tmpfs: no need to use i_lock
[linux-2.6.git] / mm / shmem.c
1 /*
2  * Resizable virtual memory filesystem for Linux.
3  *
4  * Copyright (C) 2000 Linus Torvalds.
5  *               2000 Transmeta Corp.
6  *               2000-2001 Christoph Rohland
7  *               2000-2001 SAP AG
8  *               2002 Red Hat Inc.
9  * Copyright (C) 2002-2005 Hugh Dickins.
10  * Copyright (C) 2002-2005 VERITAS Software Corporation.
11  * Copyright (C) 2004 Andi Kleen, SuSE Labs
12  *
13  * Extended attribute support for tmpfs:
14  * Copyright (c) 2004, Luke Kenneth Casson Leighton <lkcl@lkcl.net>
15  * Copyright (c) 2004 Red Hat, Inc., James Morris <jmorris@redhat.com>
16  *
17  * tiny-shmem:
18  * Copyright (c) 2004, 2008 Matt Mackall <mpm@selenic.com>
19  *
20  * This file is released under the GPL.
21  */
22
23 #include <linux/fs.h>
24 #include <linux/init.h>
25 #include <linux/vfs.h>
26 #include <linux/mount.h>
27 #include <linux/pagemap.h>
28 #include <linux/file.h>
29 #include <linux/mm.h>
30 #include <linux/module.h>
31 #include <linux/percpu_counter.h>
32 #include <linux/swap.h>
33
34 static struct vfsmount *shm_mnt;
35
36 #ifdef CONFIG_SHMEM
37 /*
38  * This virtual memory filesystem is heavily based on the ramfs. It
39  * extends ramfs by the ability to use swap and honor resource limits
40  * which makes it a completely usable filesystem.
41  */
42
43 #include <linux/xattr.h>
44 #include <linux/exportfs.h>
45 #include <linux/posix_acl.h>
46 #include <linux/generic_acl.h>
47 #include <linux/mman.h>
48 #include <linux/string.h>
49 #include <linux/slab.h>
50 #include <linux/backing-dev.h>
51 #include <linux/shmem_fs.h>
52 #include <linux/writeback.h>
53 #include <linux/blkdev.h>
54 #include <linux/security.h>
55 #include <linux/swapops.h>
56 #include <linux/mempolicy.h>
57 #include <linux/namei.h>
58 #include <linux/ctype.h>
59 #include <linux/migrate.h>
60 #include <linux/highmem.h>
61 #include <linux/seq_file.h>
62 #include <linux/magic.h>
63
64 #include <asm/uaccess.h>
65 #include <asm/div64.h>
66 #include <asm/pgtable.h>
67
68 /*
69  * The maximum size of a shmem/tmpfs file is limited by the maximum size of
70  * its triple-indirect swap vector - see illustration at shmem_swp_entry().
71  *
72  * With 4kB page size, maximum file size is just over 2TB on a 32-bit kernel,
73  * but one eighth of that on a 64-bit kernel.  With 8kB page size, maximum
74  * file size is just over 4TB on a 64-bit kernel, but 16TB on a 32-bit kernel,
75  * MAX_LFS_FILESIZE being then more restrictive than swap vector layout.
76  *
77  * We use / and * instead of shifts in the definitions below, so that the swap
78  * vector can be tested with small even values (e.g. 20) for ENTRIES_PER_PAGE.
79  */
80 #define ENTRIES_PER_PAGE (PAGE_CACHE_SIZE/sizeof(unsigned long))
81 #define ENTRIES_PER_PAGEPAGE ((unsigned long long)ENTRIES_PER_PAGE*ENTRIES_PER_PAGE)
82
83 #define SHMSWP_MAX_INDEX (SHMEM_NR_DIRECT + (ENTRIES_PER_PAGEPAGE/2) * (ENTRIES_PER_PAGE+1))
84 #define SHMSWP_MAX_BYTES (SHMSWP_MAX_INDEX << PAGE_CACHE_SHIFT)
85
86 #define SHMEM_MAX_BYTES  min_t(unsigned long long, SHMSWP_MAX_BYTES, MAX_LFS_FILESIZE)
87 #define SHMEM_MAX_INDEX  ((unsigned long)((SHMEM_MAX_BYTES+1) >> PAGE_CACHE_SHIFT))
88
89 #define BLOCKS_PER_PAGE  (PAGE_CACHE_SIZE/512)
90 #define VM_ACCT(size)    (PAGE_CACHE_ALIGN(size) >> PAGE_SHIFT)
91
92 /* info->flags needs VM_flags to handle pagein/truncate races efficiently */
93 #define SHMEM_PAGEIN     VM_READ
94 #define SHMEM_TRUNCATE   VM_WRITE
95
96 /* Definition to limit shmem_truncate's steps between cond_rescheds */
97 #define LATENCY_LIMIT    64
98
99 /* Pretend that each entry is of this size in directory's i_size */
100 #define BOGO_DIRENT_SIZE 20
101
102 struct shmem_xattr {
103         struct list_head list;  /* anchored by shmem_inode_info->xattr_list */
104         char *name;             /* xattr name */
105         size_t size;
106         char value[0];
107 };
108
109 /* Flag allocation requirements to shmem_getpage and shmem_swp_alloc */
110 enum sgp_type {
111         SGP_READ,       /* don't exceed i_size, don't allocate page */
112         SGP_CACHE,      /* don't exceed i_size, may allocate page */
113         SGP_DIRTY,      /* like SGP_CACHE, but set new page dirty */
114         SGP_WRITE,      /* may exceed i_size, may allocate page */
115 };
116
117 #ifdef CONFIG_TMPFS
118 static unsigned long shmem_default_max_blocks(void)
119 {
120         return totalram_pages / 2;
121 }
122
123 static unsigned long shmem_default_max_inodes(void)
124 {
125         return min(totalram_pages - totalhigh_pages, totalram_pages / 2);
126 }
127 #endif
128
129 static int shmem_getpage(struct inode *inode, unsigned long idx,
130                          struct page **pagep, enum sgp_type sgp, int *type);
131
132 static inline struct page *shmem_dir_alloc(gfp_t gfp_mask)
133 {
134         /*
135          * The above definition of ENTRIES_PER_PAGE, and the use of
136          * BLOCKS_PER_PAGE on indirect pages, assume PAGE_CACHE_SIZE:
137          * might be reconsidered if it ever diverges from PAGE_SIZE.
138          *
139          * Mobility flags are masked out as swap vectors cannot move
140          */
141         return alloc_pages((gfp_mask & ~GFP_MOVABLE_MASK) | __GFP_ZERO,
142                                 PAGE_CACHE_SHIFT-PAGE_SHIFT);
143 }
144
145 static inline void shmem_dir_free(struct page *page)
146 {
147         __free_pages(page, PAGE_CACHE_SHIFT-PAGE_SHIFT);
148 }
149
150 static struct page **shmem_dir_map(struct page *page)
151 {
152         return (struct page **)kmap_atomic(page, KM_USER0);
153 }
154
155 static inline void shmem_dir_unmap(struct page **dir)
156 {
157         kunmap_atomic(dir, KM_USER0);
158 }
159
160 static swp_entry_t *shmem_swp_map(struct page *page)
161 {
162         return (swp_entry_t *)kmap_atomic(page, KM_USER1);
163 }
164
165 static inline void shmem_swp_balance_unmap(void)
166 {
167         /*
168          * When passing a pointer to an i_direct entry, to code which
169          * also handles indirect entries and so will shmem_swp_unmap,
170          * we must arrange for the preempt count to remain in balance.
171          * What kmap_atomic of a lowmem page does depends on config
172          * and architecture, so pretend to kmap_atomic some lowmem page.
173          */
174         (void) kmap_atomic(ZERO_PAGE(0), KM_USER1);
175 }
176
177 static inline void shmem_swp_unmap(swp_entry_t *entry)
178 {
179         kunmap_atomic(entry, KM_USER1);
180 }
181
182 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
183 {
184         return sb->s_fs_info;
185 }
186
187 /*
188  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
189  * for shared memory and for shared anonymous (/dev/zero) mappings
190  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
191  * consistent with the pre-accounting of private mappings ...
192  */
193 static inline int shmem_acct_size(unsigned long flags, loff_t size)
194 {
195         return (flags & VM_NORESERVE) ?
196                 0 : security_vm_enough_memory_kern(VM_ACCT(size));
197 }
198
199 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
200 {
201         if (!(flags & VM_NORESERVE))
202                 vm_unacct_memory(VM_ACCT(size));
203 }
204
205 /*
206  * ... whereas tmpfs objects are accounted incrementally as
207  * pages are allocated, in order to allow huge sparse files.
208  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
209  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
210  */
211 static inline int shmem_acct_block(unsigned long flags)
212 {
213         return (flags & VM_NORESERVE) ?
214                 security_vm_enough_memory_kern(VM_ACCT(PAGE_CACHE_SIZE)) : 0;
215 }
216
217 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
218 {
219         if (flags & VM_NORESERVE)
220                 vm_unacct_memory(pages * VM_ACCT(PAGE_CACHE_SIZE));
221 }
222
223 static const struct super_operations shmem_ops;
224 static const struct address_space_operations shmem_aops;
225 static const struct file_operations shmem_file_operations;
226 static const struct inode_operations shmem_inode_operations;
227 static const struct inode_operations shmem_dir_inode_operations;
228 static const struct inode_operations shmem_special_inode_operations;
229 static const struct vm_operations_struct shmem_vm_ops;
230
231 static struct backing_dev_info shmem_backing_dev_info  __read_mostly = {
232         .ra_pages       = 0,    /* No readahead */
233         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_SWAP_BACKED,
234 };
235
236 static LIST_HEAD(shmem_swaplist);
237 static DEFINE_MUTEX(shmem_swaplist_mutex);
238
239 static void shmem_free_blocks(struct inode *inode, long pages)
240 {
241         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
242         if (sbinfo->max_blocks) {
243                 percpu_counter_add(&sbinfo->used_blocks, -pages);
244                 inode->i_blocks -= pages*BLOCKS_PER_PAGE;
245         }
246 }
247
248 static int shmem_reserve_inode(struct super_block *sb)
249 {
250         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
251         if (sbinfo->max_inodes) {
252                 spin_lock(&sbinfo->stat_lock);
253                 if (!sbinfo->free_inodes) {
254                         spin_unlock(&sbinfo->stat_lock);
255                         return -ENOSPC;
256                 }
257                 sbinfo->free_inodes--;
258                 spin_unlock(&sbinfo->stat_lock);
259         }
260         return 0;
261 }
262
263 static void shmem_free_inode(struct super_block *sb)
264 {
265         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
266         if (sbinfo->max_inodes) {
267                 spin_lock(&sbinfo->stat_lock);
268                 sbinfo->free_inodes++;
269                 spin_unlock(&sbinfo->stat_lock);
270         }
271 }
272
273 /**
274  * shmem_recalc_inode - recalculate the size of an inode
275  * @inode: inode to recalc
276  *
277  * We have to calculate the free blocks since the mm can drop
278  * undirtied hole pages behind our back.
279  *
280  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
281  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
282  *
283  * It has to be called with the spinlock held.
284  */
285 static void shmem_recalc_inode(struct inode *inode)
286 {
287         struct shmem_inode_info *info = SHMEM_I(inode);
288         long freed;
289
290         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
291         if (freed > 0) {
292                 info->alloced -= freed;
293                 shmem_unacct_blocks(info->flags, freed);
294                 shmem_free_blocks(inode, freed);
295         }
296 }
297
298 /**
299  * shmem_swp_entry - find the swap vector position in the info structure
300  * @info:  info structure for the inode
301  * @index: index of the page to find
302  * @page:  optional page to add to the structure. Has to be preset to
303  *         all zeros
304  *
305  * If there is no space allocated yet it will return NULL when
306  * page is NULL, else it will use the page for the needed block,
307  * setting it to NULL on return to indicate that it has been used.
308  *
309  * The swap vector is organized the following way:
310  *
311  * There are SHMEM_NR_DIRECT entries directly stored in the
312  * shmem_inode_info structure. So small files do not need an addional
313  * allocation.
314  *
315  * For pages with index > SHMEM_NR_DIRECT there is the pointer
316  * i_indirect which points to a page which holds in the first half
317  * doubly indirect blocks, in the second half triple indirect blocks:
318  *
319  * For an artificial ENTRIES_PER_PAGE = 4 this would lead to the
320  * following layout (for SHMEM_NR_DIRECT == 16):
321  *
322  * i_indirect -> dir --> 16-19
323  *            |      +-> 20-23
324  *            |
325  *            +-->dir2 --> 24-27
326  *            |        +-> 28-31
327  *            |        +-> 32-35
328  *            |        +-> 36-39
329  *            |
330  *            +-->dir3 --> 40-43
331  *                     +-> 44-47
332  *                     +-> 48-51
333  *                     +-> 52-55
334  */
335 static swp_entry_t *shmem_swp_entry(struct shmem_inode_info *info, unsigned long index, struct page **page)
336 {
337         unsigned long offset;
338         struct page **dir;
339         struct page *subdir;
340
341         if (index < SHMEM_NR_DIRECT) {
342                 shmem_swp_balance_unmap();
343                 return info->i_direct+index;
344         }
345         if (!info->i_indirect) {
346                 if (page) {
347                         info->i_indirect = *page;
348                         *page = NULL;
349                 }
350                 return NULL;                    /* need another page */
351         }
352
353         index -= SHMEM_NR_DIRECT;
354         offset = index % ENTRIES_PER_PAGE;
355         index /= ENTRIES_PER_PAGE;
356         dir = shmem_dir_map(info->i_indirect);
357
358         if (index >= ENTRIES_PER_PAGE/2) {
359                 index -= ENTRIES_PER_PAGE/2;
360                 dir += ENTRIES_PER_PAGE/2 + index/ENTRIES_PER_PAGE;
361                 index %= ENTRIES_PER_PAGE;
362                 subdir = *dir;
363                 if (!subdir) {
364                         if (page) {
365                                 *dir = *page;
366                                 *page = NULL;
367                         }
368                         shmem_dir_unmap(dir);
369                         return NULL;            /* need another page */
370                 }
371                 shmem_dir_unmap(dir);
372                 dir = shmem_dir_map(subdir);
373         }
374
375         dir += index;
376         subdir = *dir;
377         if (!subdir) {
378                 if (!page || !(subdir = *page)) {
379                         shmem_dir_unmap(dir);
380                         return NULL;            /* need a page */
381                 }
382                 *dir = subdir;
383                 *page = NULL;
384         }
385         shmem_dir_unmap(dir);
386         return shmem_swp_map(subdir) + offset;
387 }
388
389 static void shmem_swp_set(struct shmem_inode_info *info, swp_entry_t *entry, unsigned long value)
390 {
391         long incdec = value? 1: -1;
392
393         entry->val = value;
394         info->swapped += incdec;
395         if ((unsigned long)(entry - info->i_direct) >= SHMEM_NR_DIRECT) {
396                 struct page *page = kmap_atomic_to_page(entry);
397                 set_page_private(page, page_private(page) + incdec);
398         }
399 }
400
401 /**
402  * shmem_swp_alloc - get the position of the swap entry for the page.
