tmpfs: allow filepage alongside swappage
[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         int error;
831
832         idx = 0;
833         ptr = info->i_direct;
834         spin_lock(&info->lock);
835         limit = info->next_index;
836         size = limit;
837         if (size > SHMEM_NR_DIRECT)
838                 size = SHMEM_NR_DIRECT;
839         offset = shmem_find_swp(entry, ptr, ptr+size);
840         if (offset >= 0) {
841                 shmem_swp_balance_unmap();
842                 goto found;
843         }
844         if (!info->i_indirect)
845                 goto lost2;
846
847         dir = shmem_dir_map(info->i_indirect);
848         stage = SHMEM_NR_DIRECT + ENTRIES_PER_PAGEPAGE/2;
849
850         for (idx = SHMEM_NR_DIRECT; idx < limit; idx += ENTRIES_PER_PAGE, dir++) {
851                 if (unlikely(idx == stage)) {
852                         shmem_dir_unmap(dir-1);
853                         dir = shmem_dir_map(info->i_indirect) +
854                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
855                         while (!*dir) {
856                                 dir++;
857                                 idx += ENTRIES_PER_PAGEPAGE;
858                                 if (idx >= limit)
859                                         goto lost1;
860                         }
861                         stage = idx + ENTRIES_PER_PAGEPAGE;
862                         subdir = *dir;
863                         shmem_dir_unmap(dir);
864                         dir = shmem_dir_map(subdir);
865                 }
866                 subdir = *dir;
867                 if (subdir && page_private(subdir)) {
868                         ptr = shmem_swp_map(subdir);
869                         size = limit - idx;
870                         if (size > ENTRIES_PER_PAGE)
871                                 size = ENTRIES_PER_PAGE;
872                         offset = shmem_find_swp(entry, ptr, ptr+size);
873                         if (offset >= 0) {
874                                 shmem_dir_unmap(dir);
875                                 goto found;
876                         }
877                         shmem_swp_unmap(ptr);
878                 }
879         }
880 lost1:
881         shmem_dir_unmap(dir-1);
882 lost2:
883         spin_unlock(&info->lock);
884         return 0;
885 found:
886         idx += offset;
887         inode = &info->vfs_inode;
888         error = add_to_page_cache(page, inode->i_mapping, idx, GFP_ATOMIC);
889         if (error == -EEXIST) {
890                 struct page *filepage = find_get_page(inode->i_mapping, idx);
891                 if (filepage) {
892                         /*
893                          * There might be a more uptodate page coming down
894                          * from a stacked writepage: forget our swappage if so.
895                          */
896                         if (PageUptodate(filepage))
897                                 error = 0;
898                         page_cache_release(filepage);
899                 }
900         }
901         if (!error) {
902                 delete_from_swap_cache(page);
903                 set_page_dirty(page);
904                 info->flags |= SHMEM_PAGEIN;
905                 shmem_swp_set(info, ptr + offset, 0);
906         }
907         shmem_swp_unmap(ptr);
908         spin_unlock(&info->lock);
909         /*
910          * Decrement swap count even when the entry is left behind:
911          * try_to_unuse will skip over mms, then reincrement count.
912          */
913         swap_free(entry);
914         return 1;
915 }
916
917 /*
918  * shmem_unuse() search for an eventually swapped out shmem page.
919  */
920 int shmem_unuse(swp_entry_t entry, struct page *page)
921 {
922         struct list_head *p, *next;
923         struct shmem_inode_info *info;
924         int found = 0;
925
926         spin_lock(&shmem_swaplist_lock);
927         list_for_each_safe(p, next, &shmem_swaplist) {
928                 info = list_entry(p, struct shmem_inode_info, swaplist);
929                 if (!info->swapped)
930                         list_del_init(&info->swaplist);
931                 else if (shmem_unuse_inode(info, entry, page)) {
932                         /* move head to start search for next from here */
933                         list_move_tail(&shmem_swaplist, &info->swaplist);
934                         found = 1;
935                         break;
936                 }
937         }
938         spin_unlock(&shmem_swaplist_lock);
939         return found;
940 }
941
942 /*
943  * Move the page from the page cache to the swap cache.
944  */
945 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
946 {
947         struct shmem_inode_info *info;
948         swp_entry_t *entry, swap;
949         struct address_space *mapping;
950         unsigned long index;
951         struct inode *inode;
952
953         BUG_ON(!PageLocked(page));
954         mapping = page->mapping;
955         index = page->index;
956         inode = mapping->host;
957         info = SHMEM_I(inode);
958         if (info->flags & VM_LOCKED)
959                 goto redirty;
960         if (!total_swap_pages)
961                 goto redirty;
962
963         /*
964          * shmem_backing_dev_info's capabilities prevent regular writeback or
965          * sync from ever calling shmem_writepage; but a stacking filesystem
966          * may use the ->writepage of its underlying filesystem, in which case
967          * tmpfs should write out to swap only in response to memory pressure,
968          * and not for pdflush or sync.  However, in those cases, we do still
969          * want to check if there's a redundant swappage to be discarded.
970          */
971         if (wbc->for_reclaim)
972                 swap = get_swap_page();
973         else
974                 swap.val = 0;
975
976         spin_lock(&info->lock);
977         if (index >= info->next_index) {
978                 BUG_ON(!(info->flags & SHMEM_TRUNCATE));
979                 goto unlock;
980         }
981         entry = shmem_swp_entry(info, index, NULL);
982         if (entry->val) {
983                 /*
984                  * The more uptodate page coming down from a stacked
985                  * writepage should replace our old swappage.
