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