memcgroup: fix hang with shmem/tmpfs
[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  * This file is released under the GPL.
18  */
19
20 /*
21  * This virtual memory filesystem is heavily based on the ramfs. It
22  * extends ramfs by the ability to use swap and honor resource limits
23  * which makes it a completely usable filesystem.
24  */
25
26 #include <linux/module.h>
27 #include <linux/init.h>
28 #include <linux/fs.h>
29 #include <linux/xattr.h>
30 #include <linux/exportfs.h>
31 #include <linux/generic_acl.h>
32 #include <linux/mm.h>
33 #include <linux/mman.h>
34 #include <linux/file.h>
35 #include <linux/swap.h>
36 #include <linux/pagemap.h>
37 #include <linux/string.h>
38 #include <linux/slab.h>
39 #include <linux/backing-dev.h>
40 #include <linux/shmem_fs.h>
41 #include <linux/mount.h>
42 #include <linux/writeback.h>
43 #include <linux/vfs.h>
44 #include <linux/blkdev.h>
45 #include <linux/security.h>
46 #include <linux/swapops.h>
47 #include <linux/mempolicy.h>
48 #include <linux/namei.h>
49 #include <linux/ctype.h>
50 #include <linux/migrate.h>
51 #include <linux/highmem.h>
52
53 #include <asm/uaccess.h>
54 #include <asm/div64.h>
55 #include <asm/pgtable.h>
56
57 /* This magic number is used in glibc for posix shared memory */
58 #define TMPFS_MAGIC     0x01021994
59
60 #define ENTRIES_PER_PAGE (PAGE_CACHE_SIZE/sizeof(unsigned long))
61 #define ENTRIES_PER_PAGEPAGE (ENTRIES_PER_PAGE*ENTRIES_PER_PAGE)
62 #define BLOCKS_PER_PAGE  (PAGE_CACHE_SIZE/512)
63
64 #define SHMEM_MAX_INDEX  (SHMEM_NR_DIRECT + (ENTRIES_PER_PAGEPAGE/2) * (ENTRIES_PER_PAGE+1))
65 #define SHMEM_MAX_BYTES  ((unsigned long long)SHMEM_MAX_INDEX << PAGE_CACHE_SHIFT)
66
67 #define VM_ACCT(size)    (PAGE_CACHE_ALIGN(size) >> PAGE_SHIFT)
68
69 /* info->flags needs VM_flags to handle pagein/truncate races efficiently */
70 #define SHMEM_PAGEIN     VM_READ
71 #define SHMEM_TRUNCATE   VM_WRITE
72
73 /* Definition to limit shmem_truncate's steps between cond_rescheds */
74 #define LATENCY_LIMIT    64
75
76 /* Pretend that each entry is of this size in directory's i_size */
77 #define BOGO_DIRENT_SIZE 20
78
79 /* Flag allocation requirements to shmem_getpage and shmem_swp_alloc */
80 enum sgp_type {
81         SGP_READ,       /* don't exceed i_size, don't allocate page */
82         SGP_CACHE,      /* don't exceed i_size, may allocate page */
83         SGP_DIRTY,      /* like SGP_CACHE, but set new page dirty */
84         SGP_WRITE,      /* may exceed i_size, may allocate page */
85 };
86
87 static int shmem_getpage(struct inode *inode, unsigned long idx,
88                          struct page **pagep, enum sgp_type sgp, int *type);
89
90 static inline struct page *shmem_dir_alloc(gfp_t gfp_mask)
91 {
92         /*
93          * The above definition of ENTRIES_PER_PAGE, and the use of
94          * BLOCKS_PER_PAGE on indirect pages, assume PAGE_CACHE_SIZE:
95          * might be reconsidered if it ever diverges from PAGE_SIZE.
96          *
97          * Mobility flags are masked out as swap vectors cannot move
98          */
99         return alloc_pages((gfp_mask & ~GFP_MOVABLE_MASK) | __GFP_ZERO,
100                                 PAGE_CACHE_SHIFT-PAGE_SHIFT);
101 }
102
103 static inline void shmem_dir_free(struct page *page)
104 {
105         __free_pages(page, PAGE_CACHE_SHIFT-PAGE_SHIFT);
106 }
107
108 static struct page **shmem_dir_map(struct page *page)
109 {
110         return (struct page **)kmap_atomic(page, KM_USER0);
111 }
112
113 static inline void shmem_dir_unmap(struct page **dir)
114 {
115         kunmap_atomic(dir, KM_USER0);
116 }
117
118 static swp_entry_t *shmem_swp_map(struct page *page)
119 {
120         return (swp_entry_t *)kmap_atomic(page, KM_USER1);
121 }
122
123 static inline void shmem_swp_balance_unmap(void)
124 {
125         /*
126          * When passing a pointer to an i_direct entry, to code which
127          * also handles indirect entries and so will shmem_swp_unmap,
128          * we must arrange for the preempt count to remain in balance.
129          * What kmap_atomic of a lowmem page does depends on config
130          * and architecture, so pretend to kmap_atomic some lowmem page.
131          */
132         (void) kmap_atomic(ZERO_PAGE(0), KM_USER1);
133 }
134
135 static inline void shmem_swp_unmap(swp_entry_t *entry)
136 {
137         kunmap_atomic(entry, KM_USER1);
138 }
139
140 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
141 {
142         return sb->s_fs_info;
143 }
144
145 /*
146  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
147  * for shared memory and for shared anonymous (/dev/zero) mappings
148  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
149  * consistent with the pre-accounting of private mappings ...
150  */
151 static inline int shmem_acct_size(unsigned long flags, loff_t size)
152 {
153         return (flags & VM_ACCOUNT)?
154                 security_vm_enough_memory(VM_ACCT(size)): 0;
155 }
156
157 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
158 {
159         if (flags & VM_ACCOUNT)
160                 vm_unacct_memory(VM_ACCT(size));
161 }
162
163 /*
164  * ... whereas tmpfs objects are accounted incrementally as
165  * pages are allocated, in order to allow huge sparse files.
166  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
167  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
168  */
169 static inline int shmem_acct_block(unsigned long flags)
170 {
171         return (flags & VM_ACCOUNT)?
172                 0: security_vm_enough_memory(VM_ACCT(PAGE_CACHE_SIZE));
173 }
174
175 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
176 {
177         if (!(flags & VM_ACCOUNT))
178                 vm_unacct_memory(pages * VM_ACCT(PAGE_CACHE_SIZE));
179 }
180
181 static const struct super_operations shmem_ops;
182 static const struct address_space_operations shmem_aops;
183 static const struct file_operations shmem_file_operations;
184 static const struct inode_operations shmem_inode_operations;
185 static const struct inode_operations shmem_dir_inode_operations;
186 static const struct inode_operations shmem_special_inode_operations;
187 static struct vm_operations_struct shmem_vm_ops;
188
189 static struct backing_dev_info shmem_backing_dev_info  __read_mostly = {
190         .ra_pages       = 0,    /* No readahead */
191         .capabilities   = BDI_CAP_NO_ACCT_DIRTY | BDI_CAP_NO_WRITEBACK,
192         .unplug_io_fn   = default_unplug_io_fn,
193 };
194
195 static LIST_HEAD(shmem_swaplist);
196 static DEFINE_MUTEX(shmem_swaplist_mutex);
197
198 static void shmem_free_blocks(struct inode *inode, long pages)
199 {
200         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
201         if (sbinfo->max_blocks) {
202                 spin_lock(&sbinfo->stat_lock);
203                 sbinfo->free_blocks += pages;
204                 inode->i_blocks -= pages*BLOCKS_PER_PAGE;
205                 spin_unlock(&sbinfo->stat_lock);
206         }
207 }
208
209 static int shmem_reserve_inode(struct super_block *sb)
210 {
211         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
212         if (sbinfo->max_inodes) {
213                 spin_lock(&sbinfo->stat_lock);
214                 if (!sbinfo->free_inodes) {
215                         spin_unlock(&sbinfo->stat_lock);
216                         return -ENOSPC;
217                 }
218                 sbinfo->free_inodes--;
219                 spin_unlock(&sbinfo->stat_lock);
220         }
221         return 0;
222 }
223
224 static void shmem_free_inode(struct super_block *sb)
225 {
226         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
227         if (sbinfo->max_inodes) {
228                 spin_lock(&sbinfo->stat_lock);
229                 sbinfo->free_inodes++;
230                 spin_unlock(&sbinfo->stat_lock);
231         }
232 }
233
234 /*
235  * shmem_recalc_inode - recalculate the size of an inode
236  *
237  * @inode: inode to recalc
238  *
239  * We have to calculate the free blocks since the mm can drop
240  * undirtied hole pages behind our back.
241  *
242  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
243  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
244  *
245  * It has to be called with the spinlock held.
246  */
247 static void shmem_recalc_inode(struct inode *inode)
248 {
249         struct shmem_inode_info *info = SHMEM_I(inode);
250         long freed;
251
252         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
253         if (freed > 0) {
254                 info->alloced -= freed;
255                 shmem_unacct_blocks(info->flags, freed);
256                 shmem_free_blocks(inode, freed);
257         }
258 }
259
260 /*
261  * shmem_swp_entry - find the swap vector position in the info structure
262  *
263  * @info:  info structure for the inode
264  * @index: index of the page to find
265  * @page:  optional page to add to the structure. Has to be preset to
266  *         all zeros
267  *
268  * If there is no space allocated yet it will return NULL when
269  * page is NULL, else it will use the page for the needed block,
270  * setting it to NULL on return to indicate that it has been used.
271  *
272  * The swap vector is organized the following way:
273  *
274  * There are SHMEM_NR_DIRECT entries directly stored in the
275  * shmem_inode_info structure. So small files do not need an addional
276  * allocation.
277  *
278  * For pages with index > SHMEM_NR_DIRECT there is the pointer
279  * i_indirect which points to a page which holds in the first half
280  * doubly indirect blocks, in the second half triple indirect blocks:
281  *
282  * For an artificial ENTRIES_PER_PAGE = 4 this would lead to the
283  * following layout (for SHMEM_NR_DIRECT == 16):
284  *
285  * i_indirect -> dir --> 16-19
286  *            |      +-> 20-23
287  *            |
288  *            +-->dir2 --> 24-27
289  *            |        +-> 28-31
290  *            |        +-> 32-35
291  *            |        +-> 36-39
292  *            |
293  *            +-->dir3 --> 40-43
294  *                     +-> 44-47
295  *                     +-> 48-51
296  *                     +-> 52-55
297  */
298 static swp_entry_t *shmem_swp_entry(struct shmem_inode_info *info, unsigned long index, struct page **page)
299 {
300         unsigned long offset;
301         struct page **dir;
302         struct page *subdir;
303
304         if (index < SHMEM_NR_DIRECT) {
305                 shmem_swp_balance_unmap();
306                 return info->i_direct+index;
307         }
308         if (!info->i_indirect) {
309                 if (page) {
310                         info->i_indirect = *page;
311                         *page = NULL;
312                 }
313                 return NULL;                    /* need another page */
314         }
315
316         index -= SHMEM_NR_DIRECT;
317         offset = index % ENTRIES_PER_PAGE;
318         index /= ENTRIES_PER_PAGE;
319         dir = shmem_dir_map(info->i_indirect);
320
321         if (index >= ENTRIES_PER_PAGE/2) {
322                 index -= ENTRIES_PER_PAGE/2;
323                 dir += ENTRIES_PER_PAGE/2 + index/ENTRIES_PER_PAGE;
324                 index %= ENTRIES_PER_PAGE;
325                 subdir = *dir;
326                 if (!subdir) {
327                         if (page) {
328                                 *dir = *page;
329                                 *page = NULL;
330                         }
331                         shmem_dir_unmap(dir);
332                         return NULL;            /* need another page */
333                 }
334                 shmem_dir_unmap(dir);
335                 dir = shmem_dir_map(subdir);
336         }
337
338         dir += index;
339         subdir = *dir;
340         if (!subdir) {
341                 if (!page || !(subdir = *page)) {
342                         shmem_dir_unmap(dir);
343                         return NULL;            /* need a page */
344                 }
345                 *dir = subdir;
346                 *page = NULL;
347         }
348         shmem_dir_unmap(dir);
349         return shmem_swp_map(subdir) + offset;
350 }
351
352 static void shmem_swp_set(struct shmem_inode_info *info, swp_entry_t *entry, unsigned long value)
353 {
354         long incdec = value? 1: -1;
355
356         entry->val = value;
357         info->swapped += incdec;
358         if ((unsigned long)(entry - info->i_direct) >= SHMEM_NR_DIRECT) {
359                 struct page *page = kmap_atomic_to_page(entry);
360                 set_page_private(page, page_private(page) + incdec);
361         }
362 }
363
364 /*
365  * shmem_swp_alloc - get the position of the swap entry for the page.