986                  */
987                 free_swap_and_cache(*entry);
988                 shmem_swp_set(info, entry, 0);
989         }
990         shmem_recalc_inode(inode);
991
992         if (swap.val && add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
993                 remove_from_page_cache(page);
994                 shmem_swp_set(info, entry, swap.val);
995                 shmem_swp_unmap(entry);
996                 spin_unlock(&info->lock);
997                 if (list_empty(&info->swaplist)) {
998                         spin_lock(&shmem_swaplist_lock);
999                         /* move instead of add in case we're racing */
1000                         list_move_tail(&info->swaplist, &shmem_swaplist);
1001                         spin_unlock(&shmem_swaplist_lock);
1002                 }
1003                 swap_duplicate(swap);
1004                 BUG_ON(page_mapped(page));
1005                 page_cache_release(page);       /* pagecache ref */
1006                 set_page_dirty(page);
1007                 unlock_page(page);
1008                 return 0;
1009         }
1010
1011         shmem_swp_unmap(entry);
1012 unlock:
1013         spin_unlock(&info->lock);
1014         swap_free(swap);
1015 redirty:
1016         set_page_dirty(page);
1017         if (wbc->for_reclaim)
1018                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
1019         unlock_page(page);
1020         return 0;
1021 }
1022
1023 #ifdef CONFIG_NUMA
1024 static inline int shmem_parse_mpol(char *value, int *policy, nodemask_t *policy_nodes)
1025 {
1026         char *nodelist = strchr(value, ':');
1027         int err = 1;
1028
1029         if (nodelist) {
1030                 /* NUL-terminate policy string */
1031                 *nodelist++ = '\0';
1032                 if (nodelist_parse(nodelist, *policy_nodes))
1033                         goto out;
1034                 if (!nodes_subset(*policy_nodes, node_states[N_HIGH_MEMORY]))
1035                         goto out;
1036         }
1037         if (!strcmp(value, "default")) {
1038                 *policy = MPOL_DEFAULT;
1039                 /* Don't allow a nodelist */
1040                 if (!nodelist)
1041                         err = 0;
1042         } else if (!strcmp(value, "prefer")) {
1043                 *policy = MPOL_PREFERRED;
1044                 /* Insist on a nodelist of one node only */
1045                 if (nodelist) {
1046                         char *rest = nodelist;
1047                         while (isdigit(*rest))
1048                                 rest++;
1049                         if (!*rest)
1050                                 err = 0;
1051                 }
1052         } else if (!strcmp(value, "bind")) {
1053                 *policy = MPOL_BIND;
1054                 /* Insist on a nodelist */
1055                 if (nodelist)
1056                         err = 0;
1057         } else if (!strcmp(value, "interleave")) {
1058                 *policy = MPOL_INTERLEAVE;
1059                 /*
1060                  * Default to online nodes with memory if no nodelist
1061                  */
1062                 if (!nodelist)
1063                         *policy_nodes = node_states[N_HIGH_MEMORY];
1064                 err = 0;
1065         }
1066 out:
1067         /* Restore string for error message */
1068         if (nodelist)
1069                 *--nodelist = ':';
1070         return err;
1071 }
1072
1073 static struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1074                         struct shmem_inode_info *info, unsigned long idx)
1075 {
1076         struct vm_area_struct pvma;
1077         struct page *page;
1078
1079         /* Create a pseudo vma that just contains the policy */
1080         pvma.vm_start = 0;
1081         pvma.vm_pgoff = idx;
1082         pvma.vm_ops = NULL;
1083         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, idx);
1084         page = swapin_readahead(entry, gfp, &pvma, 0);
1085         mpol_free(pvma.vm_policy);
1086         return page;
1087 }
1088
1089 static struct page *shmem_alloc_page(gfp_t gfp,
1090                         struct shmem_inode_info *info, unsigned long idx)
1091 {
1092         struct vm_area_struct pvma;
1093         struct page *page;
1094
1095         /* Create a pseudo vma that just contains the policy */
1096         pvma.vm_start = 0;
1097         pvma.vm_pgoff = idx;
1098         pvma.vm_ops = NULL;
1099         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, idx);
1100         page = alloc_page_vma(gfp, &pvma, 0);
1101         mpol_free(pvma.vm_policy);
1102         return page;
1103 }
1104 #else
1105 static inline int shmem_parse_mpol(char *value, int *policy,
1106                                                 nodemask_t *policy_nodes)
1107 {
1108         return 1;
1109 }
1110
1111 static inline struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1112                         struct shmem_inode_info *info, unsigned long idx)
1113 {
1114         return swapin_readahead(entry, gfp, NULL, 0);
1115 }
1116
1117 static inline struct page *shmem_alloc_page(gfp_t gfp,
1118                         struct shmem_inode_info *info, unsigned long idx)
1119 {
1120         return alloc_page(gfp);
1121 }
1122 #endif
1123
1124 /*
1125  * shmem_getpage - either get the page from swap or allocate a new one
1126  *
1127  * If we allocate a new one we do not mark it dirty. That's up to the
1128  * vm. If we swap it in we mark it dirty since we also free the swap
1129  * entry since a page cannot live in both the swap and page cache
1130  */
1131 static int shmem_getpage(struct inode *inode, unsigned long idx,
1132                         struct page **pagep, enum sgp_type sgp, int *type)
1133 {
1134         struct address_space *mapping = inode->i_mapping;
1135         struct shmem_inode_info *info = SHMEM_I(inode);
1136         struct shmem_sb_info *sbinfo;
1137         struct page *filepage = *pagep;
1138         struct page *swappage;
1139         swp_entry_t *entry;
1140         swp_entry_t swap;
1141         gfp_t gfp;
1142         int error;
1143
1144         if (idx >= SHMEM_MAX_INDEX)
1145                 return -EFBIG;
1146
1147         if (type)
1148                 *type = 0;
1149
1150         /*
1151          * Normally, filepage is NULL on entry, and either found
1152          * uptodate immediately, or allocated and zeroed, or read
1153          * in under swappage, which is then assigned to filepage.
1154          * But shmem_readpage (required for splice) passes in a locked
1155          * filepage, which may be found not uptodate by other callers
1156          * too, and may need to be copied from the swappage read in.
1157          */
1158 repeat:
1159         if (!filepage)
1160                 filepage = find_lock_page(mapping, idx);
1161         if (filepage && PageUptodate(filepage))
1162                 goto done;
1163         error = 0;
1164         gfp = mapping_gfp_mask(mapping);
1165
1166         spin_lock(&info->lock);
1167         shmem_recalc_inode(inode);
1168         entry = shmem_swp_alloc(info, idx, sgp);
1169         if (IS_ERR(entry)) {
1170                 spin_unlock(&info->lock);
1171                 error = PTR_ERR(entry);
1172                 goto failed;
1173         }
1174         swap = *entry;
1175
1176         if (swap.val) {
1177                 /* Look it up and read it in.. */
1178                 swappage = lookup_swap_cache(swap);
1179                 if (!swappage) {
1180                         shmem_swp_unmap(entry);
1181                         /* here we actually do the io */
1182                         if (type && !(*type & VM_FAULT_MAJOR)) {
1183                                 __count_vm_event(PGMAJFAULT);
1184                                 *type |= VM_FAULT_MAJOR;
1185                         }
1186                         spin_unlock(&info->lock);
1187                         swappage = shmem_swapin(swap, gfp, info, idx);
1188                         if (!swappage) {
1189                                 spin_lock(&info->lock);
1190                                 entry = shmem_swp_alloc(info, idx, sgp);
1191                                 if (IS_ERR(entry))
1192                                         error = PTR_ERR(entry);
1193                                 else {
1194                                         if (entry->val == swap.val)
1195                                                 error = -ENOMEM;
1196                                         shmem_swp_unmap(entry);
1197                                 }
1198                                 spin_unlock(&info->lock);
1199                                 if (error)
1200                                         goto failed;
1201                                 goto repeat;
1202                         }
1203                         wait_on_page_locked(swappage);
1204                         page_cache_release(swappage);
1205                         goto repeat;
1206                 }
1207
1208                 /* We have to do this with page locked to prevent races */
1209                 if (TestSetPageLocked(swappage)) {
1210                         shmem_swp_unmap(entry);
1211                         spin_unlock(&info->lock);
1212                         wait_on_page_locked(swappage);
1213                         page_cache_release(swappage);
1214                         goto repeat;
1215                 }
1216                 if (PageWriteback(swappage)) {
1217                         shmem_swp_unmap(entry);
1218                         spin_unlock(&info->lock);
1219                         wait_on_page_writeback(swappage);
1220                         unlock_page(swappage);
1221                         page_cache_release(swappage);
1222                         goto repeat;
1223                 }
1224                 if (!PageUptodate(swappage)) {
1225                         shmem_swp_unmap(entry);
1226                         spin_unlock(&info->lock);
1227                         unlock_page(swappage);
1228                         page_cache_release(swappage);
1229                         error = -EIO;
1230                         goto failed;
1231                 }