403  * @info:       info structure for the inode
404  * @index:      index of the page to find
405  * @sgp:        check and recheck i_size? skip allocation?
406  *
407  * If the entry does not exist, allocate it.
408  */
409 static swp_entry_t *shmem_swp_alloc(struct shmem_inode_info *info, unsigned long index, enum sgp_type sgp)
410 {
411         struct inode *inode = &info->vfs_inode;
412         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
413         struct page *page = NULL;
414         swp_entry_t *entry;
415
416         if (sgp != SGP_WRITE &&
417             ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode))
418                 return ERR_PTR(-EINVAL);
419
420         while (!(entry = shmem_swp_entry(info, index, &page))) {
421                 if (sgp == SGP_READ)
422                         return shmem_swp_map(ZERO_PAGE(0));
423                 /*
424                  * Test used_blocks against 1 less max_blocks, since we have 1 data
425                  * page (and perhaps indirect index pages) yet to allocate:
426                  * a waste to allocate index if we cannot allocate data.
427                  */
428                 if (sbinfo->max_blocks) {
429                         if (percpu_counter_compare(&sbinfo->used_blocks,
430                                                 sbinfo->max_blocks - 1) >= 0)
431                                 return ERR_PTR(-ENOSPC);
432                         percpu_counter_inc(&sbinfo->used_blocks);
433                         inode->i_blocks += BLOCKS_PER_PAGE;
434                 }
435
436                 spin_unlock(&info->lock);
437                 page = shmem_dir_alloc(mapping_gfp_mask(inode->i_mapping));
438                 spin_lock(&info->lock);
439
440                 if (!page) {
441                         shmem_free_blocks(inode, 1);
442                         return ERR_PTR(-ENOMEM);
443                 }
444                 if (sgp != SGP_WRITE &&
445                     ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
446                         entry = ERR_PTR(-EINVAL);
447                         break;
448                 }
449                 if (info->next_index <= index)
450                         info->next_index = index + 1;
451         }
452         if (page) {
453                 /* another task gave its page, or truncated the file */
454                 shmem_free_blocks(inode, 1);
455                 shmem_dir_free(page);
456         }
457         if (info->next_index <= index && !IS_ERR(entry))
458                 info->next_index = index + 1;
459         return entry;
460 }
461
462 /**
463  * shmem_free_swp - free some swap entries in a directory
464  * @dir:        pointer to the directory
465  * @edir:       pointer after last entry of the directory
466  * @punch_lock: pointer to spinlock when needed for the holepunch case
467  */
468 static int shmem_free_swp(swp_entry_t *dir, swp_entry_t *edir,
469                                                 spinlock_t *punch_lock)
470 {
471         spinlock_t *punch_unlock = NULL;
472         swp_entry_t *ptr;
473         int freed = 0;
474
475         for (ptr = dir; ptr < edir; ptr++) {
476                 if (ptr->val) {
477                         if (unlikely(punch_lock)) {
478                                 punch_unlock = punch_lock;
479                                 punch_lock = NULL;
480                                 spin_lock(punch_unlock);
481                                 if (!ptr->val)
482                                         continue;
483                         }
484                         free_swap_and_cache(*ptr);
485                         *ptr = (swp_entry_t){0};
486                         freed++;
487                 }
488         }
489         if (punch_unlock)
490                 spin_unlock(punch_unlock);
491         return freed;
492 }
493
494 static int shmem_map_and_free_swp(struct page *subdir, int offset,
495                 int limit, struct page ***dir, spinlock_t *punch_lock)
496 {
497         swp_entry_t *ptr;
498         int freed = 0;
499
500         ptr = shmem_swp_map(subdir);
501         for (; offset < limit; offset += LATENCY_LIMIT) {
502                 int size = limit - offset;
503                 if (size > LATENCY_LIMIT)
504                         size = LATENCY_LIMIT;
505                 freed += shmem_free_swp(ptr+offset, ptr+offset+size,
506                                                         punch_lock);
507                 if (need_resched()) {
508                         shmem_swp_unmap(ptr);
509                         if (*dir) {
510                                 shmem_dir_unmap(*dir);
511                                 *dir = NULL;
512                         }
513                         cond_resched();
514                         ptr = shmem_swp_map(subdir);
515                 }
516         }
517         shmem_swp_unmap(ptr);
518         return freed;
519 }
520
521 static void shmem_free_pages(struct list_head *next)
522 {
523         struct page *page;
524         int freed = 0;
525
526         do {
527                 page = container_of(next, struct page, lru);
528                 next = next->next;
529                 shmem_dir_free(page);
530                 freed++;
531                 if (freed >= LATENCY_LIMIT) {
532                         cond_resched();
533                         freed = 0;
534                 }
535         } while (next);
536 }
537
538 void shmem_truncate_range(struct inode *inode, loff_t start, loff_t end)
539 {
540         struct shmem_inode_info *info = SHMEM_I(inode);
541         unsigned long idx;
542         unsigned long size;
543         unsigned long limit;
544         unsigned long stage;
545         unsigned long diroff;
546         struct page **dir;
547         struct page *topdir;
548         struct page *middir;
549         struct page *subdir;
550         swp_entry_t *ptr;
551         LIST_HEAD(pages_to_free);
552         long nr_pages_to_free = 0;
553         long nr_swaps_freed = 0;
554         int offset;
555         int freed;
556         int punch_hole;
557         spinlock_t *needs_lock;
558         spinlock_t *punch_lock;
559         unsigned long upper_limit;
560
561         truncate_inode_pages_range(inode->i_mapping, start, end);
562
563         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
564         idx = (start + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
565         if (idx >= info->next_index)
566                 return;
567
568         spin_lock(&info->lock);
569         info->flags |= SHMEM_TRUNCATE;
570         if (likely(end == (loff_t) -1)) {
571                 limit = info->next_index;
572                 upper_limit = SHMEM_MAX_INDEX;
573                 info->next_index = idx;
574                 needs_lock = NULL;
575                 punch_hole = 0;
576         } else {
577                 if (end + 1 >= inode->i_size) { /* we may free a little more */
578                         limit = (inode->i_size + PAGE_CACHE_SIZE - 1) >>
579                                                         PAGE_CACHE_SHIFT;
580                         upper_limit = SHMEM_MAX_INDEX;
581                 } else {
582                         limit = (end + 1) >> PAGE_CACHE_SHIFT;
583                         upper_limit = limit;
584                 }
585                 needs_lock = &info->lock;
586                 punch_hole = 1;
587         }
588
589         topdir = info->i_indirect;
590         if (topdir && idx <= SHMEM_NR_DIRECT && !punch_hole) {
591                 info->i_indirect = NULL;
592                 nr_pages_to_free++;
593                 list_add(&topdir->lru, &pages_to_free);
594         }
595         spin_unlock(&info->lock);
596
597         if (info->swapped && idx < SHMEM_NR_DIRECT) {
598                 ptr = info->i_direct;
599                 size = limit;
600                 if (size > SHMEM_NR_DIRECT)
601                         size = SHMEM_NR_DIRECT;
602                 nr_swaps_freed = shmem_free_swp(ptr+idx, ptr+size, needs_lock);
603         }
604
605         /*
606          * If there are no indirect blocks or we are punching a hole
607          * below indirect blocks, nothing to be done.
608          */
609         if (!topdir || limit <= SHMEM_NR_DIRECT)
610                 goto done2;
611
612         /*
613          * The truncation case has already dropped info->lock, and we're safe
614          * because i_size and next_index have already been lowered, preventing
615          * access beyond.  But in the punch_hole case, we still need to take
616          * the lock when updating the swap directory, because there might be
617          * racing accesses by shmem_getpage(SGP_CACHE), shmem_unuse_inode or
618          * shmem_writepage.  However, whenever we find we can remove a whole
619          * directory page (not at the misaligned start or end of the range),
620          * we first NULLify its pointer in the level above, and then have no
621          * need to take the lock when updating its contents: needs_lock and
622          * punch_lock (either pointing to info->lock or NULL) manage this.
623          */
624
625         upper_limit -= SHMEM_NR_DIRECT;
626         limit -= SHMEM_NR_DIRECT;
627         idx = (idx > SHMEM_NR_DIRECT)? (idx - SHMEM_NR_DIRECT): 0;
628         offset = idx % ENTRIES_PER_PAGE;
629         idx -= offset;
630
631         dir = shmem_dir_map(topdir);
632         stage = ENTRIES_PER_PAGEPAGE/2;
633         if (idx < ENTRIES_PER_PAGEPAGE/2) {
634                 middir = topdir;
635                 diroff = idx/ENTRIES_PER_PAGE;
636         } else {
637                 dir += ENTRIES_PER_PAGE/2;
638                 dir += (idx - ENTRIES_PER_PAGEPAGE/2)/ENTRIES_PER_PAGEPAGE;
639                 while (stage <= idx)
640                         stage += ENTRIES_PER_PAGEPAGE;
641                 middir = *dir;
642                 if (*dir) {
643                         diroff = ((idx - ENTRIES_PER_PAGEPAGE/2) %
644                                 ENTRIES_PER_PAGEPAGE) / ENTRIES_PER_PAGE;
645                         if (!diroff && !offset && upper_limit >= stage) {
646                                 if (needs_lock) {
647                                         spin_lock(needs_lock);
648                                         *dir = NULL;
649                                         spin_unlock(needs_lock);
650                                         needs_lock = NULL;
651                                 } else
652                                         *dir = NULL;
653                                 nr_pages_to_free++;
654                                 list_add(&middir->lru, &pages_to_free);
655                         }
656                         shmem_dir_unmap(dir);
657                         dir = shmem_dir_map(middir);
658                 } else {
659                         diroff = 0;
660                         offset = 0;
661                         idx = stage;
662                 }
663         }
664
665         for (; idx < limit; idx += ENTRIES_PER_PAGE, diroff++) {
666                 if (unlikely(idx == stage)) {
667                         shmem_dir_unmap(dir);
668                         dir = shmem_dir_map(topdir) +
669                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
670                         while (!*dir) {
671                                 dir++;
672                                 idx += ENTRIES_PER_PAGEPAGE;
673                                 if (idx >= limit)
674                                         goto done1;
675                         }
676                         stage = idx + ENTRIES_PER_PAGEPAGE;
677                         middir = *dir;
678                         if (punch_hole)
679                                 needs_lock = &info->lock;
680                         if (upper_limit >= stage) {
681                                 if (needs_lock) {
682                                         spin_lock(needs_lock);
683                                         *dir = NULL;
684                                         spin_unlock(needs_lock);
685                                         needs_lock = NULL;
686                                 } else
687                                         *dir = NULL;
688                                 nr_pages_to_free++;
689                                 list_add(&middir->lru, &pages_to_free);
690                         }
691                         shmem_dir_unmap(dir);
692                         cond_resched();
693                         dir = shmem_dir_map(middir);
694                         diroff = 0;
695                 }
696                 punch_lock = needs_lock;
697                 subdir = dir[diroff];
698                 if (subdir && !offset && upper_limit-idx >= ENTRIES_PER_PAGE) {
699                         if (needs_lock) {
700                                 spin_lock(needs_lock);
701                                 dir[diroff] = NULL;
702                                 spin_unlock(needs_lock);
703                                 punch_lock = NULL;
704                         } else
705                                 dir[diroff] = NULL;
706                         nr_pages_to_free++;
707                         list_add(&subdir->lru, &pages_to_free);
708                 }
709                 if (subdir && page_private(subdir) /* has swap entries */) {
710                         size = limit - idx;
711                         if (size > ENTRIES_PER_PAGE)
712                                 size = ENTRIES_PER_PAGE;
713                         freed = shmem_map_and_free_swp(subdir,
714                                         offset, size, &dir, punch_lock);
715                         if (!dir)
716                                 dir = shmem_dir_map(middir);
717                         nr_swaps_freed += freed;
718                         if (offset || punch_lock) {
719                                 spin_lock(&info->lock);
720                                 set_page_private(subdir,
721                                         page_private(subdir) - freed);
722                                 spin_unlock(&info->lock);
723                         } else
724                                 BUG_ON(page_private(subdir) != freed);
725                 }
726                 offset = 0;
727         }
728 done1:
729         shmem_dir_unmap(dir);
730 done2:
731         if (inode->i_mapping->nrpages && (info->flags & SHMEM_PAGEIN)) {
732                 /*
733                  * Call truncate_inode_pages again: racing shmem_unuse_inode
734                  * may have swizzled a page in from swap since
735                  * truncate_pagecache or generic_delete_inode did it, before we
736                  * lowered next_index.  Also, though shmem_getpage checks
737                  * i_size before adding to cache, no recheck after: so fix the
738                  * narrow window there too.