366  *                   If it does not exist allocate the entry.
367  *
368  * @info:       info structure for the inode
369  * @index:      index of the page to find
370  * @sgp:        check and recheck i_size? skip allocation?
371  */
372 static swp_entry_t *shmem_swp_alloc(struct shmem_inode_info *info, unsigned long index, enum sgp_type sgp)
373 {
374         struct inode *inode = &info->vfs_inode;
375         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
376         struct page *page = NULL;
377         swp_entry_t *entry;
378
379         if (sgp != SGP_WRITE &&
380             ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode))
381                 return ERR_PTR(-EINVAL);
382
383         while (!(entry = shmem_swp_entry(info, index, &page))) {
384                 if (sgp == SGP_READ)
385                         return shmem_swp_map(ZERO_PAGE(0));
386                 /*
387                  * Test free_blocks against 1 not 0, since we have 1 data
388                  * page (and perhaps indirect index pages) yet to allocate:
389                  * a waste to allocate index if we cannot allocate data.
390                  */
391                 if (sbinfo->max_blocks) {
392                         spin_lock(&sbinfo->stat_lock);
393                         if (sbinfo->free_blocks <= 1) {
394                                 spin_unlock(&sbinfo->stat_lock);
395                                 return ERR_PTR(-ENOSPC);
396                         }
397                         sbinfo->free_blocks--;
398                         inode->i_blocks += BLOCKS_PER_PAGE;
399                         spin_unlock(&sbinfo->stat_lock);
400                 }
401
402                 spin_unlock(&info->lock);
403                 page = shmem_dir_alloc(mapping_gfp_mask(inode->i_mapping));
404                 if (page)
405                         set_page_private(page, 0);
406                 spin_lock(&info->lock);
407
408                 if (!page) {
409                         shmem_free_blocks(inode, 1);
410                         return ERR_PTR(-ENOMEM);
411                 }
412                 if (sgp != SGP_WRITE &&
413                     ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
414                         entry = ERR_PTR(-EINVAL);
415                         break;
416                 }
417                 if (info->next_index <= index)
418                         info->next_index = index + 1;
419         }
420         if (page) {
421                 /* another task gave its page, or truncated the file */
422                 shmem_free_blocks(inode, 1);
423                 shmem_dir_free(page);
424         }
425         if (info->next_index <= index && !IS_ERR(entry))
426                 info->next_index = index + 1;
427         return entry;
428 }
429
430 /*
431  * shmem_free_swp - free some swap entries in a directory
432  *
433  * @dir:        pointer to the directory
434  * @edir:       pointer after last entry of the directory
435  * @punch_lock: pointer to spinlock when needed for the holepunch case
436  */
437 static int shmem_free_swp(swp_entry_t *dir, swp_entry_t *edir,
438                                                 spinlock_t *punch_lock)
439 {
440         spinlock_t *punch_unlock = NULL;
441         swp_entry_t *ptr;
442         int freed = 0;
443
444         for (ptr = dir; ptr < edir; ptr++) {
445                 if (ptr->val) {
446                         if (unlikely(punch_lock)) {
447                                 punch_unlock = punch_lock;
448                                 punch_lock = NULL;
449                                 spin_lock(punch_unlock);
450                                 if (!ptr->val)
451                                         continue;
452                         }
453                         free_swap_and_cache(*ptr);
454                         *ptr = (swp_entry_t){0};
455                         freed++;
456                 }
457         }
458         if (punch_unlock)
459                 spin_unlock(punch_unlock);
460         return freed;
461 }
462
463 static int shmem_map_and_free_swp(struct page *subdir, int offset,
464                 int limit, struct page ***dir, spinlock_t *punch_lock)
465 {
466         swp_entry_t *ptr;
467         int freed = 0;
468
469         ptr = shmem_swp_map(subdir);
470         for (; offset < limit; offset += LATENCY_LIMIT) {
471                 int size = limit - offset;
472                 if (size > LATENCY_LIMIT)
473                         size = LATENCY_LIMIT;
474                 freed += shmem_free_swp(ptr+offset, ptr+offset+size,
475                                                         punch_lock);
476                 if (need_resched()) {
477                         shmem_swp_unmap(ptr);
478                         if (*dir) {
479                                 shmem_dir_unmap(*dir);
480                                 *dir = NULL;
481                         }
482                         cond_resched();
483                         ptr = shmem_swp_map(subdir);
484                 }
485         }
486         shmem_swp_unmap(ptr);
487         return freed;
488 }
489
490 static void shmem_free_pages(struct list_head *next)
491 {
492         struct page *page;
493         int freed = 0;
494
495         do {
496                 page = container_of(next, struct page, lru);
497                 next = next->next;
498                 shmem_dir_free(page);
499                 freed++;
500                 if (freed >= LATENCY_LIMIT) {
501                         cond_resched();
502                         freed = 0;
503                 }
504         } while (next);
505 }
506
507 static void shmem_truncate_range(struct inode *inode, loff_t start, loff_t end)
508 {
509         struct shmem_inode_info *info = SHMEM_I(inode);
510         unsigned long idx;
511         unsigned long size;
512         unsigned long limit;
513         unsigned long stage;
514         unsigned long diroff;
515         struct page **dir;
516         struct page *topdir;
517         struct page *middir;
518         struct page *subdir;
519         swp_entry_t *ptr;
520         LIST_HEAD(pages_to_free);
521         long nr_pages_to_free = 0;
522         long nr_swaps_freed = 0;
523         int offset;
524         int freed;
525         int punch_hole;
526         spinlock_t *needs_lock;
527         spinlock_t *punch_lock;
528         unsigned long upper_limit;
529
530         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
531         idx = (start + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
532         if (idx >= info->next_index)
533                 return;
534
535         spin_lock(&info->lock);
536         info->flags |= SHMEM_TRUNCATE;
537         if (likely(end == (loff_t) -1)) {
538                 limit = info->next_index;
539                 upper_limit = SHMEM_MAX_INDEX;
540                 info->next_index = idx;
541                 needs_lock = NULL;
542                 punch_hole = 0;
543         } else {
544                 if (end + 1 >= inode->i_size) { /* we may free a little more */
545                         limit = (inode->i_size + PAGE_CACHE_SIZE - 1) >>
546                                                         PAGE_CACHE_SHIFT;
547                         upper_limit = SHMEM_MAX_INDEX;
548                 } else {
549                         limit = (end + 1) >> PAGE_CACHE_SHIFT;
550                         upper_limit = limit;
551                 }
552                 needs_lock = &info->lock;
553                 punch_hole = 1;
554         }
555
556         topdir = info->i_indirect;
557         if (topdir && idx <= SHMEM_NR_DIRECT && !punch_hole) {
558                 info->i_indirect = NULL;
559                 nr_pages_to_free++;
560                 list_add(&topdir->lru, &pages_to_free);
561         }
562         spin_unlock(&info->lock);
563
564         if (info->swapped && idx < SHMEM_NR_DIRECT) {
565                 ptr = info->i_direct;
566                 size = limit;
567                 if (size > SHMEM_NR_DIRECT)
568                         size = SHMEM_NR_DIRECT;
569                 nr_swaps_freed = shmem_free_swp(ptr+idx, ptr+size, needs_lock);
570         }
571
572         /*
573          * If there are no indirect blocks or we are punching a hole
574          * below indirect blocks, nothing to be done.
575          */
576         if (!topdir || limit <= SHMEM_NR_DIRECT)
577                 goto done2;
578
579         /*
580          * The truncation case has already dropped info->lock, and we're safe
581          * because i_size and next_index have already been lowered, preventing
582          * access beyond.  But in the punch_hole case, we still need to take
583          * the lock when updating the swap directory, because there might be
584          * racing accesses by shmem_getpage(SGP_CACHE), shmem_unuse_inode or
585          * shmem_writepage.  However, whenever we find we can remove a whole
586          * directory page (not at the misaligned start or end of the range),
587          * we first NULLify its pointer in the level above, and then have no
588          * need to take the lock when updating its contents: needs_lock and
589          * punch_lock (either pointing to info->lock or NULL) manage this.