1232
1233                 if (filepage) {
1234                         shmem_swp_set(info, entry, 0);
1235                         shmem_swp_unmap(entry);
1236                         delete_from_swap_cache(swappage);
1237                         spin_unlock(&info->lock);
1238                         copy_highpage(filepage, swappage);
1239                         unlock_page(swappage);
1240                         page_cache_release(swappage);
1241                         flush_dcache_page(filepage);
1242                         SetPageUptodate(filepage);
1243                         set_page_dirty(filepage);
1244                         swap_free(swap);
1245                 } else if (!(error = add_to_page_cache(
1246                                 swappage, mapping, idx, GFP_ATOMIC))) {
1247                         info->flags |= SHMEM_PAGEIN;
1248                         shmem_swp_set(info, entry, 0);
1249                         shmem_swp_unmap(entry);
1250                         delete_from_swap_cache(swappage);
1251                         spin_unlock(&info->lock);
1252                         filepage = swappage;
1253                         set_page_dirty(filepage);
1254                         swap_free(swap);
1255                 } else {
1256                         shmem_swp_unmap(entry);
1257                         spin_unlock(&info->lock);
1258                         unlock_page(swappage);
1259                         page_cache_release(swappage);
1260                         if (error == -ENOMEM) {
1261                                 /* let kswapd refresh zone for GFP_ATOMICs */
1262                                 congestion_wait(WRITE, HZ/50);
1263                         }
1264                         goto repeat;
1265                 }
1266         } else if (sgp == SGP_READ && !filepage) {
1267                 shmem_swp_unmap(entry);
1268                 filepage = find_get_page(mapping, idx);
1269                 if (filepage &&
1270                     (!PageUptodate(filepage) || TestSetPageLocked(filepage))) {
1271                         spin_unlock(&info->lock);
1272                         wait_on_page_locked(filepage);
1273                         page_cache_release(filepage);
1274                         filepage = NULL;
1275                         goto repeat;
1276                 }
1277                 spin_unlock(&info->lock);
1278         } else {
1279                 shmem_swp_unmap(entry);
1280                 sbinfo = SHMEM_SB(inode->i_sb);
1281                 if (sbinfo->max_blocks) {
1282                         spin_lock(&sbinfo->stat_lock);
1283                         if (sbinfo->free_blocks == 0 ||
1284                             shmem_acct_block(info->flags)) {
1285                                 spin_unlock(&sbinfo->stat_lock);
1286                                 spin_unlock(&info->lock);
1287                                 error = -ENOSPC;
1288                                 goto failed;
1289                         }
1290                         sbinfo->free_blocks--;
1291                         inode->i_blocks += BLOCKS_PER_PAGE;
1292                         spin_unlock(&sbinfo->stat_lock);
1293                 } else if (shmem_acct_block(info->flags)) {
1294                         spin_unlock(&info->lock);
1295                         error = -ENOSPC;
1296                         goto failed;
1297                 }
1298
1299                 if (!filepage) {
1300                         spin_unlock(&info->lock);
1301                         filepage = shmem_alloc_page(gfp, info, idx);
1302                         if (!filepage) {
1303                                 shmem_unacct_blocks(info->flags, 1);
1304                                 shmem_free_blocks(inode, 1);
1305                                 error = -ENOMEM;
1306                                 goto failed;
1307                         }
1308
1309                         spin_lock(&info->lock);
1310                         entry = shmem_swp_alloc(info, idx, sgp);
1311                         if (IS_ERR(entry))
1312                                 error = PTR_ERR(entry);
1313                         else {
1314                                 swap = *entry;
1315                                 shmem_swp_unmap(entry);
1316                         }
1317                         if (error || swap.val || 0 != add_to_page_cache_lru(
1318                                         filepage, mapping, idx, GFP_ATOMIC)) {
1319                                 spin_unlock(&info->lock);
1320                                 page_cache_release(filepage);
1321                                 shmem_unacct_blocks(info->flags, 1);
1322                                 shmem_free_blocks(inode, 1);
1323                                 filepage = NULL;
1324                                 if (error)
1325                                         goto failed;
1326                                 goto repeat;
1327                         }
1328                         info->flags |= SHMEM_PAGEIN;
1329                 }
1330
1331                 info->alloced++;
1332                 spin_unlock(&info->lock);
1333                 clear_highpage(filepage);
1334                 flush_dcache_page(filepage);
1335                 SetPageUptodate(filepage);
1336         }
1337 done:
1338         *pagep = filepage;
1339         return 0;
1340
1341 failed:
1342         if (*pagep != filepage) {
1343                 unlock_page(filepage);
1344                 page_cache_release(filepage);
1345         }
1346         return error;
1347 }
1348
1349 static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1350 {
1351         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1352         int error;
1353         int ret;
1354
1355         if (((loff_t)vmf->pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
1356                 return VM_FAULT_SIGBUS;
1357
1358         error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_CACHE, &ret);
1359         if (error)
1360                 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1361
1362         mark_page_accessed(vmf->page);
1363         return ret | VM_FAULT_LOCKED;
1364 }
1365
1366 #ifdef CONFIG_NUMA
1367 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *new)
1368 {
1369         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1370         return mpol_set_shared_policy(&SHMEM_I(i)->policy, vma, new);
1371 }
1372
1373 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
1374                                           unsigned long addr)
1375 {
1376         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1377         unsigned long idx;
1378
1379         idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1380         return mpol_shared_policy_lookup(&SHMEM_I(i)->policy, idx);
1381 }
1382 #endif
1383
1384 int shmem_lock(struct file *file, int lock, struct user_struct *user)
1385 {
1386         struct inode *inode = file->f_path.dentry->d_inode;
1387         struct shmem_inode_info *info = SHMEM_I(inode);
1388         int retval = -ENOMEM;
1389
1390         spin_lock(&info->lock);
1391         if (lock && !(info->flags & VM_LOCKED)) {
1392                 if (!user_shm_lock(inode->i_size, user))
1393                         goto out_nomem;
1394                 info->flags |= VM_LOCKED;
1395         }
1396         if (!lock && (info->flags & VM_LOCKED) && user) {
1397                 user_shm_unlock(inode->i_size, user);
1398                 info->flags &= ~VM_LOCKED;
1399         }
1400         retval = 0;
1401 out_nomem:
1402         spin_unlock(&info->lock);
1403         return retval;
1404 }
1405
1406 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1407 {
1408         file_accessed(file);
1409         vma->vm_ops = &shmem_vm_ops;
1410         vma->vm_flags |= VM_CAN_NONLINEAR;
1411         return 0;
1412 }
1413
1414 static struct inode *
1415 shmem_get_inode(struct super_block *sb, int mode, dev_t dev)
1416 {
1417         struct inode *inode;
1418         struct shmem_inode_info *info;
1419         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1420
1421         if (shmem_reserve_inode(sb))
1422                 return NULL;
1423
1424         inode = new_inode(sb);
1425         if (inode) {
1426                 inode->i_mode = mode;
1427                 inode->i_uid = current->fsuid;
1428                 inode->i_gid = current->fsgid;
1429                 inode->i_blocks = 0;
1430                 inode->i_mapping->a_ops = &shmem_aops;
1431                 inode->i_mapping->backing_dev_info = &shmem_backing_dev_info;
1432                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1433                 inode->i_generation = get_seconds();
1434                 info = SHMEM_I(inode);
1435                 memset(info, 0, (char *)inode - (char *)info);
1436                 spin_lock_init(&info->lock);
1437                 INIT_LIST_HEAD(&info->swaplist);
1438
1439                 switch (mode & S_IFMT) {
1440                 default:
1441                         inode->i_op = &shmem_special_inode_operations;
1442                         init_special_inode(inode, mode, dev);
1443                         break;
1444                 case S_IFREG:
1445                         inode->i_op = &shmem_inode_operations;
1446                         inode->i_fop = &shmem_file_operations;
1447                         mpol_shared_policy_init(&info->policy, sbinfo->policy,
1448                                                         &sbinfo->policy_nodes);
1449                         break;
1450                 case S_IFDIR:
1451                         inc_nlink(inode);
1452                         /* Some things misbehave if size == 0 on a directory */
1453                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
1454                         inode->i_op = &shmem_dir_inode_operations;
1455                         inode->i_fop = &simple_dir_operations;
1456                         break;
1457                 case S_IFLNK:
1458                         /*
1459                          * Must not load anything in the rbtree,
1460                          * mpol_free_shared_policy will not be called.