739                  */
740                 truncate_inode_pages_range(inode->i_mapping, start, end);
741         }
742
743         spin_lock(&info->lock);
744         info->flags &= ~SHMEM_TRUNCATE;
745         info->swapped -= nr_swaps_freed;
746         if (nr_pages_to_free)
747                 shmem_free_blocks(inode, nr_pages_to_free);
748         shmem_recalc_inode(inode);
749         spin_unlock(&info->lock);
750
751         /*
752          * Empty swap vector directory pages to be freed?
753          */
754         if (!list_empty(&pages_to_free)) {
755                 pages_to_free.prev->next = NULL;
756                 shmem_free_pages(pages_to_free.next);
757         }
758 }
759 EXPORT_SYMBOL_GPL(shmem_truncate_range);
760
761 static int shmem_setattr(struct dentry *dentry, struct iattr *attr)
762 {
763         struct inode *inode = dentry->d_inode;
764         int error;
765
766         error = inode_change_ok(inode, attr);
767         if (error)
768                 return error;
769
770         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
771                 loff_t oldsize = inode->i_size;
772                 loff_t newsize = attr->ia_size;
773                 struct page *page = NULL;
774
775                 if (newsize < oldsize) {
776                         /*
777                          * If truncating down to a partial page, then
778                          * if that page is already allocated, hold it
779                          * in memory until the truncation is over, so
780                          * truncate_partial_page cannot miss it were
781                          * it assigned to swap.
782                          */
783                         if (newsize & (PAGE_CACHE_SIZE-1)) {
784                                 (void) shmem_getpage(inode,
785                                         newsize >> PAGE_CACHE_SHIFT,
786                                                 &page, SGP_READ, NULL);
787                                 if (page)
788                                         unlock_page(page);
789                         }
790                         /*
791                          * Reset SHMEM_PAGEIN flag so that shmem_truncate can
792                          * detect if any pages might have been added to cache
793                          * after truncate_inode_pages.  But we needn't bother
794                          * if it's being fully truncated to zero-length: the
795                          * nrpages check is efficient enough in that case.
796                          */
797                         if (newsize) {
798                                 struct shmem_inode_info *info = SHMEM_I(inode);
799                                 spin_lock(&info->lock);
800                                 info->flags &= ~SHMEM_PAGEIN;
801                                 spin_unlock(&info->lock);
802                         }
803                 }
804                 if (newsize != oldsize) {
805                         i_size_write(inode, newsize);
806                         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
807                 }
808                 if (newsize < oldsize) {
809                         loff_t holebegin = round_up(newsize, PAGE_SIZE);
810                         unmap_mapping_range(inode->i_mapping, holebegin, 0, 1);
811                         shmem_truncate_range(inode, newsize, (loff_t)-1);
812                         /* unmap again to remove racily COWed private pages */
813                         unmap_mapping_range(inode->i_mapping, holebegin, 0, 1);
814                 }
815                 if (page)
816                         page_cache_release(page);
817         }
818
819         setattr_copy(inode, attr);
820 #ifdef CONFIG_TMPFS_POSIX_ACL
821         if (attr->ia_valid & ATTR_MODE)
822                 error = generic_acl_chmod(inode);
823 #endif
824         return error;
825 }
826
827 static void shmem_evict_inode(struct inode *inode)
828 {
829         struct shmem_inode_info *info = SHMEM_I(inode);
830         struct shmem_xattr *xattr, *nxattr;
831
832         if (inode->i_mapping->a_ops == &shmem_aops) {
833                 shmem_unacct_size(info->flags, inode->i_size);
834                 inode->i_size = 0;
835                 shmem_truncate_range(inode, 0, (loff_t)-1);
836                 if (!list_empty(&info->swaplist)) {
837                         mutex_lock(&shmem_swaplist_mutex);
838                         list_del_init(&info->swaplist);
839                         mutex_unlock(&shmem_swaplist_mutex);
840                 }
841         }
842
843         list_for_each_entry_safe(xattr, nxattr, &info->xattr_list, list) {
844                 kfree(xattr->name);
845                 kfree(xattr);
846         }
847         BUG_ON(inode->i_blocks);
848         shmem_free_inode(inode->i_sb);
849         end_writeback(inode);
850 }
851
852 static inline int shmem_find_swp(swp_entry_t entry, swp_entry_t *dir, swp_entry_t *edir)
853 {
854         swp_entry_t *ptr;
855
856         for (ptr = dir; ptr < edir; ptr++) {
857                 if (ptr->val == entry.val)
858                         return ptr - dir;
859         }
860         return -1;
861 }
862
863 static int shmem_unuse_inode(struct shmem_inode_info *info, swp_entry_t entry, struct page *page)
864 {
865         struct address_space *mapping;
866         unsigned long idx;
867         unsigned long size;
868         unsigned long limit;
869         unsigned long stage;
870         struct page **dir;
871         struct page *subdir;
872         swp_entry_t *ptr;
873         int offset;
874         int error;
875
876         idx = 0;
877         ptr = info->i_direct;
878         spin_lock(&info->lock);
879         if (!info->swapped) {
880                 list_del_init(&info->swaplist);
881                 goto lost2;
882         }
883         limit = info->next_index;
884         size = limit;
885         if (size > SHMEM_NR_DIRECT)
886                 size = SHMEM_NR_DIRECT;
887         offset = shmem_find_swp(entry, ptr, ptr+size);
888         if (offset >= 0) {
889                 shmem_swp_balance_unmap();
890                 goto found;
891         }
892         if (!info->i_indirect)
893                 goto lost2;
894
895         dir = shmem_dir_map(info->i_indirect);
896         stage = SHMEM_NR_DIRECT + ENTRIES_PER_PAGEPAGE/2;
897
898         for (idx = SHMEM_NR_DIRECT; idx < limit; idx += ENTRIES_PER_PAGE, dir++) {
899                 if (unlikely(idx == stage)) {
900                         shmem_dir_unmap(dir-1);
901                         if (cond_resched_lock(&info->lock)) {
902                                 /* check it has not been truncated */
903                                 if (limit > info->next_index) {
904                                         limit = info->next_index;
905                                         if (idx >= limit)
906                                                 goto lost2;
907                                 }
908                         }
909                         dir = shmem_dir_map(info->i_indirect) +
910                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
911                         while (!*dir) {
912                                 dir++;
913                                 idx += ENTRIES_PER_PAGEPAGE;
914                                 if (idx >= limit)
915                                         goto lost1;
916                         }
917                         stage = idx + ENTRIES_PER_PAGEPAGE;
918                         subdir = *dir;
919                         shmem_dir_unmap(dir);
920                         dir = shmem_dir_map(subdir);
921                 }
922                 subdir = *dir;
923                 if (subdir && page_private(subdir)) {
924                         ptr = shmem_swp_map(subdir);
925                         size = limit - idx;
926                         if (size > ENTRIES_PER_PAGE)
927                                 size = ENTRIES_PER_PAGE;
928                         offset = shmem_find_swp(entry, ptr, ptr+size);
929                         shmem_swp_unmap(ptr);
930                         if (offset >= 0) {
931                                 shmem_dir_unmap(dir);
932                                 ptr = shmem_swp_map(subdir);
933                                 goto found;
934                         }
935                 }
936         }
937 lost1:
938         shmem_dir_unmap(dir-1);
939 lost2:
940         spin_unlock(&info->lock);
941         return 0;
942 found:
943         idx += offset;
944         ptr += offset;
945
946         /*
947          * Move _head_ to start search for next from here.
948          * But be careful: shmem_evict_inode checks list_empty without taking
949          * mutex, and there's an instant in list_move_tail when info->swaplist
950          * would appear empty, if it were the only one on shmem_swaplist.  We
951          * could avoid doing it if inode NULL; or use this minor optimization.
952          */
953         if (shmem_swaplist.next != &info->swaplist)
954                 list_move_tail(&shmem_swaplist, &info->swaplist);
955
956         /*
957          * We rely on shmem_swaplist_mutex, not only to protect the swaplist,
958          * but also to hold up shmem_evict_inode(): so inode cannot be freed
959          * beneath us (pagelock doesn't help until the page is in pagecache).
960          */
961         mapping = info->vfs_inode.i_mapping;
962         error = add_to_page_cache_locked(page, mapping, idx, GFP_NOWAIT);
963         /* which does mem_cgroup_uncharge_cache_page on error */
964
965         if (error == -EEXIST) {
966                 struct page *filepage = find_get_page(mapping, idx);
967                 error = 1;
968                 if (filepage) {
969                         /*
970                          * There might be a more uptodate page coming down
971                          * from a stacked writepage: forget our swappage if so.
972                          */
973                         if (PageUptodate(filepage))
974                                 error = 0;
975                         page_cache_release(filepage);
976                 }
977         }
978         if (!error) {
979                 delete_from_swap_cache(page);
980                 set_page_dirty(page);
981                 info->flags |= SHMEM_PAGEIN;
982                 shmem_swp_set(info, ptr, 0);
983                 swap_free(entry);
984                 error = 1;      /* not an error, but entry was found */
985         }
986         shmem_swp_unmap(ptr);
987         spin_unlock(&info->lock);
988         return error;
989 }
990
991 /*
992  * shmem_unuse() search for an eventually swapped out shmem page.
993  */
994 int shmem_unuse(swp_entry_t entry, struct page *page)
995 {
996         struct list_head *p, *next;
997         struct shmem_inode_info *info;
998         int found = 0;
999         int error;
1000
1001         /*
1002          * Charge page using GFP_KERNEL while we can wait, before taking
1003          * the shmem_swaplist_mutex which might hold up shmem_writepage().
1004          * Charged back to the user (not to caller) when swap account is used.
1005          * add_to_page_cache() will be called with GFP_NOWAIT.
1006          */
1007         error = mem_cgroup_cache_charge(page, current->mm, GFP_KERNEL);
1008         if (error)
1009                 goto out;
1010         /*
1011          * Try to preload while we can wait, to not make a habit of
1012          * draining atomic reserves; but don't latch on to this cpu,
1013          * it's okay if sometimes we get rescheduled after this.
1014          */
1015         error = radix_tree_preload(GFP_KERNEL);
1016         if (error)
1017                 goto uncharge;
1018         radix_tree_preload_end();
1019
1020         mutex_lock(&shmem_swaplist_mutex);
1021         list_for_each_safe(p, next, &shmem_swaplist) {
1022                 info = list_entry(p, struct shmem_inode_info, swaplist);
1023                 found = shmem_unuse_inode(info, entry, page);
1024                 cond_resched();
1025                 if (found)
1026                         break;
1027         }
1028         mutex_unlock(&shmem_swaplist_mutex);
1029
1030 uncharge:
1031         if (!found)
1032                 mem_cgroup_uncharge_cache_page(page);
1033         if (found < 0)
1034                 error = found;
1035 out:
1036         unlock_page(page);
1037         page_cache_release(page);
1038         return error;
1039 }
1040
1041 /*
1042  * Move the page from the page cache to the swap cache.
1043  */
1044 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1045 {
1046         struct shmem_inode_info *info;
1047         swp_entry_t *entry, swap;
1048         struct address_space *mapping;
1049         unsigned long index;
1050         struct inode *inode;
1051
1052         BUG_ON(!PageLocked(page));
1053         mapping = page->mapping;
1054         index = page->index;
1055         inode = mapping->host;
1056         info = SHMEM_I(inode);
1057         if (info->flags & VM_LOCKED)
1058                 goto redirty;
1059         if (!total_swap_pages)
1060                 goto redirty;
1061
1062         /*
1063          * shmem_backing_dev_info's capabilities prevent regular writeback or
1064          * sync from ever calling shmem_writepage; but a stacking filesystem
1065          * may use the ->writepage of its underlying filesystem, in which case
1066          * tmpfs should write out to swap only in response to memory pressure,
1067          * and not for the writeback threads or sync.  However, in those cases,
1068          * we do still want to check if there's a redundant swappage to be
1069          * discarded.
1070          */
1071         if (wbc->for_reclaim)
1072                 swap = get_swap_page();
1073         else
1074                 swap.val = 0;
1075
1076         /*
1077          * Add inode to shmem_unuse()'s list of swapped-out inodes,
1078          * if it's not already there.  Do it now because we cannot take
1079          * mutex while holding spinlock, and must do so before the page
1080          * is moved to swap cache, when its pagelock no longer protects
1081          * the inode from eviction.  But don't unlock the mutex until
1082          * we've taken the spinlock, because shmem_unuse_inode() will
1083          * prune a !swapped inode from the swaplist under both locks.
1084          */
1085         if (swap.val) {
1086                 mutex_lock(&shmem_swaplist_mutex);
1087                 if (list_empty(&info->swaplist))
1088                         list_add_tail(&info->swaplist, &shmem_swaplist);
1089         }
1090
1091         spin_lock(&info->lock);
1092         if (swap.val)
1093                 mutex_unlock(&shmem_swaplist_mutex);
1094
1095         if (index >= info->next_index) {
1096                 BUG_ON(!(info->flags & SHMEM_TRUNCATE));
1097                 goto unlock;
1098         }
1099         entry = shmem_swp_entry(info, index, NULL);
1100         if (entry->val) {
1101                 /*
1102                  * The more uptodate page coming down from a stacked
1103                  * writepage should replace our old swappage.
