]> git.openfabrics.org - ~shefty/rdma-dev.git/blob - mm/shmem.c
89341b658bd06b05d908f8becfef7fe0d904b808
[~shefty/rdma-dev.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) 2002-2011 Hugh Dickins.
10  * Copyright (C) 2011 Google Inc.
11  * Copyright (C) 2002-2005 VERITAS Software Corporation.
12  * Copyright (C) 2004 Andi Kleen, SuSE Labs
13  *
14  * Extended attribute support for tmpfs:
15  * Copyright (c) 2004, Luke Kenneth Casson Leighton <lkcl@lkcl.net>
16  * Copyright (c) 2004 Red Hat, Inc., James Morris <jmorris@redhat.com>
17  *
18  * tiny-shmem:
19  * Copyright (c) 2004, 2008 Matt Mackall <mpm@selenic.com>
20  *
21  * This file is released under the GPL.
22  */
23
24 #include <linux/fs.h>
25 #include <linux/init.h>
26 #include <linux/vfs.h>
27 #include <linux/mount.h>
28 #include <linux/pagemap.h>
29 #include <linux/file.h>
30 #include <linux/mm.h>
31 #include <linux/export.h>
32 #include <linux/swap.h>
33
34 static struct vfsmount *shm_mnt;
35
36 #ifdef CONFIG_SHMEM
37 /*
38  * This virtual memory filesystem is heavily based on the ramfs. It
39  * extends ramfs by the ability to use swap and honor resource limits
40  * which makes it a completely usable filesystem.
41  */
42
43 #include <linux/xattr.h>
44 #include <linux/exportfs.h>
45 #include <linux/posix_acl.h>
46 #include <linux/generic_acl.h>
47 #include <linux/mman.h>
48 #include <linux/string.h>
49 #include <linux/slab.h>
50 #include <linux/backing-dev.h>
51 #include <linux/shmem_fs.h>
52 #include <linux/writeback.h>
53 #include <linux/blkdev.h>
54 #include <linux/pagevec.h>
55 #include <linux/percpu_counter.h>
56 #include <linux/falloc.h>
57 #include <linux/splice.h>
58 #include <linux/security.h>
59 #include <linux/swapops.h>
60 #include <linux/mempolicy.h>
61 #include <linux/namei.h>
62 #include <linux/ctype.h>
63 #include <linux/migrate.h>
64 #include <linux/highmem.h>
65 #include <linux/seq_file.h>
66 #include <linux/magic.h>
67
68 #include <asm/uaccess.h>
69 #include <asm/pgtable.h>
70
71 #define BLOCKS_PER_PAGE  (PAGE_CACHE_SIZE/512)
72 #define VM_ACCT(size)    (PAGE_CACHE_ALIGN(size) >> PAGE_SHIFT)
73
74 /* Pretend that each entry is of this size in directory's i_size */
75 #define BOGO_DIRENT_SIZE 20
76
77 /* Symlink up to this size is kmalloc'ed instead of using a swappable page */
78 #define SHORT_SYMLINK_LEN 128
79
80 /*
81  * shmem_fallocate and shmem_writepage communicate via inode->i_private
82  * (with i_mutex making sure that it has only one user at a time):
83  * we would prefer not to enlarge the shmem inode just for that.
84  */
85 struct shmem_falloc {
86         pgoff_t start;          /* start of range currently being fallocated */
87         pgoff_t next;           /* the next page offset to be fallocated */
88         pgoff_t nr_falloced;    /* how many new pages have been fallocated */
89         pgoff_t nr_unswapped;   /* how often writepage refused to swap out */
90 };
91
92 /* Flag allocation requirements to shmem_getpage */
93 enum sgp_type {
94         SGP_READ,       /* don't exceed i_size, don't allocate page */
95         SGP_CACHE,      /* don't exceed i_size, may allocate page */
96         SGP_DIRTY,      /* like SGP_CACHE, but set new page dirty */
97         SGP_WRITE,      /* may exceed i_size, may allocate !Uptodate page */
98         SGP_FALLOC,     /* like SGP_WRITE, but make existing page Uptodate */
99 };
100
101 #ifdef CONFIG_TMPFS
102 static unsigned long shmem_default_max_blocks(void)
103 {
104         return totalram_pages / 2;
105 }
106
107 static unsigned long shmem_default_max_inodes(void)
108 {
109         return min(totalram_pages - totalhigh_pages, totalram_pages / 2);
110 }
111 #endif
112
113 static bool shmem_should_replace_page(struct page *page, gfp_t gfp);
114 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
115                                 struct shmem_inode_info *info, pgoff_t index);
116 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
117         struct page **pagep, enum sgp_type sgp, gfp_t gfp, int *fault_type);
118
119 static inline int shmem_getpage(struct inode *inode, pgoff_t index,
120         struct page **pagep, enum sgp_type sgp, int *fault_type)
121 {
122         return shmem_getpage_gfp(inode, index, pagep, sgp,
123                         mapping_gfp_mask(inode->i_mapping), fault_type);
124 }
125
126 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
127 {
128         return sb->s_fs_info;
129 }
130
131 /*
132  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
133  * for shared memory and for shared anonymous (/dev/zero) mappings
134  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
135  * consistent with the pre-accounting of private mappings ...
