tmpfs: add shmem_read_mapping_page_gfp
[linux-2.6.git] / mm / shmem.c
1 /*
2  * Resizable virtual memory filesystem for Linux.
3  *
4  * Copyright (C) 2000 Linus Torvalds.
5  *               2000 Transmeta Corp.
6  *               2000-2001 Christoph Rohland
7  *               2000-2001 SAP AG
8  *               2002 Red Hat Inc.
9  * Copyright (C) 2002-2005 Hugh Dickins.
10  * Copyright (C) 2002-2005 VERITAS Software Corporation.
11  * Copyright (C) 2004 Andi Kleen, SuSE Labs
12  *
13  * Extended attribute support for tmpfs:
14  * Copyright (c) 2004, Luke Kenneth Casson Leighton <lkcl@lkcl.net>
15  * Copyright (c) 2004 Red Hat, Inc., James Morris <jmorris@redhat.com>
16  *
17  * tiny-shmem:
18  * Copyright (c) 2004, 2008 Matt Mackall <mpm@selenic.com>
19  *
20  * This file is released under the GPL.
21  */
22
23 #include <linux/fs.h>
24 #include <linux/init.h>
25 #include <linux/vfs.h>
26 #include <linux/mount.h>
27 #include <linux/pagemap.h>
28 #include <linux/file.h>
29 #include <linux/mm.h>
30 #include <linux/module.h>
31 #include <linux/percpu_counter.h>
32 #include <linux/swap.h>
33
34 static struct vfsmount *shm_mnt;
35
36 #ifdef CONFIG_SHMEM
37 /*
38  * This virtual memory filesystem is heavily based on the ramfs. It
39  * extends ramfs by the ability to use swap and honor resource limits
40  * which makes it a completely usable filesystem.
41  */
42
43 #include <linux/xattr.h>
44 #include <linux/exportfs.h>
45 #include <linux/posix_acl.h>
46 #include <linux/generic_acl.h>
47 #include <linux/mman.h>
48 #include <linux/string.h>
49 #include <linux/slab.h>
50 #include <linux/backing-dev.h>
51 #include <linux/shmem_fs.h>
52 #include <linux/writeback.h>
53 #include <linux/blkdev.h>
54 #include <linux/security.h>
55 #include <linux/swapops.h>
56 #include <linux/mempolicy.h>
57 #include <linux/namei.h>
58 #include <linux/ctype.h>
59 #include <linux/migrate.h>
60 #include <linux/highmem.h>
61 #include <linux/seq_file.h>
62 #include <linux/magic.h>
63
64 #include <asm/uaccess.h>
65 #include <asm/div64.h>
66 #include <asm/pgtable.h>
67
68 /*
69  * The maximum size of a shmem/tmpfs file is limited by the maximum size of
70  * its triple-indirect swap vector - see illustration at shmem_swp_entry().
71  *
72  * With 4kB page size, maximum file size is just over 2TB on a 32-bit kernel,
73  * but one eighth of that on a 64-bit kernel.  With 8kB page size, maximum
74  * file size is just over 4TB on a 64-bit kernel, but 16TB on a 32-bit kernel,
75  * MAX_LFS_FILESIZE being then more restrictive than swap vector layout.
76  *
77  * We use / and * instead of shifts in the definitions below, so that the swap
78  * vector can be tested with small even values (e.g. 20) for ENTRIES_PER_PAGE.
79  */
80 #define ENTRIES_PER_PAGE (PAGE_CACHE_SIZE/sizeof(unsigned long))
81 #define ENTRIES_PER_PAGEPAGE ((unsigned long long)ENTRIES_PER_PAGE*ENTRIES_PER_PAGE)
82
83 #define SHMSWP_MAX_INDEX (SHMEM_NR_DIRECT + (ENTRIES_PER_PAGEPAGE/2) * (ENTRIES_PER_PAGE+1))
84 #define SHMSWP_MAX_BYTES (SHMSWP_MAX_INDEX << PAGE_CACHE_SHIFT)
85
86 #define SHMEM_MAX_BYTES  min_t(unsigned long long, SHMSWP_MAX_BYTES, MAX_LFS_FILESIZE)
87 #define SHMEM_MAX_INDEX  ((unsigned long)((SHMEM_MAX_BYTES+1) >> PAGE_CACHE_SHIFT))
88
89 #define BLOCKS_PER_PAGE  (PAGE_CACHE_SIZE/512)
90 #define VM_ACCT(size)    (PAGE_CACHE_ALIGN(size) >> PAGE_SHIFT)
91
92 /* info->flags needs VM_flags to handle pagein/truncate races efficiently */
93 #define SHMEM_PAGEIN     VM_READ
94 #define SHMEM_TRUNCATE   VM_WRITE
95
96 /* Definition to limit shmem_truncate's steps between cond_rescheds */
97 #define LATENCY_LIMIT    64
98
99 /* Pretend that each entry is of this size in directory's i_size */
100 #define BOGO_DIRENT_SIZE 20
101
102 struct shmem_xattr {
103         struct list_head list;  /* anchored by shmem_inode_info->xattr_list */
104         char *name;             /* xattr name */
105         size_t size;
106         char value[0];
107 };
108
109 /* Flag allocation requirements to shmem_getpage and shmem_swp_alloc */
110 enum sgp_type {
111         SGP_READ,       /* don't exceed i_size, don't allocate page */
112         SGP_CACHE,      /* don't exceed i_size, may allocate page */
113         SGP_DIRTY,      /* like SGP_CACHE, but set new page dirty */
114         SGP_WRITE,      /* may exceed i_size, may allocate page */
115 };
116
117 #ifdef CONFIG_TMPFS
118 static unsigned long shmem_default_max_blocks(void)
119 {
120         return totalram_pages / 2;
121 }
122
123 static unsigned long shmem_default_max_inodes(void)
124 {
125         return min(totalram_pages - totalhigh_pages, totalram_pages / 2);
126 }
127 #endif
128
129 static int shmem_getpage(struct inode *inode, unsigned long idx,
130                          struct page **pagep, enum sgp_type sgp, int *type);
131
132 static inline struct page *shmem_dir_alloc(gfp_t gfp_mask)
133 {
134         /*
135          * The above definition of ENTRIES_PER_PAGE, and the use of
136          * BLOCKS_PER_PAGE on indirect pages, assume PAGE_CACHE_SIZE:
137          * might be reconsidered if it ever diverges from PAGE_SIZE.
138          *
139          * Mobility flags are masked out as swap vectors cannot move
140          */
141         return alloc_pages((gfp_mask & ~GFP_MOVABLE_MASK) | __GFP_ZERO,
142                                 PAGE_CACHE_SHIFT-PAGE_SHIFT);
143 }
144
145 static inline void shmem_dir_free(struct page *page)
146 {
147         __free_pages(page, PAGE_CACHE_SHIFT-PAGE_SHIFT);
148 }
149
150 static struct page **shmem_dir_map(struct page *page)
151 {
152         return (struct page **)kmap_atomic(page, KM_USER0);
153 }
154
155 static inline void shmem_dir_unmap(struct page **dir)
156 {
157         kunmap_atomic(dir, KM_USER0);
158 }
159
160 static swp_entry_t *shmem_swp_map(struct page *page)
161 {
162         return (swp_entry_t *)kmap_atomic(page, KM_USER1);
163 }
164
165 static inline void shmem_swp_balance_unmap(void)
166 {
167         /*
168          * When passing a pointer to an i_direct entry, to code which
169          * also handles indirect entries and so will shmem_swp_unmap,
170          * we must arrange for the preempt count to remain in balance.
171          * What kmap_atomic of a lowmem page does depends on config
172          * and architecture, so pretend to kmap_atomic some lowmem page.
173          */
174         (void) kmap_atomic(ZERO_PAGE(0), KM_USER1);
175 }
176
177 static inline void shmem_swp_unmap(swp_entry_t *entry)
178 {
179         kunmap_atomic(entry, KM_USER1);
180 }
181
182 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
183 {
184         return sb->s_fs_info;
185 }
186
187 /*
188  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
189  * for shared memory and for shared anonymous (/dev/zero) mappings
190  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
191  * consistent with the pre-accounting of private mappings ...
192  */
193 static inline int shmem_acct_size(unsigned long flags, loff_t size)
194 {
195         return (flags & VM_NORESERVE) ?
196                 0 : security_vm_enough_memory_kern(VM_ACCT(size));
197 }
198
199 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
200 {
201         if (!(flags & VM_NORESERVE))
202                 vm_unacct_memory(VM_ACCT(size));
203 }
204
205 /*
206  * ... whereas tmpfs objects are accounted incrementally as
207  * pages are allocated, in order to allow huge sparse files.
208  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
209  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
210  */
211 static inline int shmem_acct_block(unsigned long flags)
212 {
213         return (flags & VM_NORESERVE) ?
214                 security_vm_enough_memory_kern(VM_ACCT(PAGE_CACHE_SIZE)) : 0;
215 }
216
217 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
218 {
219         if (flags & VM_NORESERVE)
220                 vm_unacct_memory(pages * VM_ACCT(PAGE_CACHE_SIZE));
221 }
222
223 static const struct super_operations shmem_ops;
224 static const struct address_space_operations shmem_aops;
225 static const struct file_operations shmem_file_operations;
226 static const struct inode_operations shmem_inode_operations;
227 static const struct inode_operations shmem_dir_inode_operations;
228 static const struct inode_operations shmem_special_inode_operations;
229 static const struct vm_operations_struct shmem_vm_ops;
230
231 static struct backing_dev_info shmem_backing_dev_info  __read_mostly = {
232         .ra_pages       = 0,    /* No readahead */
233         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_SWAP_BACKED,
234 };
235
236 static LIST_HEAD(shmem_swaplist);
237 static DEFINE_MUTEX(shmem_swaplist_mutex);
238
239 static void shmem_free_blocks(struct inode *inode, long pages)
240 {
241         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
242         if (sbinfo->max_blocks) {
243                 percpu_counter_add(&sbinfo->used_blocks, -pages);
244                 spin_lock(&inode->i_lock);
245                 inode->i_blocks -= pages*BLOCKS_PER_PAGE;
246                 spin_unlock(&inode->i_lock);
247         }
248 }
249
250 static int shmem_reserve_inode(struct super_block *sb)
251 {
252         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
253         if (sbinfo->max_inodes) {
254                 spin_lock(&sbinfo->stat_lock);
255                 if (!sbinfo->free_inodes) {
256                         spin_unlock(&sbinfo->stat_lock);
257                         return -ENOSPC;
258                 }
259                 sbinfo->free_inodes--;
260                 spin_unlock(&sbinfo->stat_lock);
261         }
262         return 0;
263 }
264
265 static void shmem_free_inode(struct super_block *sb)
266 {
267         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
268         if (sbinfo->max_inodes) {
269                 spin_lock(&sbinfo->stat_lock);
270                 sbinfo->free_inodes++;
271                 spin_unlock(&sbinfo->stat_lock);
272         }
273 }
274
275 /**
276  * shmem_recalc_inode - recalculate the size of an inode
277  * @inode: inode to recalc
278  *
279  * We have to calculate the free blocks since the mm can drop
280  * undirtied hole pages behind our back.
281  *
282  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
283  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
284  *
285  * It has to be called with the spinlock held.
286  */
287 static void shmem_recalc_inode(struct inode *inode)
288 {
289         struct shmem_inode_info *info = SHMEM_I(inode);
290         long freed;
291
292         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
293         if (freed > 0) {
294                 info->alloced -= freed;
295                 shmem_unacct_blocks(info->flags, freed);
296                 shmem_free_blocks(inode, freed);
297         }
298 }
299
300 /**
301  * shmem_swp_entry - find the swap vector position in the info structure
302  * @info:  info structure for the inode
303  * @index: index of the page to find
304  * @page:  optional page to add to the structure. Has to be preset to
305  *         all zeros
306  *
307  * If there is no space allocated yet it will return NULL when
308  * page is NULL, else it will use the page for the needed block,
309  * setting it to NULL on return to indicate that it has been used.
310  *
311  * The swap vector is organized the following way:
312  *
313  * There are SHMEM_NR_DIRECT entries directly stored in the
314  * shmem_inode_info structure. So small files do not need an addional
315  * allocation.