1104                  */
1105                 free_swap_and_cache(*entry);
1106                 shmem_swp_set(info, entry, 0);
1107         }
1108         shmem_recalc_inode(inode);
1109
1110         if (swap.val && add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
1111                 delete_from_page_cache(page);
1112                 shmem_swp_set(info, entry, swap.val);
1113                 shmem_swp_unmap(entry);
1114                 swap_shmem_alloc(swap);
1115                 spin_unlock(&info->lock);
1116                 BUG_ON(page_mapped(page));
1117                 swap_writepage(page, wbc);
1118                 return 0;
1119         }
1120
1121         shmem_swp_unmap(entry);
1122 unlock:
1123         spin_unlock(&info->lock);
1124         /*
1125          * add_to_swap_cache() doesn't return -EEXIST, so we can safely
1126          * clear SWAP_HAS_CACHE flag.
1127          */
1128         swapcache_free(swap, NULL);
1129 redirty:
1130         set_page_dirty(page);
1131         if (wbc->for_reclaim)
1132                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
1133         unlock_page(page);
1134         return 0;
1135 }
1136
1137 #ifdef CONFIG_NUMA
1138 #ifdef CONFIG_TMPFS
1139 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1140 {
1141         char buffer[64];
1142
1143         if (!mpol || mpol->mode == MPOL_DEFAULT)
1144                 return;         /* show nothing */
1145
1146         mpol_to_str(buffer, sizeof(buffer), mpol, 1);
1147
1148         seq_printf(seq, ",mpol=%s", buffer);
1149 }
1150
1151 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1152 {
1153         struct mempolicy *mpol = NULL;
1154         if (sbinfo->mpol) {
1155                 spin_lock(&sbinfo->stat_lock);  /* prevent replace/use races */
1156                 mpol = sbinfo->mpol;
1157                 mpol_get(mpol);
1158                 spin_unlock(&sbinfo->stat_lock);
1159         }
1160         return mpol;
1161 }
1162 #endif /* CONFIG_TMPFS */
1163
1164 static struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1165                         struct shmem_inode_info *info, unsigned long idx)
1166 {
1167         struct mempolicy mpol, *spol;
1168         struct vm_area_struct pvma;
1169         struct page *page;
1170
1171         spol = mpol_cond_copy(&mpol,
1172                                 mpol_shared_policy_lookup(&info->policy, idx));
1173
1174         /* Create a pseudo vma that just contains the policy */
1175         pvma.vm_start = 0;
1176         pvma.vm_pgoff = idx;
1177         pvma.vm_ops = NULL;
1178         pvma.vm_policy = spol;
1179         page = swapin_readahead(entry, gfp, &pvma, 0);
1180         return page;
1181 }
1182
1183 static struct page *shmem_alloc_page(gfp_t gfp,
1184                         struct shmem_inode_info *info, unsigned long idx)
1185 {
1186         struct vm_area_struct pvma;
1187
1188         /* Create a pseudo vma that just contains the policy */
1189         pvma.vm_start = 0;
1190         pvma.vm_pgoff = idx;
1191         pvma.vm_ops = NULL;
1192         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, idx);
1193
1194         /*
1195          * alloc_page_vma() will drop the shared policy reference
1196          */
1197         return alloc_page_vma(gfp, &pvma, 0);
1198 }
1199 #else /* !CONFIG_NUMA */
1200 #ifdef CONFIG_TMPFS
1201 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *p)
1202 {
1203 }
1204 #endif /* CONFIG_TMPFS */
1205
1206 static inline struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1207                         struct shmem_inode_info *info, unsigned long idx)
1208 {
1209         return swapin_readahead(entry, gfp, NULL, 0);
1210 }
1211
1212 static inline struct page *shmem_alloc_page(gfp_t gfp,
1213                         struct shmem_inode_info *info, unsigned long idx)
1214 {
1215         return alloc_page(gfp);
1216 }
1217 #endif /* CONFIG_NUMA */
1218
1219 #if !defined(CONFIG_NUMA) || !defined(CONFIG_TMPFS)
1220 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1221 {
1222         return NULL;
1223 }
1224 #endif
1225
1226 /*
1227  * shmem_getpage - either get the page from swap or allocate a new one
1228  *
1229  * If we allocate a new one we do not mark it dirty. That's up to the
1230  * vm. If we swap it in we mark it dirty since we also free the swap
1231  * entry since a page cannot live in both the swap and page cache
1232  */
1233 static int shmem_getpage(struct inode *inode, unsigned long idx,
1234                         struct page **pagep, enum sgp_type sgp, int *type)
1235 {
1236         struct address_space *mapping = inode->i_mapping;
1237         struct shmem_inode_info *info = SHMEM_I(inode);
1238         struct shmem_sb_info *sbinfo;
1239         struct page *filepage = *pagep;
1240         struct page *swappage;
1241         struct page *prealloc_page = NULL;
1242         swp_entry_t *entry;
1243         swp_entry_t swap;
1244         gfp_t gfp;
1245         int error;
1246
1247         if (idx >= SHMEM_MAX_INDEX)
1248                 return -EFBIG;
1249
1250         if (type)
1251                 *type = 0;
1252
1253         /*
1254          * Normally, filepage is NULL on entry, and either found
1255          * uptodate immediately, or allocated and zeroed, or read
1256          * in under swappage, which is then assigned to filepage.
1257          * But shmem_readpage (required for splice) passes in a locked
1258          * filepage, which may be found not uptodate by other callers
1259          * too, and may need to be copied from the swappage read in.
1260          */
1261 repeat:
1262         if (!filepage)
1263                 filepage = find_lock_page(mapping, idx);
1264         if (filepage && PageUptodate(filepage))
1265                 goto done;
1266         gfp = mapping_gfp_mask(mapping);
1267         if (!filepage) {
1268                 /*
1269                  * Try to preload while we can wait, to not make a habit of
1270                  * draining atomic reserves; but don't latch on to this cpu.
1271                  */
1272                 error = radix_tree_preload(gfp & ~__GFP_HIGHMEM);
1273                 if (error)
1274                         goto failed;
1275                 radix_tree_preload_end();
1276                 if (sgp != SGP_READ && !prealloc_page) {
1277                         /* We don't care if this fails */
1278                         prealloc_page = shmem_alloc_page(gfp, info, idx);
1279                         if (prealloc_page) {
1280                                 if (mem_cgroup_cache_charge(prealloc_page,
1281                                                 current->mm, GFP_KERNEL)) {
1282                                         page_cache_release(prealloc_page);
1283                                         prealloc_page = NULL;
1284                                 }
1285                         }
1286                 }
1287         }
1288         error = 0;
1289
1290         spin_lock(&info->lock);
1291         shmem_recalc_inode(inode);
1292         entry = shmem_swp_alloc(info, idx, sgp);
1293         if (IS_ERR(entry)) {
1294                 spin_unlock(&info->lock);
1295                 error = PTR_ERR(entry);
1296                 goto failed;
1297         }
1298         swap = *entry;
1299
1300         if (swap.val) {
1301                 /* Look it up and read it in.. */
1302                 swappage = lookup_swap_cache(swap);
1303                 if (!swappage) {
1304                         shmem_swp_unmap(entry);
1305                         spin_unlock(&info->lock);
1306                         /* here we actually do the io */
1307                         if (type)
1308                                 *type |= VM_FAULT_MAJOR;
1309                         swappage = shmem_swapin(swap, gfp, info, idx);
1310                         if (!swappage) {
1311                                 spin_lock(&info->lock);
1312                                 entry = shmem_swp_alloc(info, idx, sgp);
1313                                 if (IS_ERR(entry))
1314                                         error = PTR_ERR(entry);
1315                                 else {
1316                                         if (entry->val == swap.val)
1317                                                 error = -ENOMEM;
1318                                         shmem_swp_unmap(entry);
1319                                 }
1320                                 spin_unlock(&info->lock);
1321                                 if (error)
1322                                         goto failed;
1323                                 goto repeat;
1324                         }
1325                         wait_on_page_locked(swappage);
1326                         page_cache_release(swappage);
1327                         goto repeat;
1328                 }
1329
1330                 /* We have to do this with page locked to prevent races */
1331                 if (!trylock_page(swappage)) {
1332                         shmem_swp_unmap(entry);
1333                         spin_unlock(&info->lock);
1334                         wait_on_page_locked(swappage);
1335                         page_cache_release(swappage);
1336                         goto repeat;
1337                 }
1338                 if (PageWriteback(swappage)) {
1339                         shmem_swp_unmap(entry);
1340                         spin_unlock(&info->lock);
1341                         wait_on_page_writeback(swappage);
1342                         unlock_page(swappage);
1343                         page_cache_release(swappage);
1344                         goto repeat;
1345                 }
1346                 if (!PageUptodate(swappage)) {
1347                         shmem_swp_unmap(entry);
1348                         spin_unlock(&info->lock);
1349                         unlock_page(swappage);
1350                         page_cache_release(swappage);
1351                         error = -EIO;
1352                         goto failed;
1353                 }
1354
1355                 if (filepage) {
1356                         shmem_swp_set(info, entry, 0);
1357                         shmem_swp_unmap(entry);
1358                         delete_from_swap_cache(swappage);
1359                         spin_unlock(&info->lock);
1360                         copy_highpage(filepage, swappage);
1361                         unlock_page(swappage);
1362                         page_cache_release(swappage);
1363                         flush_dcache_page(filepage);
1364                         SetPageUptodate(filepage);
1365                         set_page_dirty(filepage);
1366                         swap_free(swap);
1367                 } else if (!(error = add_to_page_cache_locked(swappage, mapping,
1368                                         idx, GFP_NOWAIT))) {
1369                         info->flags |= SHMEM_PAGEIN;
1370                         shmem_swp_set(info, entry, 0);
1371                         shmem_swp_unmap(entry);
1372                         delete_from_swap_cache(swappage);
1373                         spin_unlock(&info->lock);
1374                         filepage = swappage;
1375                         set_page_dirty(filepage);
1376                         swap_free(swap);
1377                 } else {
1378                         shmem_swp_unmap(entry);
1379                         spin_unlock(&info->lock);
1380                         if (error == -ENOMEM) {
1381                                 /*
1382                                  * reclaim from proper memory cgroup and
1383                                  * call memcg's OOM if needed.
1384                                  */
1385                                 error = mem_cgroup_shmem_charge_fallback(
1386                                                                 swappage,
1387                                                                 current->mm,
1388                                                                 gfp);
1389                                 if (error) {
1390                                         unlock_page(swappage);
1391                                         page_cache_release(swappage);
1392                                         goto failed;
1393                                 }
1394                         }
1395                         unlock_page(swappage);
1396                         page_cache_release(swappage);
1397                         goto repeat;
1398                 }
1399         } else if (sgp == SGP_READ && !filepage) {
1400                 shmem_swp_unmap(entry);
1401                 filepage = find_get_page(mapping, idx);
1402                 if (filepage &&
1403                     (!PageUptodate(filepage) || !trylock_page(filepage))) {
1404                         spin_unlock(&info->lock);
1405                         wait_on_page_locked(filepage);
1406                         page_cache_release(filepage);
1407                         filepage = NULL;
1408                         goto repeat;
1409                 }
1410                 spin_unlock(&info->lock);
1411         } else {
1412                 shmem_swp_unmap(entry);
1413                 sbinfo = SHMEM_SB(inode->i_sb);
1414                 if (sbinfo->max_blocks) {
1415                         if (percpu_counter_compare(&sbinfo->used_blocks,
1416                                                 sbinfo->max_blocks) >= 0 ||
1417                             shmem_acct_block(info->flags))
1418                                 goto nospace;
1419                         percpu_counter_inc(&sbinfo->used_blocks);
1420                         inode->i_blocks += BLOCKS_PER_PAGE;
1421                 } else if (shmem_acct_block(info->flags))
1422                         goto nospace;
1423
1424                 if (!filepage) {
1425                         int ret;
1426
1427                         if (!prealloc_page) {
1428                                 spin_unlock(&info->lock);
1429                                 filepage = shmem_alloc_page(gfp, info, idx);
1430                                 if (!filepage) {
1431                                         spin_lock(&info->lock);
1432                                         shmem_unacct_blocks(info->flags, 1);
1433                                         shmem_free_blocks(inode, 1);
1434                                         spin_unlock(&info->lock);
1435                                         error = -ENOMEM;
1436                                         goto failed;
1437                                 }
1438                                 SetPageSwapBacked(filepage);
1439
1440                                 /*
1441                                  * Precharge page while we can wait, compensate
1442                                  * after
1443                                  */
1444                                 error = mem_cgroup_cache_charge(filepage,
1445                                         current->mm, GFP_KERNEL);
1446                                 if (error) {
1447                                         page_cache_release(filepage);
1448                                         spin_lock(&info->lock);
1449                                         shmem_unacct_blocks(info->flags, 1);
1450                                         shmem_free_blocks(inode, 1);
1451                                         spin_unlock(&info->lock);
1452                                         filepage = NULL;
1453                                         goto failed;
1454                                 }
1455
1456                                 spin_lock(&info->lock);
1457                         } else {
1458                                 filepage = prealloc_page;
1459                                 prealloc_page = NULL;
1460                                 SetPageSwapBacked(filepage);
1461                         }
1462
1463                         entry = shmem_swp_alloc(info, idx, sgp);
1464                         if (IS_ERR(entry))
1465                                 error = PTR_ERR(entry);
1466                         else {
1467                                 swap = *entry;
1468                                 shmem_swp_unmap(entry);
1469                         }
1470                         ret = error || swap.val;
1471                         if (ret)
1472                                 mem_cgroup_uncharge_cache_page(filepage);
1473                         else
1474                                 ret = add_to_page_cache_lru(filepage, mapping,
1475                                                 idx, GFP_NOWAIT);
1476                         /*
1477                          * At add_to_page_cache_lru() failure, uncharge will
1478                          * be done automatically.
