acct_size(flags, size)            0
2665 #define shmem_unacct_size(flags, size)          do {} while (0)
2666 #define SHMEM_MAX_BYTES                         MAX_LFS_FILESIZE
2667
2668 #endif /* CONFIG_SHMEM */
2669
2670 /* common code */
2671
2672 /**
2673  * shmem_file_setup - get an unlinked file living in tmpfs
2674  * @name: name for dentry (to be seen in /proc/<pid>/maps
2675  * @size: size to be set for the file
2676  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
2677  */
2678 struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
2679 {
2680         int error;
2681         struct file *file;
2682         struct inode *inode;
2683         struct path path;
2684         struct dentry *root;
2685         struct qstr this;
2686
2687         if (IS_ERR(shm_mnt))
2688                 return (void *)shm_mnt;
2689
2690         if (size < 0 || size > SHMEM_MAX_BYTES)
2691                 return ERR_PTR(-EINVAL);
2692
2693         if (shmem_acct_size(flags, size))
2694                 return ERR_PTR(-ENOMEM);
2695
2696         error = -ENOMEM;
2697         this.name = name;
2698         this.len = strlen(name);
2699         this.hash = 0; /* will go */
2700         root = shm_mnt->mnt_root;
2701         path.dentry = d_alloc(root, &this);
2702         if (!path.dentry)
2703                 goto put_memory;
2704         path.mnt = mntget(shm_mnt);
2705
2706         error = -ENOSPC;
2707         inode = shmem_get_inode(root->d_sb, NULL, S_IFREG | S_IRWXUGO, 0, flags);
2708         if (!inode)
2709                 goto put_dentry;
2710
2711         d_instantiate(path.dentry, inode);
2712         inode->i_size = size;
2713         inode->i_nlink = 0;     /* It is unlinked */
2714 #ifndef CONFIG_MMU
2715         error = ramfs_nommu_expand_for_mapping(inode, size);
2716         if (error)
2717                 goto put_dentry;
2718 #endif
2719
2720         error = -ENFILE;
2721         file = alloc_file(&path, FMODE_WRITE | FMODE_READ,
2722                   &shmem_file_operations);
2723         if (!file)
2724                 goto put_dentry;
2725
2726         return file;
2727
2728 put_dentry:
2729         path_put(&path);
2730 put_memory:
2731         shmem_unacct_size(flags, size);
2732         return ERR_PTR(error);
2733 }
2734 EXPORT_SYMBOL_GPL(shmem_file_setup);
2735
2736 /**
2737  * shmem_zero_setup - setup a shared anonymous mapping
2738  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
2739  */
2740 int shmem_zero_setup(struct vm_area_struct *vma)
2741 {
2742         struct file *file;
2743         loff_t size = vma->vm_end - vma->vm_start;
2744
2745         file = shmem_file_setup("dev/zero", size, vma->vm_flags);
2746         if (IS_ERR(file))
2747                 return PTR_ERR(file);
2748
2749         if (vma->vm_file)
2750                 fput(vma->vm_file);
2751         vma->vm_file = file;
2752         vma->vm_ops = &shmem_vm_ops;
2753         return 0;
2754 }