136  */
137 static inline int shmem_acct_size(unsigned long flags, loff_t size)
138 {
139         return (flags & VM_NORESERVE) ?
140                 0 : security_vm_enough_memory_mm(current->mm, VM_ACCT(size));
141 }
142
143 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
144 {
145         if (!(flags & VM_NORESERVE))
146                 vm_unacct_memory(VM_ACCT(size));
147 }
148
149 /*
150  * ... whereas tmpfs objects are accounted incrementally as
151  * pages are allocated, in order to allow huge sparse files.
152  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
153  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
154  */
155 static inline int shmem_acct_block(unsigned long flags)
156 {
157         return (flags & VM_NORESERVE) ?
158                 security_vm_enough_memory_mm(current->mm, VM_ACCT(PAGE_CACHE_SIZE)) : 0;
159 }
160
161 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
162 {
163         if (flags & VM_NORESERVE)
164                 vm_unacct_memory(pages * VM_ACCT(PAGE_CACHE_SIZE));
165 }
166
167 static const struct super_operations shmem_ops;
168 static const struct address_space_operations shmem_aops;
169 static const struct file_operations shmem_file_operations;
170 static const struct inode_operations shmem_inode_operations;
171 static const struct inode_operations shmem_dir_inode_operations;
172 static const struct inode_operations shmem_special_inode_operations;
173 static const struct vm_operations_struct shmem_vm_ops;
174
175 static struct backing_dev_info shmem_backing_dev_info  __read_mostly = {
176         .ra_pages       = 0,    /* No readahead */
177         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_SWAP_BACKED,
178 };
179
180 static LIST_HEAD(shmem_swaplist);
181 static DEFINE_MUTEX(shmem_swaplist_mutex);
182
183 static int shmem_reserve_inode(struct super_block *sb)
184 {
185         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
186         if (sbinfo->max_inodes) {
187                 spin_lock(&sbinfo->stat_lock);
188                 if (!sbinfo->free_inodes) {
189                         spin_unlock(&sbinfo->stat_lock);
190                         return -ENOSPC;
191                 }
192                 sbinfo->free_inodes--;
193                 spin_unlock(&sbinfo->stat_lock);
194         }
195         return 0;
196 }
197
198 static void shmem_free_inode(struct super_block *sb)
199 {
200         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
201         if (sbinfo->max_inodes) {
202                 spin_lock(&sbinfo->stat_lock);
203                 sbinfo->free_inodes++;
204                 spin_unlock(&sbinfo->stat_lock);
205         }
206 }
207
208 /**
209  * shmem_recalc_inode - recalculate the block usage of an inode
210  * @inode: inode to recalc
211  *
212  * We have to calculate the free blocks since the mm can drop
213  * undirtied hole pages behind our back.
214  *
215  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
216  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
217  *
218  * It has to be called with the spinlock held.
219  */
220 static void shmem_recalc_inode(struct inode *inode)
221 {
222         struct shmem_inode_info *info = SHMEM_I(inode);
223         long freed;
224
225         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
226         if (freed > 0) {
227                 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
228                 if (sbinfo->max_blocks)
229                         percpu_counter_add(&sbinfo->used_blocks, -freed);
230                 info->alloced -= freed;
231                 inode->i_blocks -= freed * BLOCKS_PER_PAGE;
232                 shmem_unacct_blocks(info->flags, freed);
233         }
234 }
235
236 /*
237  * Replace item expected in radix tree by a new item, while holding tree lock.
238  */
239 static int shmem_radix_tree_replace(struct address_space *mapping,
240                         pgoff_t index, void *expected, void *replacement)
241 {
242         void **pslot;
243         void *item = NULL;
244
245         VM_BUG_ON(!expected);
246         pslot = radix_tree_lookup_slot(&mapping->page_tree, index);
247         if (pslot)
248                 item = radix_tree_deref_slot_protected(pslot,
249                                                         &mapping->tree_lock);
250         if (item != expected)
251                 return -ENOENT;
252         if (replacement)
253                 radix_tree_replace_slot(pslot, replacement);
254         else
255                 radix_tree_delete(&mapping->page_tree, index);
256         return 0;
257 }
258
259 /*
260  * Sometimes, before we decide whether to proceed or to fail, we must check
261  * that an entry was not already brought back from swap by a racing thread.