316  *
317  * For pages with index > SHMEM_NR_DIRECT there is the pointer
318  * i_indirect which points to a page which holds in the first half
319  * doubly indirect blocks, in the second half triple indirect blocks:
320  *
321  * For an artificial ENTRIES_PER_PAGE = 4 this would lead to the
322  * following layout (for SHMEM_NR_DIRECT == 16):
323  *
324  * i_indirect -> dir --> 16-19
325  *            |      +-> 20-23
326  *            |
327  *            +-->dir2 --> 24-27
328  *            |        +-> 28-31
329  *            |        +-> 32-35
330  *            |        +-> 36-39
331  *            |
332  *            +-->dir3 --> 40-43
333  *                     +-> 44-47
334  *                     +-> 48-51
335  *                     +-> 52-55
336  */
337 static swp_entry_t *shmem_swp_entry(struct shmem_inode_info *info, unsigned long index, struct page **page)
338 {
339         unsigned long offset;
340         struct page **dir;
341         struct page *subdir;
342
343         if (index < SHMEM_NR_DIRECT) {
344                 shmem_swp_balance_unmap();
345                 return info->i_direct+index;
346         }
347         if (!info->i_indirect) {
348                 if (page) {
349                         info->i_indirect = *page;
350                         *page = NULL;
351                 }
352                 return NULL;                    /* need another page */
353         }
354
355         index -= SHMEM_NR_DIRECT;
356         offset = index % ENTRIES_PER_PAGE;
357         index /= ENTRIES_PER_PAGE;
358         dir = shmem_dir_map(info->i_indirect);
359
360         if (index >= ENTRIES_PER_PAGE/2) {
361                 index -= ENTRIES_PER_PAGE/2;
362                 dir += ENTRIES_PER_PAGE/2 + index/ENTRIES_PER_PAGE;
363                 index %= ENTRIES_PER_PAGE;
364                 subdir = *dir;
365                 if (!subdir) {
366                         if (page) {
367                                 *dir = *page;
368                                 *page = NULL;
369                         }
370                         shmem_dir_unmap(dir);
371                         return NULL;            /* need another page */
372                 }
373                 shmem_dir_unmap(dir);
374                 dir = shmem_dir_map(subdir);
375         }
376
377         dir += index;
378         subdir = *dir;
379         if (!subdir) {
380                 if (!page || !(subdir = *page)) {
381                         shmem_dir_unmap(dir);
382                         return NULL;            /* need a page */
383                 }
384                 *dir = subdir;
385                 *page = NULL;
386         }
387         shmem_dir_unmap(dir);
388         return shmem_swp_map(subdir) + offset;
389 }
390
391 static void shmem_swp_set(struct shmem_inode_info *info, swp_entry_t *entry, unsigned long value)
392 {
393         long incdec = value? 1: -1;
394
395         entry->val = value;
396         info->swapped += incdec;
397         if ((unsigned long)(entry - info->i_direct) >= SHMEM_NR_DIRECT) {
398                 struct page *page = kmap_atomic_to_page(entry);
399                 set_page_private(page, page_private(page) + incdec);
400         }
401 }
402
403 /**
404  * shmem_swp_alloc - get the position of the swap entry for the page.
405  * @info:       info structure for the inode
406  * @index:      index of the page to find
407  * @sgp:        check and recheck i_size? skip allocation?
408  *
409  * If the entry does not exist, allocate it.
410  */
411 static swp_entry_t *shmem_swp_alloc(struct shmem_inode_info *info, unsigned long index, enum sgp_type sgp)
412 {
413         struct inode *inode = &info->vfs_inode;
414         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
415         struct page *page = NULL;
416         swp_entry_t *entry;
417
418         if (sgp != SGP_WRITE &&
419             ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode))
420                 return ERR_PTR(-EINVAL);
421
422         while (!(entry = shmem_swp_entry(info, index, &page))) {
423                 if (sgp == SGP_READ)
424                         return shmem_swp_map(ZERO_PAGE(0));
425                 /*
426                  * Test used_blocks against 1 less max_blocks, since we have 1 data
427                  * page (and perhaps indirect index pages) yet to allocate:
428                  * a waste to allocate index if we cannot allocate data.
429                  */
430                 if (sbinfo->max_blocks) {
431                         if (percpu_counter_compare(&sbinfo->used_blocks,
432                                                 sbinfo->max_blocks - 1) >= 0)
433                                 return ERR_PTR(-ENOSPC);
434                         percpu_counter_inc(&sbinfo->used_blocks);
435                         spin_lock(&inode->i_lock);
436                         inode->i_blocks += BLOCKS_PER_PAGE;
437                         spin_unlock(&inode->i_lock);
438                 }
439
440                 spin_unlock(&info->lock);
441                 page = shmem_dir_alloc(mapping_gfp_mask(inode->i_mapping));
442                 spin_lock(&info->lock);
443
444                 if (!page) {
445                         shmem_free_blocks(inode, 1);
446                         return ERR_PTR(-ENOMEM);
447                 }
448                 if (sgp != SGP_WRITE &&
449                     ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
450                         entry = ERR_PTR(-EINVAL);
451                         break;
452                 }
453                 if (info->next_index <= index)
454                         info->next_index = index + 1;
455         }
456         if (page) {
457                 /* another task gave its page, or truncated the file */
458                 shmem_free_blocks(inode, 1);
459                 shmem_dir_free(page);
460         }
461         if (info->next_index <= index && !IS_ERR(entry))
462                 info->next_index = index + 1;
463         return entry;
464 }
465
466 /**
467  * shmem_free_swp - free some swap entries in a directory
468  * @dir:        pointer to the directory
469  * @edir:       pointer after last entry of the directory
470  * @punch_lock: pointer to spinlock when needed for the holepunch case
471  */
472 static int shmem_free_swp(swp_entry_t *dir, swp_entry_t *edir,
473                                                 spinlock_t *punch_lock)
474 {
475         spinlock_t *punch_unlock = NULL;
476         swp_entry_t *ptr;
477         int freed = 0;
478
479         for (ptr = dir; ptr < edir; ptr++) {
480                 if (ptr->val) {
481                         if (unlikely(punch_lock)) {
482                                 punch_unlock = punch_lock;
483                                 punch_lock = NULL;
484                                 spin_lock(punch_unlock);
485                                 if (!ptr->val)
486                                         continue;
487                         }
488                         free_swap_and_cache(*ptr);
489                         *ptr = (swp_entry_t){0};
490                         freed++;
491                 }
492         }
493         if (punch_unlock)
494                 spin_unlock(punch_unlock);
495         return freed;
496 }
497
498 static int shmem_map_and_free_swp(struct page *subdir, int offset,
499                 int limit, struct page ***dir, spinlock_t *punch_lock)
500 {
501         swp_entry_t *ptr;
502         int freed = 0;
503
504         ptr = shmem_swp_map(subdir);
505         for (; offset < limit; offset += LATENCY_LIMIT) {
506                 int size = limit - offset;
507                 if (size > LATENCY_LIMIT)
508                         size = LATENCY_LIMIT;
509                 freed += shmem_free_swp(ptr+offset, ptr+offset+size,
510                                                         punch_lock);
511                 if (need_resched()) {
512                         shmem_swp_unmap(ptr);
513                         if (*dir) {
514                                 shmem_dir_unmap(*dir);
515                                 *dir = NULL;
516                         }
517                         cond_resched();
518                         ptr = shmem_swp_map(subdir);
519                 }
520         }
521         shmem_swp_unmap(ptr);
522         return freed;
523 }
524
525 static void shmem_free_pages(struct list_head *next)
526 {
527         struct page *page;
528         int freed = 0;
529
530         do {
531                 page = container_of(next, struct page, lru);
532                 next = next->next;
533                 shmem_dir_free(page);
534                 freed++;
535                 if (freed >= LATENCY_LIMIT) {
536                         cond_resched();
537                         freed = 0;
538                 }
539         } while (next);
540 }
541
542 void shmem_truncate_range(struct inode *inode, loff_t start, loff_t end)
543 {
544         struct shmem_inode_info *info = SHMEM_I(inode);
545         unsigned long idx;
546         unsigned long size;
547         unsigned long limit;
548         unsigned long stage;
549         unsigned long diroff;
550         struct page **dir;
551         struct page *topdir;
552         struct page *middir;
553         struct page *subdir;
554         swp_entry_t *ptr;
555         LIST_HEAD(pages_to_free);
556         long nr_pages_to_free = 0;
557         long nr_swaps_freed = 0;
558         int offset;
559         int freed;
560         int punch_hole;
561         spinlock_t *needs_lock;
562         spinlock_t *punch_lock;
563         unsigned long upper_limit;
564
565         truncate_inode_pages_range(inode->i_mapping, start, end);
566
567         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
568         idx = (start + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
569         if (idx >= info->next_index)
570                 return;
571
572         spin_lock(&info->lock);
573         info->flags |= SHMEM_TRUNCATE;
574         if (likely(end == (loff_t) -1)) {
575                 limit = info->next_index;
576                 upper_limit = SHMEM_MAX_INDEX;
577                 info->next_index = idx;
578                 needs_lock = NULL;
579                 punch_hole = 0;
580         } else {
581                 if (end + 1 >= inode->i_size) { /* we may free a little more */
582                         limit = (inode->i_size + PAGE_CACHE_SIZE - 1) >>
583                                                         PAGE_CACHE_SHIFT;
584                         upper_limit = SHMEM_MAX_INDEX;
585                 } else {
586                         limit = (end + 1) >> PAGE_CACHE_SHIFT;
587                         upper_limit = limit;
588                 }
589                 needs_lock = &info->lock;
590                 punch_hole = 1;
591         }
592
593         topdir = info->i_indirect;
594         if (topdir && idx <= SHMEM_NR_DIRECT && !punch_hole) {
595                 info->i_indirect = NULL;
596                 nr_pages_to_free++;
597                 list_add(&topdir->lru, &pages_to_free);
598         }
599         spin_unlock(&info->lock);
600
601         if (info->swapped && idx < SHMEM_NR_DIRECT) {
602                 ptr = info->i_direct;
603                 size = limit;
604                 if (size > SHMEM_NR_DIRECT)
605                         size = SHMEM_NR_DIRECT;
606                 nr_swaps_freed = shmem_free_swp(ptr+idx, ptr+size, needs_lock);
607         }
608
609         /*
610          * If there are no indirect blocks or we are punching a hole
611          * below indirect blocks, nothing to be done.
612          */
613         if (!topdir || limit <= SHMEM_NR_DIRECT)
614                 goto done2;
615
616         /*
617          * The truncation case has already dropped info->lock, and we're safe
618          * because i_size and next_index have already been lowered, preventing
619          * access beyond.  But in the punch_hole case, we still need to take
620          * the lock when updating the swap directory, because there might be
621          * racing accesses by shmem_getpage(SGP_CACHE), shmem_unuse_inode or
622          * shmem_writepage.  However, whenever we find we can remove a whole
623          * directory page (not at the misaligned start or end of the range),
624          * we first NULLify its pointer in the level above, and then have no
625          * need to take the lock when updating its contents: needs_lock and
626          * punch_lock (either pointing to info->lock or NULL) manage this.
627          */
628
629         upper_limit -= SHMEM_NR_DIRECT;
630         limit -= SHMEM_NR_DIRECT;
631         idx = (idx > SHMEM_NR_DIRECT)? (idx - SHMEM_NR_DIRECT): 0;
632         offset = idx % ENTRIES_PER_PAGE;
633         idx -= offset;
634
635         dir = shmem_dir_map(topdir);
636         stage = ENTRIES_PER_PAGEPAGE/2;
637         if (idx < ENTRIES_PER_PAGEPAGE/2) {
638                 middir = topdir;
639                 diroff = idx/ENTRIES_PER_PAGE;
640         } else {
641                 dir += ENTRIES_PER_PAGE/2;
642                 dir += (idx - ENTRIES_PER_PAGEPAGE/2)/ENTRIES_PER_PAGEPAGE;
643                 while (stage <= idx)
644                         stage += ENTRIES_PER_PAGEPAGE;
645                 middir = *dir;
646                 if (*dir) {
647                         diroff = ((idx - ENTRIES_PER_PAGEPAGE/2) %
648                                 ENTRIES_PER_PAGEPAGE) / ENTRIES_PER_PAGE;
649                         if (!diroff && !offset && upper_limit >= stage) {
650                                 if (needs_lock) {
651                                         spin_lock(needs_lock);
652                                         *dir = NULL;
653                                         spin_unlock(needs_lock);
654                                         needs_lock = NULL;
655                                 } else
656                                         *dir = NULL;
657                                 nr_pages_to_free++;
658                                 list_add(&middir->lru, &pages_to_free);
659                         }
660                         shmem_dir_unmap(dir);
661                         dir = shmem_dir_map(middir);
662                 } else {
663                         diroff = 0;
664                         offset = 0;
665                         idx = stage;
666                 }
667         }
668
669         for (; idx < limit; idx += ENTRIES_PER_PAGE, diroff++) {
670                 if (unlikely(idx == stage)) {
671                         shmem_dir_unmap(dir);
672                         dir = shmem_dir_map(topdir) +
673                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
674                         while (!*dir) {
675                                 dir++;
676                                 idx += ENTRIES_PER_PAGEPAGE;
677                                 if (idx >= limit)
678                                         goto done1;
679                         }
680                         stage = idx + ENTRIES_PER_PAGEPAGE;
681                         middir = *dir;
682                         if (punch_hole)
683                                 needs_lock = &info->lock;
684                         if (upper_limit >= stage) {
685                                 if (needs_lock) {
686                                         spin_lock(needs_lock);
687                                         *dir = NULL;
688                                         spin_unlock(needs_lock);
689                                         needs_lock = NULL;
690                                 } else
691                                         *dir = NULL;
692                                 nr_pages_to_free++;
693                                 list_add(&middir->lru, &pages_to_free);
694                         }
695                         shmem_dir_unmap(dir);
696                         cond_resched();
697                         dir = shmem_dir_map(middir);
698                         diroff = 0;
699                 }
700                 punch_lock = needs_lock;
701                 subdir = dir[diroff];
702                 if (subdir && !offset && upper_limit-idx >= ENTRIES_PER_PAGE) {
703                         if (needs_lock) {
704                                 spin_lock(needs_lock);
705                                 dir[diroff] = NULL;
706                                 spin_unlock(needs_lock);
707                                 punch_lock = NULL;
708                         } else
709                                 dir[diroff] = NULL;
710                         nr_pages_to_free++;
711                         list_add(&subdir->lru, &pages_to_free);
712                 }
713                 if (subdir && page_private(subdir) /* has swap entries */) {
714                         size = limit - idx;
715                         if (size > ENTRIES_PER_PAGE)
716                                 size = ENTRIES_PER_PAGE;
717                         freed = shmem_map_and_free_swp(subdir,
718                                         offset, size, &dir, punch_lock);
719                         if (!dir)
720                                 dir = shmem_dir_map(middir);
721                         nr_swaps_freed += freed;
722                         if (offset || punch_lock) {
723                                 spin_lock(&info->lock);
724                                 set_page_private(subdir,
725                                         page_private(subdir) - freed);
726                                 spin_unlock(&info->lock);
727                         } else
728                                 BUG_ON(page_private(subdir) != freed);
729                 }
730                 offset = 0;
731         }
732 done1:
733         shmem_dir_unmap(dir);
734 done2:
735         if (inode->i_mapping->nrpages && (info->flags & SHMEM_PAGEIN)) {
736                 /*
737                  * Call truncate_inode_pages again: racing shmem_unuse_inode
738                  * may have swizzled a page in from swap since
739                  * truncate_pagecache or generic_delete_inode did it, before we
740                  * lowered next_index.  Also, though shmem_getpage checks
741                  * i_size before adding to cache, no recheck after: so fix the
742                  * narrow window there too.