1461                          */
1462                         mpol_shared_policy_init(&info->policy, MPOL_DEFAULT,
1463                                                 NULL);
1464                         break;
1465                 }
1466         } else
1467                 shmem_free_inode(sb);
1468         return inode;
1469 }
1470
1471 #ifdef CONFIG_TMPFS
1472 static const struct inode_operations shmem_symlink_inode_operations;
1473 static const struct inode_operations shmem_symlink_inline_operations;
1474
1475 /*
1476  * Normally tmpfs avoids the use of shmem_readpage and shmem_write_begin;
1477  * but providing them allows a tmpfs file to be used for splice, sendfile, and
1478  * below the loop driver, in the generic fashion that many filesystems support.
1479  */
1480 static int shmem_readpage(struct file *file, struct page *page)
1481 {
1482         struct inode *inode = page->mapping->host;
1483         int error = shmem_getpage(inode, page->index, &page, SGP_CACHE, NULL);
1484         unlock_page(page);
1485         return error;
1486 }
1487
1488 static int
1489 shmem_write_begin(struct file *file, struct address_space *mapping,
1490                         loff_t pos, unsigned len, unsigned flags,
1491                         struct page **pagep, void **fsdata)
1492 {
1493         struct inode *inode = mapping->host;
1494         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1495         *pagep = NULL;
1496         return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1497 }
1498
1499 static int
1500 shmem_write_end(struct file *file, struct address_space *mapping,
1501                         loff_t pos, unsigned len, unsigned copied,
1502                         struct page *page, void *fsdata)
1503 {
1504         struct inode *inode = mapping->host;
1505
1506         if (pos + copied > inode->i_size)
1507                 i_size_write(inode, pos + copied);
1508
1509         unlock_page(page);
1510         set_page_dirty(page);
1511         page_cache_release(page);
1512
1513         return copied;
1514 }
1515
1516 static void do_shmem_file_read(struct file *filp, loff_t *ppos, read_descriptor_t *desc, read_actor_t actor)
1517 {
1518         struct inode *inode = filp->f_path.dentry->d_inode;
1519         struct address_space *mapping = inode->i_mapping;
1520         unsigned long index, offset;
1521
1522         index = *ppos >> PAGE_CACHE_SHIFT;
1523         offset = *ppos & ~PAGE_CACHE_MASK;
1524
1525         for (;;) {
1526                 struct page *page = NULL;
1527                 unsigned long end_index, nr, ret;
1528                 loff_t i_size = i_size_read(inode);
1529
1530                 end_index = i_size >> PAGE_CACHE_SHIFT;
1531                 if (index > end_index)
1532                         break;
1533                 if (index == end_index) {
1534                         nr = i_size & ~PAGE_CACHE_MASK;
1535                         if (nr <= offset)
1536                                 break;
1537                 }
1538
1539                 desc->error = shmem_getpage(inode, index, &page, SGP_READ, NULL);
1540                 if (desc->error) {
1541                         if (desc->error == -EINVAL)
1542                                 desc->error = 0;
1543                         break;
1544                 }
1545                 if (page)
1546                         unlock_page(page);
1547
1548                 /*
1549                  * We must evaluate after, since reads (unlike writes)
1550                  * are called without i_mutex protection against truncate
1551                  */
1552                 nr = PAGE_CACHE_SIZE;
1553                 i_size = i_size_read(inode);
1554                 end_index = i_size >> PAGE_CACHE_SHIFT;
1555                 if (index == end_index) {
1556                         nr = i_size & ~PAGE_CACHE_MASK;
1557                         if (nr <= offset) {
1558                                 if (page)
1559                                         page_cache_release(page);
1560                                 break;
1561                         }
1562                 }
1563                 nr -= offset;
1564
1565                 if (page) {
1566                         /*
1567                          * If users can be writing to this page using arbitrary
1568                          * virtual addresses, take care about potential aliasing
1569                          * before reading the page on the kernel side.
1570                          */
1571                         if (mapping_writably_mapped(mapping))
1572                                 flush_dcache_page(page);
1573                         /*
1574                          * Mark the page accessed if we read the beginning.
1575                          */
1576                         if (!offset)
1577                                 mark_page_accessed(page);
1578                 } else {
1579                         page = ZERO_PAGE(0);
1580                         page_cache_get(page);
1581                 }
1582
1583                 /*
1584                  * Ok, we have the page, and it's up-to-date, so
1585                  * now we can copy it to user space...
1586                  *
1587                  * The actor routine returns how many bytes were actually used..
1588                  * NOTE! This may not be the same as how much of a user buffer
1589                  * we filled up (we may be padding etc), so we can only update
1590                  * "pos" here (the actor routine has to update the user buffer
1591                  * pointers and the remaining count).
1592                  */
1593                 ret = actor(desc, page, offset, nr);
1594                 offset += ret;
1595                 index += offset >> PAGE_CACHE_SHIFT;
1596                 offset &= ~PAGE_CACHE_MASK;
1597
1598                 page_cache_release(page);
1599                 if (ret != nr || !desc->count)
1600                         break;
1601
1602                 cond_resched();
1603         }
1604
1605         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1606         file_accessed(filp);
1607 }
1608
1609 static ssize_t shmem_file_read(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
1610 {
1611         read_descriptor_t desc;
1612
1613         if ((ssize_t) count < 0)
1614                 return -EINVAL;
1615         if (!access_ok(VERIFY_WRITE, buf, count))
1616                 return -EFAULT;
1617         if (!count)
1618                 return 0;
1619
1620         desc.written = 0;
1621         desc.count = count;
1622         desc.arg.buf = buf;
1623         desc.error = 0;
1624
1625         do_shmem_file_read(filp, ppos, &desc, file_read_actor);
1626         if (desc.written)
1627                 return desc.written;
1628         return desc.error;
1629 }
1630
1631 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
1632 {
1633         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
1634
1635         buf->f_type = TMPFS_MAGIC;
1636         buf->f_bsize = PAGE_CACHE_SIZE;
1637         buf->f_namelen = NAME_MAX;
1638         spin_lock(&sbinfo->stat_lock);
1639         if (sbinfo->max_blocks) {
1640                 buf->f_blocks = sbinfo->max_blocks;
1641                 buf->f_bavail = buf->f_bfree = sbinfo->free_blocks;
1642         }
1643         if (sbinfo->max_inodes) {
1644                 buf->f_files = sbinfo->max_inodes;
1645                 buf->f_ffree = sbinfo->free_inodes;
1646         }
1647         /* else leave those fields 0 like simple_statfs */
1648         spin_unlock(&sbinfo->stat_lock);
1649         return 0;
1650 }
1651
1652 /*
1653  * File creation. Allocate an inode, and we're done..