262  *
263  * Checking page is not enough: by the time a SwapCache page is locked, it
264  * might be reused, and again be SwapCache, using the same swap as before.
265  */
266 static bool shmem_confirm_swap(struct address_space *mapping,
267                                pgoff_t index, swp_entry_t swap)
268 {
269         void *item;
270
271         rcu_read_lock();
272         item = radix_tree_lookup(&mapping->page_tree, index);
273         rcu_read_unlock();
274         return item == swp_to_radix_entry(swap);
275 }
276
277 /*
278  * Like add_to_page_cache_locked, but error if expected item has gone.
279  */
280 static int shmem_add_to_page_cache(struct page *page,
281                                    struct address_space *mapping,
282                                    pgoff_t index, gfp_t gfp, void *expected)
283 {
284         int error;
285
286         VM_BUG_ON(!PageLocked(page));
287         VM_BUG_ON(!PageSwapBacked(page));
288
289         page_cache_get(page);
290         page->mapping = mapping;
291         page->index = index;
292
293         spin_lock_irq(&mapping->tree_lock);
294         if (!expected)
295                 error = radix_tree_insert(&mapping->page_tree, index, page);
296         else
297                 error = shmem_radix_tree_replace(mapping, index, expected,
298                                                                  page);
299         if (!error) {
300                 mapping->nrpages++;
301                 __inc_zone_page_state(page, NR_FILE_PAGES);
302                 __inc_zone_page_state(page, NR_SHMEM);
303                 spin_unlock_irq(&mapping->tree_lock);
304         } else {
305                 page->mapping = NULL;
306                 spin_unlock_irq(&mapping->tree_lock);
307                 page_cache_release(page);
308         }
309         return error;
310 }
311
312 /*
313  * Like delete_from_page_cache, but substitutes swap for page.
314  */
315 static void shmem_delete_from_page_cache(struct page *page, void *radswap)
316 {
317         struct address_space *mapping = page->mapping;
318         int error;
319
320         spin_lock_irq(&mapping->tree_lock);
321         error = shmem_radix_tree_replace(mapping, page->index, page, radswap);
322         page->mapping = NULL;
323         mapping->nrpages--;
324         __dec_zone_page_state(page, NR_FILE_PAGES);
325         __dec_zone_page_state(page, NR_SHMEM);
326         spin_unlock_irq(&mapping->tree_lock);
327         page_cache_release(page);
328         BUG_ON(error);
329 }
330
331 /*
332  * Like find_get_pages, but collecting swap entries as well as pages.
333  */
334 static unsigned shmem_find_get_pages_and_swap(struct address_space *mapping,
335                                         pgoff_t start, unsigned int nr_pages,
336                                         struct page **pages, pgoff_t *indices)
337 {
338         unsigned int i;
339         unsigned int ret;
340         unsigned int nr_found;
341
342         rcu_read_lock();
343 restart:
344         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
345                                 (void ***)pages, indices, start, nr_pages);
346         ret = 0;
347         for (i = 0; i < nr_found; i++) {
348                 struct page *page;
349 repeat:
350                 page = radix_tree_deref_slot((void **)pages[i]);
351                 if (unlikely(!page))
352                         continue;
353                 if (radix_tree_exception(page)) {
354                         if (radix_tree_deref_retry(page))
355                                 goto restart;
356                         /*
357                          * Otherwise, we must be storing a swap entry
358                          * here as an exceptional entry: so return it
359                          * without attempting to raise page count.
360                          */
361                         goto export;
362                 }
363                 if (!page_cache_get_speculative(page))
364                         goto repeat;
365
366                 /* Has the page moved? */
367                 if (unlikely(page != *((void **)pages[i]))) {
368                         page_cache_release(page);
369                         goto repeat;
370                 }
371 export:
372                 indices[ret] = indices[i];
373                 pages[ret] = page;
374                 ret++;
375         }
376         if (unlikely(!ret && nr_found))
377                 goto restart;
378         rcu_read_unlock();
379         return ret;
380 }
381
382 /*
383  * Remove swap entry from radix tree, free the swap and its page cache.
384  */
385 static int shmem_free_swap(struct address_space *mapping,
386                            pgoff_t index, void *radswap)
387 {
388         int error;
389
390         spin_lock_irq(&mapping->tree_lock);
391         error = shmem_radix_tree_replace(mapping, index, radswap, NULL);
392         spin_unlock_irq(&mapping->tree_lock);
393         if (!error)
394                 free_swap_and_cache(radix_to_swp_entry(radswap));
395         return error;
396 }
397
398 /*
399  * Pagevec may contain swap entries, so shuffle up pages before releasing.