1479                          */
1480                         if (ret) {
1481                                 shmem_unacct_blocks(info->flags, 1);
1482                                 shmem_free_blocks(inode, 1);
1483                                 spin_unlock(&info->lock);
1484                                 page_cache_release(filepage);
1485                                 filepage = NULL;
1486                                 if (error)
1487                                         goto failed;
1488                                 goto repeat;
1489                         }
1490                         info->flags |= SHMEM_PAGEIN;
1491                 }
1492
1493                 info->alloced++;
1494                 spin_unlock(&info->lock);
1495                 clear_highpage(filepage);
1496                 flush_dcache_page(filepage);
1497                 SetPageUptodate(filepage);
1498                 if (sgp == SGP_DIRTY)
1499                         set_page_dirty(filepage);
1500         }
1501 done:
1502         *pagep = filepage;
1503         error = 0;
1504         goto out;
1505
1506 nospace:
1507         /*
1508          * Perhaps the page was brought in from swap between find_lock_page
1509          * and taking info->lock?  We allow for that at add_to_page_cache_lru,
1510          * but must also avoid reporting a spurious ENOSPC while working on a
1511          * full tmpfs.  (When filepage has been passed in to shmem_getpage, it
1512          * is already in page cache, which prevents this race from occurring.)
1513          */
1514         if (!filepage) {
1515                 struct page *page = find_get_page(mapping, idx);
1516                 if (page) {
1517                         spin_unlock(&info->lock);
1518                         page_cache_release(page);
1519                         goto repeat;
1520                 }
1521         }
1522         spin_unlock(&info->lock);
1523         error = -ENOSPC;
1524 failed:
1525         if (*pagep != filepage) {
1526                 unlock_page(filepage);
1527                 page_cache_release(filepage);
1528         }
1529 out:
1530         if (prealloc_page) {
1531                 mem_cgroup_uncharge_cache_page(prealloc_page);
1532                 page_cache_release(prealloc_page);
1533         }
1534         return error;
1535 }
1536
1537 static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1538 {
1539         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1540         int error;
1541         int ret;
1542
1543         if (((loff_t)vmf->pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
1544                 return VM_FAULT_SIGBUS;
1545
1546         error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_CACHE, &ret);
1547         if (error)
1548                 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1549         if (ret & VM_FAULT_MAJOR) {
1550                 count_vm_event(PGMAJFAULT);
1551                 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1552         }
1553         return ret | VM_FAULT_LOCKED;
1554 }
1555
1556 #ifdef CONFIG_NUMA
1557 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *new)
1558 {
1559         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1560         return mpol_set_shared_policy(&SHMEM_I(i)->policy, vma, new);
1561 }
1562
1563 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
1564                                           unsigned long addr)
1565 {
1566         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1567         unsigned long idx;
1568
1569         idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1570         return mpol_shared_policy_lookup(&SHMEM_I(i)->policy, idx);
1571 }
1572 #endif
1573
1574 int shmem_lock(struct file *file, int lock, struct user_struct *user)
1575 {
1576         struct inode *inode = file->f_path.dentry->d_inode;
1577         struct shmem_inode_info *info = SHMEM_I(inode);
1578         int retval = -ENOMEM;
1579
1580         spin_lock(&info->lock);
1581         if (lock && !(info->flags & VM_LOCKED)) {
1582                 if (!user_shm_lock(inode->i_size, user))
1583                         goto out_nomem;
1584                 info->flags |= VM_LOCKED;
1585                 mapping_set_unevictable(file->f_mapping);
1586         }
1587         if (!lock && (info->flags & VM_LOCKED) && user) {
1588                 user_shm_unlock(inode->i_size, user);
1589                 info->flags &= ~VM_LOCKED;
1590                 mapping_clear_unevictable(file->f_mapping);
1591                 scan_mapping_unevictable_pages(file->f_mapping);
1592         }
1593         retval = 0;
1594
1595 out_nomem:
1596         spin_unlock(&info->lock);
1597         return retval;
1598 }
1599
1600 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1601 {
1602         file_accessed(file);
1603         vma->vm_ops = &shmem_vm_ops;
1604         vma->vm_flags |= VM_CAN_NONLINEAR;
1605         return 0;
1606 }
1607
1608 static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
1609                                      int mode, dev_t dev, unsigned long flags)
1610 {
1611         struct inode *inode;
1612         struct shmem_inode_info *info;
1613         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1614
1615         if (shmem_reserve_inode(sb))
1616                 return NULL;
1617
1618         inode = new_inode(sb);
1619         if (inode) {
1620                 inode->i_ino = get_next_ino();
1621                 inode_init_owner(inode, dir, mode);
1622                 inode->i_blocks = 0;
1623                 inode->i_mapping->backing_dev_info = &shmem_backing_dev_info;
1624                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1625                 inode->i_generation = get_seconds();
1626                 info = SHMEM_I(inode);
1627                 memset(info, 0, (char *)inode - (char *)info);
1628                 spin_lock_init(&info->lock);
1629                 info->flags = flags & VM_NORESERVE;
1630                 INIT_LIST_HEAD(&info->swaplist);
1631                 INIT_LIST_HEAD(&info->xattr_list);
1632                 cache_no_acl(inode);
1633
1634                 switch (mode & S_IFMT) {
1635                 default:
1636                         inode->i_op = &shmem_special_inode_operations;
1637                         init_special_inode(inode, mode, dev);
1638                         break;
1639                 case S_IFREG:
1640                         inode->i_mapping->a_ops = &shmem_aops;
1641                         inode->i_op = &shmem_inode_operations;
1642                         inode->i_fop = &shmem_file_operations;
1643                         mpol_shared_policy_init(&info->policy,
1644                                                  shmem_get_sbmpol(sbinfo));
1645                         break;
1646                 case S_IFDIR:
1647                         inc_nlink(inode);
1648                         /* Some things misbehave if size == 0 on a directory */
1649                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
1650                         inode->i_op = &shmem_dir_inode_operations;
1651                         inode->i_fop = &simple_dir_operations;
1652                         break;
1653                 case S_IFLNK:
1654                         /*
1655                          * Must not load anything in the rbtree,
1656                          * mpol_free_shared_policy will not be called.
1657                          */
1658                         mpol_shared_policy_init(&info->policy, NULL);
1659                         break;
1660                 }
1661         } else
1662                 shmem_free_inode(sb);
1663         return inode;
1664 }
1665
1666 #ifdef CONFIG_TMPFS
1667 static const struct inode_operations shmem_symlink_inode_operations;
1668 static const struct inode_operations shmem_symlink_inline_operations;
1669
1670 /*
1671  * Normally tmpfs avoids the use of shmem_readpage and shmem_write_begin;
1672  * but providing them allows a tmpfs file to be used for splice, sendfile, and
1673  * below the loop driver, in the generic fashion that many filesystems support.
1674  */
1675 static int shmem_readpage(struct file *file, struct page *page)
1676 {
1677         struct inode *inode = page->mapping->host;
1678         int error = shmem_getpage(inode, page->index, &page, SGP_CACHE, NULL);
1679         unlock_page(page);
1680         return error;
1681 }
1682
1683 static int
1684 shmem_write_begin(struct file *file, struct address_space *mapping,
1685                         loff_t pos, unsigned len, unsigned flags,
1686                         struct page **pagep, void **fsdata)
1687 {
1688         struct inode *inode = mapping->host;
1689         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1690         *pagep = NULL;
1691         return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1692 }
1693
1694 static int
1695 shmem_write_end(struct file *file, struct address_space *mapping,
1696                         loff_t pos, unsigned len, unsigned copied,
1697                         struct page *page, void *fsdata)
1698 {
1699         struct inode *inode = mapping->host;
1700
1701         if (pos + copied > inode->i_size)
1702                 i_size_write(inode, pos + copied);
1703
1704         set_page_dirty(page);
1705         unlock_page(page);
1706         page_cache_release(page);
1707
1708         return copied;
1709 }
1710
1711 static void do_shmem_file_read(struct file *filp, loff_t *ppos, read_descriptor_t *desc, read_actor_t actor)
1712 {
1713         struct inode *inode = filp->f_path.dentry->d_inode;
1714         struct address_space *mapping = inode->i_mapping;
1715         unsigned long index, offset;
1716         enum sgp_type sgp = SGP_READ;
1717
1718         /*
1719          * Might this read be for a stacking filesystem?  Then when reading
1720          * holes of a sparse file, we actually need to allocate those pages,
1721          * and even mark them dirty, so it cannot exceed the max_blocks limit.
1722          */
1723         if (segment_eq(get_fs(), KERNEL_DS))
1724                 sgp = SGP_DIRTY;
1725
1726         index = *ppos >> PAGE_CACHE_SHIFT;
1727         offset = *ppos & ~PAGE_CACHE_MASK;
1728
1729         for (;;) {
1730                 struct page *page = NULL;
1731                 unsigned long end_index, nr, ret;
1732                 loff_t i_size = i_size_read(inode);
1733
1734                 end_index = i_size >> PAGE_CACHE_SHIFT;
1735                 if (index > end_index)
1736                         break;
1737                 if (index == end_index) {
1738                         nr = i_size & ~PAGE_CACHE_MASK;
1739                         if (nr <= offset)
1740                                 break;
1741                 }
1742
1743                 desc->error = shmem_getpage(inode, index, &page, sgp, NULL);
1744                 if (desc->error) {
1745                         if (desc->error == -EINVAL)
1746                                 desc->error = 0;
1747                         break;
1748                 }
1749                 if (page)
1750                         unlock_page(page);
1751
1752                 /*
1753                  * We must evaluate after, since reads (unlike writes)
1754                  * are called without i_mutex protection against truncate
1755                  */
1756                 nr = PAGE_CACHE_SIZE;
1757                 i_size = i_size_read(inode);
1758                 end_index = i_size >> PAGE_CACHE_SHIFT;
1759                 if (index == end_index) {
1760                         nr = i_size & ~PAGE_CACHE_MASK;
1761                         if (nr <= offset) {
1762                                 if (page)
1763                                         page_cache_release(page);
1764                                 break;
1765                         }
1766                 }
1767                 nr -= offset;
1768
1769                 if (page) {
1770                         /*
1771                          * If users can be writing to this page using arbitrary
1772                          * virtual addresses, take care about potential aliasing
1773                          * before reading the page on the kernel side.
1774                          */
1775                         if (mapping_writably_mapped(mapping))
1776                                 flush_dcache_page(page);
1777                         /*
1778                          * Mark the page accessed if we read the beginning.
1779                          */
1780                         if (!offset)
1781                                 mark_page_accessed(page);
1782                 } else {
1783                         page = ZERO_PAGE(0);
1784                         page_cache_get(page);
1785                 }
1786
1787                 /*
1788                  * Ok, we have the page, and it's up-to-date, so
1789                  * now we can copy it to user space...
1790                  *
1791                  * The actor routine returns how many bytes were actually used..
1792                  * NOTE! This may not be the same as how much of a user buffer
1793                  * we filled up (we may be padding etc), so we can only update
1794                  * "pos" here (the actor routine has to update the user buffer
1795                  * pointers and the remaining count).
1796                  */
1797                 ret = actor(desc, page, offset, nr);
1798                 offset += ret;
1799                 index += offset >> PAGE_CACHE_SHIFT;
1800                 offset &= ~PAGE_CACHE_MASK;
1801
1802                 page_cache_release(page);
1803                 if (ret != nr || !desc->count)
1804                         break;
1805
1806                 cond_resched();
1807         }
1808
1809         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1810         file_accessed(filp);
1811 }
1812
1813 static ssize_t shmem_file_aio_read(struct kiocb *iocb,
1814                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
1815 {
1816         struct file *filp = iocb->ki_filp;
1817         ssize_t retval;
1818         unsigned long seg;
1819         size_t count;
1820         loff_t *ppos = &iocb->ki_pos;
1821
1822         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1823         if (retval)
1824                 return retval;
1825
1826         for (seg = 0; seg < nr_segs; seg++) {
1827                 read_descriptor_t desc;
1828
1829                 desc.written = 0;
1830                 desc.arg.buf = iov[seg].iov_base;
1831                 desc.count = iov[seg].iov_len;
1832                 if (desc.count == 0)
1833                         continue;
1834                 desc.error = 0;
1835                 do_shmem_file_read(filp, ppos, &desc, file_read_actor);
1836                 retval += desc.written;
1837                 if (desc.error) {
1838                         retval = retval ?: desc.error;
1839                         break;
1840                 }
1841                 if (desc.count > 0)
1842                         break;
1843         }
1844         return retval;
1845 }
1846
1847 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
1848 {
1849         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
1850
1851         buf->f_type = TMPFS_MAGIC;
1852         buf->f_bsize = PAGE_CACHE_SIZE;
1853         buf->f_namelen = NAME_MAX;
1854         if (sbinfo->max_blocks) {
1855                 buf->f_blocks = sbinfo->max_blocks;
1856                 buf->f_bavail = buf->f_bfree =
1857                                 sbinfo->max_blocks - percpu_counter_sum(&sbinfo->used_blocks);
1858         }
1859         if (sbinfo->max_inodes) {
1860                 buf->f_files = sbinfo->max_inodes;
1861                 buf->f_ffree = sbinfo->free_inodes;
1862         }
1863         /* else leave those fields 0 like simple_statfs */
1864         return 0;
1865 }
1866
1867 /*
1868  * File creation. Allocate an inode, and we're done..