590          */
591
592         upper_limit -= SHMEM_NR_DIRECT;
593         limit -= SHMEM_NR_DIRECT;
594         idx = (idx > SHMEM_NR_DIRECT)? (idx - SHMEM_NR_DIRECT): 0;
595         offset = idx % ENTRIES_PER_PAGE;
596         idx -= offset;
597
598         dir = shmem_dir_map(topdir);
599         stage = ENTRIES_PER_PAGEPAGE/2;
600         if (idx < ENTRIES_PER_PAGEPAGE/2) {
601                 middir = topdir;
602                 diroff = idx/ENTRIES_PER_PAGE;
603         } else {
604                 dir += ENTRIES_PER_PAGE/2;
605                 dir += (idx - ENTRIES_PER_PAGEPAGE/2)/ENTRIES_PER_PAGEPAGE;
606                 while (stage <= idx)
607                         stage += ENTRIES_PER_PAGEPAGE;
608                 middir = *dir;
609                 if (*dir) {
610                         diroff = ((idx - ENTRIES_PER_PAGEPAGE/2) %
611                                 ENTRIES_PER_PAGEPAGE) / ENTRIES_PER_PAGE;
612                         if (!diroff && !offset && upper_limit >= stage) {
613                                 if (needs_lock) {
614                                         spin_lock(needs_lock);
615                                         *dir = NULL;
616                                         spin_unlock(needs_lock);
617                                         needs_lock = NULL;
618                                 } else
619                                         *dir = NULL;
620                                 nr_pages_to_free++;
621                                 list_add(&middir->lru, &pages_to_free);
622                         }
623                         shmem_dir_unmap(dir);
624                         dir = shmem_dir_map(middir);
625                 } else {
626                         diroff = 0;
627                         offset = 0;
628                         idx = stage;
629                 }
630         }
631
632         for (; idx < limit; idx += ENTRIES_PER_PAGE, diroff++) {
633                 if (unlikely(idx == stage)) {
634                         shmem_dir_unmap(dir);
635                         dir = shmem_dir_map(topdir) +
636                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
637                         while (!*dir) {
638                                 dir++;
639                                 idx += ENTRIES_PER_PAGEPAGE;
640                                 if (idx >= limit)
641                                         goto done1;
642                         }
643                         stage = idx + ENTRIES_PER_PAGEPAGE;
644                         middir = *dir;
645                         if (punch_hole)
646                                 needs_lock = &info->lock;
647                         if (upper_limit >= stage) {
648                                 if (needs_lock) {
649                                         spin_lock(needs_lock);
650                                         *dir = NULL;
651                                         spin_unlock(needs_lock);
652                                         needs_lock = NULL;
653                                 } else
654                                         *dir = NULL;
655                                 nr_pages_to_free++;
656                                 list_add(&middir->lru, &pages_to_free);
657                         }
658                         shmem_dir_unmap(dir);
659                         cond_resched();
660                         dir = shmem_dir_map(middir);
661                         diroff = 0;
662                 }
663                 punch_lock = needs_lock;
664                 subdir = dir[diroff];
665                 if (subdir && !offset && upper_limit-idx >= ENTRIES_PER_PAGE) {
666                         if (needs_lock) {
667                                 spin_lock(needs_lock);
668                                 dir[diroff] = NULL;
669                                 spin_unlock(needs_lock);
670                                 punch_lock = NULL;
671                         } else
672                                 dir[diroff] = NULL;
673                         nr_pages_to_free++;
674                         list_add(&subdir->lru, &pages_to_free);
675                 }
676                 if (subdir && page_private(subdir) /* has swap entries */) {
677                         size = limit - idx;
678                         if (size > ENTRIES_PER_PAGE)
679                                 size = ENTRIES_PER_PAGE;
680                         freed = shmem_map_and_free_swp(subdir,
681                                         offset, size, &dir, punch_lock);
682                         if (!dir)
683                                 dir = shmem_dir_map(middir);
684                         nr_swaps_freed += freed;
685                         if (offset || punch_lock) {
686                                 spin_lock(&info->lock);
687                                 set_page_private(subdir,
688                                         page_private(subdir) - freed);
689                                 spin_unlock(&info->lock);
690                         } else
691                                 BUG_ON(page_private(subdir) != freed);
692                 }
693                 offset = 0;
694         }
695 done1:
696         shmem_dir_unmap(dir);
697 done2:
698         if (inode->i_mapping->nrpages && (info->flags & SHMEM_PAGEIN)) {
699                 /*
700                  * Call truncate_inode_pages again: racing shmem_unuse_inode
701                  * may have swizzled a page in from swap since vmtruncate or
702                  * generic_delete_inode did it, before we lowered next_index.
703                  * Also, though shmem_getpage checks i_size before adding to
704                  * cache, no recheck after: so fix the narrow window there too.
705                  *
706                  * Recalling truncate_inode_pages_range and unmap_mapping_range
707                  * every time for punch_hole (which never got a chance to clear
708                  * SHMEM_PAGEIN at the start of vmtruncate_range) is expensive,
709                  * yet hardly ever necessary: try to optimize them out later.
710                  */
711                 truncate_inode_pages_range(inode->i_mapping, start, end);
712                 if (punch_hole)
713                         unmap_mapping_range(inode->i_mapping, start,
714                                                         end - start, 1);
715         }
716
717         spin_lock(&info->lock);
718         info->flags &= ~SHMEM_TRUNCATE;
719         info->swapped -= nr_swaps_freed;
720         if (nr_pages_to_free)
721                 shmem_free_blocks(inode, nr_pages_to_free);
722         shmem_recalc_inode(inode);
723         spin_unlock(&info->lock);
724
725         /*
726          * Empty swap vector directory pages to be freed?
727          */
728         if (!list_empty(&pages_to_free)) {
729                 pages_to_free.prev->next = NULL;
730                 shmem_free_pages(pages_to_free.next);
731         }
732 }
733
734 static void shmem_truncate(struct inode *inode)
735 {
736         shmem_truncate_range(inode, inode->i_size, (loff_t)-1);
737 }
738
739 static int shmem_notify_change(struct dentry *dentry, struct iattr *attr)
740 {
741         struct inode *inode = dentry->d_inode;
742         struct page *page = NULL;
743         int error;
744
745         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
746                 if (attr->ia_size < inode->i_size) {
747                         /*
748                          * If truncating down to a partial page, then
749                          * if that page is already allocated, hold it
750                          * in memory until the truncation is over, so
751                          * truncate_partial_page cannnot miss it were
752                          * it assigned to swap.
753                          */
754                         if (attr->ia_size & (PAGE_CACHE_SIZE-1)) {
755                                 (void) shmem_getpage(inode,
756                                         attr->ia_size>>PAGE_CACHE_SHIFT,
757                                                 &page, SGP_READ, NULL);
758                                 if (page)
759                                         unlock_page(page);
760                         }
761                         /*
762                          * Reset SHMEM_PAGEIN flag so that shmem_truncate can
763                          * detect if any pages might have been added to cache
764                          * after truncate_inode_pages.  But we needn't bother
765                          * if it's being fully truncated to zero-length: the
766                          * nrpages check is efficient enough in that case.
767                          */
768                         if (attr->ia_size) {
769                                 struct shmem_inode_info *info = SHMEM_I(inode);
770                                 spin_lock(&info->lock);
771                                 info->flags &= ~SHMEM_PAGEIN;
772                                 spin_unlock(&info->lock);
773                         }
774                 }
775         }
776
777         error = inode_change_ok(inode, attr);
778         if (!error)
779                 error = inode_setattr(inode, attr);
780 #ifdef CONFIG_TMPFS_POSIX_ACL
781         if (!error && (attr->ia_valid & ATTR_MODE))
782                 error = generic_acl_chmod(inode, &shmem_acl_ops);
783 #endif
784         if (page)
785                 page_cache_release(page);
786         return error;
787 }
788
789 static void shmem_delete_inode(struct inode *inode)
790 {
791         struct shmem_inode_info *info = SHMEM_I(inode);
792
793         if (inode->i_op->truncate == shmem_truncate) {
794                 truncate_inode_pages(inode->i_mapping, 0);
795                 shmem_unacct_size(info->flags, inode->i_size);
796                 inode->i_size = 0;
797                 shmem_truncate(inode);
798                 if (!list_empty(&info->swaplist)) {
799                         mutex_lock(&shmem_swaplist_mutex);
800                         list_del_init(&info->swaplist);
801                         mutex_unlock(&shmem_swaplist_mutex);
802                 }
803         }
804         BUG_ON(inode->i_blocks);
805         shmem_free_inode(inode->i_sb);
806         clear_inode(inode);
807 }
808
809 static inline int shmem_find_swp(swp_entry_t entry, swp_entry_t *dir, swp_entry_t *edir)
810 {
811         swp_entry_t *ptr;
812
813         for (ptr = dir; ptr < edir; ptr++) {
814                 if (ptr->val == entry.val)
815                         return ptr - dir;
816         }
817         return -1;
818 }
819
820 static int shmem_unuse_inode(struct shmem_inode_info *info, swp_entry_t entry, struct page *page)
821 {
822         struct inode *inode;
823         unsigned long idx;
824         unsigned long size;
825         unsigned long limit;
826         unsigned long stage;
827         struct page **dir;
828         struct page *subdir;
829         swp_entry_t *ptr;
830         int offset;
831         int error;
832
833         idx = 0;
834         ptr = info->i_direct;
835         spin_lock(&info->lock);
836         if (!info->swapped) {
837                 list_del_init(&info->swaplist);
838                 goto lost2;
839         }
840         limit = info->next_index;
841         size = limit;
842         if (size > SHMEM_NR_DIRECT)
843                 size = SHMEM_NR_DIRECT;
844         offset = shmem_find_swp(entry, ptr, ptr+size);
845         if (offset >= 0)
846                 goto found;
847         if (!info->i_indirect)
848                 goto lost2;
849
850         dir = shmem_dir_map(info->i_indirect);
851         stage = SHMEM_NR_DIRECT + ENTRIES_PER_PAGEPAGE/2;
852
853         for (idx = SHMEM_NR_DIRECT; idx < limit; idx += ENTRIES_PER_PAGE, dir++) {
854                 if (unlikely(idx == stage)) {
855                         shmem_dir_unmap(dir-1);
856                         if (cond_resched_lock(&info->lock)) {
857                                 /* check it has not been truncated */
858                                 if (limit > info->next_index) {
859                                         limit = info->next_index;
860                                         if (idx >= limit)
861                                                 goto lost2;
862                                 }
863                         }
864                         dir = shmem_dir_map(info->i_indirect) +
865                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
866                         while (!*dir) {
867                                 dir++;
868                                 idx += ENTRIES_PER_PAGEPAGE;
869                                 if (idx >= limit)
870                                         goto lost1;
871                         }
872                         stage = idx + ENTRIES_PER_PAGEPAGE;
873                         subdir = *dir;
874                         shmem_dir_unmap(dir);
875                         dir = shmem_dir_map(subdir);
876                 }
877                 subdir = *dir;
878                 if (subdir && page_private(subdir)) {
879                         ptr = shmem_swp_map(subdir);
880                         size = limit - idx;
881                         if (size > ENTRIES_PER_PAGE)
882                                 size = ENTRIES_PER_PAGE;
883                         offset = shmem_find_swp(entry, ptr, ptr+size);
884                         shmem_swp_unmap(ptr);
885                         if (offset >= 0) {
886                                 shmem_dir_unmap(dir);
887                                 goto found;
888                         }
889                 }
890         }
891 lost1:
892         shmem_dir_unmap(dir-1);
893 lost2:
894         spin_unlock(&info->lock);
895         return 0;
896 found:
897         idx += offset;
898         inode = igrab(&info->vfs_inode);
899         spin_unlock(&info->lock);
900
901         /*
902          * Move _head_ to start search for next from here.
903          * But be careful: shmem_delete_inode checks list_empty without taking
904          * mutex, and there's an instant in list_move_tail when info->swaplist
905          * would appear empty, if it were the only one on shmem_swaplist.  We
906          * could avoid doing it if inode NULL; or use this minor optimization.
907          */
908         if (shmem_swaplist.next != &info->swaplist)
909                 list_move_tail(&shmem_swaplist, &info->swaplist);
910         mutex_unlock(&shmem_swaplist_mutex);
911
912         error = 1;
913         if (!inode)
914                 goto out;
915         /* Precharge page while we can wait, compensate afterwards */
916         error = mem_cgroup_cache_charge(page, current->mm, GFP_KERNEL);
917         if (error)
918                 goto out;
919         error = radix_tree_preload(GFP_KERNEL);
920         if (error)
921                 goto uncharge;
922         error = 1;
923
924         spin_lock(&info->lock);
925         ptr = shmem_swp_entry(info, idx, NULL);
926         if (ptr && ptr->val == entry.val)
927                 error = add_to_page_cache(page, inode->i_mapping,
928                                                 idx, GFP_NOWAIT);
929         if (error == -EEXIST) {
930                 struct page *filepage = find_get_page(inode->i_mapping, idx);
931                 error = 1;
932                 if (filepage) {
933                         /*
934                          * There might be a more uptodate page coming down
935                          * from a stacked writepage: forget our swappage if so.
936                          */
937                         if (PageUptodate(filepage))
938                                 error = 0;
939                         page_cache_release(filepage);
940                 }
941         }
942         if (!error) {
943                 delete_from_swap_cache(page);
944                 set_page_dirty(page);
945                 info->flags |= SHMEM_PAGEIN;
946                 shmem_swp_set(info, ptr, 0);
947                 swap_free(entry);
948                 error = 1;      /* not an error, but entry was found */
949         }
950         if (ptr)
951                 shmem_swp_unmap(ptr);
952         spin_unlock(&info->lock);
953         radix_tree_preload_end();
954 uncharge:
955         mem_cgroup_uncharge_page(page);
956 out:
957         unlock_page(page);
958         page_cache_release(page);
959         iput(inode);            /* allows for NULL */
960         return error;
961 }
962
963 /*
964  * shmem_unuse() search for an eventually swapped out shmem page.