400  */
401 static void shmem_deswap_pagevec(struct pagevec *pvec)
402 {
403         int i, j;
404
405         for (i = 0, j = 0; i < pagevec_count(pvec); i++) {
406                 struct page *page = pvec->pages[i];
407                 if (!radix_tree_exceptional_entry(page))
408                         pvec->pages[j++] = page;
409         }
410         pvec->nr = j;
411 }
412
413 /*
414  * SysV IPC SHM_UNLOCK restore Unevictable pages to their evictable lists.
415  */
416 void shmem_unlock_mapping(struct address_space *mapping)
417 {
418         struct pagevec pvec;
419         pgoff_t indices[PAGEVEC_SIZE];
420         pgoff_t index = 0;
421
422         pagevec_init(&pvec, 0);
423         /*
424          * Minor point, but we might as well stop if someone else SHM_LOCKs it.
425          */
426         while (!mapping_unevictable(mapping)) {
427                 /*
428                  * Avoid pagevec_lookup(): find_get_pages() returns 0 as if it
429                  * has finished, if it hits a row of PAGEVEC_SIZE swap entries.
430                  */
431                 pvec.nr = shmem_find_get_pages_and_swap(mapping, index,
432                                         PAGEVEC_SIZE, pvec.pages, indices);
433                 if (!pvec.nr)
434                         break;
435                 index = indices[pvec.nr - 1] + 1;
436                 shmem_deswap_pagevec(&pvec);
437                 check_move_unevictable_pages(pvec.pages, pvec.nr);
438                 pagevec_release(&pvec);
439                 cond_resched();
440         }
441 }
442
443 /*
444  * Remove range of pages and swap entries from radix tree, and free them.
445  * If !unfalloc, truncate or punch hole; if unfalloc, undo failed fallocate.
446  */
447 static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
448                                                                  bool unfalloc)
449 {
450         struct address_space *mapping = inode->i_mapping;
451         struct shmem_inode_info *info = SHMEM_I(inode);
452         pgoff_t start = (lstart + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
453         pgoff_t end = (lend + 1) >> PAGE_CACHE_SHIFT;
454         unsigned int partial_start = lstart & (PAGE_CACHE_SIZE - 1);
455         unsigned int partial_end = (lend + 1) & (PAGE_CACHE_SIZE - 1);
456         struct pagevec pvec;
457         pgoff_t indices[PAGEVEC_SIZE];
458         long nr_swaps_freed = 0;
459         pgoff_t index;
460         int i;
461
462         if (lend == -1)
463                 end = -1;       /* unsigned, so actually very big */
464
465         pagevec_init(&pvec, 0);
466         index = start;
467         while (index < end) {
468                 pvec.nr = shmem_find_get_pages_and_swap(mapping, index,
469                                 min(end - index, (pgoff_t)PAGEVEC_SIZE),
470                                                         pvec.pages, indices);
471                 if (!pvec.nr)
472                         break;
473                 mem_cgroup_uncharge_start();
474                 for (i = 0; i < pagevec_count(&pvec); i++) {
475                         struct page *page = pvec.pages[i];
476
477                         index = indices[i];
478                         if (index >= end)
479                                 break;
480
481                         if (radix_tree_exceptional_entry(page)) {
482                                 if (unfalloc)
483                                         continue;
484                                 nr_swaps_freed += !shmem_free_swap(mapping,
485                                                                 index, page);
486                                 continue;
487                         }
488
489                         if (!trylock_page(page))
490                                 continue;
491                         if (!unfalloc || !PageUptodate(page)) {
492                                 if (page->mapping == mapping) {
493                                         VM_BUG_ON(PageWriteback(page));
494                                         truncate_inode_page(mapping, page);
495                                 }
496                         }
497                         unlock_page(page);
498                 }
499                 shmem_deswap_pagevec(&pvec);
500                 pagevec_release(&pvec);
501                 mem_cgroup_uncharge_end();
502                 cond_resched();
503                 index++;
504         }
505
506         if (partial_start) {
507                 struct page *page = NULL;
508                 shmem_getpage(inode, start - 1, &page, SGP_READ, NULL);
509                 if (page) {
510                         unsigned int top = PAGE_CACHE_SIZE;
511                         if (start > end) {
512                                 top = partial_end;
513                                 partial_end = 0;
514                         }
515                         zero_user_segment(page, partial_start, top);
516                         set_page_dirty(page);
517                         unlock_page(page);
518                         page_cache_release(page);
519                 }
520         }
521         if (partial_end) {
522                 struct page *page = NULL;
523                 shmem_getpage(inode, end, &page, SGP_READ, NULL);
524                 if (page) {
525                         zero_user_segment(page, 0, partial_end);
526                         set_page_dirty(page);
527                         unlock_page(page);
528                         page_cache_release(page);
529                 }
530         }
531         if (start >= end)
532                 return;
533
534         index = start;
535         for ( ; ; ) {