743                  */
744                 truncate_inode_pages_range(inode->i_mapping, start, end);
745         }
746
747         spin_lock(&info->lock);
748         info->flags &= ~SHMEM_TRUNCATE;
749         info->swapped -= nr_swaps_freed;
750         if (nr_pages_to_free)
751                 shmem_free_blocks(inode, nr_pages_to_free);
752         shmem_recalc_inode(inode);
753         spin_unlock(&info->lock);
754
755         /*
756          * Empty swap vector directory pages to be freed?
757          */
758         if (!list_empty(&pages_to_free)) {
759                 pages_to_free.prev->next = NULL;
760                 shmem_free_pages(pages_to_free.next);
761         }
762 }
763 EXPORT_SYMBOL_GPL(shmem_truncate_range);
764
765 static int shmem_setattr(struct dentry *dentry, struct iattr *attr)
766 {
767         struct inode *inode = dentry->d_inode;
768         int error;
769
770         error = inode_change_ok(inode, attr);
771         if (error)
772                 return error;
773
774         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
775                 loff_t oldsize = inode->i_size;
776                 loff_t newsize = attr->ia_size;
777                 struct page *page = NULL;
778
779                 if (newsize < oldsize) {
780                         /*
781                          * If truncating down to a partial page, then
782                          * if that page is already allocated, hold it
783                          * in memory until the truncation is over, so
784                          * truncate_partial_page cannot miss it were
785                          * it assigned to swap.
786                          */
787                         if (newsize & (PAGE_CACHE_SIZE-1)) {
788                                 (void) shmem_getpage(inode,
789                                         newsize >> PAGE_CACHE_SHIFT,
790                                                 &page, SGP_READ, NULL);
791                                 if (page)
792                                         unlock_page(page);
793                         }
794                         /*
795                          * Reset SHMEM_PAGEIN flag so that shmem_truncate can
796                          * detect if any pages might have been added to cache
797                          * after truncate_inode_pages.  But we needn't bother
798                          * if it's being fully truncated to zero-length: the
799                          * nrpages check is efficient enough in that case.
800                          */
801                         if (newsize) {
802                                 struct shmem_inode_info *info = SHMEM_I(inode);
803                                 spin_lock(&info->lock);
804                                 info->flags &= ~SHMEM_PAGEIN;
805                                 spin_unlock(&info->lock);
806                         }
807                 }
808                 if (newsize != oldsize) {
809                         i_size_write(inode, newsize);
810                         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
811                 }
812                 if (newsize < oldsize) {
813                         loff_t holebegin = round_up(newsize, PAGE_SIZE);
814                         unmap_mapping_range(inode->i_mapping, holebegin, 0, 1);
815                         shmem_truncate_range(inode, newsize, (loff_t)-1);
816                         /* unmap again to remove racily COWed private pages */
817                         unmap_mapping_range(inode->i_mapping, holebegin, 0, 1);
818                 }
819                 if (page)
820                         page_cache_release(page);
821         }
822
823         setattr_copy(inode, attr);
824 #ifdef CONFIG_TMPFS_POSIX_ACL
825         if (attr->ia_valid & ATTR_MODE)
826                 error = generic_acl_chmod(inode);
827 #endif
828         return error;
829 }
830
831 static void shmem_evict_inode(struct inode *inode)
832 {
833         struct shmem_inode_info *info = SHMEM_I(inode);
834         struct shmem_xattr *xattr, *nxattr;
835
836         if (inode->i_mapping->a_ops == &shmem_aops) {
837                 shmem_unacct_size(info->flags, inode->i_size);
838                 inode->i_size = 0;
839                 shmem_truncate_range(inode, 0, (loff_t)-1);
840                 if (!list_empty(&info->swaplist)) {
841                         mutex_lock(&shmem_swaplist_mutex);
842                         list_del_init(&info->swaplist);
843                         mutex_unlock(&shmem_swaplist_mutex);
844                 }
845         }
846
847         list_for_each_entry_safe(xattr, nxattr, &info->xattr_list, list) {
848                 kfree(xattr->name);
849                 kfree(xattr);
850         }
851         BUG_ON(inode->i_blocks);
852         shmem_free_inode(inode->i_sb);
853         end_writeback(inode);
854 }
855
856 static inline int shmem_find_swp(swp_entry_t entry, swp_entry_t *dir, swp_entry_t *edir)
857 {
858         swp_entry_t *ptr;
859
860         for (ptr = dir; ptr < edir; ptr++) {
861                 if (ptr->val == entry.val)
862                         return ptr - dir;
863         }
864         return -1;
865 }
866
867 static int shmem_unuse_inode(struct shmem_inode_info *info, swp_entry_t entry, struct page *page)
868 {
869         struct address_space *mapping;
870         unsigned long idx;
871         unsigned long size;
872         unsigned long limit;
873         unsigned long stage;
874         struct page **dir;
875         struct page *subdir;
876         swp_entry_t *ptr;
877         int offset;
878         int error;
879
880         idx = 0;
881         ptr = info->i_direct;
882         spin_lock(&info->lock);
883         if (!info->swapped) {
884                 list_del_init(&info->swaplist);
885                 goto lost2;
886         }
887         limit = info->next_index;
888         size = limit;
889         if (size > SHMEM_NR_DIRECT)
890                 size = SHMEM_NR_DIRECT;
891         offset = shmem_find_swp(entry, ptr, ptr+size);
892         if (offset >= 0) {
893                 shmem_swp_balance_unmap();
894                 goto found;
895         }
896         if (!info->i_indirect)
897                 goto lost2;
898
899         dir = shmem_dir_map(info->i_indirect);
900         stage = SHMEM_NR_DIRECT + ENTRIES_PER_PAGEPAGE/2;
901
902         for (idx = SHMEM_NR_DIRECT; idx < limit; idx += ENTRIES_PER_PAGE, dir++) {
903                 if (unlikely(idx == stage)) {
904                         shmem_dir_unmap(dir-1);
905                         if (cond_resched_lock(&info->lock)) {
906                                 /* check it has not been truncated */
907                                 if (limit > info->next_index) {
908                                         limit = info->next_index;
909                                         if (idx >= limit)
910                                                 goto lost2;
911                                 }
912                         }
913                         dir = shmem_dir_map(info->i_indirect) +
914                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
915                         while (!*dir) {
916                                 dir++;
917                                 idx += ENTRIES_PER_PAGEPAGE;
918                                 if (idx >= limit)
919                                         goto lost1;
920                         }
921                         stage = idx + ENTRIES_PER_PAGEPAGE;
922                         subdir = *dir;
923                         shmem_dir_unmap(dir);
924                         dir = shmem_dir_map(subdir);
925                 }
926                 subdir = *dir;
927                 if (subdir && page_private(subdir)) {
928                         ptr = shmem_swp_map(subdir);
929                         size = limit - idx;
930                         if (size > ENTRIES_PER_PAGE)
931                                 size = ENTRIES_PER_PAGE;
932                         offset = shmem_find_swp(entry, ptr, ptr+size);
933                         shmem_swp_unmap(ptr);
934                         if (offset >= 0) {
935                                 shmem_dir_unmap(dir);
936                                 ptr = shmem_swp_map(subdir);
937                                 goto found;
938                         }
939                 }
940         }
941 lost1:
942         shmem_dir_unmap(dir-1);
943 lost2:
944         spin_unlock(&info->lock);
945         return 0;
946 found:
947         idx += offset;
948         ptr += offset;
949
950         /*
951          * Move _head_ to start search for next from here.
952          * But be careful: shmem_evict_inode checks list_empty without taking
953          * mutex, and there's an instant in list_move_tail when info->swaplist
954          * would appear empty, if it were the only one on shmem_swaplist.  We
955          * could avoid doing it if inode NULL; or use this minor optimization.
956          */
957         if (shmem_swaplist.next != &info->swaplist)
958                 list_move_tail(&shmem_swaplist, &info->swaplist);
959
960         /*
961          * We rely on shmem_swaplist_mutex, not only to protect the swaplist,
962          * but also to hold up shmem_evict_inode(): so inode cannot be freed
963          * beneath us (pagelock doesn't help until the page is in pagecache).
964          */
965         mapping = info->vfs_inode.i_mapping;
966         error = add_to_page_cache_locked(page, mapping, idx, GFP_NOWAIT);
967         /* which does mem_cgroup_uncharge_cache_page on error */
968
969         if (error == -EEXIST) {
970                 struct page *filepage = find_get_page(mapping, idx);
971                 error = 1;
972                 if (filepage) {
973                         /*
974                          * There might be a more uptodate page coming down
975                          * from a stacked writepage: forget our swappage if so.
976                          */
977                         if (PageUptodate(filepage))
978                                 error = 0;
979                         page_cache_release(filepage);
980                 }
981         }
982         if (!error) {
983                 delete_from_swap_cache(page);
984                 set_page_dirty(page);
985                 info->flags |= SHMEM_PAGEIN;
986                 shmem_swp_set(info, ptr, 0);
987                 swap_free(entry);
988                 error = 1;      /* not an error, but entry was found */
989         }
990         shmem_swp_unmap(ptr);
991         spin_unlock(&info->lock);
992         return error;
993 }
994
995 /*
996  * shmem_unuse() search for an eventually swapped out shmem page.
997  */
998 int shmem_unuse(swp_entry_t entry, struct page *page)
999 {
1000         struct list_head *p, *next;
1001         struct shmem_inode_info *info;
1002         int found = 0;
1003         int error;
1004
1005         /*
1006          * Charge page using GFP_KERNEL while we can wait, before taking
1007          * the shmem_swaplist_mutex which might hold up shmem_writepage().
1008          * Charged back to the user (not to caller) when swap account is used.
1009          * add_to_page_cache() will be called with GFP_NOWAIT.
1010          */
1011         error = mem_cgroup_cache_charge(page, current->mm, GFP_KERNEL);
1012         if (error)
1013                 goto out;
1014         /*
1015          * Try to preload while we can wait, to not make a habit of
1016          * draining atomic reserves; but don't latch on to this cpu,
1017          * it's okay if sometimes we get rescheduled after this.
1018          */
1019         error = radix_tree_preload(GFP_KERNEL);
1020         if (error)
1021                 goto uncharge;
1022         radix_tree_preload_end();
1023
1024         mutex_lock(&shmem_swaplist_mutex);
1025         list_for_each_safe(p, next, &shmem_swaplist) {
1026                 info = list_entry(p, struct shmem_inode_info, swaplist);
1027                 found = shmem_unuse_inode(info, entry, page);
1028                 cond_resched();
1029                 if (found)
1030                         break;
1031         }
1032         mutex_unlock(&shmem_swaplist_mutex);
1033
1034 uncharge:
1035         if (!found)
1036                 mem_cgroup_uncharge_cache_page(page);
1037         if (found < 0)
1038                 error = found;
1039 out:
1040         unlock_page(page);
1041         page_cache_release(page);
1042         return error;
1043 }
1044
1045 /*
1046  * Move the page from the page cache to the swap cache.
1047  */
1048 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1049 {
1050         struct shmem_inode_info *info;
1051         swp_entry_t *entry, swap;
1052         struct address_space *mapping;
1053         unsigned long index;
1054         struct inode *inode;
1055
1056         BUG_ON(!PageLocked(page));
1057         mapping = page->mapping;
1058         index = page->index;
1059         inode = mapping->host;
1060         info = SHMEM_I(inode);
1061         if (info->flags & VM_LOCKED)
1062                 goto redirty;
1063         if (!total_swap_pages)
1064                 goto redirty;
1065
1066         /*
1067          * shmem_backing_dev_info's capabilities prevent regular writeback or
1068          * sync from ever calling shmem_writepage; but a stacking filesystem
1069          * may use the ->writepage of its underlying filesystem, in which case
1070          * tmpfs should write out to swap only in response to memory pressure,
1071          * and not for the writeback threads or sync.  However, in those cases,
1072          * we do still want to check if there's a redundant swappage to be
1073          * discarded.
1074          */
1075         if (wbc->for_reclaim)
1076                 swap = get_swap_page();
1077         else
1078                 swap.val = 0;
1079
1080         /*
1081          * Add inode to shmem_unuse()'s list of swapped-out inodes,
1082          * if it's not already there.  Do it now because we cannot take
1083          * mutex while holding spinlock, and must do so before the page
1084          * is moved to swap cache, when its pagelock no longer protects
1085          * the inode from eviction.  But don't unlock the mutex until
1086          * we've taken the spinlock, because shmem_unuse_inode() will
1087          * prune a !swapped inode from the swaplist under both locks.
1088          */
1089         if (swap.val) {
1090                 mutex_lock(&shmem_swaplist_mutex);
1091                 if (list_empty(&info->swaplist))
1092                         list_add_tail(&info->swaplist, &shmem_swaplist);
1093         }
1094
1095         spin_lock(&info->lock);
1096         if (swap.val)
1097                 mutex_unlock(&shmem_swaplist_mutex);
1098
1099         if (index >= info->next_index) {
1100                 BUG_ON(!(info->flags & SHMEM_TRUNCATE));
1101                 goto unlock;
1102         }
1103         entry = shmem_swp_entry(info, index, NULL);
1104         if (entry->val) {
1105                 /*
1106                  * The more uptodate page coming down from a stacked
1107                  * writepage should replace our old swappage.