1654  */
1655 static int
1656 shmem_mknod(struct inode *dir, struct dentry *dentry, int mode, dev_t dev)
1657 {
1658         struct inode *inode = shmem_get_inode(dir->i_sb, mode, dev);
1659         int error = -ENOSPC;
1660
1661         if (inode) {
1662                 error = security_inode_init_security(inode, dir, NULL, NULL,
1663                                                      NULL);
1664                 if (error) {
1665                         if (error != -EOPNOTSUPP) {
1666                                 iput(inode);
1667                                 return error;
1668                         }
1669                 }
1670                 error = shmem_acl_init(inode, dir);
1671                 if (error) {
1672                         iput(inode);
1673                         return error;
1674                 }
1675                 if (dir->i_mode & S_ISGID) {
1676                         inode->i_gid = dir->i_gid;
1677                         if (S_ISDIR(mode))
1678                                 inode->i_mode |= S_ISGID;
1679                 }
1680                 dir->i_size += BOGO_DIRENT_SIZE;
1681                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1682                 d_instantiate(dentry, inode);
1683                 dget(dentry); /* Extra count - pin the dentry in core */
1684         }
1685         return error;
1686 }
1687
1688 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, int mode)
1689 {
1690         int error;
1691
1692         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
1693                 return error;
1694         inc_nlink(dir);
1695         return 0;
1696 }
1697
1698 static int shmem_create(struct inode *dir, struct dentry *dentry, int mode,
1699                 struct nameidata *nd)
1700 {
1701         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
1702 }
1703
1704 /*
1705  * Link a file..
1706  */
1707 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
1708 {
1709         struct inode *inode = old_dentry->d_inode;
1710         int ret;
1711
1712         /*
1713          * No ordinary (disk based) filesystem counts links as inodes;
1714          * but each new link needs a new dentry, pinning lowmem, and
1715          * tmpfs dentries cannot be pruned until they are unlinked.
1716          */
1717         ret = shmem_reserve_inode(inode->i_sb);
1718         if (ret)
1719                 goto out;
1720
1721         dir->i_size += BOGO_DIRENT_SIZE;
1722         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1723         inc_nlink(inode);
1724         atomic_inc(&inode->i_count);    /* New dentry reference */
1725         dget(dentry);           /* Extra pinning count for the created dentry */
1726         d_instantiate(dentry, inode);
1727 out:
1728         return ret;
1729 }
1730
1731 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
1732 {
1733         struct inode *inode = dentry->d_inode;
1734
1735         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
1736                 shmem_free_inode(inode->i_sb);
1737
1738         dir->i_size -= BOGO_DIRENT_SIZE;
1739         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1740         drop_nlink(inode);
1741         dput(dentry);   /* Undo the count from "create" - this does all the work */
1742         return 0;
1743 }
1744
1745 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
1746 {
1747         if (!simple_empty(dentry))
1748                 return -ENOTEMPTY;
1749
1750         drop_nlink(dentry->d_inode);
1751         drop_nlink(dir);
1752         return shmem_unlink(dir, dentry);
1753 }
1754
1755 /*
1756  * The VFS layer already does all the dentry stuff for rename,
1757  * we just have to decrement the usage count for the target if
1758  * it exists so that the VFS layer correctly free's it when it
1759  * gets overwritten.
1760  */
1761 static int shmem_rename(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
1762 {
1763         struct inode *inode = old_dentry->d_inode;
1764         int they_are_dirs = S_ISDIR(inode->i_mode);
1765
1766         if (!simple_empty(new_dentry))
1767                 return -ENOTEMPTY;
1768
1769         if (new_dentry->d_inode) {
1770                 (void) shmem_unlink(new_dir, new_dentry);
1771                 if (they_are_dirs)
1772                         drop_nlink(old_dir);
1773         } else if (they_are_dirs) {
1774                 drop_nlink(old_dir);
1775                 inc_nlink(new_dir);
1776         }
1777
1778         old_dir->i_size -= BOGO_DIRENT_SIZE;
1779         new_dir->i_size += BOGO_DIRENT_SIZE;
1780         old_dir->i_ctime = old_dir->i_mtime =
1781         new_dir->i_ctime = new_dir->i_mtime =
1782         inode->i_ctime = CURRENT_TIME;
1783         return 0;
1784 }
1785
1786 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
1787 {
1788         int error;
1789         int len;
1790         struct inode *inode;
1791         struct page *page = NULL;
1792         char *kaddr;
1793         struct shmem_inode_info *info;
1794
1795         len = strlen(symname) + 1;
1796         if (len > PAGE_CACHE_SIZE)
1797                 return -ENAMETOOLONG;
1798
1799         inode = shmem_get_inode(dir->i_sb, S_IFLNK|S_IRWXUGO, 0);
1800         if (!inode)
1801                 return -ENOSPC;
1802
1803         error = security_inode_init_security(inode, dir, NULL, NULL,
1804                                              NULL);
1805         if (error) {
1806                 if (error != -EOPNOTSUPP) {
1807                         iput(inode);
1808                         return error;
1809                 }
1810                 error = 0;
1811         }
1812
1813         info = SHMEM_I(inode);
1814         inode->i_size = len-1;
1815         if (len <= (char *)inode - (char *)info) {
1816                 /* do it inline */
1817                 memcpy(info, symname, len);
1818                 inode->i_op = &shmem_symlink_inline_operations;
1819         } else {
1820                 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
1821                 if (error) {
1822                         iput(inode);
1823                         return error;
1824                 }
1825                 unlock_page(page);
1826                 inode->i_op = &shmem_symlink_inode_operations;
1827                 kaddr = kmap_atomic(page, KM_USER0);
1828                 memcpy(kaddr, symname, len);
1829                 kunmap_atomic(kaddr, KM_USER0);
1830                 set_page_dirty(page);
1831                 page_cache_release(page);
1832         }
1833         if (dir->i_mode & S_ISGID)
1834                 inode->i_gid = dir->i_gid;
1835         dir->i_size += BOGO_DIRENT_SIZE;
1836         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1837         d_instantiate(dentry, inode);
1838         dget(dentry);
1839         return 0;
1840 }
1841
1842 static void *shmem_follow_link_inline(struct dentry *dentry, struct nameidata *nd)
1843 {
1844         nd_set_link(nd, (char *)SHMEM_I(dentry->d_inode));
1845         return NULL;
1846 }
1847
1848 static void *shmem_follow_link(struct dentry *dentry, struct nameidata *nd)
1849 {
1850         struct page *page = NULL;
1851         int res = shmem_getpage(dentry->d_inode, 0, &page, SGP_READ, NULL);
1852         nd_set_link(nd, res ? ERR_PTR(res) : kmap(page));
1853         if (page)
1854                 unlock_page(page);
1855         return page;
1856 }
1857
1858 static void shmem_put_link(struct dentry *dentry, struct nameidata *nd, void *cookie)
1859 {
1860         if (!IS_ERR(nd_get_link(nd))) {
1861                 struct page *page = cookie;
1862                 kunmap(page);
1863                 mark_page_accessed(page);
1864                 page_cache_release(page);
1865         }
1866 }
1867
1868 static const struct inode_operations shmem_symlink_inline_operations = {
1869         .readlink       = generic_readlink,
1870         .follow_link    = shmem_follow_link_inline,
1871 };
1872
1873 static const struct inode_operations shmem_symlink_inode_operations = {
1874         .truncate       = shmem_truncate,
1875         .readlink       = generic_readlink,
1876         .follow_link    = shmem_follow_link,
1877         .put_link       = shmem_put_link,
1878 };
1879
1880 #ifdef CONFIG_TMPFS_POSIX_ACL
1881 /**
1882  * Superblocks without xattr inode operations will get security.* xattr
1883  * support from the VFS "for free". As soon as we have any other xattrs
1884  * like ACLs, we also need to implement the security.* handlers at
1885  * filesystem level, though.