965  */
966 int shmem_unuse(swp_entry_t entry, struct page *page)
967 {
968         struct list_head *p, *next;
969         struct shmem_inode_info *info;
970         int found = 0;
971
972         mutex_lock(&shmem_swaplist_mutex);
973         list_for_each_safe(p, next, &shmem_swaplist) {
974                 info = list_entry(p, struct shmem_inode_info, swaplist);
975                 found = shmem_unuse_inode(info, entry, page);
976                 cond_resched();
977                 if (found)
978                         goto out;
979         }
980         mutex_unlock(&shmem_swaplist_mutex);
981 out:    return found;   /* 0 or 1 or -ENOMEM */
982 }
983
984 /*
985  * Move the page from the page cache to the swap cache.
986  */
987 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
988 {
989         struct shmem_inode_info *info;
990         swp_entry_t *entry, swap;
991         struct address_space *mapping;
992         unsigned long index;
993         struct inode *inode;
994
995         BUG_ON(!PageLocked(page));
996         mapping = page->mapping;
997         index = page->index;
998         inode = mapping->host;
999         info = SHMEM_I(inode);
1000         if (info->flags & VM_LOCKED)
1001                 goto redirty;
1002         if (!total_swap_pages)
1003                 goto redirty;
1004
1005         /*
1006          * shmem_backing_dev_info's capabilities prevent regular writeback or
1007          * sync from ever calling shmem_writepage; but a stacking filesystem
1008          * may use the ->writepage of its underlying filesystem, in which case
1009          * tmpfs should write out to swap only in response to memory pressure,
1010          * and not for pdflush or sync.  However, in those cases, we do still
1011          * want to check if there's a redundant swappage to be discarded.
1012          */
1013         if (wbc->for_reclaim)
1014                 swap = get_swap_page();
1015         else
1016                 swap.val = 0;
1017
1018         spin_lock(&info->lock);
1019         if (index >= info->next_index) {
1020                 BUG_ON(!(info->flags & SHMEM_TRUNCATE));
1021                 goto unlock;
1022         }
1023         entry = shmem_swp_entry(info, index, NULL);
1024         if (entry->val) {
1025                 /*
1026                  * The more uptodate page coming down from a stacked
1027                  * writepage should replace our old swappage.
1028                  */
1029                 free_swap_and_cache(*entry);
1030                 shmem_swp_set(info, entry, 0);
1031         }
1032         shmem_recalc_inode(inode);
1033
1034         if (swap.val && add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
1035                 remove_from_page_cache(page);
1036                 shmem_swp_set(info, entry, swap.val);
1037                 shmem_swp_unmap(entry);
1038                 if (list_empty(&info->swaplist))
1039                         inode = igrab(inode);
1040                 else
1041                         inode = NULL;
1042                 spin_unlock(&info->lock);
1043                 swap_duplicate(swap);
1044                 BUG_ON(page_mapped(page));
1045                 page_cache_release(page);       /* pagecache ref */
1046                 set_page_dirty(page);
1047                 unlock_page(page);
1048                 if (inode) {
1049                         mutex_lock(&shmem_swaplist_mutex);
1050                         /* move instead of add in case we're racing */
1051                         list_move_tail(&info->swaplist, &shmem_swaplist);
1052                         mutex_unlock(&shmem_swaplist_mutex);
1053                         iput(inode);
1054                 }
1055                 return 0;
1056         }
1057
1058         shmem_swp_unmap(entry);
1059 unlock:
1060         spin_unlock(&info->lock);
1061         swap_free(swap);
1062 redirty:
1063         set_page_dirty(page);
1064         if (wbc->for_reclaim)
1065                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
1066         unlock_page(page);
1067         return 0;
1068 }
1069
1070 #ifdef CONFIG_NUMA
1071 static inline int shmem_parse_mpol(char *value, int *policy, nodemask_t *policy_nodes)
1072 {
1073         char *nodelist = strchr(value, ':');
1074         int err = 1;
1075
1076         if (nodelist) {
1077                 /* NUL-terminate policy string */
1078                 *nodelist++ = '\0';
1079                 if (nodelist_parse(nodelist, *policy_nodes))
1080                         goto out;
1081                 if (!nodes_subset(*policy_nodes, node_states[N_HIGH_MEMORY]))
1082                         goto out;
1083         }
1084         if (!strcmp(value, "default")) {
1085                 *policy = MPOL_DEFAULT;
1086                 /* Don't allow a nodelist */
1087                 if (!nodelist)
1088                         err = 0;
1089         } else if (!strcmp(value, "prefer")) {
1090                 *policy = MPOL_PREFERRED;
1091                 /* Insist on a nodelist of one node only */
1092                 if (nodelist) {
1093                         char *rest = nodelist;
1094                         while (isdigit(*rest))
1095                                 rest++;
1096                         if (!*rest)
1097                                 err = 0;
1098                 }
1099         } else if (!strcmp(value, "bind")) {
1100                 *policy = MPOL_BIND;
1101                 /* Insist on a nodelist */
1102                 if (nodelist)
1103                         err = 0;
1104         } else if (!strcmp(value, "interleave")) {
1105                 *policy = MPOL_INTERLEAVE;
1106                 /*
1107                  * Default to online nodes with memory if no nodelist
1108                  */
1109                 if (!nodelist)
1110                         *policy_nodes = node_states[N_HIGH_MEMORY];
1111                 err = 0;
1112         }
1113 out:
1114         /* Restore string for error message */
1115         if (nodelist)
1116                 *--nodelist = ':';
1117         return err;
1118 }
1119
1120 static struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1121                         struct shmem_inode_info *info, unsigned long idx)
1122 {
1123         struct vm_area_struct pvma;
1124         struct page *page;
1125
1126         /* Create a pseudo vma that just contains the policy */
1127         pvma.vm_start = 0;
1128         pvma.vm_pgoff = idx;
1129         pvma.vm_ops = NULL;
1130         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, idx);
1131         page = swapin_readahead(entry, gfp, &pvma, 0);
1132         mpol_free(pvma.vm_policy);
1133         return page;
1134 }
1135
1136 static struct page *shmem_alloc_page(gfp_t gfp,
1137                         struct shmem_inode_info *info, unsigned long idx)
1138 {
1139         struct vm_area_struct pvma;
1140         struct page *page;
1141
1142         /* Create a pseudo vma that just contains the policy */
1143         pvma.vm_start = 0;
1144         pvma.vm_pgoff = idx;
1145         pvma.vm_ops = NULL;
1146         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, idx);
1147         page = alloc_page_vma(gfp, &pvma, 0);
1148         mpol_free(pvma.vm_policy);
1149         return page;
1150 }
1151 #else
1152 static inline int shmem_parse_mpol(char *value, int *policy,
1153                                                 nodemask_t *policy_nodes)
1154 {
1155         return 1;
1156 }
1157
1158 static inline struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1159                         struct shmem_inode_info *info, unsigned long idx)
1160 {
1161         return swapin_readahead(entry, gfp, NULL, 0);
1162 }
1163
1164 static inline struct page *shmem_alloc_page(gfp_t gfp,
1165                         struct shmem_inode_info *info, unsigned long idx)
1166 {
1167         return alloc_page(gfp);
1168 }
1169 #endif
1170
1171 /*
1172  * shmem_getpage - either get the page from swap or allocate a new one
1173  *
1174  * If we allocate a new one we do not mark it dirty. That's up to the
1175  * vm. If we swap it in we mark it dirty since we also free the swap
1176  * entry since a page cannot live in both the swap and page cache
1177  */
1178 static int shmem_getpage(struct inode *inode, unsigned long idx,
1179                         struct page **pagep, enum sgp_type sgp, int *type)
1180 {
1181         struct address_space *mapping = inode->i_mapping;
1182         struct shmem_inode_info *info = SHMEM_I(inode);
1183         struct shmem_sb_info *sbinfo;
1184         struct page *filepage = *pagep;
1185         struct page *swappage;
1186         swp_entry_t *entry;
1187         swp_entry_t swap;
1188         gfp_t gfp;
1189         int error;
1190
1191         if (idx >= SHMEM_MAX_INDEX)
1192                 return -EFBIG;
1193
1194         if (type)
1195                 *type = 0;
1196
1197         /*
1198          * Normally, filepage is NULL on entry, and either found
1199          * uptodate immediately, or allocated and zeroed, or read
1200          * in under swappage, which is then assigned to filepage.
1201          * But shmem_readpage (required for splice) passes in a locked
1202          * filepage, which may be found not uptodate by other callers
1203          * too, and may need to be copied from the swappage read in.
1204          */
1205 repeat:
1206         if (!filepage)
1207                 filepage = find_lock_page(mapping, idx);
1208         if (filepage && PageUptodate(filepage))
1209                 goto done;
1210         error = 0;
1211         gfp = mapping_gfp_mask(mapping);
1212         if (!filepage) {
1213                 /*
1214                  * Try to preload while we can wait, to not make a habit of
1215                  * draining atomic reserves; but don't latch on to this cpu.