1108                  */
1109                 free_swap_and_cache(*entry);
1110                 shmem_swp_set(info, entry, 0);
1111         }
1112         shmem_recalc_inode(inode);
1113
1114         if (swap.val && add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
1115                 delete_from_page_cache(page);
1116                 shmem_swp_set(info, entry, swap.val);
1117                 shmem_swp_unmap(entry);
1118                 swap_shmem_alloc(swap);
1119                 spin_unlock(&info->lock);
1120                 BUG_ON(page_mapped(page));
1121                 swap_writepage(page, wbc);
1122                 return 0;
1123         }
1124
1125         shmem_swp_unmap(entry);
1126 unlock:
1127         spin_unlock(&info->lock);
1128         /*
1129          * add_to_swap_cache() doesn't return -EEXIST, so we can safely
1130          * clear SWAP_HAS_CACHE flag.
1131          */
1132         swapcache_free(swap, NULL);
1133 redirty:
1134         set_page_dirty(page);
1135         if (wbc->for_reclaim)
1136                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
1137         unlock_page(page);
1138         return 0;
1139 }
1140
1141 #ifdef CONFIG_NUMA
1142 #ifdef CONFIG_TMPFS
1143 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1144 {
1145         char buffer[64];
1146
1147         if (!mpol || mpol->mode == MPOL_DEFAULT)
1148                 return;         /* show nothing */
1149
1150         mpol_to_str(buffer, sizeof(buffer), mpol, 1);
1151
1152         seq_printf(seq, ",mpol=%s", buffer);
1153 }
1154
1155 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1156 {
1157         struct mempolicy *mpol = NULL;
1158         if (sbinfo->mpol) {
1159                 spin_lock(&sbinfo->stat_lock);  /* prevent replace/use races */
1160                 mpol = sbinfo->mpol;
1161                 mpol_get(mpol);
1162                 spin_unlock(&sbinfo->stat_lock);
1163         }
1164         return mpol;
1165 }
1166 #endif /* CONFIG_TMPFS */
1167
1168 static struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1169                         struct shmem_inode_info *info, unsigned long idx)
1170 {
1171         struct mempolicy mpol, *spol;
1172         struct vm_area_struct pvma;
1173         struct page *page;
1174
1175         spol = mpol_cond_copy(&mpol,
1176                                 mpol_shared_policy_lookup(&info->policy, idx));
1177
1178         /* Create a pseudo vma that just contains the policy */
1179         pvma.vm_start = 0;
1180         pvma.vm_pgoff = idx;
1181         pvma.vm_ops = NULL;
1182         pvma.vm_policy = spol;
1183         page = swapin_readahead(entry, gfp, &pvma, 0);
1184         return page;
1185 }
1186
1187 static struct page *shmem_alloc_page(gfp_t gfp,
1188                         struct shmem_inode_info *info, unsigned long idx)
1189 {
1190         struct vm_area_struct pvma;
1191
1192         /* Create a pseudo vma that just contains the policy */
1193         pvma.vm_start = 0;
1194         pvma.vm_pgoff = idx;
1195         pvma.vm_ops = NULL;
1196         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, idx);
1197
1198         /*
1199          * alloc_page_vma() will drop the shared policy reference
1200          */
1201         return alloc_page_vma(gfp, &pvma, 0);
1202 }
1203 #else /* !CONFIG_NUMA */
1204 #ifdef CONFIG_TMPFS
1205 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *p)
1206 {
1207 }
1208 #endif /* CONFIG_TMPFS */
1209
1210 static inline struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1211                         struct shmem_inode_info *info, unsigned long idx)
1212 {
1213         return swapin_readahead(entry, gfp, NULL, 0);
1214 }
1215
1216 static inline struct page *shmem_alloc_page(gfp_t gfp,
1217                         struct shmem_inode_info *info, unsigned long idx)
1218 {
1219         return alloc_page(gfp);
1220 }
1221 #endif /* CONFIG_NUMA */
1222
1223 #if !defined(CONFIG_NUMA) || !defined(CONFIG_TMPFS)
1224 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1225 {
1226         return NULL;
1227 }
1228 #endif
1229
1230 /*
1231  * shmem_getpage - either get the page from swap or allocate a new one
1232  *
1233  * If we allocate a new one we do not mark it dirty. That's up to the
1234  * vm. If we swap it in we mark it dirty since we also free the swap
1235  * entry since a page cannot live in both the swap and page cache
1236  */
1237 static int shmem_getpage(struct inode *inode, unsigned long idx,
1238                         struct page **pagep, enum sgp_type sgp, int *type)
1239 {
1240         struct address_space *mapping = inode->i_mapping;
1241         struct shmem_inode_info *info = SHMEM_I(inode);
1242         struct shmem_sb_info *sbinfo;
1243         struct page *filepage = *pagep;
1244         struct page *swappage;
1245         struct page *prealloc_page = NULL;
1246         swp_entry_t *entry;
1247         swp_entry_t swap;
1248         gfp_t gfp;
1249         int error;
1250
1251         if (idx >= SHMEM_MAX_INDEX)
1252                 return -EFBIG;
1253
1254         if (type)
1255                 *type = 0;
1256
1257         /*
1258          * Normally, filepage is NULL on entry, and either found
1259          * uptodate immediately, or allocated and zeroed, or read
1260          * in under swappage, which is then assigned to filepage.
1261          * But shmem_readpage (required for splice) passes in a locked
1262          * filepage, which may be found not uptodate by other callers
1263          * too, and may need to be copied from the swappage read in.
1264          */
1265 repeat:
1266         if (!filepage)
1267                 filepage = find_lock_page(mapping, idx);
1268         if (filepage && PageUptodate(filepage))
1269                 goto done;
1270         gfp = mapping_gfp_mask(mapping);
1271         if (!filepage) {
1272                 /*
1273                  * Try to preload while we can wait, to not make a habit of
1274                  * draining atomic reserves; but don't latch on to this cpu.
1275                  */
1276                 error = radix_tree_preload(gfp & ~__GFP_HIGHMEM);
1277                 if (error)
1278                         goto failed;
1279                 radix_tree_preload_end();
1280                 if (sgp != SGP_READ && !prealloc_page) {
1281                         /* We don't care if this fails */
1282                         prealloc_page = shmem_alloc_page(gfp, info, idx);
1283                         if (prealloc_page) {
1284                                 if (mem_cgroup_cache_charge(prealloc_page,
1285                                                 current->mm, GFP_KERNEL)) {
1286                                         page_cache_release(prealloc_page);
1287                                         prealloc_page = NULL;
1288                                 }
1289                         }
1290                 }
1291         }
1292         error = 0;
1293
1294         spin_lock(&info->lock);
1295         shmem_recalc_inode(inode);
1296         entry = shmem_swp_alloc(info, idx, sgp);
1297         if (IS_ERR(entry)) {
1298                 spin_unlock(&info->lock);
1299                 error = PTR_ERR(entry);
1300                 goto failed;
1301         }
1302         swap = *entry;
1303
1304         if (swap.val) {
1305                 /* Look it up and read it in.. */
1306                 swappage = lookup_swap_cache(swap);
1307                 if (!swappage) {
1308                         shmem_swp_unmap(entry);
1309                         spin_unlock(&info->lock);
1310                         /* here we actually do the io */
1311                         if (type)
1312                                 *type |= VM_FAULT_MAJOR;
1313                         swappage = shmem_swapin(swap, gfp, info, idx);
1314                         if (!swappage) {
1315                                 spin_lock(&info->lock);
1316                                 entry = shmem_swp_alloc(info, idx, sgp);
1317                                 if (IS_ERR(entry))
1318                                         error = PTR_ERR(entry);
1319                                 else {
1320                                         if (entry->val == swap.val)
1321                                                 error = -ENOMEM;
1322                                         shmem_swp_unmap(entry);
1323                                 }
1324                                 spin_unlock(&info->lock);
1325                                 if (error)
1326                                         goto failed;
1327                                 goto repeat;
1328                         }
1329                         wait_on_page_locked(swappage);
1330                         page_cache_release(swappage);
1331                         goto repeat;
1332                 }
1333
1334                 /* We have to do this with page locked to prevent races */
1335                 if (!trylock_page(swappage)) {
1336                         shmem_swp_unmap(entry);
1337                         spin_unlock(&info->lock);
1338                         wait_on_page_locked(swappage);
1339                         page_cache_release(swappage);
1340                         goto repeat;
1341                 }
1342                 if (PageWriteback(swappage)) {
1343                         shmem_swp_unmap(entry);
1344                         spin_unlock(&info->lock);
1345                         wait_on_page_writeback(swappage);
1346                         unlock_page(swappage);
1347                         page_cache_release(swappage);
1348                         goto repeat;
1349                 }
1350                 if (!PageUptodate(swappage)) {
1351                         shmem_swp_unmap(entry);
1352                         spin_unlock(&info->lock);
1353                         unlock_page(swappage);
1354                         page_cache_release(swappage);
1355                         error = -EIO;
1356                         goto failed;
1357                 }
1358
1359                 if (filepage) {
1360                         shmem_swp_set(info, entry, 0);
1361                         shmem_swp_unmap(entry);
1362                         delete_from_swap_cache(swappage);
1363                         spin_unlock(&info->lock);
1364                         copy_highpage(filepage, swappage);
1365                         unlock_page(swappage);
1366                         page_cache_release(swappage);
1367                         flush_dcache_page(filepage);
1368                         SetPageUptodate(filepage);
1369                         set_page_dirty(filepage);
1370                         swap_free(swap);
1371                 } else if (!(error = add_to_page_cache_locked(swappage, mapping,
1372                                         idx, GFP_NOWAIT))) {
1373                         info->flags |= SHMEM_PAGEIN;
1374                         shmem_swp_set(info, entry, 0);
1375                         shmem_swp_unmap(entry);
1376                         delete_from_swap_cache(swappage);
1377                         spin_unlock(&info->lock);
1378                         filepage = swappage;
1379                         set_page_dirty(filepage);
1380                         swap_free(swap);
1381                 } else {
1382                         shmem_swp_unmap(entry);
1383                         spin_unlock(&info->lock);
1384                         if (error == -ENOMEM) {
1385                                 /*
1386                                  * reclaim from proper memory cgroup and
1387                                  * call memcg's OOM if needed.
1388                                  */
1389                                 error = mem_cgroup_shmem_charge_fallback(
1390                                                                 swappage,
1391                                                                 current->mm,
1392                                                                 gfp);
1393                                 if (error) {
1394                                         unlock_page(swappage);
1395                                         page_cache_release(swappage);
1396                                         goto failed;
1397                                 }
1398                         }
1399                         unlock_page(swappage);
1400                         page_cache_release(swappage);
1401                         goto repeat;
1402                 }
1403         } else if (sgp == SGP_READ && !filepage) {
1404                 shmem_swp_unmap(entry);
1405                 filepage = find_get_page(mapping, idx);
1406                 if (filepage &&
1407                     (!PageUptodate(filepage) || !trylock_page(filepage))) {
1408                         spin_unlock(&info->lock);
1409                         wait_on_page_locked(filepage);
1410                         page_cache_release(filepage);
1411                         filepage = NULL;
1412                         goto repeat;
1413                 }
1414                 spin_unlock(&info->lock);
1415         } else {
1416                 shmem_swp_unmap(entry);
1417                 sbinfo = SHMEM_SB(inode->i_sb);
1418                 if (sbinfo->max_blocks) {
1419                         if (percpu_counter_compare(&sbinfo->used_blocks,
1420                                                 sbinfo->max_blocks) >= 0 ||
1421                             shmem_acct_block(info->flags))
1422                                 goto nospace;
1423                         percpu_counter_inc(&sbinfo->used_blocks);
1424                         spin_lock(&inode->i_lock);
1425                         inode->i_blocks += BLOCKS_PER_PAGE;
1426                         spin_unlock(&inode->i_lock);
1427                 } else if (shmem_acct_block(info->flags))
1428                         goto nospace;
1429
1430                 if (!filepage) {
1431                         int ret;
1432
1433                         if (!prealloc_page) {
1434                                 spin_unlock(&info->lock);
1435                                 filepage = shmem_alloc_page(gfp, info, idx);
1436                                 if (!filepage) {
1437                                         shmem_unacct_blocks(info->flags, 1);
1438                                         shmem_free_blocks(inode, 1);
1439                                         error = -ENOMEM;
1440                                         goto failed;
1441                                 }
1442                                 SetPageSwapBacked(filepage);
1443
1444                                 /*
1445                                  * Precharge page while we can wait, compensate
1446                                  * after
1447                                  */
1448                                 error = mem_cgroup_cache_charge(filepage,
1449                                         current->mm, GFP_KERNEL);
1450                                 if (error) {
1451                                         page_cache_release(filepage);
1452                                         shmem_unacct_blocks(info->flags, 1);
1453                                         shmem_free_blocks(inode, 1);
1454                                         filepage = NULL;
1455                                         goto failed;
1456                                 }
1457
1458                                 spin_lock(&info->lock);
1459                         } else {
1460                                 filepage = prealloc_page;
1461                                 prealloc_page = NULL;
1462                                 SetPageSwapBacked(filepage);
1463                         }
1464
1465                         entry = shmem_swp_alloc(info, idx, sgp);
1466                         if (IS_ERR(entry))
1467                                 error = PTR_ERR(entry);
1468                         else {
1469                                 swap = *entry;
1470                                 shmem_swp_unmap(entry);
1471                         }
1472                         ret = error || swap.val;
1473                         if (ret)
1474                                 mem_cgroup_uncharge_cache_page(filepage);
1475                         else
1476                                 ret = add_to_page_cache_lru(filepage, mapping,
1477                                                 idx, GFP_NOWAIT);
1478                         /*
1479                          * At add_to_page_cache_lru() failure, uncharge will
1480                          * be done automatically.