536                 cond_resched();
537                 pvec.nr = shmem_find_get_pages_and_swap(mapping, index,
538                                 min(end - index, (pgoff_t)PAGEVEC_SIZE),
539                                                         pvec.pages, indices);
540                 if (!pvec.nr) {
541                         if (index == start || unfalloc)
542                                 break;
543                         index = start;
544                         continue;
545                 }
546                 if ((index == start || unfalloc) && indices[0] >= end) {
547                         shmem_deswap_pagevec(&pvec);
548                         pagevec_release(&pvec);
549                         break;
550                 }
551                 mem_cgroup_uncharge_start();
552                 for (i = 0; i < pagevec_count(&pvec); i++) {
553                         struct page *page = pvec.pages[i];
554
555                         index = indices[i];
556                         if (index >= end)
557                                 break;
558
559                         if (radix_tree_exceptional_entry(page)) {
560                                 if (unfalloc)
561                                         continue;
562                                 nr_swaps_freed += !shmem_free_swap(mapping,
563                                                                 index, page);
564                                 continue;
565                         }
566
567                         lock_page(page);
568                         if (!unfalloc || !PageUptodate(page)) {
569                                 if (page->mapping == mapping) {
570                                         VM_BUG_ON(PageWriteback(page));
571                                         truncate_inode_page(mapping, page);
572                                 }
573                         }
574                         unlock_page(page);
575                 }
576                 shmem_deswap_pagevec(&pvec);
577                 pagevec_release(&pvec);
578                 mem_cgroup_uncharge_end();
579                 index++;
580         }
581
582         spin_lock(&info->lock);
583         info->swapped -= nr_swaps_freed;
584         shmem_recalc_inode(inode);
585         spin_unlock(&info->lock);
586 }
587
588 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
589 {
590         shmem_undo_range(inode, lstart, lend, false);
591         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
592 }
593 EXPORT_SYMBOL_GPL(shmem_truncate_range);
594
595 static int shmem_setattr(struct dentry *dentry, struct iattr *attr)
596 {
597         struct inode *inode = dentry->d_inode;
598         int error;
599
600         error = inode_change_ok(inode, attr);
601         if (error)
602                 return error;
603
604         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
605                 loff_t oldsize = inode->i_size;
606                 loff_t newsize = attr->ia_size;
607
608                 if (newsize != oldsize) {
609                         i_size_write(inode, newsize);
610                         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
611                 }
612                 if (newsize < oldsize) {
613                         loff_t holebegin = round_up(newsize, PAGE_SIZE);
614                         unmap_mapping_range(inode->i_mapping, holebegin, 0, 1);
615                         shmem_truncate_range(inode, newsize, (loff_t)-1);
616                         /* unmap again to remove racily COWed private pages */
617                         unmap_mapping_range(inode->i_mapping, holebegin, 0, 1);
618                 }
619         }
620
621         setattr_copy(inode, attr);
622 #ifdef CONFIG_TMPFS_POSIX_ACL
623         if (attr->ia_valid & ATTR_MODE)
624                 error = generic_acl_chmod(inode);
625 #endif
626         return error;
627 }
628
629 static void shmem_evict_inode(struct inode *inode)
630 {
631         struct shmem_inode_info *info = SHMEM_I(inode);
632
633         if (inode->i_mapping->a_ops == &shmem_aops) {
634                 shmem_unacct_size(info->flags, inode->i_size);
635                 inode->i_size = 0;
636                 shmem_truncate_range(inode, 0, (loff_t)-1);
637                 if (!list_empty(&info->swaplist)) {
638                         mutex_lock(&shmem_swaplist_mutex);
639                         list_del_init(&info->swaplist);
640                         mutex_unlock(&shmem_swaplist_mutex);
641                 }
642         } else
643                 kfree(info->symlink);
644
645         simple_xattrs_free(&info->xattrs);
646         WARN_ON(inode->i_blocks);
647         shmem_free_inode(inode->i_sb);
648         clear_inode(inode);
649 }
650
651 /*
652  * If swap found in inode, free it and move page from swapcache to filecache.
653  */
654 static int shmem_unuse_inode(struct shmem_inode_info *info,
655                              swp_entry_t swap, struct page **pagep)
656 {
657         struct address_space *mapping = info->vfs_inode.i_mapping;
658         void *radswap;
659         pgoff_t index;
660         gfp_t gfp;
661         int error = 0;
662
663         radswap = swp_to_radix_entry(swap);
664         index = radix_tree_locate_item(&mapping->page_tree, radswap);
665         if (index == -1)
666                 return 0;
667
668         /*
669          * Move _head_ to start search for next from here.
670          * But be careful: shmem_evict_inode checks list_empty without taking
671          * mutex, and there's an instant in list_move_tail when info->swaplist
672          * would appear empty, if it were the only one on shmem_swaplist.