1216                  */
1217                 error = radix_tree_preload(gfp & ~__GFP_HIGHMEM);
1218                 if (error)
1219                         goto failed;
1220                 radix_tree_preload_end();
1221         }
1222
1223         spin_lock(&info->lock);
1224         shmem_recalc_inode(inode);
1225         entry = shmem_swp_alloc(info, idx, sgp);
1226         if (IS_ERR(entry)) {
1227                 spin_unlock(&info->lock);
1228                 error = PTR_ERR(entry);
1229                 goto failed;
1230         }
1231         swap = *entry;
1232
1233         if (swap.val) {
1234                 /* Look it up and read it in.. */
1235                 swappage = lookup_swap_cache(swap);
1236                 if (!swappage) {
1237                         shmem_swp_unmap(entry);
1238                         /* here we actually do the io */
1239                         if (type && !(*type & VM_FAULT_MAJOR)) {
1240                                 __count_vm_event(PGMAJFAULT);
1241                                 *type |= VM_FAULT_MAJOR;
1242                         }
1243                         spin_unlock(&info->lock);
1244                         swappage = shmem_swapin(swap, gfp, info, idx);
1245                         if (!swappage) {
1246                                 spin_lock(&info->lock);
1247                                 entry = shmem_swp_alloc(info, idx, sgp);
1248                                 if (IS_ERR(entry))
1249                                         error = PTR_ERR(entry);
1250                                 else {
1251                                         if (entry->val == swap.val)
1252                                                 error = -ENOMEM;
1253                                         shmem_swp_unmap(entry);
1254                                 }
1255                                 spin_unlock(&info->lock);
1256                                 if (error)
1257                                         goto failed;
1258                                 goto repeat;
1259                         }
1260                         wait_on_page_locked(swappage);
1261                         page_cache_release(swappage);
1262                         goto repeat;
1263                 }
1264
1265                 /* We have to do this with page locked to prevent races */
1266                 if (TestSetPageLocked(swappage)) {
1267                         shmem_swp_unmap(entry);
1268                         spin_unlock(&info->lock);
1269                         wait_on_page_locked(swappage);
1270                         page_cache_release(swappage);
1271                         goto repeat;
1272                 }
1273                 if (PageWriteback(swappage)) {
1274                         shmem_swp_unmap(entry);
1275                         spin_unlock(&info->lock);
1276                         wait_on_page_writeback(swappage);
1277                         unlock_page(swappage);
1278                         page_cache_release(swappage);
1279                         goto repeat;
1280                 }
1281                 if (!PageUptodate(swappage)) {
1282                         shmem_swp_unmap(entry);
1283                         spin_unlock(&info->lock);
1284                         unlock_page(swappage);
1285                         page_cache_release(swappage);
1286                         error = -EIO;
1287                         goto failed;
1288                 }
1289
1290                 if (filepage) {
1291                         shmem_swp_set(info, entry, 0);
1292                         shmem_swp_unmap(entry);
1293                         delete_from_swap_cache(swappage);
1294                         spin_unlock(&info->lock);
1295                         copy_highpage(filepage, swappage);
1296                         unlock_page(swappage);
1297                         page_cache_release(swappage);
1298                         flush_dcache_page(filepage);
1299                         SetPageUptodate(filepage);
1300                         set_page_dirty(filepage);
1301                         swap_free(swap);
1302                 } else if (!(error = add_to_page_cache(
1303                                 swappage, mapping, idx, GFP_NOWAIT))) {
1304                         info->flags |= SHMEM_PAGEIN;
1305                         shmem_swp_set(info, entry, 0);
1306                         shmem_swp_unmap(entry);
1307                         delete_from_swap_cache(swappage);
1308                         spin_unlock(&info->lock);
1309                         filepage = swappage;
1310                         set_page_dirty(filepage);
1311                         swap_free(swap);
1312                 } else {
1313                         shmem_swp_unmap(entry);
1314                         spin_unlock(&info->lock);
1315                         unlock_page(swappage);
1316                         page_cache_release(swappage);
1317                         if (error == -ENOMEM) {
1318                                 /* allow reclaim from this memory cgroup */
1319                                 error = mem_cgroup_cache_charge(NULL,
1320                                         current->mm, gfp & ~__GFP_HIGHMEM);
1321                                 if (error)
1322                                         goto failed;
1323                         }
1324                         goto repeat;
1325                 }
1326         } else if (sgp == SGP_READ && !filepage) {
1327                 shmem_swp_unmap(entry);
1328                 filepage = find_get_page(mapping, idx);
1329                 if (filepage &&
1330                     (!PageUptodate(filepage) || TestSetPageLocked(filepage))) {
1331                         spin_unlock(&info->lock);
1332                         wait_on_page_locked(filepage);
1333                         page_cache_release(filepage);
1334                         filepage = NULL;
1335                         goto repeat;
1336                 }
1337                 spin_unlock(&info->lock);
1338         } else {
1339                 shmem_swp_unmap(entry);
1340                 sbinfo = SHMEM_SB(inode->i_sb);
1341                 if (sbinfo->max_blocks) {
1342                         spin_lock(&sbinfo->stat_lock);
1343                         if (sbinfo->free_blocks == 0 ||
1344                             shmem_acct_block(info->flags)) {
1345                                 spin_unlock(&sbinfo->stat_lock);
1346                                 spin_unlock(&info->lock);
1347                                 error = -ENOSPC;
1348                                 goto failed;
1349                         }
1350                         sbinfo->free_blocks--;
1351                         inode->i_blocks += BLOCKS_PER_PAGE;
1352                         spin_unlock(&sbinfo->stat_lock);
1353                 } else if (shmem_acct_block(info->flags)) {
1354                         spin_unlock(&info->lock);
1355                         error = -ENOSPC;
1356                         goto failed;
1357                 }
1358
1359                 if (!filepage) {
1360                         spin_unlock(&info->lock);
1361                         filepage = shmem_alloc_page(gfp, info, idx);
1362                         if (!filepage) {
1363                                 shmem_unacct_blocks(info->flags, 1);
1364                                 shmem_free_blocks(inode, 1);
1365                                 error = -ENOMEM;
1366                                 goto failed;
1367                         }
1368
1369                         /* Precharge page while we can wait, compensate after */
1370                         error = mem_cgroup_cache_charge(filepage, current->mm,
1371                                                         gfp & ~__GFP_HIGHMEM);
1372                         if (error) {
1373                                 page_cache_release(filepage);
1374                                 shmem_unacct_blocks(info->flags, 1);
1375                                 shmem_free_blocks(inode, 1);
1376                                 filepage = NULL;
1377                                 goto failed;
1378                         }
1379
1380                         spin_lock(&info->lock);
1381                         entry = shmem_swp_alloc(info, idx, sgp);
1382                         if (IS_ERR(entry))
1383                                 error = PTR_ERR(entry);
1384                         else {
1385                                 swap = *entry;
1386                                 shmem_swp_unmap(entry);
1387                         }
1388                         if (error || swap.val || 0 != add_to_page_cache_lru(
1389                                         filepage, mapping, idx, GFP_NOWAIT)) {
1390                                 spin_unlock(&info->lock);
1391                                 mem_cgroup_uncharge_page(filepage);
1392                                 page_cache_release(filepage);
1393                                 shmem_unacct_blocks(info->flags, 1);
1394                                 shmem_free_blocks(inode, 1);
1395                                 filepage = NULL;
1396                                 if (error)
1397                                         goto failed;
1398                                 goto repeat;
1399                         }
1400                         mem_cgroup_uncharge_page(filepage);
1401                         info->flags |= SHMEM_PAGEIN;
1402                 }
1403
1404                 info->alloced++;
1405                 spin_unlock(&info->lock);
1406                 clear_highpage(filepage);
1407                 flush_dcache_page(filepage);
1408                 SetPageUptodate(filepage);
1409                 if (sgp == SGP_DIRTY)
1410                         set_page_dirty(filepage);
1411         }
1412 done:
1413         *pagep = filepage;
1414         return 0;
1415
1416 failed:
1417         if (*pagep != filepage) {
1418                 unlock_page(filepage);
1419                 page_cache_release(filepage);
1420         }
1421         return error;
1422 }
1423
1424 static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1425 {
1426         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1427         int error;
1428         int ret;
1429
1430         if (((loff_t)vmf->pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
1431                 return VM_FAULT_SIGBUS;
1432
1433         error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_CACHE, &ret);
1434         if (error)
1435                 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1436
1437         mark_page_accessed(vmf->page);
1438         return ret | VM_FAULT_LOCKED;
1439 }
1440
1441 #ifdef CONFIG_NUMA
1442 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *new)
1443 {
1444         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1445         return mpol_set_shared_policy(&SHMEM_I(i)->policy, vma, new);
1446 }
1447
1448 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
1449                                           unsigned long addr)
1450 {
1451         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1452         unsigned long idx;
1453
1454         idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1455         return mpol_shared_policy_lookup(&SHMEM_I(i)->policy, idx);
1456 }
1457 #endif
1458
1459 int shmem_lock(struct file *file, int lock, struct user_struct *user)
1460 {
1461         struct inode *inode = file->f_path.dentry->d_inode;
1462         struct shmem_inode_info *info = SHMEM_I(inode);
1463         int retval = -ENOMEM;
1464
1465         spin_lock(&info->lock);
1466         if (lock && !(info->flags & VM_LOCKED)) {
1467                 if (!user_shm_lock(inode->i_size, user))
1468                         goto out_nomem;
1469                 info->flags |= VM_LOCKED;
1470         }
1471         if (!lock && (info->flags & VM_LOCKED) && user) {
1472                 user_shm_unlock(inode->i_size, user);
1473                 info->flags &= ~VM_LOCKED;
1474         }
1475         retval = 0;
1476 out_nomem:
1477         spin_unlock(&info->lock);
1478         return retval;
1479 }
1480
1481 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1482 {
1483         file_accessed(file);
1484         vma->vm_ops = &shmem_vm_ops;
1485         vma->vm_flags |= VM_CAN_NONLINEAR;
1486         return 0;
1487 }
1488
1489 static struct inode *
1490 shmem_get_inode(struct super_block *sb, int mode, dev_t dev)
1491 {
1492         struct inode *inode;
1493         struct shmem_inode_info *info;
1494         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1495
1496         if (shmem_reserve_inode(sb))
1497                 return NULL;
1498
1499         inode = new_inode(sb);
1500         if (inode) {
1501                 inode->i_mode = mode;
1502                 inode->i_uid = current->fsuid;
1503                 inode->i_gid = current->fsgid;
1504                 inode->i_blocks = 0;
1505                 inode->i_mapping->a_ops = &shmem_aops;
1506                 inode->i_mapping->backing_dev_info = &shmem_backing_dev_info;
1507                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1508                 inode->i_generation = get_seconds();
1509                 info = SHMEM_I(inode);
1510                 memset(info, 0, (char *)inode - (char *)info);
1511                 spin_lock_init(&info->lock);
1512                 INIT_LIST_HEAD(&info->swaplist);
1513
1514                 switch (mode & S_IFMT) {
1515                 default:
1516                         inode->i_op = &shmem_special_inode_operations;
1517                         init_special_inode(inode, mode, dev);
1518                         break;
1519                 case S_IFREG:
1520                         inode->i_op = &shmem_inode_operations;
1521                         inode->i_fop = &shmem_file_operations;
1522                         mpol_shared_policy_init(&info->policy, sbinfo->policy,
1523                                                         &sbinfo->policy_nodes);
1524                         break;
1525                 case S_IFDIR:
1526                         inc_nlink(inode);
1527                         /* Some things misbehave if size == 0 on a directory */
1528                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
1529                         inode->i_op = &shmem_dir_inode_operations;
1530                         inode->i_fop = &simple_dir_operations;
1531                         break;
1532                 case S_IFLNK:
1533                         /*
1534                          * Must not load anything in the rbtree,
1535                          * mpol_free_shared_policy will not be called.
1536                          */
1537                         mpol_shared_policy_init(&info->policy, MPOL_DEFAULT,
1538                                                 NULL);
1539                         break;
1540                 }
1541         } else
1542                 shmem_free_inode(sb);
1543         return inode;
1544 }
1545
1546 #ifdef CONFIG_TMPFS
1547 static const struct inode_operations shmem_symlink_inode_operations;
1548 static const struct inode_operations shmem_symlink_inline_operations;
1549
1550 /*
1551  * Normally tmpfs avoids the use of shmem_readpage and shmem_write_begin;
1552  * but providing them allows a tmpfs file to be used for splice, sendfile, and
1553  * below the loop driver, in the generic fashion that many filesystems support.
1554  */
1555 static int shmem_readpage(struct file *file, struct page *page)
1556 {
1557         struct inode *inode = page->mapping->host;
1558         int error = shmem_getpage(inode, page->index, &page, SGP_CACHE, NULL);
1559         unlock_page(page);
1560         return error;
1561 }
1562
1563 static int
1564 shmem_write_begin(struct file *file, struct address_space *mapping,
1565                         loff_t pos, unsigned len, unsigned flags,
1566                         struct page **pagep, void **fsdata)
1567 {
1568         struct inode *inode = mapping->host;
1569         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1570         *pagep = NULL;
1571         return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1572 }
1573
1574 static int
1575 shmem_write_end(struct file *file, struct address_space *mapping,
1576                         loff_t pos, unsigned len, unsigned copied,
1577                         struct page *page, void *fsdata)
1578 {
1579         struct inode *inode = mapping->host;
1580
1581         if (pos + copied > inode->i_size)
1582                 i_size_write(inode, pos + copied);
1583
1584         unlock_page(page);
1585         set_page_dirty(page);
1586         page_cache_release(page);
1587
1588         return copied;
1589 }
1590
1591 static void do_shmem_file_read(struct file *filp, loff_t *ppos, read_descriptor_t *desc, read_actor_t actor)
1592 {
1593         struct inode *inode = filp->f_path.dentry->d_inode;
1594         struct address_space *mapping = inode->i_mapping;
1595         unsigned long index, offset;
1596         enum sgp_type sgp = SGP_READ;
1597
1598         /*
1599          * Might this read be for a stacking filesystem?  Then when reading
1600          * holes of a sparse file, we actually need to allocate those pages,
1601          * and even mark them dirty, so it cannot exceed the max_blocks limit.