1481                          */
1482                         if (ret) {
1483                                 spin_unlock(&info->lock);
1484                                 page_cache_release(filepage);
1485                                 shmem_unacct_blocks(info->flags, 1);
1486                                 shmem_free_blocks(inode, 1);
1487                                 filepage = NULL;
1488                                 if (error)
1489                                         goto failed;
1490                                 goto repeat;
1491                         }
1492                         info->flags |= SHMEM_PAGEIN;
1493                 }
1494
1495                 info->alloced++;
1496                 spin_unlock(&info->lock);
1497                 clear_highpage(filepage);
1498                 flush_dcache_page(filepage);
1499                 SetPageUptodate(filepage);
1500                 if (sgp == SGP_DIRTY)
1501                         set_page_dirty(filepage);
1502         }
1503 done:
1504         *pagep = filepage;
1505         error = 0;
1506         goto out;
1507
1508 nospace:
1509         /*
1510          * Perhaps the page was brought in from swap between find_lock_page
1511          * and taking info->lock?  We allow for that at add_to_page_cache_lru,
1512          * but must also avoid reporting a spurious ENOSPC while working on a
1513          * full tmpfs.  (When filepage has been passed in to shmem_getpage, it
1514          * is already in page cache, which prevents this race from occurring.)
1515          */
1516         if (!filepage) {
1517                 struct page *page = find_get_page(mapping, idx);
1518                 if (page) {
1519                         spin_unlock(&info->lock);
1520                         page_cache_release(page);
1521                         goto repeat;
1522                 }
1523         }
1524         spin_unlock(&info->lock);
1525         error = -ENOSPC;
1526 failed:
1527         if (*pagep != filepage) {
1528                 unlock_page(filepage);
1529                 page_cache_release(filepage);
1530         }
1531 out:
1532         if (prealloc_page) {
1533                 mem_cgroup_uncharge_cache_page(prealloc_page);
1534                 page_cache_release(prealloc_page);
1535         }
1536         return error;
1537 }
1538
1539 static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1540 {
1541         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1542         int error;
1543         int ret;
1544
1545         if (((loff_t)vmf->pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
1546                 return VM_FAULT_SIGBUS;
1547
1548         error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_CACHE, &ret);
1549         if (error)
1550                 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1551         if (ret & VM_FAULT_MAJOR) {
1552                 count_vm_event(PGMAJFAULT);
1553                 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1554         }
1555         return ret | VM_FAULT_LOCKED;
1556 }
1557
1558 #ifdef CONFIG_NUMA
1559 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *new)
1560 {
1561         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1562         return mpol_set_shared_policy(&SHMEM_I(i)->policy, vma, new);
1563 }
1564
1565 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
1566                                           unsigned long addr)
1567 {
1568         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1569         unsigned long idx;
1570
1571         idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1572         return mpol_shared_policy_lookup(&SHMEM_I(i)->policy, idx);
1573 }
1574 #endif
1575
1576 int shmem_lock(struct file *file, int lock, struct user_struct *user)
1577 {
1578         struct inode *inode = file->f_path.dentry->d_inode;
1579         struct shmem_inode_info *info = SHMEM_I(inode);
1580         int retval = -ENOMEM;
1581
1582         spin_lock(&info->lock);
1583         if (lock && !(info->flags & VM_LOCKED)) {
1584                 if (!user_shm_lock(inode->i_size, user))
1585                         goto out_nomem;
1586                 info->flags |= VM_LOCKED;
1587                 mapping_set_unevictable(file->f_mapping);
1588         }
1589         if (!lock && (info->flags & VM_LOCKED) && user) {
1590                 user_shm_unlock(inode->i_size, user);
1591                 info->flags &= ~VM_LOCKED;
1592                 mapping_clear_unevictable(file->f_mapping);
1593                 scan_mapping_unevictable_pages(file->f_mapping);
1594         }
1595         retval = 0;
1596
1597 out_nomem:
1598         spin_unlock(&info->lock);
1599         return retval;
1600 }
1601
1602 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1603 {
1604         file_accessed(file);
1605         vma->vm_ops = &shmem_vm_ops;
1606         vma->vm_flags |= VM_CAN_NONLINEAR;
1607         return 0;
1608 }
1609
1610 static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
1611                                      int mode, dev_t dev, unsigned long flags)
1612 {
1613         struct inode *inode;
1614         struct shmem_inode_info *info;
1615         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1616
1617         if (shmem_reserve_inode(sb))
1618                 return NULL;
1619
1620         inode = new_inode(sb);
1621         if (inode) {
1622                 inode->i_ino = get_next_ino();
1623                 inode_init_owner(inode, dir, mode);
1624                 inode->i_blocks = 0;
1625                 inode->i_mapping->backing_dev_info = &shmem_backing_dev_info;
1626                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1627                 inode->i_generation = get_seconds();
1628                 info = SHMEM_I(inode);
1629                 memset(info, 0, (char *)inode - (char *)info);
1630                 spin_lock_init(&info->lock);
1631                 info->flags = flags & VM_NORESERVE;
1632                 INIT_LIST_HEAD(&info->swaplist);
1633                 INIT_LIST_HEAD(&info->xattr_list);
1634                 cache_no_acl(inode);
1635
1636                 switch (mode & S_IFMT) {
1637                 default:
1638                         inode->i_op = &shmem_special_inode_operations;
1639                         init_special_inode(inode, mode, dev);
1640                         break;
1641                 case S_IFREG:
1642                         inode->i_mapping->a_ops = &shmem_aops;
1643                         inode->i_op = &shmem_inode_operations;
1644                         inode->i_fop = &shmem_file_operations;
1645                         mpol_shared_policy_init(&info->policy,
1646                                                  shmem_get_sbmpol(sbinfo));
1647                         break;
1648                 case S_IFDIR:
1649                         inc_nlink(inode);
1650                         /* Some things misbehave if size == 0 on a directory */
1651                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
1652                         inode->i_op = &shmem_dir_inode_operations;
1653                         inode->i_fop = &simple_dir_operations;
1654                         break;
1655                 case S_IFLNK:
1656                         /*
1657                          * Must not load anything in the rbtree,
1658                          * mpol_free_shared_policy will not be called.
1659                          */
1660                         mpol_shared_policy_init(&info->policy, NULL);
1661                         break;
1662                 }
1663         } else
1664                 shmem_free_inode(sb);
1665         return inode;
1666 }
1667
1668 #ifdef CONFIG_TMPFS
1669 static const struct inode_operations shmem_symlink_inode_operations;
1670 static const struct inode_operations shmem_symlink_inline_operations;
1671
1672 /*
1673  * Normally tmpfs avoids the use of shmem_readpage and shmem_write_begin;
1674  * but providing them allows a tmpfs file to be used for splice, sendfile, and
1675  * below the loop driver, in the generic fashion that many filesystems support.
1676  */
1677 static int shmem_readpage(struct file *file, struct page *page)
1678 {
1679         struct inode *inode = page->mapping->host;
1680         int error = shmem_getpage(inode, page->index, &page, SGP_CACHE, NULL);
1681         unlock_page(page);
1682         return error;
1683 }
1684
1685 static int
1686 shmem_write_begin(struct file *file, struct address_space *mapping,
1687                         loff_t pos, unsigned len, unsigned flags,
1688                         struct page **pagep, void **fsdata)
1689 {
1690         struct inode *inode = mapping->host;
1691         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1692         *pagep = NULL;
1693         return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1694 }
1695
1696 static int
1697 shmem_write_end(struct file *file, struct address_space *mapping,
1698                         loff_t pos, unsigned len, unsigned copied,
1699                         struct page *page, void *fsdata)
1700 {
1701         struct inode *inode = mapping->host;
1702
1703         if (pos + copied > inode->i_size)
1704                 i_size_write(inode, pos + copied);
1705
1706         set_page_dirty(page);
1707         unlock_page(page);
1708         page_cache_release(page);
1709
1710         return copied;
1711 }
1712
1713 static void do_shmem_file_read(struct file *filp, loff_t *ppos, read_descriptor_t *desc, read_actor_t actor)
1714 {
1715         struct inode *inode = filp->f_path.dentry->d_inode;
1716         struct address_space *mapping = inode->i_mapping;
1717         unsigned long index, offset;
1718         enum sgp_type sgp = SGP_READ;
1719
1720         /*
1721          * Might this read be for a stacking filesystem?  Then when reading
1722          * holes of a sparse file, we actually need to allocate those pages,
1723          * and even mark them dirty, so it cannot exceed the max_blocks limit.
1724          */
1725         if (segment_eq(get_fs(), KERNEL_DS))
1726                 sgp = SGP_DIRTY;
1727
1728         index = *ppos >> PAGE_CACHE_SHIFT;
1729         offset = *ppos & ~PAGE_CACHE_MASK;
1730
1731         for (;;) {
1732                 struct page *page = NULL;
1733                 unsigned long end_index, nr, ret;
1734                 loff_t i_size = i_size_read(inode);
1735
1736                 end_index = i_size >> PAGE_CACHE_SHIFT;
1737                 if (index > end_index)
1738                         break;
1739                 if (index == end_index) {
1740                         nr = i_size & ~PAGE_CACHE_MASK;
1741                         if (nr <= offset)
1742                                 break;
1743                 }
1744
1745                 desc->error = shmem_getpage(inode, index, &page, sgp, NULL);
1746                 if (desc->error) {
1747                         if (desc->error == -EINVAL)
1748                                 desc->error = 0;
1749                         break;
1750                 }
1751                 if (page)
1752                         unlock_page(page);
1753
1754                 /*
1755                  * We must evaluate after, since reads (unlike writes)
1756                  * are called without i_mutex protection against truncate
1757                  */
1758                 nr = PAGE_CACHE_SIZE;
1759                 i_size = i_size_read(inode);
1760                 end_index = i_size >> PAGE_CACHE_SHIFT;
1761                 if (index == end_index) {
1762                         nr = i_size & ~PAGE_CACHE_MASK;
1763                         if (nr <= offset) {
1764                                 if (page)
1765                                         page_cache_release(page);
1766                                 break;
1767                         }
1768                 }
1769                 nr -= offset;
1770
1771                 if (page) {
1772                         /*
1773                          * If users can be writing to this page using arbitrary
1774                          * virtual addresses, take care about potential aliasing
1775                          * before reading the page on the kernel side.
1776                          */
1777                         if (mapping_writably_mapped(mapping))
1778                                 flush_dcache_page(page);
1779                         /*
1780                          * Mark the page accessed if we read the beginning.
1781                          */
1782                         if (!offset)
1783                                 mark_page_accessed(page);
1784                 } else {
1785                         page = ZERO_PAGE(0);
1786                         page_cache_get(page);
1787                 }
1788
1789                 /*
1790                  * Ok, we have the page, and it's up-to-date, so
1791                  * now we can copy it to user space...
1792                  *
1793                  * The actor routine returns how many bytes were actually used..
1794                  * NOTE! This may not be the same as how much of a user buffer
1795                  * we filled up (we may be padding etc), so we can only update
1796                  * "pos" here (the actor routine has to update the user buffer
1797                  * pointers and the remaining count).
1798                  */
1799                 ret = actor(desc, page, offset, nr);
1800                 offset += ret;
1801                 index += offset >> PAGE_CACHE_SHIFT;
1802                 offset &= ~PAGE_CACHE_MASK;
1803
1804                 page_cache_release(page);
1805                 if (ret != nr || !desc->count)
1806                         break;
1807
1808                 cond_resched();
1809         }
1810
1811         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1812         file_accessed(filp);
1813 }
1814
1815 static ssize_t shmem_file_aio_read(struct kiocb *iocb,
1816                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
1817 {
1818         struct file *filp = iocb->ki_filp;
1819         ssize_t retval;
1820         unsigned long seg;
1821         size_t count;
1822         loff_t *ppos = &iocb->ki_pos;
1823
1824         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1825         if (retval)
1826                 return retval;
1827
1828         for (seg = 0; seg < nr_segs; seg++) {
1829                 read_descriptor_t desc;
1830
1831                 desc.written = 0;
1832                 desc.arg.buf = iov[seg].iov_base;
1833                 desc.count = iov[seg].iov_len;
1834                 if (desc.count == 0)
1835                         continue;
1836                 desc.error = 0;
1837                 do_shmem_file_read(filp, ppos, &desc, file_read_actor);
1838                 retval += desc.written;
1839                 if (desc.error) {
1840                         retval = retval ?: desc.error;
1841                         break;
1842                 }
1843                 if (desc.count > 0)
1844                         break;
1845         }
1846         return retval;
1847 }
1848
1849 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
1850 {
1851         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
1852
1853         buf->f_type = TMPFS_MAGIC;
1854         buf->f_bsize = PAGE_CACHE_SIZE;
1855         buf->f_namelen = NAME_MAX;
1856         if (sbinfo->max_blocks) {
1857                 buf->f_blocks = sbinfo->max_blocks;
1858                 buf->f_bavail = buf->f_bfree =
1859                                 sbinfo->max_blocks - percpu_counter_sum(&sbinfo->used_blocks);
1860         }
1861         if (sbinfo->max_inodes) {
1862                 buf->f_files = sbinfo->max_inodes;
1863                 buf->f_ffree = sbinfo->free_inodes;
1864         }
1865         /* else leave those fields 0 like simple_statfs */
1866         return 0;
1867 }
1868
1869 /*
1870  * File creation. Allocate an inode, and we're done..