673          */
674         if (shmem_swaplist.next != &info->swaplist)
675                 list_move_tail(&shmem_swaplist, &info->swaplist);
676
677         gfp = mapping_gfp_mask(mapping);
678         if (shmem_should_replace_page(*pagep, gfp)) {
679                 mutex_unlock(&shmem_swaplist_mutex);
680                 error = shmem_replace_page(pagep, gfp, info, index);
681                 mutex_lock(&shmem_swaplist_mutex);
682                 /*
683                  * We needed to drop mutex to make that restrictive page
684                  * allocation, but the inode might have been freed while we
685                  * dropped it: although a racing shmem_evict_inode() cannot
686                  * complete without emptying the radix_tree, our page lock
687                  * on this swapcache page is not enough to prevent that -
688                  * free_swap_and_cache() of our swap entry will only
689                  * trylock_page(), removing swap from radix_tree whatever.
690                  *
691                  * We must not proceed to shmem_add_to_page_cache() if the
692                  * inode has been freed, but of course we cannot rely on
693                  * inode or mapping or info to check that.  However, we can
694                  * safely check if our swap entry is still in use (and here
695                  * it can't have got reused for another page): if it's still
696                  * in use, then the inode cannot have been freed yet, and we
697                  * can safely proceed (if it's no longer in use, that tells
698                  * nothing about the inode, but we don't need to unuse swap).
699                  */
700                 if (!page_swapcount(*pagep))
701                         error = -ENOENT;
702         }
703
704         /*
705          * We rely on shmem_swaplist_mutex, not only to protect the swaplist,
706          * but also to hold up shmem_evict_inode(): so inode cannot be freed
707          * beneath us (pagelock doesn't help until the page is in pagecache).
708          */
709         if (!error)
710                 error = shmem_add_to_page_cache(*pagep, mapping, index,
711                                                 GFP_NOWAIT, radswap);
712         if (error != -ENOMEM) {
713                 /*
714                  * Truncation and eviction use free_swap_and_cache(), which
715                  * only does trylock page: if we raced, best clean up here.
716                  */
717                 delete_from_swap_cache(*pagep);
718                 set_page_dirty(*pagep);
719                 if (!error) {
720                         spin_lock(&info->lock);
721                         info->swapped--;
722                         spin_unlock(&info->lock);
723                         swap_free(swap);
724                 }
725                 error = 1;      /* not an error, but entry was found */
726         }
727         return error;
728 }
729
730 /*
731  * Search through swapped inodes to find and replace swap by page.
732  */
733 int shmem_unuse(swp_entry_t swap, struct page *page)
734 {
735         struct list_head *this, *next;
736         struct shmem_inode_info *info;
737         int found = 0;
738         int error = 0;
739
740         /*
741          * There's a faint possibility that swap page was replaced before
742          * caller locked it: caller will come back later with the right page.
743          */
744         if (unlikely(!PageSwapCache(page) || page_private(page) != swap.val))
745                 goto out;
746
747         /*
748          * Charge page using GFP_KERNEL while we can wait, before taking
749          * the shmem_swaplist_mutex which might hold up shmem_writepage().
750          * Charged back to the user (not to caller) when swap account is used.
751          */
752         error = mem_cgroup_cache_charge(page, current->mm, GFP_KERNEL);
753         if (error)
754                 goto out;
755         /* No radix_tree_preload: swap entry keeps a place for page in tree */
756
757         mutex_lock(&shmem_swaplist_mutex);
758         list_for_each_safe(this, next, &shmem_swaplist) {
759                 info = list_entry(this, struct shmem_inode_info, swaplist);
760                 if (info->swapped)
761                         found = shmem_unuse_inode(info, swap, &page);
762                 else
763                         list_del_init(&info->swaplist);
764                 cond_resched();
765                 if (found)
766                         break;
767         }
768         mutex_unlock(&shmem_swaplist_mutex);
769
770         if (found < 0)
771                 error = found;
772 out:
773         unlock_page(page);
774         page_cache_release(page);
775         return error;
776 }
777
778 /*
779  * Move the page from the page cache to the swap cache.
780  */
781 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
782 {
783         struct shmem_inode_info *info;
784         struct address_space *mapping;
785         struct inode *inode;
786         swp_entry_t swap;
787         pgoff_t index;
788
789         BUG_ON(!PageLocked(page));
790         mapping = page->mapping;
791         index = page->index;
792         inode = mapping->host;
793         info = SHMEM_I(inode);
794         if (info->flags & VM_LOCKED)
795                 goto redirty;
796         if (!total_swap_pages)
797                 goto redirty;
798
799         /*
800          * shmem_backing_dev_info's capabilities prevent regular writeback or
801          * sync from ever calling shmem_writepage; but a stacking filesystem
802          * might use ->writepage of its underlying filesystem, in which case
803          * tmpfs should write out to swap only in response to memory pressure,
804          * and not for the writeback threads or sync.