1886  */
1887
1888 static size_t shmem_xattr_security_list(struct inode *inode, char *list,
1889                                         size_t list_len, const char *name,
1890                                         size_t name_len)
1891 {
1892         return security_inode_listsecurity(inode, list, list_len);
1893 }
1894
1895 static int shmem_xattr_security_get(struct inode *inode, const char *name,
1896                                     void *buffer, size_t size)
1897 {
1898         if (strcmp(name, "") == 0)
1899                 return -EINVAL;
1900         return security_inode_getsecurity(inode, name, buffer, size,
1901                                           -EOPNOTSUPP);
1902 }
1903
1904 static int shmem_xattr_security_set(struct inode *inode, const char *name,
1905                                     const void *value, size_t size, int flags)
1906 {
1907         if (strcmp(name, "") == 0)
1908                 return -EINVAL;
1909         return security_inode_setsecurity(inode, name, value, size, flags);
1910 }
1911
1912 static struct xattr_handler shmem_xattr_security_handler = {
1913         .prefix = XATTR_SECURITY_PREFIX,
1914         .list   = shmem_xattr_security_list,
1915         .get    = shmem_xattr_security_get,
1916         .set    = shmem_xattr_security_set,
1917 };
1918
1919 static struct xattr_handler *shmem_xattr_handlers[] = {
1920         &shmem_xattr_acl_access_handler,
1921         &shmem_xattr_acl_default_handler,
1922         &shmem_xattr_security_handler,
1923         NULL
1924 };
1925 #endif
1926
1927 static struct dentry *shmem_get_parent(struct dentry *child)
1928 {
1929         return ERR_PTR(-ESTALE);
1930 }
1931
1932 static int shmem_match(struct inode *ino, void *vfh)
1933 {
1934         __u32 *fh = vfh;
1935         __u64 inum = fh[2];
1936         inum = (inum << 32) | fh[1];
1937         return ino->i_ino == inum && fh[0] == ino->i_generation;
1938 }
1939
1940 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
1941                 struct fid *fid, int fh_len, int fh_type)
1942 {
1943         struct inode *inode;
1944         struct dentry *dentry = NULL;
1945         u64 inum = fid->raw[2];
1946         inum = (inum << 32) | fid->raw[1];
1947
1948         if (fh_len < 3)
1949                 return NULL;
1950
1951         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
1952                         shmem_match, fid->raw);
1953         if (inode) {
1954                 dentry = d_find_alias(inode);
1955                 iput(inode);
1956         }
1957
1958         return dentry;
1959 }
1960
1961 static int shmem_encode_fh(struct dentry *dentry, __u32 *fh, int *len,
1962                                 int connectable)
1963 {
1964         struct inode *inode = dentry->d_inode;
1965
1966         if (*len < 3)
1967                 return 255;
1968
1969         if (hlist_unhashed(&inode->i_hash)) {
1970                 /* Unfortunately insert_inode_hash is not idempotent,
1971                  * so as we hash inodes here rather than at creation
1972                  * time, we need a lock to ensure we only try
1973                  * to do it once
1974                  */
1975                 static DEFINE_SPINLOCK(lock);
1976                 spin_lock(&lock);
1977                 if (hlist_unhashed(&inode->i_hash))
1978                         __insert_inode_hash(inode,
1979                                             inode->i_ino + inode->i_generation);
1980                 spin_unlock(&lock);
1981         }
1982
1983         fh[0] = inode->i_generation;
1984         fh[1] = inode->i_ino;
1985         fh[2] = ((__u64)inode->i_ino) >> 32;
1986
1987         *len = 3;
1988         return 1;
1989 }
1990
1991 static const struct export_operations shmem_export_ops = {
1992         .get_parent     = shmem_get_parent,
1993         .encode_fh      = shmem_encode_fh,
1994         .fh_to_dentry   = shmem_fh_to_dentry,
1995 };
1996
1997 static int shmem_parse_options(char *options, int *mode, uid_t *uid,
1998         gid_t *gid, unsigned long *blocks, unsigned long *inodes,
1999         int *policy, nodemask_t *policy_nodes)
2000 {
2001         char *this_char, *value, *rest;
2002
2003         while (options != NULL) {
2004                 this_char = options;
2005                 for (;;) {
2006                         /*
2007                          * NUL-terminate this option: unfortunately,
2008                          * mount options form a comma-separated list,
2009                          * but mpol's nodelist may also contain commas.
2010                          */
2011                         options = strchr(options, ',');
2012                         if (options == NULL)
2013                                 break;
2014                         options++;
2015                         if (!isdigit(*options)) {
2016                                 options[-1] = '\0';
2017                                 break;
2018                         }
2019                 }
2020                 if (!*this_char)
2021                         continue;
2022                 if ((value = strchr(this_char,'=')) != NULL) {
2023                         *value++ = 0;
2024                 } else {
2025                         printk(KERN_ERR
2026                             "tmpfs: No value for mount option '%s'\n",
2027                             this_char);
2028                         return 1;
2029                 }
2030
2031                 if (!strcmp(this_char,"size")) {
2032                         unsigned long long size;
2033                         size = memparse(value,&rest);
2034                         if (*rest == '%') {
2035                                 size <<= PAGE_SHIFT;
2036                                 size *= totalram_pages;
2037                                 do_div(size, 100);
2038                                 rest++;
2039                         }
2040                         if (*rest)
2041                                 goto bad_val;
2042                         *blocks = DIV_ROUND_UP(size, PAGE_CACHE_SIZE);
2043                 } else if (!strcmp(this_char,"nr_blocks")) {
2044                         *blocks = memparse(value,&rest);
2045                         if (*rest)
2046                                 goto bad_val;
2047                 } else if (!strcmp(this_char,"nr_inodes")) {
2048                         *inodes = memparse(value,&rest);
2049                         if (*rest)
2050                                 goto bad_val;
2051                 } else if (!strcmp(this_char,"mode")) {
2052                         if (!mode)
2053                                 continue;
2054                         *mode = simple_strtoul(value,&rest,8);
2055                         if (*rest)
2056                                 goto bad_val;
2057                 } else if (!strcmp(this_char,"uid")) {
2058                         if (!uid)
2059                                 continue;
2060                         *uid = simple_strtoul(value,&rest,0);
2061                         if (*rest)
2062                                 goto bad_val;
2063                 } else if (!strcmp(this_char,"gid")) {
2064                         if (!gid)
2065                                 continue;
2066                         *gid = simple_strtoul(value,&rest,0);
2067                         if (*rest)
2068                                 goto bad_val;
2069                 } else if (!strcmp(this_char,"mpol")) {
2070                         if (shmem_parse_mpol(value,policy,policy_nodes))
2071                                 goto bad_val;
2072                 } else {
2073                         printk(KERN_ERR "tmpfs: Bad mount option %s\n",
2074                                this_char);
2075                         return 1;
2076                 }
2077         }
2078         return 0;
2079
2080 bad_val:
2081         printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
2082                value, this_char);
2083         return 1;
2084
2085 }
2086
2087 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2088 {
2089         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2090         unsigned long max_blocks = sbinfo->max_blocks;
2091         unsigned long max_inodes = sbinfo->max_inodes;
2092         int policy = sbinfo->policy;
2093         nodemask_t policy_nodes = sbinfo->policy_nodes;
2094         unsigned long blocks;
2095         unsigned long inodes;
2096         int error = -EINVAL;
2097
2098         if (shmem_parse_options(data, NULL, NULL, NULL, &max_blocks,
2099                                 &max_inodes, &policy, &policy_nodes))
2100                 return error;
2101
2102         spin_lock(&sbinfo->stat_lock);
2103         blocks = sbinfo->max_blocks - sbinfo->free_blocks;
2104         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2105         if (max_blocks < blocks)
2106                 goto out;
2107         if (max_inodes < inodes)
2108                 goto out;
2109         /*
2110          * Those tests also disallow limited->unlimited while any are in
2111          * use, so i_blocks will always be zero when max_blocks is zero;
2112          * but we must separately disallow unlimited->limited, because
2113          * in that case we have no record of how much is already in use.