1871  */
1872 static int
1873 shmem_mknod(struct inode *dir, struct dentry *dentry, int mode, dev_t dev)
1874 {
1875         struct inode *inode;
1876         int error = -ENOSPC;
1877
1878         inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
1879         if (inode) {
1880                 error = security_inode_init_security(inode, dir,
1881                                                      &dentry->d_name, NULL,
1882                                                      NULL, NULL);
1883                 if (error) {
1884                         if (error != -EOPNOTSUPP) {
1885                                 iput(inode);
1886                                 return error;
1887                         }
1888                 }
1889 #ifdef CONFIG_TMPFS_POSIX_ACL
1890                 error = generic_acl_init(inode, dir);
1891                 if (error) {
1892                         iput(inode);
1893                         return error;
1894                 }
1895 #else
1896                 error = 0;
1897 #endif
1898                 dir->i_size += BOGO_DIRENT_SIZE;
1899                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1900                 d_instantiate(dentry, inode);
1901                 dget(dentry); /* Extra count - pin the dentry in core */
1902         }
1903         return error;
1904 }
1905
1906 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, int mode)
1907 {
1908         int error;
1909
1910         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
1911                 return error;
1912         inc_nlink(dir);
1913         return 0;
1914 }
1915
1916 static int shmem_create(struct inode *dir, struct dentry *dentry, int mode,
1917                 struct nameidata *nd)
1918 {
1919         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
1920 }
1921
1922 /*
1923  * Link a file..
1924  */
1925 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
1926 {
1927         struct inode *inode = old_dentry->d_inode;
1928         int ret;
1929
1930         /*
1931          * No ordinary (disk based) filesystem counts links as inodes;
1932          * but each new link needs a new dentry, pinning lowmem, and
1933          * tmpfs dentries cannot be pruned until they are unlinked.
1934          */
1935         ret = shmem_reserve_inode(inode->i_sb);
1936         if (ret)
1937                 goto out;
1938
1939         dir->i_size += BOGO_DIRENT_SIZE;
1940         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1941         inc_nlink(inode);
1942         ihold(inode);   /* New dentry reference */
1943         dget(dentry);           /* Extra pinning count for the created dentry */
1944         d_instantiate(dentry, inode);
1945 out:
1946         return ret;
1947 }
1948
1949 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
1950 {
1951         struct inode *inode = dentry->d_inode;
1952
1953         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
1954                 shmem_free_inode(inode->i_sb);
1955
1956         dir->i_size -= BOGO_DIRENT_SIZE;
1957         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1958         drop_nlink(inode);
1959         dput(dentry);   /* Undo the count from "create" - this does all the work */
1960         return 0;
1961 }
1962
1963 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
1964 {
1965         if (!simple_empty(dentry))
1966                 return -ENOTEMPTY;
1967
1968         drop_nlink(dentry->d_inode);
1969         drop_nlink(dir);
1970         return shmem_unlink(dir, dentry);
1971 }
1972
1973 /*
1974  * The VFS layer already does all the dentry stuff for rename,
1975  * we just have to decrement the usage count for the target if
1976  * it exists so that the VFS layer correctly free's it when it
1977  * gets overwritten.
1978  */
1979 static int shmem_rename(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
1980 {
1981         struct inode *inode = old_dentry->d_inode;
1982         int they_are_dirs = S_ISDIR(inode->i_mode);
1983
1984         if (!simple_empty(new_dentry))
1985                 return -ENOTEMPTY;
1986
1987         if (new_dentry->d_inode) {
1988                 (void) shmem_unlink(new_dir, new_dentry);
1989                 if (they_are_dirs)
1990                         drop_nlink(old_dir);
1991         } else if (they_are_dirs) {
1992                 drop_nlink(old_dir);
1993                 inc_nlink(new_dir);
1994         }
1995
1996         old_dir->i_size -= BOGO_DIRENT_SIZE;
1997         new_dir->i_size += BOGO_DIRENT_SIZE;
1998         old_dir->i_ctime = old_dir->i_mtime =
1999         new_dir->i_ctime = new_dir->i_mtime =
2000         inode->i_ctime = CURRENT_TIME;
2001         return 0;
2002 }
2003
2004 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
2005 {
2006         int error;
2007         int len;
2008         struct inode *inode;
2009         struct page *page = NULL;
2010         char *kaddr;
2011         struct shmem_inode_info *info;
2012
2013         len = strlen(symname) + 1;
2014         if (len > PAGE_CACHE_SIZE)
2015                 return -ENAMETOOLONG;
2016
2017         inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0, VM_NORESERVE);
2018         if (!inode)
2019                 return -ENOSPC;
2020
2021         error = security_inode_init_security(inode, dir, &dentry->d_name, NULL,
2022                                              NULL, NULL);
2023         if (error) {
2024                 if (error != -EOPNOTSUPP) {
2025                         iput(inode);
2026                         return error;
2027                 }
2028                 error = 0;
2029         }
2030
2031         info = SHMEM_I(inode);
2032         inode->i_size = len-1;
2033         if (len <= SHMEM_SYMLINK_INLINE_LEN) {
2034                 /* do it inline */
2035                 memcpy(info->inline_symlink, symname, len);
2036                 inode->i_op = &shmem_symlink_inline_operations;
2037         } else {
2038                 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
2039                 if (error) {
2040                         iput(inode);
2041                         return error;
2042                 }
2043                 inode->i_mapping->a_ops = &shmem_aops;
2044                 inode->i_op = &shmem_symlink_inode_operations;
2045                 kaddr = kmap_atomic(page, KM_USER0);
2046                 memcpy(kaddr, symname, len);
2047                 kunmap_atomic(kaddr, KM_USER0);
2048                 set_page_dirty(page);
2049                 unlock_page(page);
2050                 page_cache_release(page);
2051         }
2052         dir->i_size += BOGO_DIRENT_SIZE;
2053         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2054         d_instantiate(dentry, inode);
2055         dget(dentry);
2056         return 0;
2057 }
2058
2059 static void *shmem_follow_link_inline(struct dentry *dentry, struct nameidata *nd)
2060 {
2061         nd_set_link(nd, SHMEM_I(dentry->d_inode)->inline_symlink);
2062         return NULL;
2063 }
2064
2065 static void *shmem_follow_link(struct dentry *dentry, struct nameidata *nd)
2066 {
2067         struct page *page = NULL;
2068         int res = shmem_getpage(dentry->d_inode, 0, &page, SGP_READ, NULL);
2069         nd_set_link(nd, res ? ERR_PTR(res) : kmap(page));
2070         if (page)
2071                 unlock_page(page);
2072         return page;
2073 }
2074
2075 static void shmem_put_link(struct dentry *dentry, struct nameidata *nd, void *cookie)
2076 {
2077         if (!IS_ERR(nd_get_link(nd))) {
2078                 struct page *page = cookie;
2079                 kunmap(page);
2080                 mark_page_accessed(page);
2081                 page_cache_release(page);
2082         }
2083 }
2084
2085 #ifdef CONFIG_TMPFS_XATTR
2086 /*
2087  * Superblocks without xattr inode operations may get some security.* xattr
2088  * support from the LSM "for free". As soon as we have any other xattrs
2089  * like ACLs, we also need to implement the security.* handlers at
2090  * filesystem level, though.
2091  */
2092
2093 static int shmem_xattr_get(struct dentry *dentry, const char *name,
2094                            void *buffer, size_t size)
2095 {
2096         struct shmem_inode_info *info;
2097         struct shmem_xattr *xattr;
2098         int ret = -ENODATA;
2099
2100         info = SHMEM_I(dentry->d_inode);
2101
2102         spin_lock(&info->lock);
2103         list_for_each_entry(xattr, &info->xattr_list, list) {
2104                 if (strcmp(name, xattr->name))
2105                         continue;
2106
2107                 ret = xattr->size;
2108                 if (buffer) {
2109                         if (size < xattr->size)
2110                                 ret = -ERANGE;
2111                         else
2112                                 memcpy(buffer, xattr->value, xattr->size);
2113                 }
2114                 break;
2115         }
2116         spin_unlock(&info->lock);
2117         return ret;
2118 }
2119
2120 static int shmem_xattr_set(struct dentry *dentry, const char *name,
2121                            const void *value, size_t size, int flags)
2122 {
2123         struct inode *inode = dentry->d_inode;
2124         struct shmem_inode_info *info = SHMEM_I(inode);
2125         struct shmem_xattr *xattr;
2126         struct shmem_xattr *new_xattr = NULL;
2127         size_t len;
2128         int err = 0;
2129
2130         /* value == NULL means remove */
2131         if (value) {
2132                 /* wrap around? */
2133                 len = sizeof(*new_xattr) + size;
2134                 if (len <= sizeof(*new_xattr))
2135                         return -ENOMEM;
2136
2137                 new_xattr = kmalloc(len, GFP_KERNEL);
2138                 if (!new_xattr)
2139                         return -ENOMEM;
2140
2141                 new_xattr->name = kstrdup(name, GFP_KERNEL);
2142                 if (!new_xattr->name) {
2143                         kfree(new_xattr);
2144                         return -ENOMEM;
2145                 }
2146
2147                 new_xattr->size = size;
2148                 memcpy(new_xattr->value, value, size);
2149         }
2150
2151         spin_lock(&info->lock);
2152         list_for_each_entry(xattr, &info->xattr_list, list) {
2153                 if (!strcmp(name, xattr->name)) {
2154                         if (flags & XATTR_CREATE) {
2155                                 xattr = new_xattr;
2156                                 err = -EEXIST;
2157                         } else if (new_xattr) {
2158                                 list_replace(&xattr->list, &new_xattr->list);
2159                         } else {
2160                                 list_del(&xattr->list);
2161                         }
2162                         goto out;
2163                 }
2164         }
2165         if (flags & XATTR_REPLACE) {
2166                 xattr = new_xattr;
2167                 err = -ENODATA;
2168         } else {
2169                 list_add(&new_xattr->list, &info->xattr_list);
2170                 xattr = NULL;
2171         }
2172 out:
2173         spin_unlock(&info->lock);
2174         if (xattr)
2175                 kfree(xattr->name);
2176         kfree(xattr);
2177         return err;
2178 }
2179
2180
2181 static const struct xattr_handler *shmem_xattr_handlers[] = {
2182 #ifdef CONFIG_TMPFS_POSIX_ACL
2183         &generic_acl_access_handler,
2184         &generic_acl_default_handler,
2185 #endif
2186         NULL
2187 };
2188
2189 static int shmem_xattr_validate(const char *name)
2190 {
2191         struct { const char *prefix; size_t len; } arr[] = {
2192                 { XATTR_SECURITY_PREFIX, XATTR_SECURITY_PREFIX_LEN },
2193                 { XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN }
2194         };
2195         int i;
2196
2197         for (i = 0; i < ARRAY_SIZE(arr); i++) {
2198                 size_t preflen = arr[i].len;
2199                 if (strncmp(name, arr[i].prefix, preflen) == 0) {
2200                         if (!name[preflen])
2201                                 return -EINVAL;
2202                         return 0;
2203                 }
2204         }
2205         return -EOPNOTSUPP;
2206 }
2207
2208 static ssize_t shmem_getxattr(struct dentry *dentry, const char *name,
2209                               void *buffer, size_t size)
2210 {
2211         int err;
2212
2213         /*
2214          * If this is a request for a synthetic attribute in the system.*
2215          * namespace use the generic infrastructure to resolve a handler
2216          * for it via sb->s_xattr.
2217          */
2218         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2219                 return generic_getxattr(dentry, name, buffer, size);
2220
2221         err = shmem_xattr_validate(name);
2222         if (err)
2223                 return err;
2224
2225         return shmem_xattr_get(dentry, name, buffer, size);
2226 }
2227
2228 static int shmem_setxattr(struct dentry *dentry, const char *name,
2229                           const void *value, size_t size, int flags)
2230 {
2231         int err;
2232
2233         /*
2234          * If this is a request for a synthetic attribute in the system.*
2235          * namespace use the generic infrastructure to resolve a handler
2236          * for it via sb->s_xattr.
2237          */
2238         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2239                 return generic_setxattr(dentry, name, value, size, flags);
2240
2241         err = shmem_xattr_validate(name);
2242         if (err)
2243                 return err;
2244
2245         if (size == 0)
2246                 value = "";  /* empty EA, do not remove */
2247
2248         return shmem_xattr_set(dentry, name, value, size, flags);
2249
2250 }
2251
2252 static int shmem_removexattr(struct dentry *dentry, const char *name)
2253 {
2254         int err;
2255
2256         /*
2257          * If this is a request for a synthetic attribute in the system.*
2258          * namespace use the generic infrastructure to resolve a handler
2259          * for it via sb->s_xattr.