1869  */
1870 static int
1871 shmem_mknod(struct inode *dir, struct dentry *dentry, int mode, dev_t dev)
1872 {
1873         struct inode *inode;
1874         int error = -ENOSPC;
1875
1876         inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
1877         if (inode) {
1878                 error = security_inode_init_security(inode, dir,
1879                                                      &dentry->d_name, NULL,
1880                                                      NULL, NULL);
1881                 if (error) {
1882                         if (error != -EOPNOTSUPP) {
1883                                 iput(inode);
1884                                 return error;
1885                         }
1886                 }
1887 #ifdef CONFIG_TMPFS_POSIX_ACL
1888                 error = generic_acl_init(inode, dir);
1889                 if (error) {
1890                         iput(inode);
1891                         return error;
1892                 }
1893 #else
1894                 error = 0;
1895 #endif
1896                 dir->i_size += BOGO_DIRENT_SIZE;
1897                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1898                 d_instantiate(dentry, inode);
1899                 dget(dentry); /* Extra count - pin the dentry in core */
1900         }
1901         return error;
1902 }
1903
1904 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, int mode)
1905 {
1906         int error;
1907
1908         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
1909                 return error;
1910         inc_nlink(dir);
1911         return 0;
1912 }
1913
1914 static int shmem_create(struct inode *dir, struct dentry *dentry, int mode,
1915                 struct nameidata *nd)
1916 {
1917         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
1918 }
1919
1920 /*
1921  * Link a file..
1922  */
1923 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
1924 {
1925         struct inode *inode = old_dentry->d_inode;
1926         int ret;
1927
1928         /*
1929          * No ordinary (disk based) filesystem counts links as inodes;
1930          * but each new link needs a new dentry, pinning lowmem, and
1931          * tmpfs dentries cannot be pruned until they are unlinked.
1932          */
1933         ret = shmem_reserve_inode(inode->i_sb);
1934         if (ret)
1935                 goto out;
1936
1937         dir->i_size += BOGO_DIRENT_SIZE;
1938         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1939         inc_nlink(inode);
1940         ihold(inode);   /* New dentry reference */
1941         dget(dentry);           /* Extra pinning count for the created dentry */
1942         d_instantiate(dentry, inode);
1943 out:
1944         return ret;
1945 }
1946
1947 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
1948 {
1949         struct inode *inode = dentry->d_inode;
1950
1951         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
1952                 shmem_free_inode(inode->i_sb);
1953
1954         dir->i_size -= BOGO_DIRENT_SIZE;
1955         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1956         drop_nlink(inode);
1957         dput(dentry);   /* Undo the count from "create" - this does all the work */
1958         return 0;
1959 }
1960
1961 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
1962 {
1963         if (!simple_empty(dentry))
1964                 return -ENOTEMPTY;
1965
1966         drop_nlink(dentry->d_inode);
1967         drop_nlink(dir);
1968         return shmem_unlink(dir, dentry);
1969 }
1970
1971 /*
1972  * The VFS layer already does all the dentry stuff for rename,
1973  * we just have to decrement the usage count for the target if
1974  * it exists so that the VFS layer correctly free's it when it
1975  * gets overwritten.
1976  */
1977 static int shmem_rename(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
1978 {
1979         struct inode *inode = old_dentry->d_inode;
1980         int they_are_dirs = S_ISDIR(inode->i_mode);
1981
1982         if (!simple_empty(new_dentry))
1983                 return -ENOTEMPTY;
1984
1985         if (new_dentry->d_inode) {
1986                 (void) shmem_unlink(new_dir, new_dentry);
1987                 if (they_are_dirs)
1988                         drop_nlink(old_dir);
1989         } else if (they_are_dirs) {
1990                 drop_nlink(old_dir);
1991                 inc_nlink(new_dir);
1992         }
1993
1994         old_dir->i_size -= BOGO_DIRENT_SIZE;
1995         new_dir->i_size += BOGO_DIRENT_SIZE;
1996         old_dir->i_ctime = old_dir->i_mtime =
1997         new_dir->i_ctime = new_dir->i_mtime =
1998         inode->i_ctime = CURRENT_TIME;
1999         return 0;
2000 }
2001
2002 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
2003 {
2004         int error;
2005         int len;
2006         struct inode *inode;
2007         struct page *page = NULL;
2008         char *kaddr;
2009         struct shmem_inode_info *info;
2010
2011         len = strlen(symname) + 1;
2012         if (len > PAGE_CACHE_SIZE)
2013                 return -ENAMETOOLONG;
2014
2015         inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0, VM_NORESERVE);
2016         if (!inode)
2017                 return -ENOSPC;
2018
2019         error = security_inode_init_security(inode, dir, &dentry->d_name, NULL,
2020                                              NULL, NULL);
2021         if (error) {
2022                 if (error != -EOPNOTSUPP) {
2023                         iput(inode);
2024                         return error;
2025                 }
2026                 error = 0;
2027         }
2028
2029         info = SHMEM_I(inode);
2030         inode->i_size = len-1;
2031         if (len <= SHMEM_SYMLINK_INLINE_LEN) {
2032                 /* do it inline */
2033                 memcpy(info->inline_symlink, symname, len);
2034                 inode->i_op = &shmem_symlink_inline_operations;
2035         } else {
2036                 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
2037                 if (error) {
2038                         iput(inode);
2039                         return error;
2040                 }
2041                 inode->i_mapping->a_ops = &shmem_aops;
2042                 inode->i_op = &shmem_symlink_inode_operations;
2043                 kaddr = kmap_atomic(page, KM_USER0);
2044                 memcpy(kaddr, symname, len);
2045                 kunmap_atomic(kaddr, KM_USER0);
2046                 set_page_dirty(page);
2047                 unlock_page(page);
2048                 page_cache_release(page);
2049         }
2050         dir->i_size += BOGO_DIRENT_SIZE;
2051         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2052         d_instantiate(dentry, inode);
2053         dget(dentry);
2054         return 0;
2055 }
2056
2057 static void *shmem_follow_link_inline(struct dentry *dentry, struct nameidata *nd)
2058 {
2059         nd_set_link(nd, SHMEM_I(dentry->d_inode)->inline_symlink);
2060         return NULL;
2061 }
2062
2063 static void *shmem_follow_link(struct dentry *dentry, struct nameidata *nd)
2064 {
2065         struct page *page = NULL;
2066         int res = shmem_getpage(dentry->d_inode, 0, &page, SGP_READ, NULL);
2067         nd_set_link(nd, res ? ERR_PTR(res) : kmap(page));
2068         if (page)
2069                 unlock_page(page);
2070         return page;
2071 }
2072
2073 static void shmem_put_link(struct dentry *dentry, struct nameidata *nd, void *cookie)
2074 {
2075         if (!IS_ERR(nd_get_link(nd))) {
2076                 struct page *page = cookie;
2077                 kunmap(page);
2078                 mark_page_accessed(page);
2079                 page_cache_release(page);
2080         }
2081 }
2082
2083 #ifdef CONFIG_TMPFS_XATTR
2084 /*
2085  * Superblocks without xattr inode operations may get some security.* xattr
2086  * support from the LSM "for free". As soon as we have any other xattrs
2087  * like ACLs, we also need to implement the security.* handlers at
2088  * filesystem level, though.
2089  */
2090
2091 static int shmem_xattr_get(struct dentry *dentry, const char *name,
2092                            void *buffer, size_t size)
2093 {
2094         struct shmem_inode_info *info;
2095         struct shmem_xattr *xattr;
2096         int ret = -ENODATA;
2097
2098         info = SHMEM_I(dentry->d_inode);
2099
2100         spin_lock(&info->lock);
2101         list_for_each_entry(xattr, &info->xattr_list, list) {
2102                 if (strcmp(name, xattr->name))
2103                         continue;
2104
2105                 ret = xattr->size;
2106                 if (buffer) {
2107                         if (size < xattr->size)
2108                                 ret = -ERANGE;
2109                         else
2110                                 memcpy(buffer, xattr->value, xattr->size);
2111                 }
2112                 break;
2113         }
2114         spin_unlock(&info->lock);
2115         return ret;
2116 }
2117
2118 static int shmem_xattr_set(struct dentry *dentry, const char *name,
2119                            const void *value, size_t size, int flags)
2120 {
2121         struct inode *inode = dentry->d_inode;
2122         struct shmem_inode_info *info = SHMEM_I(inode);
2123         struct shmem_xattr *xattr;
2124         struct shmem_xattr *new_xattr = NULL;
2125         size_t len;
2126         int err = 0;
2127
2128         /* value == NULL means remove */
2129         if (value) {
2130                 /* wrap around? */
2131                 len = sizeof(*new_xattr) + size;
2132                 if (len <= sizeof(*new_xattr))
2133                         return -ENOMEM;
2134
2135                 new_xattr = kmalloc(len, GFP_KERNEL);
2136                 if (!new_xattr)
2137                         return -ENOMEM;
2138
2139                 new_xattr->name = kstrdup(name, GFP_KERNEL);
2140                 if (!new_xattr->name) {
2141                         kfree(new_xattr);
2142                         return -ENOMEM;
2143                 }
2144
2145                 new_xattr->size = size;
2146                 memcpy(new_xattr->value, value, size);
2147         }
2148
2149         spin_lock(&info->lock);
2150         list_for_each_entry(xattr, &info->xattr_list, list) {
2151                 if (!strcmp(name, xattr->name)) {
2152                         if (flags & XATTR_CREATE) {
2153                                 xattr = new_xattr;
2154                                 err = -EEXIST;
2155                         } else if (new_xattr) {
2156                                 list_replace(&xattr->list, &new_xattr->list);
2157                         } else {
2158                                 list_del(&xattr->list);
2159                         }
2160                         goto out;
2161                 }
2162         }
2163         if (flags & XATTR_REPLACE) {
2164                 xattr = new_xattr;
2165                 err = -ENODATA;
2166         } else {
2167                 list_add(&new_xattr->list, &info->xattr_list);
2168                 xattr = NULL;
2169         }
2170 out:
2171         spin_unlock(&info->lock);
2172         if (xattr)
2173                 kfree(xattr->name);
2174         kfree(xattr);
2175         return err;
2176 }
2177
2178
2179 static const struct xattr_handler *shmem_xattr_handlers[] = {
2180 #ifdef CONFIG_TMPFS_POSIX_ACL
2181         &generic_acl_access_handler,
2182         &generic_acl_default_handler,
2183 #endif
2184         NULL
2185 };
2186
2187 static int shmem_xattr_validate(const char *name)
2188 {
2189         struct { const char *prefix; size_t len; } arr[] = {
2190                 { XATTR_SECURITY_PREFIX, XATTR_SECURITY_PREFIX_LEN },
2191                 { XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN }
2192         };
2193         int i;
2194
2195         for (i = 0; i < ARRAY_SIZE(arr); i++) {
2196                 size_t preflen = arr[i].len;
2197                 if (strncmp(name, arr[i].prefix, preflen) == 0) {
2198                         if (!name[preflen])
2199                                 return -EINVAL;
2200                         return 0;
2201                 }
2202         }
2203         return -EOPNOTSUPP;
2204 }
2205
2206 static ssize_t shmem_getxattr(struct dentry *dentry, const char *name,
2207                               void *buffer, size_t size)
2208 {
2209         int err;
2210
2211         /*
2212          * If this is a request for a synthetic attribute in the system.*
2213          * namespace use the generic infrastructure to resolve a handler
2214          * for it via sb->s_xattr.
2215          */
2216         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2217                 return generic_getxattr(dentry, name, buffer, size);
2218
2219         err = shmem_xattr_validate(name);
2220         if (err)
2221                 return err;
2222
2223         return shmem_xattr_get(dentry, name, buffer, size);
2224 }
2225
2226 static int shmem_setxattr(struct dentry *dentry, const char *name,
2227                           const void *value, size_t size, int flags)
2228 {
2229         int err;
2230
2231         /*
2232          * If this is a request for a synthetic attribute in the system.*
2233          * namespace use the generic infrastructure to resolve a handler
2234          * for it via sb->s_xattr.
2235          */
2236         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2237                 return generic_setxattr(dentry, name, value, size, flags);
2238
2239         err = shmem_xattr_validate(name);
2240         if (err)
2241                 return err;
2242
2243         if (size == 0)
2244                 value = "";  /* empty EA, do not remove */
2245
2246         return shmem_xattr_set(dentry, name, value, size, flags);
2247
2248 }
2249
2250 static int shmem_removexattr(struct dentry *dentry, const char *name)
2251 {
2252         int err;
2253
2254         /*
2255          * If this is a request for a synthetic attribute in the system.*
2256          * namespace use the generic infrastructure to resolve a handler
2257          * for it via sb->s_xattr.