1602          */
1603         if (segment_eq(get_fs(), KERNEL_DS))
1604                 sgp = SGP_DIRTY;
1605
1606         index = *ppos >> PAGE_CACHE_SHIFT;
1607         offset = *ppos & ~PAGE_CACHE_MASK;
1608
1609         for (;;) {
1610                 struct page *page = NULL;
1611                 unsigned long end_index, nr, ret;
1612                 loff_t i_size = i_size_read(inode);
1613
1614                 end_index = i_size >> PAGE_CACHE_SHIFT;
1615                 if (index > end_index)
1616                         break;
1617                 if (index == end_index) {
1618                         nr = i_size & ~PAGE_CACHE_MASK;
1619                         if (nr <= offset)
1620                                 break;
1621                 }
1622
1623                 desc->error = shmem_getpage(inode, index, &page, sgp, NULL);
1624                 if (desc->error) {
1625                         if (desc->error == -EINVAL)
1626                                 desc->error = 0;
1627                         break;
1628                 }
1629                 if (page)
1630                         unlock_page(page);
1631
1632                 /*
1633                  * We must evaluate after, since reads (unlike writes)
1634                  * are called without i_mutex protection against truncate
1635                  */
1636                 nr = PAGE_CACHE_SIZE;
1637                 i_size = i_size_read(inode);
1638                 end_index = i_size >> PAGE_CACHE_SHIFT;
1639                 if (index == end_index) {
1640                         nr = i_size & ~PAGE_CACHE_MASK;
1641                         if (nr <= offset) {
1642                                 if (page)
1643                                         page_cache_release(page);
1644                                 break;
1645                         }
1646                 }
1647                 nr -= offset;
1648
1649                 if (page) {
1650                         /*
1651                          * If users can be writing to this page using arbitrary
1652                          * virtual addresses, take care about potential aliasing
1653                          * before reading the page on the kernel side.
1654                          */
1655                         if (mapping_writably_mapped(mapping))
1656                                 flush_dcache_page(page);
1657                         /*
1658                          * Mark the page accessed if we read the beginning.
1659                          */
1660                         if (!offset)
1661                                 mark_page_accessed(page);
1662                 } else {
1663                         page = ZERO_PAGE(0);
1664                         page_cache_get(page);
1665                 }
1666
1667                 /*
1668                  * Ok, we have the page, and it's up-to-date, so
1669                  * now we can copy it to user space...
1670                  *
1671                  * The actor routine returns how many bytes were actually used..
1672                  * NOTE! This may not be the same as how much of a user buffer
1673                  * we filled up (we may be padding etc), so we can only update
1674                  * "pos" here (the actor routine has to update the user buffer
1675                  * pointers and the remaining count).
1676                  */
1677                 ret = actor(desc, page, offset, nr);
1678                 offset += ret;
1679                 index += offset >> PAGE_CACHE_SHIFT;
1680                 offset &= ~PAGE_CACHE_MASK;
1681
1682                 page_cache_release(page);
1683                 if (ret != nr || !desc->count)
1684                         break;
1685
1686                 cond_resched();
1687         }
1688
1689         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1690         file_accessed(filp);
1691 }
1692
1693 static ssize_t shmem_file_read(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
1694 {
1695         read_descriptor_t desc;
1696
1697         if ((ssize_t) count < 0)
1698                 return -EINVAL;
1699         if (!access_ok(VERIFY_WRITE, buf, count))
1700                 return -EFAULT;
1701         if (!count)
1702                 return 0;
1703
1704         desc.written = 0;
1705         desc.count = count;
1706         desc.arg.buf = buf;
1707         desc.error = 0;
1708
1709         do_shmem_file_read(filp, ppos, &desc, file_read_actor);
1710         if (desc.written)
1711                 return desc.written;
1712         return desc.error;
1713 }
1714
1715 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
1716 {
1717         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
1718
1719         buf->f_type = TMPFS_MAGIC;
1720         buf->f_bsize = PAGE_CACHE_SIZE;
1721         buf->f_namelen = NAME_MAX;
1722         spin_lock(&sbinfo->stat_lock);
1723         if (sbinfo->max_blocks) {
1724                 buf->f_blocks = sbinfo->max_blocks;
1725                 buf->f_bavail = buf->f_bfree = sbinfo->free_blocks;
1726         }
1727         if (sbinfo->max_inodes) {
1728                 buf->f_files = sbinfo->max_inodes;
1729                 buf->f_ffree = sbinfo->free_inodes;
1730         }
1731         /* else leave those fields 0 like simple_statfs */
1732         spin_unlock(&sbinfo->stat_lock);
1733         return 0;
1734 }
1735
1736 /*
1737  * File creation. Allocate an inode, and we're done..
1738  */
1739 static int
1740 shmem_mknod(struct inode *dir, struct dentry *dentry, int mode, dev_t dev)
1741 {
1742         struct inode *inode = shmem_get_inode(dir->i_sb, mode, dev);
1743         int error = -ENOSPC;
1744
1745         if (inode) {
1746                 error = security_inode_init_security(inode, dir, NULL, NULL,
1747                                                      NULL);
1748                 if (error) {
1749                         if (error != -EOPNOTSUPP) {
1750                                 iput(inode);
1751                                 return error;
1752                         }
1753                 }
1754                 error = shmem_acl_init(inode, dir);
1755                 if (error) {
1756                         iput(inode);
1757                         return error;
1758                 }
1759                 if (dir->i_mode & S_ISGID) {
1760                         inode->i_gid = dir->i_gid;
1761                         if (S_ISDIR(mode))
1762                                 inode->i_mode |= S_ISGID;
1763                 }
1764                 dir->i_size += BOGO_DIRENT_SIZE;
1765                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1766                 d_instantiate(dentry, inode);
1767                 dget(dentry); /* Extra count - pin the dentry in core */
1768         }
1769         return error;
1770 }
1771
1772 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, int mode)
1773 {
1774         int error;
1775
1776         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
1777                 return error;
1778         inc_nlink(dir);
1779         return 0;
1780 }
1781
1782 static int shmem_create(struct inode *dir, struct dentry *dentry, int mode,
1783                 struct nameidata *nd)
1784 {
1785         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
1786 }
1787
1788 /*
1789  * Link a file..
1790  */
1791 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
1792 {
1793         struct inode *inode = old_dentry->d_inode;
1794         int ret;
1795
1796         /*
1797          * No ordinary (disk based) filesystem counts links as inodes;
1798          * but each new link needs a new dentry, pinning lowmem, and
1799          * tmpfs dentries cannot be pruned until they are unlinked.
1800          */
1801         ret = shmem_reserve_inode(inode->i_sb);
1802         if (ret)
1803                 goto out;
1804
1805         dir->i_size += BOGO_DIRENT_SIZE;
1806         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1807         inc_nlink(inode);
1808         atomic_inc(&inode->i_count);    /* New dentry reference */
1809         dget(dentry);           /* Extra pinning count for the created dentry */
1810         d_instantiate(dentry, inode);
1811 out:
1812         return ret;
1813 }
1814
1815 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
1816 {
1817         struct inode *inode = dentry->d_inode;
1818
1819         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
1820                 shmem_free_inode(inode->i_sb);
1821
1822         dir->i_size -= BOGO_DIRENT_SIZE;
1823         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1824         drop_nlink(inode);
1825         dput(dentry);   /* Undo the count from "create" - this does all the work */
1826         return 0;
1827 }
1828
1829 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
1830 {
1831         if (!simple_empty(dentry))
1832                 return -ENOTEMPTY;
1833
1834         drop_nlink(dentry->d_inode);
1835         drop_nlink(dir);
1836         return shmem_unlink(dir, dentry);
1837 }
1838
1839 /*
1840  * The VFS layer already does all the dentry stuff for rename,
1841  * we just have to decrement the usage count for the target if
1842  * it exists so that the VFS layer correctly free's it when it
1843  * gets overwritten.
1844  */
1845 static int shmem_rename(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
1846 {
1847         struct inode *inode = old_dentry->d_inode;
1848         int they_are_dirs = S_ISDIR(inode->i_mode);
1849
1850         if (!simple_empty(new_dentry))
1851                 return -ENOTEMPTY;
1852
1853         if (new_dentry->d_inode) {
1854                 (void) shmem_unlink(new_dir, new_dentry);
1855                 if (they_are_dirs)
1856                         drop_nlink(old_dir);
1857         } else if (they_are_dirs) {
1858                 drop_nlink(old_dir);
1859                 inc_nlink(new_dir);
1860         }
1861
1862         old_dir->i_size -= BOGO_DIRENT_SIZE;
1863         new_dir->i_size += BOGO_DIRENT_SIZE;
1864         old_dir->i_ctime = old_dir->i_mtime =
1865         new_dir->i_ctime = new_dir->i_mtime =
1866         inode->i_ctime = CURRENT_TIME;
1867         return 0;
1868 }
1869
1870 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
1871 {
1872         int error;
1873         int len;
1874         struct inode *inode;
1875         struct page *page = NULL;
1876         char *kaddr;
1877         struct shmem_inode_info *info;
1878
1879         len = strlen(symname) + 1;
1880         if (len > PAGE_CACHE_SIZE)
1881                 return -ENAMETOOLONG;
1882
1883         inode = shmem_get_inode(dir->i_sb, S_IFLNK|S_IRWXUGO, 0);
1884         if (!inode)
1885                 return -ENOSPC;
1886
1887         error = security_inode_init_security(inode, dir, NULL, NULL,
1888                                              NULL);
1889         if (error) {
1890                 if (error != -EOPNOTSUPP) {
1891                         iput(inode);
1892                         return error;
1893                 }
1894                 error = 0;
1895         }
1896
1897         info = SHMEM_I(inode);
1898         inode->i_size = len-1;
1899         if (len <= (char *)inode - (char *)info) {
1900                 /* do it inline */
1901                 memcpy(info, symname, len);
1902                 inode->i_op = &shmem_symlink_inline_operations;
1903         } else {
1904                 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
1905                 if (error) {
1906                         iput(inode);
1907                         return error;
1908                 }
1909                 unlock_page(page);
1910                 inode->i_op = &shmem_symlink_inode_operations;
1911                 kaddr = kmap_atomic(page, KM_USER0);
1912                 memcpy(kaddr, symname, len);
1913                 kunmap_atomic(kaddr, KM_USER0);
1914                 set_page_dirty(page);
1915                 page_cache_release(page);
1916         }
1917         if (dir->i_mode & S_ISGID)
1918                 inode->i_gid = dir->i_gid;
1919         dir->i_size += BOGO_DIRENT_SIZE;
1920         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1921         d_instantiate(dentry, inode);
1922         dget(dentry);
1923         return 0;
1924 }
1925
1926 static void *shmem_follow_link_inline(struct dentry *dentry, struct nameidata *nd)
1927 {
1928         nd_set_link(nd, (char *)SHMEM_I(dentry->d_inode));
1929         return NULL;
1930 }
1931
1932 static void *shmem_follow_link(struct dentry *dentry, struct nameidata *nd)
1933 {
1934         struct page *page = NULL;
1935         int res = shmem_getpage(dentry->d_inode, 0, &page, SGP_READ, NULL);
1936         nd_set_link(nd, res ? ERR_PTR(res) : kmap(page));
1937         if (page)
1938                 unlock_page(page);
1939         return page;
1940 }
1941
1942 static void shmem_put_link(struct dentry *dentry, struct nameidata *nd, void *cookie)
1943 {
1944         if (!IS_ERR(nd_get_link(nd))) {
1945                 struct page *page = cookie;
1946                 kunmap(page);
1947                 mark_page_accessed(page);
1948                 page_cache_release(page);
1949         }
1950 }
1951
1952 static const struct inode_operations shmem_symlink_inline_operations = {
1953         .readlink       = generic_readlink,
1954         .follow_link    = shmem_follow_link_inline,
1955 };
1956
1957 static const struct inode_operations shmem_symlink_inode_operations = {
1958         .truncate       = shmem_truncate,
1959         .readlink       = generic_readlink,
1960         .follow_link    = shmem_follow_link,
1961         .put_link       = shmem_put_link,
1962 };
1963
1964 #ifdef CONFIG_TMPFS_POSIX_ACL
1965 /**
1966  * Superblocks without xattr inode operations will get security.* xattr
1967  * support from the VFS "for free". As soon as we have any other xattrs
1968  * like ACLs, we also need to implement the security.* handlers at
1969  * filesystem level, though.