2114          */
2115         if (max_blocks && !sbinfo->max_blocks)
2116                 goto out;
2117         if (max_inodes && !sbinfo->max_inodes)
2118                 goto out;
2119
2120         error = 0;
2121         sbinfo->max_blocks  = max_blocks;
2122         sbinfo->free_blocks = max_blocks - blocks;
2123         sbinfo->max_inodes  = max_inodes;
2124         sbinfo->free_inodes = max_inodes - inodes;
2125         sbinfo->policy = policy;
2126         sbinfo->policy_nodes = policy_nodes;
2127 out:
2128         spin_unlock(&sbinfo->stat_lock);
2129         return error;
2130 }
2131 #endif
2132
2133 static void shmem_put_super(struct super_block *sb)
2134 {
2135         kfree(sb->s_fs_info);
2136         sb->s_fs_info = NULL;
2137 }
2138
2139 static int shmem_fill_super(struct super_block *sb,
2140                             void *data, int silent)
2141 {
2142         struct inode *inode;
2143         struct dentry *root;
2144         int mode   = S_IRWXUGO | S_ISVTX;
2145         uid_t uid = current->fsuid;
2146         gid_t gid = current->fsgid;
2147         int err = -ENOMEM;
2148         struct shmem_sb_info *sbinfo;
2149         unsigned long blocks = 0;
2150         unsigned long inodes = 0;
2151         int policy = MPOL_DEFAULT;
2152         nodemask_t policy_nodes = node_states[N_HIGH_MEMORY];
2153
2154 #ifdef CONFIG_TMPFS
2155         /*
2156          * Per default we only allow half of the physical ram per
2157          * tmpfs instance, limiting inodes to one per page of lowmem;
2158          * but the internal instance is left unlimited.
2159          */
2160         if (!(sb->s_flags & MS_NOUSER)) {
2161                 blocks = totalram_pages / 2;
2162                 inodes = totalram_pages - totalhigh_pages;
2163                 if (inodes > blocks)
2164                         inodes = blocks;
2165                 if (shmem_parse_options(data, &mode, &uid, &gid, &blocks,
2166                                         &inodes, &policy, &policy_nodes))
2167                         return -EINVAL;
2168         }
2169         sb->s_export_op = &shmem_export_ops;
2170 #else
2171         sb->s_flags |= MS_NOUSER;
2172 #endif
2173
2174         /* Round up to L1_CACHE_BYTES to resist false sharing */
2175         sbinfo = kmalloc(max((int)sizeof(struct shmem_sb_info),
2176                                 L1_CACHE_BYTES), GFP_KERNEL);
2177         if (!sbinfo)
2178                 return -ENOMEM;
2179
2180         spin_lock_init(&sbinfo->stat_lock);
2181         sbinfo->max_blocks = blocks;
2182         sbinfo->free_blocks = blocks;
2183         sbinfo->max_inodes = inodes;
2184         sbinfo->free_inodes = inodes;
2185         sbinfo->policy = policy;
2186         sbinfo->policy_nodes = policy_nodes;
2187
2188         sb->s_fs_info = sbinfo;
2189         sb->s_maxbytes = SHMEM_MAX_BYTES;
2190         sb->s_blocksize = PAGE_CACHE_SIZE;
2191         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
2192         sb->s_magic = TMPFS_MAGIC;
2193         sb->s_op = &shmem_ops;
2194         sb->s_time_gran = 1;
2195 #ifdef CONFIG_TMPFS_POSIX_ACL
2196         sb->s_xattr = shmem_xattr_handlers;
2197         sb->s_flags |= MS_POSIXACL;
2198 #endif
2199
2200         inode = shmem_get_inode(sb, S_IFDIR | mode, 0);
2201         if (!inode)
2202                 goto failed;
2203         inode->i_uid = uid;
2204         inode->i_gid = gid;
2205         root = d_alloc_root(inode);
2206         if (!root)
2207                 goto failed_iput;
2208         sb->s_root = root;
2209         return 0;
2210
2211 failed_iput:
2212         iput(inode);
2213 failed:
2214         shmem_put_super(sb);
2215         return err;
2216 }
2217
2218 static struct kmem_cache *shmem_inode_cachep;
2219
2220 static struct inode *shmem_alloc_inode(struct super_block *sb)
2221 {
2222         struct shmem_inode_info *p;
2223         p = (struct shmem_inode_info *)kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
2224         if (!p)
2225                 return NULL;
2226         return &p->vfs_inode;
2227 }
2228
2229 static void shmem_destroy_inode(struct inode *inode)
2230 {
2231         if ((inode->i_mode & S_IFMT) == S_IFREG) {
2232                 /* only struct inode is valid if it's an inline symlink */
2233                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
2234         }
2235         shmem_acl_destroy_inode(inode);
2236         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
2237 }
2238
2239 static void init_once(struct kmem_cache *cachep, void *foo)
2240 {
2241         struct shmem_inode_info *p = (struct shmem_inode_info *) foo;
2242
2243         inode_init_once(&p->vfs_inode);
2244 #ifdef CONFIG_TMPFS_POSIX_ACL
2245         p->i_acl = NULL;
2246         p->i_default_acl = NULL;
2247 #endif
2248 }
2249
2250 static int init_inodecache(void)
2251 {
2252         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
2253                                 sizeof(struct shmem_inode_info),
2254                                 0, SLAB_PANIC, init_once);
2255         return 0;
2256 }
2257
2258 static void destroy_inodecache(void)
2259 {
2260         kmem_cache_destroy(shmem_inode_cachep);
2261 }
2262
2263 static const struct address_space_operations shmem_aops = {
2264         .writepage      = shmem_writepage,
2265         .set_page_dirty = __set_page_dirty_no_writeback,
2266 #ifdef CONFIG_TMPFS
2267         .readpage       = shmem_readpage,
2268         .write_begin    = shmem_write_begin,
2269         .write_end      = shmem_write_end,
2270 #endif
2271         .migratepage    = migrate_page,
2272 };
2273
2274 static const struct file_operations shmem_file_operations = {
2275         .mmap           = shmem_mmap,
2276 #ifdef CONFIG_TMPFS
2277         .llseek         = generic_file_llseek,
2278         .read           = shmem_file_read,
2279         .write          = do_sync_write,
2280         .aio_write      = generic_file_aio_write,
2281         .fsync          = simple_sync_file,
2282         .splice_read    = generic_file_splice_read,
2283         .splice_write   = generic_file_splice_write,
2284 #endif
2285 };
2286
2287 static const struct inode_operations shmem_inode_operations = {
2288         .truncate       = shmem_truncate,
2289         .setattr        = shmem_notify_change,
2290         .truncate_range = shmem_truncate_range,
2291 #ifdef CONFIG_TMPFS_POSIX_ACL
2292         .setxattr       = generic_setxattr,