805          */
806         if (!wbc->for_reclaim) {
807                 WARN_ON_ONCE(1);        /* Still happens? Tell us about it! */
808                 goto redirty;
809         }
810
811         /*
812          * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
813          * value into swapfile.c, the only way we can correctly account for a
814          * fallocated page arriving here is now to initialize it and write it.
815          *
816          * That's okay for a page already fallocated earlier, but if we have
817          * not yet completed the fallocation, then (a) we want to keep track
818          * of this page in case we have to undo it, and (b) it may not be a
819          * good idea to continue anyway, once we're pushing into swap.  So
820          * reactivate the page, and let shmem_fallocate() quit when too many.
821          */
822         if (!PageUptodate(page)) {
823                 if (inode->i_private) {
824                         struct shmem_falloc *shmem_falloc;
825                         spin_lock(&inode->i_lock);
826                         shmem_falloc = inode->i_private;
827                         if (shmem_falloc &&
828                             index >= shmem_falloc->start &&
829                             index < shmem_falloc->next)
830                                 shmem_falloc->nr_unswapped++;
831                         else
832                                 shmem_falloc = NULL;
833                         spin_unlock(&inode->i_lock);
834                         if (shmem_falloc)
835                                 goto redirty;
836                 }
837                 clear_highpage(page);
838                 flush_dcache_page(page);
839                 SetPageUptodate(page);
840         }
841
842         swap = get_swap_page();
843         if (!swap.val)
844                 goto redirty;
845
846         /*
847          * Add inode to shmem_unuse()'s list of swapped-out inodes,
848          * if it's not already there.  Do it now before the page is
849          * moved to swap cache, when its pagelock no longer protects
850          * the inode from eviction.  But don't unlock the mutex until
851          * we've incremented swapped, because shmem_unuse_inode() will
852          * prune a !swapped inode from the swaplist under this mutex.
853          */
854         mutex_lock(&shmem_swaplist_mutex);
855         if (list_empty(&info->swaplist))
856                 list_add_tail(&info->swaplist, &shmem_swaplist);
857
858         if (add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
859                 swap_shmem_alloc(swap);
860                 shmem_delete_from_page_cache(page, swp_to_radix_entry(swap));
861
862                 spin_lock(&info->lock);
863                 info->swapped++;
864                 shmem_recalc_inode(inode);
865                 spin_unlock(&info->lock);
866
867                 mutex_unlock(&shmem_swaplist_mutex);
868                 BUG_ON(page_mapped(page));
869                 swap_writepage(page, wbc);
870                 return 0;
871         }
872
873         mutex_unlock(&shmem_swaplist_mutex);
874         swapcache_free(swap, NULL);
875 redirty:
876         set_page_dirty(page);
877         if (wbc->for_reclaim)
878                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
879         unlock_page(page);
880         return 0;
881 }
882
883 #ifdef CONFIG_NUMA
884 #ifdef CONFIG_TMPFS
885 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
886 {
887         char buffer[64];
888
889         if (!mpol || mpol->mode == MPOL_DEFAULT)
890                 return;         /* show nothing */
891
892         mpol_to_str(buffer, sizeof(buffer), mpol, 1);
893
894         seq_printf(seq, ",mpol=%s", buffer);
895 }
896
897 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
898 {
899         struct mempolicy *mpol = NULL;
900         if (sbinfo->mpol) {
901                 spin_lock(&sbinfo->stat_lock);  /* prevent replace/use races */
902                 mpol = sbinfo->mpol;
903                 mpol_get(mpol);
904                 spin_unlock(&sbinfo->stat_lock);
905         }
906         return mpol;
907 }
908 #endif /* CONFIG_TMPFS */
909
910 static struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
911                         struct shmem_inode_info *info, pgoff_t index)
912 {
913         struct mempolicy mpol, *spol;
914         struct vm_area_struct pvma;
915
916         spol = mpol_cond_copy(&mpol,
917                         mpol_shared_policy_lookup(&info->policy, index));
918
919         /* Create a pseudo vma that just contains the policy */
920         pvma.vm_start = 0;
921         /* Bias interleave by inode number to distribute better across nodes */
922         pvma.vm_pgoff = index + info->vfs_inode.i_ino;
923         pvma.vm_ops = NULL;
924         pvma.vm_policy = spol;
925         return swapin_readahead(swap, gfp, &pvma, 0);
926 }
927
928 static struct page *shmem_alloc_page(gfp_t gfp,
929                         struct shmem_inode_info *info, pgoff_t index)
930 {
931         struct vm_area_struct pvma;
932