1970  */
1971
1972 static size_t shmem_xattr_security_list(struct inode *inode, char *list,
1973                                         size_t list_len, const char *name,
1974                                         size_t name_len)
1975 {
1976         return security_inode_listsecurity(inode, list, list_len);
1977 }
1978
1979 static int shmem_xattr_security_get(struct inode *inode, const char *name,
1980                                     void *buffer, size_t size)
1981 {
1982         if (strcmp(name, "") == 0)
1983                 return -EINVAL;
1984         return xattr_getsecurity(inode, name, buffer, size);
1985 }
1986
1987 static int shmem_xattr_security_set(struct inode *inode, const char *name,
1988                                     const void *value, size_t size, int flags)
1989 {
1990         if (strcmp(name, "") == 0)
1991                 return -EINVAL;
1992         return security_inode_setsecurity(inode, name, value, size, flags);
1993 }
1994
1995 static struct xattr_handler shmem_xattr_security_handler = {
1996         .prefix = XATTR_SECURITY_PREFIX,
1997         .list   = shmem_xattr_security_list,
1998         .get    = shmem_xattr_security_get,
1999         .set    = shmem_xattr_security_set,
2000 };
2001
2002 static struct xattr_handler *shmem_xattr_handlers[] = {
2003         &shmem_xattr_acl_access_handler,
2004         &shmem_xattr_acl_default_handler,
2005         &shmem_xattr_security_handler,
2006         NULL
2007 };
2008 #endif
2009
2010 static struct dentry *shmem_get_parent(struct dentry *child)
2011 {
2012         return ERR_PTR(-ESTALE);
2013 }
2014
2015 static int shmem_match(struct inode *ino, void *vfh)
2016 {
2017         __u32 *fh = vfh;
2018         __u64 inum = fh[2];
2019         inum = (inum << 32) | fh[1];
2020         return ino->i_ino == inum && fh[0] == ino->i_generation;
2021 }
2022
2023 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
2024                 struct fid *fid, int fh_len, int fh_type)
2025 {
2026         struct inode *inode;
2027         struct dentry *dentry = NULL;
2028         u64 inum = fid->raw[2];
2029         inum = (inum << 32) | fid->raw[1];
2030
2031         if (fh_len < 3)
2032                 return NULL;
2033
2034         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
2035                         shmem_match, fid->raw);
2036         if (inode) {
2037                 dentry = d_find_alias(inode);
2038                 iput(inode);
2039         }
2040
2041         return dentry;
2042 }
2043
2044 static int shmem_encode_fh(struct dentry *dentry, __u32 *fh, int *len,
2045                                 int connectable)
2046 {
2047         struct inode *inode = dentry->d_inode;
2048
2049         if (*len < 3)
2050                 return 255;
2051
2052         if (hlist_unhashed(&inode->i_hash)) {
2053                 /* Unfortunately insert_inode_hash is not idempotent,
2054                  * so as we hash inodes here rather than at creation
2055                  * time, we need a lock to ensure we only try
2056                  * to do it once
2057                  */
2058                 static DEFINE_SPINLOCK(lock);
2059                 spin_lock(&lock);
2060                 if (hlist_unhashed(&inode->i_hash))
2061                         __insert_inode_hash(inode,
2062                                             inode->i_ino + inode->i_generation);
2063                 spin_unlock(&lock);
2064         }
2065
2066         fh[0] = inode->i_generation;
2067         fh[1] = inode->i_ino;
2068         fh[2] = ((__u64)inode->i_ino) >> 32;
2069
2070         *len = 3;
2071         return 1;
2072 }
2073
2074 static const struct export_operations shmem_export_ops = {
2075         .get_parent     = shmem_get_parent,
2076         .encode_fh      = shmem_encode_fh,
2077         .fh_to_dentry   = shmem_fh_to_dentry,
2078 };
2079
2080 static int shmem_parse_options(char *options, int *mode, uid_t *uid,
2081         gid_t *gid, unsigned long *blocks, unsigned long *inodes,
2082         int *policy, nodemask_t *policy_nodes)
2083 {
2084         char *this_char, *value, *rest;
2085
2086         while (options != NULL) {
2087                 this_char = options;
2088                 for (;;) {
2089                         /*
2090                          * NUL-terminate this option: unfortunately,
2091                          * mount options form a comma-separated list,
2092                          * but mpol's nodelist may also contain commas.
2093                          */
2094                         options = strchr(options, ',');
2095                         if (options == NULL)
2096                                 break;
2097                         options++;
2098                         if (!isdigit(*options)) {
2099                                 options[-1] = '\0';
2100                                 break;
2101                         }
2102                 }
2103                 if (!*this_char)
2104                         continue;
2105                 if ((value = strchr(this_char,'=')) != NULL) {
2106                         *value++ = 0;
2107                 } else {
2108                         printk(KERN_ERR
2109                             "tmpfs: No value for mount option '%s'\n",
2110                             this_char);
2111                         return 1;
2112                 }
2113
2114                 if (!strcmp(this_char,"size")) {
2115                         unsigned long long size;
2116                         size = memparse(value,&rest);
2117                         if (*rest == '%') {
2118                                 size <<= PAGE_SHIFT;
2119                                 size *= totalram_pages;
2120                                 do_div(size, 100);
2121                                 rest++;
2122                         }
2123                         if (*rest)
2124                                 goto bad_val;
2125                         *blocks = DIV_ROUND_UP(size, PAGE_CACHE_SIZE);
2126                 } else if (!strcmp(this_char,"nr_blocks")) {
2127                         *blocks = memparse(value,&rest);
2128                         if (*rest)
2129                                 goto bad_val;
2130                 } else if (!strcmp(this_char,"nr_inodes")) {
2131                         *inodes = memparse(value,&rest);
2132                         if (*rest)
2133                                 goto bad_val;
2134                 } else if (!strcmp(this_char,"mode")) {
2135                         if (!mode)
2136                                 continue;
2137                         *mode = simple_strtoul(value,&rest,8);
2138                         if (*rest)
2139                                 goto bad_val;
2140                 } else if (!strcmp(this_char,"uid")) {
2141                         if (!uid)
2142                                 continue;
2143                         *uid = simple_strtoul(value,&rest,0);
2144                         if (*rest)
2145                                 goto bad_val;
2146                 } else if (!strcmp(this_char,"gid")) {
2147                         if (!gid)
2148                                 continue;
2149                         *gid = simple_strtoul(value,&rest,0);
2150                         if (*rest)
2151                                 goto bad_val;
2152                 } else if (!strcmp(this_char,"mpol")) {
2153                         if (shmem_parse_mpol(value,policy,policy_nodes))
2154                                 goto bad_val;
2155                 } else {
2156                         printk(KERN_ERR "tmpfs: Bad mount option %s\n",
2157                                this_char);
2158                         return 1;
2159                 }
2160         }
2161         return 0;
2162
2163 bad_val:
2164         printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
2165                value, this_char);
2166         return 1;
2167
2168 }
2169
2170 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2171 {
2172         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2173         unsigned long max_blocks = sbinfo->max_blocks;
2174         unsigned long max_inodes = sbinfo->max_inodes;
2175         int policy = sbinfo->policy;
2176         nodemask_t policy_nodes = sbinfo->policy_nodes;
2177         unsigned long blocks;
2178         unsigned long inodes;
2179         int error = -EINVAL;
2180
2181         if (shmem_parse_options(data, NULL, NULL, NULL, &max_blocks,
2182                                 &max_inodes, &policy, &policy_nodes))
2183                 return error;
2184
2185         spin_lock(&sbinfo->stat_lock);
2186         blocks = sbinfo->max_blocks - sbinfo->free_blocks;
2187         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2188         if (max_blocks < blocks)
2189                 goto out;
2190         if (max_inodes < inodes)
2191                 goto out;
2192         /*
2193          * Those tests also disallow limited->unlimited while any are in
2194          * use, so i_blocks will always be zero when max_blocks is zero;
2195          * but we must separately disallow unlimited->limited, because
2196          * in that case we have no record of how much is already in use.
2197          */
2198         if (max_blocks && !sbinfo->max_blocks)
2199                 goto out;
2200         if (max_inodes && !sbinfo->max_inodes)
2201                 goto out;
2202
2203         error = 0;
2204         sbinfo->max_blocks  = max_blocks;
2205         sbinfo->free_blocks = max_blocks - blocks;
2206         sbinfo->max_inodes  = max_inodes;
2207         sbinfo->free_inodes = max_inodes - inodes;
2208         sbinfo->policy = policy;
2209         sbinfo->policy_nodes = policy_nodes;
2210 out:
2211         spin_unlock(&sbinfo->stat_lock);
2212         return error;
2213 }
2214 #endif
2215
2216 static void shmem_put_super(struct super_block *sb)
2217 {
2218         kfree(sb->s_fs_info);
2219         sb->s_fs_info = NULL;
2220 }
2221
2222 static int shmem_fill_super(struct super_block *sb,
2223                             void *data, int silent)
2224 {
2225         struct inode *inode;
2226         struct dentry *root;
2227         int mode   = S_IRWXUGO | S_ISVTX;
2228         uid_t uid = current->fsuid;
2229         gid_t gid = current->fsgid;
2230         int err = -ENOMEM;
2231         struct shmem_sb_info *sbinfo;
2232         unsigned long blocks = 0;
2233         unsigned long inodes = 0;
2234         int policy = MPOL_DEFAULT;
2235         nodemask_t policy_nodes = node_states[N_HIGH_MEMORY];
2236
2237 #ifdef CONFIG_TMPFS
2238         /*
2239          * Per default we only allow half of the physical ram per
2240          * tmpfs instance, limiting inodes to one per page of lowmem;
2241          * but the internal instance is left unlimited.