2258          */
2259         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2260                 return generic_removexattr(dentry, name);
2261
2262         err = shmem_xattr_validate(name);
2263         if (err)
2264                 return err;
2265
2266         return shmem_xattr_set(dentry, name, NULL, 0, XATTR_REPLACE);
2267 }
2268
2269 static bool xattr_is_trusted(const char *name)
2270 {
2271         return !strncmp(name, XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN);
2272 }
2273
2274 static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
2275 {
2276         bool trusted = capable(CAP_SYS_ADMIN);
2277         struct shmem_xattr *xattr;
2278         struct shmem_inode_info *info;
2279         size_t used = 0;
2280
2281         info = SHMEM_I(dentry->d_inode);
2282
2283         spin_lock(&info->lock);
2284         list_for_each_entry(xattr, &info->xattr_list, list) {
2285                 size_t len;
2286
2287                 /* skip "trusted." attributes for unprivileged callers */
2288                 if (!trusted && xattr_is_trusted(xattr->name))
2289                         continue;
2290
2291                 len = strlen(xattr->name) + 1;
2292                 used += len;
2293                 if (buffer) {
2294                         if (size < used) {
2295                                 used = -ERANGE;
2296                                 break;
2297                         }
2298                         memcpy(buffer, xattr->name, len);
2299                         buffer += len;
2300                 }
2301         }
2302         spin_unlock(&info->lock);
2303
2304         return used;
2305 }
2306 #endif /* CONFIG_TMPFS_XATTR */
2307
2308 static const struct inode_operations shmem_symlink_inline_operations = {
2309         .readlink       = generic_readlink,
2310         .follow_link    = shmem_follow_link_inline,
2311 #ifdef CONFIG_TMPFS_XATTR
2312         .setxattr       = shmem_setxattr,
2313         .getxattr       = shmem_getxattr,
2314         .listxattr      = shmem_listxattr,
2315         .removexattr    = shmem_removexattr,
2316 #endif
2317 };
2318
2319 static const struct inode_operations shmem_symlink_inode_operations = {
2320         .readlink       = generic_readlink,
2321         .follow_link    = shmem_follow_link,
2322         .put_link       = shmem_put_link,
2323 #ifdef CONFIG_TMPFS_XATTR
2324         .setxattr       = shmem_setxattr,
2325         .getxattr       = shmem_getxattr,
2326         .listxattr      = shmem_listxattr,
2327         .removexattr    = shmem_removexattr,
2328 #endif
2329 };
2330
2331 static struct dentry *shmem_get_parent(struct dentry *child)
2332 {
2333         return ERR_PTR(-ESTALE);
2334 }
2335
2336 static int shmem_match(struct inode *ino, void *vfh)
2337 {
2338         __u32 *fh = vfh;
2339         __u64 inum = fh[2];
2340         inum = (inum << 32) | fh[1];
2341         return ino->i_ino == inum && fh[0] == ino->i_generation;
2342 }
2343
2344 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
2345                 struct fid *fid, int fh_len, int fh_type)
2346 {
2347         struct inode *inode;
2348         struct dentry *dentry = NULL;
2349         u64 inum = fid->raw[2];
2350         inum = (inum << 32) | fid->raw[1];
2351
2352         if (fh_len < 3)
2353                 return NULL;
2354
2355         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
2356                         shmem_match, fid->raw);
2357         if (inode) {
2358                 dentry = d_find_alias(inode);
2359                 iput(inode);
2360         }
2361
2362         return dentry;
2363 }
2364
2365 static int shmem_encode_fh(struct dentry *dentry, __u32 *fh, int *len,
2366                                 int connectable)
2367 {
2368         struct inode *inode = dentry->d_inode;
2369
2370         if (*len < 3) {
2371                 *len = 3;
2372                 return 255;
2373         }
2374
2375         if (inode_unhashed(inode)) {
2376                 /* Unfortunately insert_inode_hash is not idempotent,
2377                  * so as we hash inodes here rather than at creation
2378                  * time, we need a lock to ensure we only try
2379                  * to do it once
2380                  */
2381                 static DEFINE_SPINLOCK(lock);
2382                 spin_lock(&lock);
2383                 if (inode_unhashed(inode))
2384                         __insert_inode_hash(inode,
2385                                             inode->i_ino + inode->i_generation);
2386                 spin_unlock(&lock);
2387         }
2388
2389         fh[0] = inode->i_generation;
2390         fh[1] = inode->i_ino;
2391         fh[2] = ((__u64)inode->i_ino) >> 32;
2392
2393         *len = 3;
2394         return 1;
2395 }
2396
2397 static const struct export_operations shmem_export_ops = {
2398         .get_parent     = shmem_get_parent,
2399         .encode_fh      = shmem_encode_fh,
2400         .fh_to_dentry   = shmem_fh_to_dentry,
2401 };
2402
2403 static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
2404                                bool remount)
2405 {
2406         char *this_char, *value, *rest;
2407
2408         while (options != NULL) {
2409                 this_char = options;
2410                 for (;;) {
2411                         /*
2412                          * NUL-terminate this option: unfortunately,
2413                          * mount options form a comma-separated list,
2414                          * but mpol's nodelist may also contain commas.
2415                          */
2416                         options = strchr(options, ',');
2417                         if (options == NULL)
2418                                 break;
2419                         options++;
2420                         if (!isdigit(*options)) {
2421                                 options[-1] = '\0';
2422                                 break;
2423                         }
2424                 }
2425                 if (!*this_char)
2426                         continue;
2427                 if ((value = strchr(this_char,'=')) != NULL) {
2428                         *value++ = 0;
2429                 } else {
2430                         printk(KERN_ERR
2431                             "tmpfs: No value for mount option '%s'\n",
2432                             this_char);
2433                         return 1;
2434                 }
2435
2436                 if (!strcmp(this_char,"size")) {
2437                         unsigned long long size;
2438                         size = memparse(value,&rest);
2439                         if (*rest == '%') {
2440                                 size <<= PAGE_SHIFT;
2441                                 size *= totalram_pages;
2442                                 do_div(size, 100);
2443                                 rest++;
2444                         }
2445                         if (*rest)
2446                                 goto bad_val;
2447                         sbinfo->max_blocks =
2448                                 DIV_ROUND_UP(size, PAGE_CACHE_SIZE);
2449                 } else if (!strcmp(this_char,"nr_blocks")) {
2450                         sbinfo->max_blocks = memparse(value, &rest);
2451                         if (*rest)
2452                                 goto bad_val;
2453                 } else if (!strcmp(this_char,"nr_inodes")) {
2454                         sbinfo->max_inodes = memparse(value, &rest);
2455                         if (*rest)
2456                                 goto bad_val;
2457                 } else if (!strcmp(this_char,"mode")) {
2458                         if (remount)
2459                                 continue;
2460                         sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
2461                         if (*rest)
2462                                 goto bad_val;
2463                 } else if (!strcmp(this_char,"uid")) {
2464                         if (remount)
2465                                 continue;
2466                         sbinfo->uid = simple_strtoul(value, &rest, 0);
2467                         if (*rest)
2468                                 goto bad_val;
2469                 } else if (!strcmp(this_char,"gid")) {
2470                         if (remount)
2471                                 continue;
2472                         sbinfo->gid = simple_strtoul(value, &rest, 0);
2473                         if (*rest)
2474                                 goto bad_val;
2475                 } else if (!strcmp(this_char,"mpol")) {
2476                         if (mpol_parse_str(value, &sbinfo->mpol, 1))
2477                                 goto bad_val;
2478                 } else {
2479                         printk(KERN_ERR "tmpfs: Bad mount option %s\n",
2480                                this_char);
2481                         return 1;
2482                 }
2483         }
2484         return 0;
2485
2486 bad_val:
2487         printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
2488                value, this_char);
2489         return 1;
2490
2491 }
2492
2493 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2494 {
2495         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2496         struct shmem_sb_info config = *sbinfo;
2497         unsigned long inodes;
2498         int error = -EINVAL;
2499
2500         if (shmem_parse_options(data, &config, true))
2501                 return error;
2502
2503         spin_lock(&sbinfo->stat_lock);
2504         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2505         if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
2506                 goto out;
2507         if (config.max_inodes < inodes)
2508                 goto out;
2509         /*
2510          * Those tests also disallow limited->unlimited while any are in
2511          * use, so i_blocks will always be zero when max_blocks is zero;
2512          * but we must separately disallow unlimited->limited, because
2513          * in that case we have no record of how much is already in use.
2514          */
2515         if (config.max_blocks && !sbinfo->max_blocks)
2516                 goto out;
2517         if (config.max_inodes && !sbinfo->max_inodes)
2518                 goto out;
2519
2520         error = 0;
2521         sbinfo->max_blocks  = config.max_blocks;
2522         sbinfo->max_inodes  = config.max_inodes;
2523         sbinfo->free_inodes = config.max_inodes - inodes;
2524
2525         mpol_put(sbinfo->mpol);
2526         sbinfo->mpol        = config.mpol;      /* transfers initial ref */
2527 out:
2528         spin_unlock(&sbinfo->stat_lock);
2529         return error;
2530 }
2531
2532 static int shmem_show_options(struct seq_file *seq, struct vfsmount *vfs)
2533 {
2534         struct shmem_sb_info *sbinfo = SHMEM_SB(vfs->mnt_sb);
2535
2536         if (sbinfo->max_blocks != shmem_default_max_blocks())
2537                 seq_printf(seq, ",size=%luk",
2538                         sbinfo->max_blocks << (PAGE_CACHE_SHIFT - 10));
2539         if (sbinfo->max_inodes != shmem_default_max_inodes())
2540                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
2541         if (sbinfo->mode != (S_IRWXUGO | S_ISVTX))
2542                 seq_printf(seq, ",mode=%03o", sbinfo->mode);
2543         if (sbinfo->uid != 0)
2544                 seq_printf(seq, ",uid=%u", sbinfo->uid);
2545         if (sbinfo->gid != 0)
2546                 seq_printf(seq, ",gid=%u", sbinfo->gid);
2547         shmem_show_mpol(seq, sbinfo->mpol);
2548         return 0;
2549 }
2550 #endif /* CONFIG_TMPFS */
2551
2552 static void shmem_put_super(struct super_block *sb)
2553 {
2554         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2555
2556         percpu_counter_destroy(&sbinfo->used_blocks);
2557         kfree(sbinfo);
2558         sb->s_fs_info = NULL;
2559 }
2560
2561 int shmem_fill_super(struct super_block *sb, void *data, int silent)
2562 {
2563         struct inode *inode;
2564         struct dentry *root;
2565         struct shmem_sb_info *sbinfo;
2566         int err = -ENOMEM;
2567
2568         /* Round up to L1_CACHE_BYTES to resist false sharing */
2569         sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
2570                                 L1_CACHE_BYTES), GFP_KERNEL);
2571         if (!sbinfo)
2572                 return -ENOMEM;
2573
2574         sbinfo->mode = S_IRWXUGO | S_ISVTX;
2575         sbinfo->uid = current_fsuid();
2576         sbinfo->gid = current_fsgid();
2577         sb->s_fs_info = sbinfo;
2578
2579 #ifdef CONFIG_TMPFS
2580         /*
2581          * Per default we only allow half of the physical ram per
2582          * tmpfs instance, limiting inodes to one per page of lowmem;
2583          * but the internal instance is left unlimited.