2260          */
2261         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2262                 return generic_removexattr(dentry, name);
2263
2264         err = shmem_xattr_validate(name);
2265         if (err)
2266                 return err;
2267
2268         return shmem_xattr_set(dentry, name, NULL, 0, XATTR_REPLACE);
2269 }
2270
2271 static bool xattr_is_trusted(const char *name)
2272 {
2273         return !strncmp(name, XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN);
2274 }
2275
2276 static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
2277 {
2278         bool trusted = capable(CAP_SYS_ADMIN);
2279         struct shmem_xattr *xattr;
2280         struct shmem_inode_info *info;
2281         size_t used = 0;
2282
2283         info = SHMEM_I(dentry->d_inode);
2284
2285         spin_lock(&info->lock);
2286         list_for_each_entry(xattr, &info->xattr_list, list) {
2287                 size_t len;
2288
2289                 /* skip "trusted." attributes for unprivileged callers */
2290                 if (!trusted && xattr_is_trusted(xattr->name))
2291                         continue;
2292
2293                 len = strlen(xattr->name) + 1;
2294                 used += len;
2295                 if (buffer) {
2296                         if (size < used) {
2297                                 used = -ERANGE;
2298                                 break;
2299                         }
2300                         memcpy(buffer, xattr->name, len);
2301                         buffer += len;
2302                 }
2303         }
2304         spin_unlock(&info->lock);
2305
2306         return used;
2307 }
2308 #endif /* CONFIG_TMPFS_XATTR */
2309
2310 static const struct inode_operations shmem_symlink_inline_operations = {
2311         .readlink       = generic_readlink,
2312         .follow_link    = shmem_follow_link_inline,
2313 #ifdef CONFIG_TMPFS_XATTR
2314         .setxattr       = shmem_setxattr,
2315         .getxattr       = shmem_getxattr,
2316         .listxattr      = shmem_listxattr,
2317         .removexattr    = shmem_removexattr,
2318 #endif
2319 };
2320
2321 static const struct inode_operations shmem_symlink_inode_operations = {
2322         .readlink       = generic_readlink,
2323         .follow_link    = shmem_follow_link,
2324         .put_link       = shmem_put_link,
2325 #ifdef CONFIG_TMPFS_XATTR
2326         .setxattr       = shmem_setxattr,
2327         .getxattr       = shmem_getxattr,
2328         .listxattr      = shmem_listxattr,
2329         .removexattr    = shmem_removexattr,
2330 #endif
2331 };
2332
2333 static struct dentry *shmem_get_parent(struct dentry *child)
2334 {
2335         return ERR_PTR(-ESTALE);
2336 }
2337
2338 static int shmem_match(struct inode *ino, void *vfh)
2339 {
2340         __u32 *fh = vfh;
2341         __u64 inum = fh[2];
2342         inum = (inum << 32) | fh[1];
2343         return ino->i_ino == inum && fh[0] == ino->i_generation;
2344 }
2345
2346 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
2347                 struct fid *fid, int fh_len, int fh_type)
2348 {
2349         struct inode *inode;
2350         struct dentry *dentry = NULL;
2351         u64 inum = fid->raw[2];
2352         inum = (inum << 32) | fid->raw[1];
2353
2354         if (fh_len < 3)
2355                 return NULL;
2356
2357         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
2358                         shmem_match, fid->raw);
2359         if (inode) {
2360                 dentry = d_find_alias(inode);
2361                 iput(inode);
2362         }
2363
2364         return dentry;
2365 }
2366
2367 static int shmem_encode_fh(struct dentry *dentry, __u32 *fh, int *len,
2368                                 int connectable)
2369 {
2370         struct inode *inode = dentry->d_inode;
2371
2372         if (*len < 3) {
2373                 *len = 3;
2374                 return 255;
2375         }
2376
2377         if (inode_unhashed(inode)) {
2378                 /* Unfortunately insert_inode_hash is not idempotent,
2379                  * so as we hash inodes here rather than at creation
2380                  * time, we need a lock to ensure we only try
2381                  * to do it once
2382                  */
2383                 static DEFINE_SPINLOCK(lock);
2384                 spin_lock(&lock);
2385                 if (inode_unhashed(inode))
2386                         __insert_inode_hash(inode,
2387                                             inode->i_ino + inode->i_generation);
2388                 spin_unlock(&lock);
2389         }
2390
2391         fh[0] = inode->i_generation;
2392         fh[1] = inode->i_ino;
2393         fh[2] = ((__u64)inode->i_ino) >> 32;
2394
2395         *len = 3;
2396         return 1;
2397 }
2398
2399 static const struct export_operations shmem_export_ops = {
2400         .get_parent     = shmem_get_parent,
2401         .encode_fh      = shmem_encode_fh,
2402         .fh_to_dentry   = shmem_fh_to_dentry,
2403 };
2404
2405 static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
2406                                bool remount)
2407 {
2408         char *this_char, *value, *rest;
2409
2410         while (options != NULL) {
2411                 this_char = options;
2412                 for (;;) {
2413                         /*
2414                          * NUL-terminate this option: unfortunately,
2415                          * mount options form a comma-separated list,
2416                          * but mpol's nodelist may also contain commas.
2417                          */
2418                         options = strchr(options, ',');
2419                         if (options == NULL)
2420                                 break;
2421                         options++;
2422                         if (!isdigit(*options)) {
2423                                 options[-1] = '\0';
2424                                 break;
2425                         }
2426                 }
2427                 if (!*this_char)
2428                         continue;
2429                 if ((value = strchr(this_char,'=')) != NULL) {
2430                         *value++ = 0;
2431                 } else {
2432                         printk(KERN_ERR
2433                             "tmpfs: No value for mount option '%s'\n",
2434                             this_char);
2435                         return 1;
2436                 }
2437
2438                 if (!strcmp(this_char,"size")) {
2439                         unsigned long long size;
2440                         size = memparse(value,&rest);
2441                         if (*rest == '%') {
2442                                 size <<= PAGE_SHIFT;
2443                                 size *= totalram_pages;
2444                                 do_div(size, 100);
2445                                 rest++;
2446                         }
2447                         if (*rest)
2448                                 goto bad_val;
2449                         sbinfo->max_blocks =
2450                                 DIV_ROUND_UP(size, PAGE_CACHE_SIZE);
2451                 } else if (!strcmp(this_char,"nr_blocks")) {
2452                         sbinfo->max_blocks = memparse(value, &rest);
2453                         if (*rest)
2454                                 goto bad_val;
2455                 } else if (!strcmp(this_char,"nr_inodes")) {
2456                         sbinfo->max_inodes = memparse(value, &rest);
2457                         if (*rest)
2458                                 goto bad_val;
2459                 } else if (!strcmp(this_char,"mode")) {
2460                         if (remount)
2461                                 continue;
2462                         sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
2463                         if (*rest)
2464                                 goto bad_val;
2465                 } else if (!strcmp(this_char,"uid")) {
2466                         if (remount)
2467                                 continue;
2468                         sbinfo->uid = simple_strtoul(value, &rest, 0);
2469                         if (*rest)
2470                                 goto bad_val;
2471                 } else if (!strcmp(this_char,"gid")) {
2472                         if (remount)
2473                                 continue;
2474                         sbinfo->gid = simple_strtoul(value, &rest, 0);
2475                         if (*rest)
2476                                 goto bad_val;
2477                 } else if (!strcmp(this_char,"mpol")) {
2478                         if (mpol_parse_str(value, &sbinfo->mpol, 1))
2479                                 goto bad_val;
2480                 } else {
2481                         printk(KERN_ERR "tmpfs: Bad mount option %s\n",
2482                                this_char);
2483                         return 1;
2484                 }
2485         }
2486         return 0;
2487
2488 bad_val:
2489         printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
2490                value, this_char);
2491         return 1;
2492
2493 }
2494
2495 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2496 {
2497         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2498         struct shmem_sb_info config = *sbinfo;
2499         unsigned long inodes;
2500         int error = -EINVAL;
2501
2502         if (shmem_parse_options(data, &config, true))
2503                 return error;
2504
2505         spin_lock(&sbinfo->stat_lock);
2506         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2507         if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
2508                 goto out;
2509         if (config.max_inodes < inodes)
2510                 goto out;
2511         /*
2512          * Those tests also disallow limited->unlimited while any are in
2513          * use, so i_blocks will always be zero when max_blocks is zero;
2514          * but we must separately disallow unlimited->limited, because
2515          * in that case we have no record of how much is already in use.
2516          */
2517         if (config.max_blocks && !sbinfo->max_blocks)
2518                 goto out;
2519         if (config.max_inodes && !sbinfo->max_inodes)
2520                 goto out;
2521
2522         error = 0;
2523         sbinfo->max_blocks  = config.max_blocks;
2524         sbinfo->max_inodes  = config.max_inodes;
2525         sbinfo->free_inodes = config.max_inodes - inodes;
2526
2527         mpol_put(sbinfo->mpol);
2528         sbinfo->mpol        = config.mpol;      /* transfers initial ref */
2529 out:
2530         spin_unlock(&sbinfo->stat_lock);
2531         return error;
2532 }
2533
2534 static int shmem_show_options(struct seq_file *seq, struct vfsmount *vfs)
2535 {
2536         struct shmem_sb_info *sbinfo = SHMEM_SB(vfs->mnt_sb);
2537
2538         if (sbinfo->max_blocks != shmem_default_max_blocks())
2539                 seq_printf(seq, ",size=%luk",
2540                         sbinfo->max_blocks << (PAGE_CACHE_SHIFT - 10));
2541         if (sbinfo->max_inodes != shmem_default_max_inodes())
2542                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
2543         if (sbinfo->mode != (S_IRWXUGO | S_ISVTX))
2544                 seq_printf(seq, ",mode=%03o", sbinfo->mode);
2545         if (sbinfo->uid != 0)
2546                 seq_printf(seq, ",uid=%u", sbinfo->uid);
2547         if (sbinfo->gid != 0)
2548                 seq_printf(seq, ",gid=%u", sbinfo->gid);
2549         shmem_show_mpol(seq, sbinfo->mpol);
2550         return 0;
2551 }
2552 #endif /* CONFIG_TMPFS */
2553
2554 static void shmem_put_super(struct super_block *sb)
2555 {
2556         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2557
2558         percpu_counter_destroy(&sbinfo->used_blocks);
2559         kfree(sbinfo);
2560         sb->s_fs_info = NULL;
2561 }
2562
2563 int shmem_fill_super(struct super_block *sb, void *data, int silent)
2564 {
2565         struct inode *inode;
2566         struct dentry *root;
2567         struct shmem_sb_info *sbinfo;
2568         int err = -ENOMEM;
2569
2570         /* Round up to L1_CACHE_BYTES to resist false sharing */
2571         sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
2572                                 L1_CACHE_BYTES), GFP_KERNEL);
2573         if (!sbinfo)
2574                 return -ENOMEM;
2575
2576         sbinfo->mode = S_IRWXUGO | S_ISVTX;
2577         sbinfo->uid = current_fsuid();
2578         sbinfo->gid = current_fsgid();
2579         sb->s_fs_info = sbinfo;
2580
2581 #ifdef CONFIG_TMPFS
2582         /*
2583          * Per default we only allow half of the physical ram per
2584          * tmpfs instance, limiting inodes to one per page of lowmem;
2585          * but the internal instance is left unlimited.