2242          */
2243         if (!(sb->s_flags & MS_NOUSER)) {
2244                 blocks = totalram_pages / 2;
2245                 inodes = totalram_pages - totalhigh_pages;
2246                 if (inodes > blocks)
2247                         inodes = blocks;
2248                 if (shmem_parse_options(data, &mode, &uid, &gid, &blocks,
2249                                         &inodes, &policy, &policy_nodes))
2250                         return -EINVAL;
2251         }
2252         sb->s_export_op = &shmem_export_ops;
2253 #else
2254         sb->s_flags |= MS_NOUSER;
2255 #endif
2256
2257         /* Round up to L1_CACHE_BYTES to resist false sharing */
2258         sbinfo = kmalloc(max((int)sizeof(struct shmem_sb_info),
2259                                 L1_CACHE_BYTES), GFP_KERNEL);
2260         if (!sbinfo)
2261                 return -ENOMEM;
2262
2263         spin_lock_init(&sbinfo->stat_lock);
2264         sbinfo->max_blocks = blocks;
2265         sbinfo->free_blocks = blocks;
2266         sbinfo->max_inodes = inodes;
2267         sbinfo->free_inodes = inodes;
2268         sbinfo->policy = policy;
2269         sbinfo->policy_nodes = policy_nodes;
2270
2271         sb->s_fs_info = sbinfo;
2272         sb->s_maxbytes = SHMEM_MAX_BYTES;
2273         sb->s_blocksize = PAGE_CACHE_SIZE;
2274         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
2275         sb->s_magic = TMPFS_MAGIC;
2276         sb->s_op = &shmem_ops;
2277         sb->s_time_gran = 1;
2278 #ifdef CONFIG_TMPFS_POSIX_ACL
2279         sb->s_xattr = shmem_xattr_handlers;
2280         sb->s_flags |= MS_POSIXACL;
2281 #endif
2282
2283         inode = shmem_get_inode(sb, S_IFDIR | mode, 0);
2284         if (!inode)
2285                 goto failed;
2286         inode->i_uid = uid;
2287         inode->i_gid = gid;
2288         root = d_alloc_root(inode);
2289         if (!root)
2290                 goto failed_iput;
2291         sb->s_root = root;
2292         return 0;
2293
2294 failed_iput:
2295         iput(inode);
2296 failed:
2297         shmem_put_super(sb);
2298         return err;
2299 }
2300
2301 static struct kmem_cache *shmem_inode_cachep;
2302
2303 static struct inode *shmem_alloc_inode(struct super_block *sb)
2304 {
2305         struct shmem_inode_info *p;
2306         p = (struct shmem_inode_info *)kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
2307         if (!p)
2308                 return NULL;
2309         return &p->vfs_inode;
2310 }
2311
2312 static void shmem_destroy_inode(struct inode *inode)
2313 {
2314         if ((inode->i_mode & S_IFMT) == S_IFREG) {
2315                 /* only struct inode is valid if it's an inline symlink */
2316                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
2317         }
2318         shmem_acl_destroy_inode(inode);
2319         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
2320 }
2321
2322 static void init_once(struct kmem_cache *cachep, void *foo)
2323 {
2324         struct shmem_inode_info *p = (struct shmem_inode_info *) foo;
2325
2326         inode_init_once(&p->vfs_inode);
2327 #ifdef CONFIG_TMPFS_POSIX_ACL
2328         p->i_acl = NULL;
2329         p->i_default_acl = NULL;
2330 #endif
2331 }
2332
2333 static int init_inodecache(void)
2334 {
2335         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
2336                                 sizeof(struct shmem_inode_info),
2337                                 0, SLAB_PANIC, init_once);
2338         return 0;
2339 }
2340
2341 static void destroy_inodecache(void)
2342 {
2343         kmem_cache_destroy(shmem_inode_cachep);
2344 }
2345
2346 static const struct address_space_operations shmem_aops = {
2347         .writepage      = shmem_writepage,
2348         .set_page_dirty = __set_page_dirty_no_writeback,
2349 #ifdef CONFIG_TMPFS
2350         .readpage       = shmem_readpage,
2351         .write_begin    = shmem_write_begin,
2352         .write_end      = shmem_write_end,
2353 #endif
2354         .migratepage    = migrate_page,
2355 };
2356
2357 static const struct file_operations shmem_file_operations = {
2358         .mmap           = shmem_mmap,
2359 #ifdef CONFIG_TMPFS
2360         .llseek         = generic_file_llseek,
2361         .read           = shmem_file_read,
2362         .write          = do_sync_write,
2363         .aio_write      = generic_file_aio_write,
2364         .fsync          = simple_sync_file,
2365         .splice_read    = generic_file_splice_read,
2366         .splice_write   = generic_file_splice_write,
2367 #endif
2368 };
2369
2370 static const struct inode_operations shmem_inode_operations = {
2371         .truncate       = shmem_truncate,
2372         .setattr        = shmem_notify_change,
2373         .truncate_range = shmem_truncate_range,
2374 #ifdef CONFIG_TMPFS_POSIX_ACL
2375         .setxattr       = generic_setxattr,
2376         .getxattr       = generic_getxattr,
2377         .listxattr      = generic_listxattr,
2378         .removexattr    = generic_removexattr,
2379         .permission     = shmem_permission,
2380 #endif
2381
2382 };
2383
2384 static const struct inode_operations shmem_dir_inode_operations = {
2385 #ifdef CONFIG_TMPFS
2386         .create         = shmem_create,
2387         .lookup         = simple_lookup,
2388         .link           = shmem_link,
2389         .unlink         = shmem_unlink,
2390         .symlink        = shmem_symlink,
2391         .mkdir          = shmem_mkdir,
2392         .rmdir          = shmem_rmdir,
2393         .mknod          = shmem_mknod,
2394         .rename         = shmem_rename,
2395 #endif
2396 #ifdef CONFIG_TMPFS_POSIX_ACL
2397         .setattr        = shmem_notify_change,
2398         .setxattr       = generic_setxattr,
2399         .getxattr       = generic_getxattr,
2400         .listxattr      = generic_listxattr,
2401         .removexattr    = generic_removexattr,
2402         .permission     = shmem_permission,
2403 #endif
2404 };
2405
2406 static const struct inode_operations shmem_special_inode_operations = {
2407 #ifdef CONFIG_TMPFS_POSIX_ACL
2408         .setattr        = shmem_notify_change,
2409         .setxattr       = generic_setxattr,
2410         .getxattr       = generic_getxattr,
2411         .listxattr      = generic_listxattr,
2412         .removexattr    = generic_removexattr,
2413         .permission     = shmem_permission,
2414 #endif
2415 };
2416
2417 static const struct super_operations shmem_ops = {
2418         .alloc_inode    = shmem_alloc_inode,
2419         .destroy_inode  = shmem_destroy_inode,
2420 #ifdef CONFIG_TMPFS
2421         .statfs         = shmem_statfs,
2422         .remount_fs     = shmem_remount_fs,
2423 #endif
2424         .delete_inode   = shmem_delete_inode,
2425         .drop_inode     = generic_delete_inode,
2426         .put_super      = shmem_put_super,
2427 };
2428
2429 static struct vm_operations_struct shmem_vm_ops = {
2430         .fault          = shmem_fault,
2431 #ifdef CONFIG_NUMA
2432         .set_policy     = shmem_set_policy,
2433         .get_policy     = shmem_get_policy,
2434 #endif
2435 };
2436
2437
2438 static int shmem_get_sb(struct file_system_type *fs_type,
2439         int flags, const char *dev_name, void *data, struct vfsmount *mnt)
2440 {
2441         return get_sb_nodev(fs_type, flags, data, shmem_fill_super, mnt);
2442 }
2443
2444 static struct file_system_type tmpfs_fs_type = {
2445         .owner          = THIS_MODULE,
2446         .name           = "tmpfs",
2447         .get_sb         = shmem_get_sb,
2448         .kill_sb        = kill_litter_super,
2449 };
2450 static struct vfsmount *shm_mnt;
2451
2452 static int __init init_tmpfs(void)
2453 {
2454         int error;
2455
2456         error = bdi_init(&shmem_backing_dev_info);
2457         if (error)
2458                 goto out4;
2459
2460         error = init_inodecache();
2461         if (error)
2462                 goto out3;
2463
2464         error = register_filesystem(&tmpfs_fs_type);
2465         if (error) {
2466                 printk(KERN_ERR "Could not register tmpfs\n");
2467                 goto out2;
2468         }
2469
2470         shm_mnt = vfs_kern_mount(&tmpfs_fs_type, MS_NOUSER,
2471                                 tmpfs_fs_type.name, NULL);
2472         if (IS_ERR(shm_mnt)) {
2473                 error = PTR_ERR(shm_mnt);
2474                 printk(KERN_ERR "Could not kern_mount tmpfs\n");
2475                 goto out1;
2476         }
2477         return 0;
2478
2479 out1:
2480         unregister_filesystem(&tmpfs_fs_type);
2481 out2:
2482         destroy_inodecache();
2483 out3:
2484         bdi_destroy(&shmem_backing_dev_info);
2485 out4:
2486         shm_mnt = ERR_PTR(error);
2487         return error;
2488 }
2489 module_init(init_tmpfs)
2490
2491 /*
2492  * shmem_file_setup - get an unlinked file living in tmpfs
2493  *
2494  * @name: name for dentry (to be seen in /proc/<pid>/maps
2495  * @size: size to be set for the file
2496  *
2497  */
2498 struct file *shmem_file_setup(char *name, loff_t size, unsigned long flags)
2499 {
2500         int error;
2501         struct file *file;
2502         struct inode *inode;
2503         struct dentry *dentry, *root;
2504         struct qstr this;
2505
2506         if (IS_ERR(shm_mnt))
2507                 return (void *)shm_mnt;
2508
2509         if (size < 0 || size > SHMEM_MAX_BYTES)
2510                 return ERR_PTR(-EINVAL);
2511
2512         if (shmem_acct_size(flags, size))
2513                 return ERR_PTR(-ENOMEM);
2514
2515         error = -ENOMEM;
2516         this.name = name;
2517         this.len = strlen(name);
2518         this.hash = 0; /* will go */
2519         root = shm_mnt->mnt_root;
2520         dentry = d_alloc(root, &this);
2521         if (!dentry)
2522                 goto put_memory;
2523
2524         error = -ENFILE;
2525         file = get_empty_filp();
2526         if (!file)
2527                 goto put_dentry;
2528
2529         error = -ENOSPC;
2530         inode = shmem_get_inode(root->d_sb, S_IFREG | S_IRWXUGO, 0);
2531         if (!inode)
2532                 goto close_file;
2533
2534         SHMEM_I(inode)->flags = flags & VM_ACCOUNT;
2535         d_instantiate(dentry, inode);
2536         inode->i_size = size;
2537         inode->i_nlink = 0;     /* It is unlinked */
2538         init_file(file, shm_mnt, dentry, FMODE_WRITE | FMODE_READ,
2539                         &shmem_file_operations);
2540         return file;
2541
2542 close_file:
2543         put_filp(file);
2544 put_dentry:
2545         dput(dentry);
2546 put_memory:
2547         shmem_unacct_size(flags, size);
2548         return ERR_PTR(error);
2549 }
2550
2551 /*
2552  * shmem_zero_setup - setup a shared anonymous mapping
2553  *
2554  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
2555  */
2556 int shmem_zero_setup(struct vm_area_struct *vma)
2557 {
2558         struct file *file;
2559         loff_t size = vma->vm_end - vma->vm_start;
2560
2561         file = shmem_file_setup("dev/zero", size, vma->vm_flags);
2562         if (IS_ERR(file))
2563                 return PTR_ERR(file);
2564
2565         if (vma->vm_file)
2566                 fput(vma->vm_file);
2567         vma->vm_file = file;
2568         vma->vm_ops = &shmem_vm_ops;
2569         return 0;
2570 }