2584          */
2585         if (!(sb->s_flags & MS_NOUSER)) {
2586                 sbinfo->max_blocks = shmem_default_max_blocks();
2587                 sbinfo->max_inodes = shmem_default_max_inodes();
2588                 if (shmem_parse_options(data, sbinfo, false)) {
2589                         err = -EINVAL;
2590                         goto failed;
2591                 }
2592         }
2593         sb->s_export_op = &shmem_export_ops;
2594 #else
2595         sb->s_flags |= MS_NOUSER;
2596 #endif
2597
2598         spin_lock_init(&sbinfo->stat_lock);
2599         if (percpu_counter_init(&sbinfo->used_blocks, 0))
2600                 goto failed;
2601         sbinfo->free_inodes = sbinfo->max_inodes;
2602
2603         sb->s_maxbytes = SHMEM_MAX_BYTES;
2604         sb->s_blocksize = PAGE_CACHE_SIZE;
2605         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
2606         sb->s_magic = TMPFS_MAGIC;
2607         sb->s_op = &shmem_ops;
2608         sb->s_time_gran = 1;
2609 #ifdef CONFIG_TMPFS_XATTR
2610         sb->s_xattr = shmem_xattr_handlers;
2611 #endif
2612 #ifdef CONFIG_TMPFS_POSIX_ACL
2613         sb->s_flags |= MS_POSIXACL;
2614 #endif
2615
2616         inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
2617         if (!inode)
2618                 goto failed;
2619         inode->i_uid = sbinfo->uid;
2620         inode->i_gid = sbinfo->gid;
2621         root = d_alloc_root(inode);
2622         if (!root)
2623                 goto failed_iput;
2624         sb->s_root = root;
2625         return 0;
2626
2627 failed_iput:
2628         iput(inode);
2629 failed:
2630         shmem_put_super(sb);
2631         return err;
2632 }
2633
2634 static struct kmem_cache *shmem_inode_cachep;
2635
2636 static struct inode *shmem_alloc_inode(struct super_block *sb)
2637 {
2638         struct shmem_inode_info *p;
2639         p = (struct shmem_inode_info *)kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
2640         if (!p)
2641                 return NULL;
2642         return &p->vfs_inode;
2643 }
2644
2645 static void shmem_i_callback(struct rcu_head *head)
2646 {
2647         struct inode *inode = container_of(head, struct inode, i_rcu);
2648         INIT_LIST_HEAD(&inode->i_dentry);
2649         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
2650 }
2651
2652 static void shmem_destroy_inode(struct inode *inode)
2653 {
2654         if ((inode->i_mode & S_IFMT) == S_IFREG) {
2655                 /* only struct inode is valid if it's an inline symlink */
2656                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
2657         }
2658         call_rcu(&inode->i_rcu, shmem_i_callback);
2659 }
2660
2661 static void init_once(void *foo)
2662 {
2663         struct shmem_inode_info *p = (struct shmem_inode_info *) foo;
2664
2665         inode_init_once(&p->vfs_inode);
2666 }
2667
2668 static int init_inodecache(void)
2669 {
2670         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
2671                                 sizeof(struct shmem_inode_info),
2672                                 0, SLAB_PANIC, init_once);
2673         return 0;
2674 }
2675
2676 static void destroy_inodecache(void)
2677 {
2678         kmem_cache_destroy(shmem_inode_cachep);
2679 }
2680
2681 static const struct address_space_operations shmem_aops = {
2682         .writepage      = shmem_writepage,
2683         .set_page_dirty = __set_page_dirty_no_writeback,
2684 #ifdef CONFIG_TMPFS
2685         .readpage       = shmem_readpage,
2686         .write_begin    = shmem_write_begin,
2687         .write_end      = shmem_write_end,
2688 #endif
2689         .migratepage    = migrate_page,
2690         .error_remove_page = generic_error_remove_page,
2691 };
2692
2693 static const struct file_operations shmem_file_operations = {
2694         .mmap           = shmem_mmap,
2695 #ifdef CONFIG_TMPFS
2696         .llseek         = generic_file_llseek,
2697         .read           = do_sync_read,
2698         .write          = do_sync_write,
2699         .aio_read       = shmem_file_aio_read,
2700         .aio_write      = generic_file_aio_write,
2701         .fsync          = noop_fsync,
2702         .splice_read    = generic_file_splice_read,
2703         .splice_write   = generic_file_splice_write,
2704 #endif
2705 };
2706
2707 static const struct inode_operations shmem_inode_operations = {
2708         .setattr        = shmem_setattr,
2709         .truncate_range = shmem_truncate_range,
2710 #ifdef CONFIG_TMPFS_XATTR
2711         .setxattr       = shmem_setxattr,
2712         .getxattr       = shmem_getxattr,
2713         .listxattr      = shmem_listxattr,
2714         .removexattr    = shmem_removexattr,
2715 #endif
2716 #ifdef CONFIG_TMPFS_POSIX_ACL
2717         .check_acl      = generic_check_acl,
2718 #endif
2719
2720 };
2721
2722 static const struct inode_operations shmem_dir_inode_operations = {
2723 #ifdef CONFIG_TMPFS
2724         .create         = shmem_create,
2725         .lookup         = simple_lookup,
2726         .link           = shmem_link,
2727         .unlink         = shmem_unlink,
2728         .symlink        = shmem_symlink,
2729         .mkdir          = shmem_mkdir,
2730         .rmdir          = shmem_rmdir,
2731         .mknod          = shmem_mknod,
2732         .rename         = shmem_rename,
2733 #endif
2734 #ifdef CONFIG_TMPFS_XATTR
2735         .setxattr       = shmem_setxattr,
2736         .getxattr       = shmem_getxattr,
2737         .listxattr      = shmem_listxattr,
2738         .removexattr    = shmem_removexattr,
2739 #endif
2740 #ifdef CONFIG_TMPFS_POSIX_ACL
2741         .setattr        = shmem_setattr,
2742         .check_acl      = generic_check_acl,
2743 #endif
2744 };
2745
2746 static const struct inode_operations shmem_special_inode_operations = {
2747 #ifdef CONFIG_TMPFS_XATTR
2748         .setxattr       = shmem_setxattr,
2749         .getxattr       = shmem_getxattr,
2750         .listxattr      = shmem_listxattr,
2751         .removexattr    = shmem_removexattr,
2752 #endif
2753 #ifdef CONFIG_TMPFS_POSIX_ACL
2754         .setattr        = shmem_setattr,
2755         .check_acl      = generic_check_acl,
2756 #endif
2757 };
2758
2759 static const struct super_operations shmem_ops = {
2760         .alloc_inode    = shmem_alloc_inode,
2761         .destroy_inode  = shmem_destroy_inode,
2762 #ifdef CONFIG_TMPFS
2763         .statfs         = shmem_statfs,
2764         .remount_fs     = shmem_remount_fs,
2765         .show_options   = shmem_show_options,
2766 #endif
2767         .evict_inode    = shmem_evict_inode,
2768         .drop_inode     = generic_delete_inode,
2769         .put_super      = shmem_put_super,
2770 };
2771
2772 static const struct vm_operations_struct shmem_vm_ops = {
2773         .fault          = shmem_fault,
2774 #ifdef CONFIG_NUMA
2775         .set_policy     = shmem_set_policy,
2776         .get_policy     = shmem_get_policy,
2777 #endif
2778 };
2779
2780
2781 static struct dentry *shmem_mount(struct file_system_type *fs_type,
2782         int flags, const char *dev_name, void *data)
2783 {
2784         return mount_nodev(fs_type, flags, data, shmem_fill_super);
2785 }
2786
2787 static struct file_system_type tmpfs_fs_type = {
2788         .owner          = THIS_MODULE,
2789         .name           = "tmpfs",
2790         .mount          = shmem_mount,
2791         .kill_sb        = kill_litter_super,
2792 };
2793
2794 int __init init_tmpfs(void)
2795 {
2796         int error;
2797
2798         error = bdi_init(&shmem_backing_dev_info);
2799         if (error)
2800                 goto out4;
2801
2802         error = init_inodecache();
2803         if (error)
2804                 goto out3;
2805
2806         error = register_filesystem(&tmpfs_fs_type);
2807         if (error) {
2808                 printk(KERN_ERR "Could not register tmpfs\n");
2809                 goto out2;
2810         }
2811
2812         shm_mnt = vfs_kern_mount(&tmpfs_fs_type, MS_NOUSER,
2813                                 tmpfs_fs_type.name, NULL);
2814         if (IS_ERR(shm_mnt)) {
2815                 error = PTR_ERR(shm_mnt);
2816                 printk(KERN_ERR "Could not kern_mount tmpfs\n");
2817                 goto out1;
2818         }
2819         return 0;
2820
2821 out1:
2822         unregister_filesystem(&tmpfs_fs_type);
2823 out2:
2824         destroy_inodecache();
2825 out3:
2826         bdi_destroy(&shmem_backing_dev_info);
2827 out4:
2828         shm_mnt = ERR_PTR(error);
2829         return error;
2830 }
2831
2832 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2833 /**
2834  * mem_cgroup_get_shmem_target - find a page or entry assigned to the shmem file
2835  * @inode: the inode to be searched
2836  * @pgoff: the offset to be searched
2837  * @pagep: the pointer for the found page to be stored
2838  * @ent: the pointer for the found swap entry to be stored
2839  *
2840  * If a page is found, refcount of it is incremented. Callers should handle
2841  * these refcount.
2842  */
2843 void mem_cgroup_get_shmem_target(struct inode *inode, pgoff_t pgoff,
2844                                         struct page **pagep, swp_entry_t *ent)
2845 {
2846         swp_entry_t entry = { .val = 0 }, *ptr;
2847         struct page *page = NULL;
2848         struct shmem_inode_info *info = SHMEM_I(inode);
2849
2850         if ((pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
2851                 goto out;
2852
2853         spin_lock(&info->lock);
2854         ptr = shmem_swp_entry(info, pgoff, NULL);
2855 #ifdef CONFIG_SWAP
2856         if (ptr && ptr->val) {
2857                 entry.val = ptr->val;
2858                 page = find_get_page(&swapper_space, entry.val);
2859         } else
2860 #endif
2861                 page = find_get_page(inode->i_mapping, pgoff);
2862         if (ptr)
2863                 shmem_swp_unmap(ptr);
2864         spin_unlock(&info->lock);
2865 out:
2866         *pagep = page;
2867         *ent = entry;
2868 }
2869 #endif
2870
2871 #else /* !CONFIG_SHMEM */
2872
2873 /*
2874  * tiny-shmem: simple shmemfs and tmpfs using ramfs code
2875  *
2876  * This is intended for small system where the benefits of the full
2877  * shmem code (swap-backed and resource-limited) are outweighed by
2878  * their complexity. On systems without swap this code should be
2879  * effectively equivalent, but much lighter weight.
2880  */
2881
2882 #include <linux/ramfs.h>
2883
2884 static struct file_system_type tmpfs_fs_type = {
2885         .name           = "tmpfs",
2886         .mount          = ramfs_mount,
2887         .kill_sb        = kill_litter_super,
2888 };
2889
2890 int __init init_tmpfs(void)
2891 {
2892         BUG_ON(register_filesystem(&tmpfs_fs_type) != 0);
2893
2894         shm_mnt = kern_mount(&tmpfs_fs_type);
2895         BUG_ON(IS_ERR(shm_mnt));
2896
2897         return 0;
2898 }
2899
2900 int shmem_unuse(swp_entry_t entry, struct page *page)
2901 {
2902         return 0;
2903 }
2904
2905 int shmem_lock(struct file *file, int lock, struct user_struct *user)
2906 {
2907         return 0;
2908 }
2909
2910 void shmem_truncate_range(struct inode *inode, loff_t start, loff_t end)
2911 {
2912         truncate_inode_pages_range(inode->i_mapping, start, end);
2913 }
2914 EXPORT_SYMBOL_GPL(shmem_truncate_range);
2915
2916 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2917 /**
2918  * mem_cgroup_get_shmem_target - find a page or entry assigned to the shmem file
2919  * @inode: the inode to be searched
2920  * @pgoff: the offset to be searched
2921  * @pagep: the pointer for the found page to be stored
2922  * @ent: the pointer for the found swap entry to be stored
2923  *
2924  * If a page is found, refcount of it is incremented. Callers should handle
2925  * these refcount.
2926  */
2927 void mem_cgroup_get_shmem_target(struct inode *inode, pgoff_t pgoff,
2928                                         struct page **pagep, swp_entry_t *ent)
2929 {
2930         struct page *page = NULL;
2931
2932         if ((pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
2933                 goto out;
2934         page = find_get_page(inode->i_mapping, pgoff);
2935 out:
2936         *pagep = page;
2937         *ent = (swp_entry_t){ .val = 0 };
2938 }
2939 #endif
2940
2941 #define shmem_vm_ops                            generic_file_vm_ops
2942 #define shmem_file_operations                   ramfs_file_operations
2943 #define shmem_get_inode(sb, dir, mode, dev, flags)      ramfs_get_inode(sb, dir, mode, dev)
2944 #define shmem_acct_size(flags, size)            0
2945 #define shmem_unacct_size(flags, size)          do {} while (0)
2946 #define SHMEM_MAX_BYTES                         MAX_LFS_FILESIZE
2947
2948 #endif /* CONFIG_SHMEM */
2949
2950 /* common code */
2951
2952 /**
2953  * shmem_file_setup - get an unlinked file living in tmpfs
2954  * @name: name for dentry (to be seen in /proc/<pid>/maps
2955  * @size: size to be set for the file
2956  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
2957  */
2958 struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
2959 {
2960         int error;
2961         struct file *file;
2962         struct inode *inode;
2963         struct path path;
2964         struct dentry *root;
2965         struct qstr this;
2966
2967         if (IS_ERR(shm_mnt))
2968                 return (void *)shm_mnt;
2969
2970         if (size < 0 || size > SHMEM_MAX_BYTES)
2971                 return ERR_PTR(-EINVAL);
2972
2973         if (shmem_acct_size(flags, size))
2974                 return ERR_PTR(-ENOMEM);
2975
2976         error = -ENOMEM;
2977         this.name = name;
2978         this.len = strlen(name);
2979         this.hash = 0; /* will go */
2980         root = shm_mnt->mnt_root;
2981         path.dentry = d_alloc(root, &this);
2982         if (!path.dentry)
2983                 goto put_memory;
2984         path.mnt = mntget(shm_mnt);
2985
2986         error = -ENOSPC;
2987         inode = shmem_get_inode(root->d_sb, NULL, S_IFREG | S_IRWXUGO, 0, flags);
2988         if (!inode)
2989                 goto put_dentry;
2990
2991         d_instantiate(path.dentry, inode);
2992         inode->i_size = size;
2993         inode->i_nlink = 0;     /* It is unlinked */
2994 #ifndef CONFIG_MMU
2995         error = ramfs_nommu_expand_for_mapping(inode, size);
2996         if (error)
2997                 goto put_dentry;
2998 #endif
2999
3000         error = -ENFILE;
3001         file = alloc_file(&path, FMODE_WRITE | FMODE_READ,
3002                   &shmem_file_operations);
3003         if (!file)
3004                 goto put_dentry;
3005
3006         return file;
3007
3008 put_dentry:
3009         path_put(&path);
3010 put_memory:
3011         shmem_unacct_size(flags, size);
3012         return ERR_PTR(error);
3013 }
3014 EXPORT_SYMBOL_GPL(shmem_file_setup);
3015
3016 /**
3017  * shmem_zero_setup - setup a shared anonymous mapping
3018  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
3019  */
3020 int shmem_zero_setup(struct vm_area_struct *vma)
3021 {
3022         struct file *file;
3023         loff_t size = vma->vm_end - vma->vm_start;
3024
3025         file = shmem_file_setup("dev/zero", size, vma->vm_flags);
3026         if (IS_ERR(file))
3027                 return PTR_ERR(file);
3028
3029         if (vma->vm_file)
3030                 fput(vma->vm_file);
3031         vma->vm_file = file;
3032         vma->vm_ops = &shmem_vm_ops;
3033         vma->vm_flags |= VM_CAN_NONLINEAR;
3034         return 0;
3035 }
3036
3037 /**
3038  * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
3039  * @mapping:    the page's address_space
3040  * @index:      the page index
3041  * @gfp:        the page allocator flags to use if allocating
3042  *
3043  * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
3044  * with any new page allocations done using the specified allocation flags.
3045  * But read_cache_page_gfp() uses the ->readpage() method: which does not
3046  * suit tmpfs, since it may have pages in swapcache, and needs to find those
3047  * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
3048  *
3049  * Provide a stub for those callers to start using now, then later
3050  * flesh it out to call shmem_getpage() with additional gfp mask, when
3051  * shmem_file_splice_read() is added and shmem_readpage() is removed.
3052  */
3053 struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
3054                                          pgoff_t index, gfp_t gfp)
3055 {
3056         return read_cache_page_gfp(mapping, index, gfp);
3057 }
3058 EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);