2293         .getxattr       = generic_getxattr,
2294         .listxattr      = generic_listxattr,
2295         .removexattr    = generic_removexattr,
2296         .permission     = shmem_permission,
2297 #endif
2298
2299 };
2300
2301 static const struct inode_operations shmem_dir_inode_operations = {
2302 #ifdef CONFIG_TMPFS
2303         .create         = shmem_create,
2304         .lookup         = simple_lookup,
2305         .link           = shmem_link,
2306         .unlink         = shmem_unlink,
2307         .symlink        = shmem_symlink,
2308         .mkdir          = shmem_mkdir,
2309         .rmdir          = shmem_rmdir,
2310         .mknod          = shmem_mknod,
2311         .rename         = shmem_rename,
2312 #endif
2313 #ifdef CONFIG_TMPFS_POSIX_ACL
2314         .setattr        = shmem_notify_change,
2315         .setxattr       = generic_setxattr,
2316         .getxattr       = generic_getxattr,
2317         .listxattr      = generic_listxattr,
2318         .removexattr    = generic_removexattr,
2319         .permission     = shmem_permission,
2320 #endif
2321 };
2322
2323 static const struct inode_operations shmem_special_inode_operations = {
2324 #ifdef CONFIG_TMPFS_POSIX_ACL
2325         .setattr        = shmem_notify_change,
2326         .setxattr       = generic_setxattr,
2327         .getxattr       = generic_getxattr,
2328         .listxattr      = generic_listxattr,
2329         .removexattr    = generic_removexattr,
2330         .permission     = shmem_permission,
2331 #endif
2332 };
2333
2334 static const struct super_operations shmem_ops = {
2335         .alloc_inode    = shmem_alloc_inode,
2336         .destroy_inode  = shmem_destroy_inode,
2337 #ifdef CONFIG_TMPFS
2338         .statfs         = shmem_statfs,
2339         .remount_fs     = shmem_remount_fs,
2340 #endif
2341         .delete_inode   = shmem_delete_inode,
2342         .drop_inode     = generic_delete_inode,
2343         .put_super      = shmem_put_super,
2344 };
2345
2346 static struct vm_operations_struct shmem_vm_ops = {
2347         .fault          = shmem_fault,
2348 #ifdef CONFIG_NUMA
2349         .set_policy     = shmem_set_policy,
2350         .get_policy     = shmem_get_policy,
2351 #endif
2352 };
2353
2354
2355 static int shmem_get_sb(struct file_system_type *fs_type,
2356         int flags, const char *dev_name, void *data, struct vfsmount *mnt)
2357 {
2358         return get_sb_nodev(fs_type, flags, data, shmem_fill_super, mnt);
2359 }
2360
2361 static struct file_system_type tmpfs_fs_type = {
2362         .owner          = THIS_MODULE,
2363         .name           = "tmpfs",
2364         .get_sb         = shmem_get_sb,
2365         .kill_sb        = kill_litter_super,
2366 };
2367 static struct vfsmount *shm_mnt;
2368
2369 static int __init init_tmpfs(void)
2370 {
2371         int error;
2372
2373         error = bdi_init(&shmem_backing_dev_info);
2374         if (error)
2375                 goto out4;
2376
2377         error = init_inodecache();
2378         if (error)
2379                 goto out3;
2380
2381         error = register_filesystem(&tmpfs_fs_type);
2382         if (error) {
2383                 printk(KERN_ERR "Could not register tmpfs\n");
2384                 goto out2;
2385         }
2386
2387         shm_mnt = vfs_kern_mount(&tmpfs_fs_type, MS_NOUSER,
2388                                 tmpfs_fs_type.name, NULL);
2389         if (IS_ERR(shm_mnt)) {
2390                 error = PTR_ERR(shm_mnt);
2391                 printk(KERN_ERR "Could not kern_mount tmpfs\n");
2392                 goto out1;
2393         }
2394         return 0;
2395
2396 out1:
2397         unregister_filesystem(&tmpfs_fs_type);
2398 out2:
2399         destroy_inodecache();
2400 out3:
2401         bdi_destroy(&shmem_backing_dev_info);
2402 out4:
2403         shm_mnt = ERR_PTR(error);
2404         return error;
2405 }
2406 module_init(init_tmpfs)
2407
2408 /*
2409  * shmem_file_setup - get an unlinked file living in tmpfs
2410  *
2411  * @name: name for dentry (to be seen in /proc/<pid>/maps
2412  * @size: size to be set for the file
2413  *
2414  */
2415 struct file *shmem_file_setup(char *name, loff_t size, unsigned long flags)
2416 {
2417         int error;
2418         struct file *file;
2419         struct inode *inode;
2420         struct dentry *dentry, *root;
2421         struct qstr this;
2422
2423         if (IS_ERR(shm_mnt))
2424                 return (void *)shm_mnt;
2425
2426         if (size < 0 || size > SHMEM_MAX_BYTES)
2427                 return ERR_PTR(-EINVAL);
2428
2429         if (shmem_acct_size(flags, size))
2430                 return ERR_PTR(-ENOMEM);
2431
2432         error = -ENOMEM;
2433         this.name = name;
2434         this.len = strlen(name);
2435         this.hash = 0; /* will go */
2436         root = shm_mnt->mnt_root;
2437         dentry = d_alloc(root, &this);
2438         if (!dentry)
2439                 goto put_memory;
2440
2441         error = -ENFILE;
2442         file = get_empty_filp();
2443         if (!file)
2444                 goto put_dentry;
2445
2446         error = -ENOSPC;
2447         inode = shmem_get_inode(root->d_sb, S_IFREG | S_IRWXUGO, 0);
2448         if (!inode)
2449                 goto close_file;
2450
2451         SHMEM_I(inode)->flags = flags & VM_ACCOUNT;
2452         d_instantiate(dentry, inode);
2453         inode->i_size = size;
2454         inode->i_nlink = 0;     /* It is unlinked */
2455         init_file(file, shm_mnt, dentry, FMODE_WRITE | FMODE_READ,
2456                         &shmem_file_operations);
2457         return file;
2458
2459 close_file:
2460         put_filp(file);
2461 put_dentry:
2462         dput(dentry);
2463 put_memory:
2464         shmem_unacct_size(flags, size);
2465         return ERR_PTR(error);
2466 }
2467
2468 /*
2469  * shmem_zero_setup - setup a shared anonymous mapping
2470  *
2471  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
2472  */
2473 int shmem_zero_setup(struct vm_area_struct *vma)
2474 {
2475         struct file *file;
2476         loff_t size = vma->vm_end - vma->vm_start;
2477
2478         file = shmem_file_setup("dev/zero", size, vma->vm_flags);
2479         if (IS_ERR(file))
2480                 return PTR_ERR(file);
2481
2482         if (vma->vm_file)
2483                 fput(vma->vm_file);
2484         vma->vm_file = file;
2485         vma->vm_ops = &shmem_vm_ops;
2486         return 0;
2487 }