2586          */
2587         if (!(sb->s_flags & MS_NOUSER)) {
2588                 sbinfo->max_blocks = shmem_default_max_blocks();
2589                 sbinfo->max_inodes = shmem_default_max_inodes();
2590                 if (shmem_parse_options(data, sbinfo, false)) {
2591                         err = -EINVAL;
2592                         goto failed;
2593                 }
2594         }
2595         sb->s_export_op = &shmem_export_ops;
2596 #else
2597         sb->s_flags |= MS_NOUSER;
2598 #endif
2599
2600         spin_lock_init(&sbinfo->stat_lock);
2601         if (percpu_counter_init(&sbinfo->used_blocks, 0))
2602                 goto failed;
2603         sbinfo->free_inodes = sbinfo->max_inodes;
2604
2605         sb->s_maxbytes = SHMEM_MAX_BYTES;
2606         sb->s_blocksize = PAGE_CACHE_SIZE;
2607         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
2608         sb->s_magic = TMPFS_MAGIC;
2609         sb->s_op = &shmem_ops;
2610         sb->s_time_gran = 1;
2611 #ifdef CONFIG_TMPFS_XATTR
2612         sb->s_xattr = shmem_xattr_handlers;
2613 #endif
2614 #ifdef CONFIG_TMPFS_POSIX_ACL
2615         sb->s_flags |= MS_POSIXACL;
2616 #endif
2617
2618         inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
2619         if (!inode)
2620                 goto failed;
2621         inode->i_uid = sbinfo->uid;
2622         inode->i_gid = sbinfo->gid;
2623         root = d_alloc_root(inode);
2624         if (!root)
2625                 goto failed_iput;
2626         sb->s_root = root;
2627         return 0;
2628
2629 failed_iput:
2630         iput(inode);
2631 failed:
2632         shmem_put_super(sb);
2633         return err;
2634 }
2635
2636 static struct kmem_cache *shmem_inode_cachep;
2637
2638 static struct inode *shmem_alloc_inode(struct super_block *sb)
2639 {
2640         struct shmem_inode_info *p;
2641         p = (struct shmem_inode_info *)kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
2642         if (!p)
2643                 return NULL;
2644         return &p->vfs_inode;
2645 }
2646
2647 static void shmem_i_callback(struct rcu_head *head)
2648 {
2649         struct inode *inode = container_of(head, struct inode, i_rcu);
2650         INIT_LIST_HEAD(&inode->i_dentry);
2651         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
2652 }
2653
2654 static void shmem_destroy_inode(struct inode *inode)
2655 {
2656         if ((inode->i_mode & S_IFMT) == S_IFREG) {
2657                 /* only struct inode is valid if it's an inline symlink */
2658                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
2659         }
2660         call_rcu(&inode->i_rcu, shmem_i_callback);
2661 }
2662
2663 static void init_once(void *foo)
2664 {
2665         struct shmem_inode_info *p = (struct shmem_inode_info *) foo;
2666
2667         inode_init_once(&p->vfs_inode);
2668 }
2669
2670 static int init_inodecache(void)
2671 {
2672         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
2673                                 sizeof(struct shmem_inode_info),
2674                                 0, SLAB_PANIC, init_once);
2675         return 0;
2676 }
2677
2678 static void destroy_inodecache(void)
2679 {
2680         kmem_cache_destroy(shmem_inode_cachep);
2681 }
2682
2683 static const struct address_space_operations shmem_aops = {
2684         .writepage      = shmem_writepage,
2685         .set_page_dirty = __set_page_dirty_no_writeback,
2686 #ifdef CONFIG_TMPFS
2687         .readpage       = shmem_readpage,
2688         .write_begin    = shmem_write_begin,
2689         .write_end      = shmem_write_end,
2690 #endif
2691         .migratepage    = migrate_page,
2692         .error_remove_page = generic_error_remove_page,
2693 };
2694
2695 static const struct file_operations shmem_file_operations = {
2696         .mmap           = shmem_mmap,
2697 #ifdef CONFIG_TMPFS
2698         .llseek         = generic_file_llseek,
2699         .read           = do_sync_read,
2700         .write          = do_sync_write,
2701         .aio_read       = shmem_file_aio_read,
2702         .aio_write      = generic_file_aio_write,
2703         .fsync          = noop_fsync,
2704         .splice_read    = generic_file_splice_read,
2705         .splice_write   = generic_file_splice_write,
2706 #endif
2707 };
2708
2709 static const struct inode_operations shmem_inode_operations = {
2710         .setattr        = shmem_setattr,
2711         .truncate_range = shmem_truncate_range,
2712 #ifdef CONFIG_TMPFS_XATTR
2713         .setxattr       = shmem_setxattr,
2714         .getxattr       = shmem_getxattr,
2715         .listxattr      = shmem_listxattr,
2716         .removexattr    = shmem_removexattr,
2717 #endif
2718 #ifdef CONFIG_TMPFS_POSIX_ACL
2719         .check_acl      = generic_check_acl,
2720 #endif
2721
2722 };
2723
2724 static const struct inode_operations shmem_dir_inode_operations = {
2725 #ifdef CONFIG_TMPFS
2726         .create         = shmem_create,
2727         .lookup         = simple_lookup,
2728         .link           = shmem_link,
2729         .unlink         = shmem_unlink,
2730         .symlink        = shmem_symlink,
2731         .mkdir          = shmem_mkdir,
2732         .rmdir          = shmem_rmdir,
2733         .mknod          = shmem_mknod,
2734         .rename         = shmem_rename,
2735 #endif
2736 #ifdef CONFIG_TMPFS_XATTR
2737         .setxattr       = shmem_setxattr,
2738         .getxattr       = shmem_getxattr,
2739         .listxattr      = shmem_listxattr,
2740         .removexattr    = shmem_removexattr,
2741 #endif
2742 #ifdef CONFIG_TMPFS_POSIX_ACL
2743         .setattr        = shmem_setattr,
2744         .check_acl      = generic_check_acl,
2745 #endif
2746 };
2747
2748 static const struct inode_operations shmem_special_inode_operations = {
2749 #ifdef CONFIG_TMPFS_XATTR
2750         .setxattr       = shmem_setxattr,
2751         .getxattr       = shmem_getxattr,
2752         .listxattr      = shmem_listxattr,
2753         .removexattr    = shmem_removexattr,
2754 #endif
2755 #ifdef CONFIG_TMPFS_POSIX_ACL
2756         .setattr        = shmem_setattr,
2757         .check_acl      = generic_check_acl,
2758 #endif
2759 };
2760
2761 static const struct super_operations shmem_ops = {
2762         .alloc_inode    = shmem_alloc_inode,
2763         .destroy_inode  = shmem_destroy_inode,
2764 #ifdef CONFIG_TMPFS
2765         .statfs         = shmem_statfs,
2766         .remount_fs     = shmem_remount_fs,
2767         .show_options   = shmem_show_options,
2768 #endif
2769         .evict_inode    = shmem_evict_inode,
2770         .drop_inode     = generic_delete_inode,
2771         .put_super      = shmem_put_super,
2772 };
2773
2774 static const struct vm_operations_struct shmem_vm_ops = {
2775         .fault          = shmem_fault,
2776 #ifdef CONFIG_NUMA
2777         .set_policy     = shmem_set_policy,
2778         .get_policy     = shmem_get_policy,
2779 #endif
2780 };
2781
2782
2783 static struct dentry *shmem_mount(struct file_system_type *fs_type,
2784         int flags, const char *dev_name, void *data)
2785 {
2786         return mount_nodev(fs_type, flags, data, shmem_fill_super);
2787 }
2788
2789 static struct file_system_type tmpfs_fs_type = {
2790         .owner          = THIS_MODULE,
2791         .name           = "tmpfs",
2792         .mount          = shmem_mount,
2793         .kill_sb        = kill_litter_super,
2794 };
2795
2796 int __init init_tmpfs(void)
2797 {
2798         int error;
2799
2800         error = bdi_init(&shmem_backing_dev_info);
2801         if (error)
2802                 goto out4;
2803
2804         error = init_inodecache();
2805         if (error)
2806                 goto out3;
2807
2808         error = register_filesystem(&tmpfs_fs_type);
2809         if (error) {
2810                 printk(KERN_ERR "Could not register tmpfs\n");
2811                 goto out2;
2812         }
2813
2814         shm_mnt = vfs_kern_mount(&tmpfs_fs_type, MS_NOUSER,
2815                                 tmpfs_fs_type.name, NULL);
2816         if (IS_ERR(shm_mnt)) {
2817                 error = PTR_ERR(shm_mnt);
2818                 printk(KERN_ERR "Could not kern_mount tmpfs\n");
2819                 goto out1;
2820         }
2821         return 0;
2822
2823 out1:
2824         unregister_filesystem(&tmpfs_fs_type);
2825 out2:
2826         destroy_inodecache();
2827 out3:
2828         bdi_destroy(&shmem_backing_dev_info);
2829 out4:
2830         shm_mnt = ERR_PTR(error);
2831         return error;
2832 }
2833
2834 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2835 /**
2836  * mem_cgroup_get_shmem_target - find a page or entry assigned to the shmem file
2837  * @inode: the inode to be searched
2838  * @pgoff: the offset to be searched
2839  * @pagep: the pointer for the found page to be stored
2840  * @ent: the pointer for the found swap entry to be stored
2841  *
2842  * If a page is found, refcount of it is incremented. Callers should handle
2843  * these refcount.
2844  */
2845 void mem_cgroup_get_shmem_target(struct inode *inode, pgoff_t pgoff,
2846                                         struct page **pagep, swp_entry_t *ent)
2847 {
2848         swp_entry_t entry = { .val = 0 }, *ptr;
2849         struct page *page = NULL;
2850         struct shmem_inode_info *info = SHMEM_I(inode);
2851
2852         if ((pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
2853                 goto out;
2854
2855         spin_lock(&info->lock);
2856         ptr = shmem_swp_entry(info, pgoff, NULL);
2857 #ifdef CONFIG_SWAP
2858         if (ptr && ptr->val) {
2859                 entry.val = ptr->val;
2860                 page = find_get_page(&swapper_space, entry.val);
2861         } else
2862 #endif
2863                 page = find_get_page(inode->i_mapping, pgoff);
2864         if (ptr)
2865                 shmem_swp_unmap(ptr);
2866         spin_unlock(&info->lock);
2867 out:
2868         *pagep = page;
2869         *ent = entry;
2870 }
2871 #endif
2872
2873 #else /* !CONFIG_SHMEM */
2874
2875 /*
2876  * tiny-shmem: simple shmemfs and tmpfs using ramfs code
2877  *
2878  * This is intended for small system where the benefits of the full
2879  * shmem code (swap-backed and resource-limited) are outweighed by
2880  * their complexity. On systems without swap this code should be
2881  * effectively equivalent, but much lighter weight.
2882  */
2883
2884 #include <linux/ramfs.h>
2885
2886 static struct file_system_type tmpfs_fs_type = {
2887         .name           = "tmpfs",
2888         .mount          = ramfs_mount,
2889         .kill_sb        = kill_litter_super,
2890 };
2891
2892 int __init init_tmpfs(void)
2893 {
2894         BUG_ON(register_filesystem(&tmpfs_fs_type) != 0);
2895
2896         shm_mnt = kern_mount(&tmpfs_fs_type);
2897         BUG_ON(IS_ERR(shm_mnt));
2898
2899         return 0;
2900 }
2901
2902 int shmem_unuse(swp_entry_t entry, struct page *page)
2903 {
2904         return 0;
2905 }
2906
2907 int shmem_lock(struct file *file, int lock, struct user_struct *user)
2908 {
2909         return 0;
2910 }
2911
2912 void shmem_truncate_range(struct inode *inode, loff_t start, loff_t end)
2913 {
2914         truncate_inode_pages_range(inode->i_mapping, start, end);
2915 }
2916 EXPORT_SYMBOL_GPL(shmem_truncate_range);
2917
2918 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2919 /**
2920  * mem_cgroup_get_shmem_target - find a page or entry assigned to the shmem file
2921  * @inode: the inode to be searched
2922  * @pgoff: the offset to be searched
2923  * @pagep: the pointer for the found page to be stored
2924  * @ent: the pointer for the found swap entry to be stored
2925  *
2926  * If a page is found, refcount of it is incremented. Callers should handle
2927  * these refcount.
2928  */
2929 void mem_cgroup_get_shmem_target(struct inode *inode, pgoff_t pgoff,
2930                                         struct page **pagep, swp_entry_t *ent)
2931 {
2932         struct page *page = NULL;
2933
2934         if ((pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
2935                 goto out;
2936         page = find_get_page(inode->i_mapping, pgoff);
2937 out:
2938         *pagep = page;
2939         *ent = (swp_entry_t){ .val = 0 };
2940 }
2941 #endif
2942
2943 #define shmem_vm_ops                            generic_file_vm_ops
2944 #define shmem_file_operations                   ramfs_file_operations
2945 #define shmem_get_inode(sb, dir, mode, dev, flags)      ramfs_get_inode(sb, dir, mode, dev)
2946 #define shmem_acct_size(flags, size)            0
2947 #define shmem_unacct_size(flags, size)          do {} while (0)
2948 #define SHMEM_MAX_BYTES                         MAX_LFS_FILESIZE
2949
2950 #endif /* CONFIG_SHMEM */
2951
2952 /* common code */
2953
2954 /**
2955  * shmem_file_setup - get an unlinked file living in tmpfs
2956  * @name: name for dentry (to be seen in /proc/<pid>/maps
2957  * @size: size to be set for the file
2958  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
2959  */
2960 struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
2961 {
2962         int error;
2963         struct file *file;
2964         struct inode *inode;
2965         struct path path;
2966         struct dentry *root;
2967         struct qstr this;
2968
2969         if (IS_ERR(shm_mnt))
2970                 return (void *)shm_mnt;
2971
2972         if (size < 0 || size > SHMEM_MAX_BYTES)
2973                 return ERR_PTR(-EINVAL);
2974
2975         if (shmem_acct_size(flags, size))
2976                 return ERR_PTR(-ENOMEM);
2977
2978         error = -ENOMEM;
2979         this.name = name;
2980         this.len = strlen(name);
2981         this.hash = 0; /* will go */
2982         root = shm_mnt->mnt_root;
2983         path.dentry = d_alloc(root, &this);
2984         if (!path.dentry)
2985                 goto put_memory;
2986         path.mnt = mntget(shm_mnt);
2987
2988         error = -ENOSPC;
2989         inode = shmem_get_inode(root->d_sb, NULL, S_IFREG | S_IRWXUGO, 0, flags);
2990         if (!inode)
2991                 goto put_dentry;
2992
2993         d_instantiate(path.dentry, inode);
2994         inode->i_size = size;
2995         inode->i_nlink = 0;     /* It is unlinked */
2996 #ifndef CONFIG_MMU
2997         error = ramfs_nommu_expand_for_mapping(inode, size);
2998         if (error)
2999                 goto put_dentry;
3000 #endif
3001
3002         error = -ENFILE;
3003         file = alloc_file(&path, FMODE_WRITE | FMODE_READ,
3004                   &shmem_file_operations);
3005         if (!file)
3006                 goto put_dentry;
3007
3008         return file;
3009
3010 put_dentry:
3011         path_put(&path);
3012 put_memory:
3013         shmem_unacct_size(flags, size);
3014         return ERR_PTR(error);
3015 }
3016 EXPORT_SYMBOL_GPL(shmem_file_setup);
3017
3018 /**
3019  * shmem_zero_setup - setup a shared anonymous mapping
3020  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
3021  */
3022 int shmem_zero_setup(struct vm_area_struct *vma)
3023 {
3024         struct file *file;
3025         loff_t size = vma->vm_end - vma->vm_start;
3026
3027         file = shmem_file_setup("dev/zero", size, vma->vm_flags);
3028         if (IS_ERR(file))
3029                 return PTR_ERR(file);
3030
3031         if (vma->vm_file)
3032                 fput(vma->vm_file);
3033         vma->vm_file = file;
3034         vma->vm_ops = &shmem_vm_ops;
3035         vma->vm_flags |= VM_CAN_NONLINEAR;
3036         return 0;
3037 }
3038
3039 /**
3040  * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
3041  * @mapping:    the page's address_space
3042  * @index:      the page index
3043  * @gfp:        the page allocator flags to use if allocating
3044  *
3045  * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
3046  * with any new page allocations done using the specified allocation flags.
3047  * But read_cache_page_gfp() uses the ->readpage() method: which does not
3048  * suit tmpfs, since it may have pages in swapcache, and needs to find those
3049  * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
3050  *
3051  * Provide a stub for those callers to start using now, then later
3052  * flesh it out to call shmem_getpage() with additional gfp mask, when
3053  * shmem_file_splice_read() is added and shmem_readpage() is removed.
3054  */
3055 struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
3056                                          pgoff_t index, gfp_t gfp)
3057 {
3058         return read_cache_page_gfp(mapping, index, gfp);
3059 }
3060 EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);