933         /* Create a pseudo vma that just contains the policy */
934         pvma.vm_start = 0;
935         /* Bias interleave by inode number to distribute better across nodes */
936         pvma.vm_pgoff = index + info->vfs_inode.i_ino;
937         pvma.vm_ops = NULL;
938         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, index);
939
940         /*
941          * alloc_page_vma() will drop the shared policy reference
942          */
943         return alloc_page_vma(gfp, &pvma, 0);
944 }
945 #else /* !CONFIG_NUMA */
946 #ifdef CONFIG_TMPFS
947 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
948 {
949 }
950 #endif /* CONFIG_TMPFS */
951
952 static inline struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
953                         struct shmem_inode_info *info, pgoff_t index)
954 {
955         return swapin_readahead(swap, gfp, NULL, 0);
956 }
957
958 static inline struct page *shmem_alloc_page(gfp_t gfp,
959                         struct shmem_inode_info *info, pgoff_t index)
960 {
961         return alloc_page(gfp);
962 }
963 #endif /* CONFIG_NUMA */
964
965 #if !defined(CONFIG_NUMA) || !defined(CONFIG_TMPFS)
966 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
967 {
968         return NULL;
969 }
970 #endif
971
972 /*
973  * When a page is moved from swapcache to shmem filecache (either by the
974  * usual swapin of shmem_getpage_gfp(), or by the less common swapoff of
975  * shmem_unuse_inode()), it may have been read in earlier from swap, in
976  * ignorance of the mapping it belongs to.  If that mapping has special
977  * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
978  * we may need to copy to a suitable page before moving to filecache.
979  *
980  * In a future release, this may well be extended to respect cpuset and
981  * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
982  * but for now it is a simple matter of zone.
983  */
984 static bool shmem_should_replace_page(struct page *page, gfp_t gfp)
985 {
986         return page_zonenum(page) > gfp_zone(gfp);
987 }
988
989 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
990                                 struct shmem_inode_info *info, pgoff_t index)
991 {
992         struct page *oldpage, *newpage;
993         struct address_space *swap_mapping;
994         pgoff_t swap_index;
995         int error;
996
997         oldpage = *pagep;
998         swap_index = page_private(oldpage);
999         swap_mapping = page_mapping(oldpage);
1000
1001         /*
1002          * We have arrived here because our zones are constrained, so don't
1003          * limit chance of success by further cpuset and node constraints.
1004          */
1005         gfp &= ~GFP_CONSTRAINT_MASK;
1006         newpage = shmem_alloc_page(gfp, info, index);
1007         if (!newpage)
1008                 return -ENOMEM;
1009
1010         page_cache_get(newpage);
1011         copy_highpage(newpage, oldpage);
1012         flush_dcache_page(newpage);
1013
1014         __set_page_locked(newpage);
1015         SetPageUptodate(newpage);
1016         SetPageSwapBacked(newpage);
1017         set_page_private(newpage, swap_index);
1018         SetPageSwapCache(newpage);
1019
1020         /*
1021          * Our caller will very soon move newpage out of swapcache, but it's
1022          * a nice clean interface for us to replace oldpage by newpage there.
1023          */
1024         spin_lock_irq(&swap_mapping->tree_lock);
1025         error = shmem_radix_tree_replace(swap_mapping, swap_index, oldpage,
1026                                                                    newpage);
1027         if (!error) {
1028                 __inc_zone_page_state(newpage, NR_FILE_PAGES);
1029                 __dec_zone_page_state(oldpage, NR_FILE_PAGES);
1030         }
1031         spin_unlock_irq(&swap_mapping->tree_lock);
1032
1033         if (unlikely(error)) {
1034                 /*
1035                  * Is this possible?  I think not, now that our callers check
1036                  * both PageSwapCache and page_private after getting page lock;
1037                  * but be defensive.  Reverse old to newpage for clear and free.
1038                  */
1039                 oldpage = newpage;
1040         } else {
1041                 mem_cgroup_replace_page_cache(oldpage, newpage);
1042                 lru_cache_add_anon(newpage);
1043                 *pagep = newpage;
1044         }
1045
1046         ClearPageSwapCache(oldpage);
1047         set_page_private(oldpage, 0);
1048
1049         unlock_page(oldpage);
1050         page_cache_release(oldpage);
1051         page_cache_release(oldpage);
1052         return error;
1053 }
1054
1055 /*
1056  * shmem_getpage_gfp - find page in cache, or get from swap, or allocate
1057  *
1058  * If we allocate a new one we do not mark it dirty. That's up to the
1059  * vm. If we swap it in we mark it dirty since we also free the swap
1060  * entry since a page cannot live in both the swap and page cache
1061  */
1062 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
1063         struct page **pagep, enum sgp_type sgp, gfp_t gfp, int *fault_type)
1064 {
1065         struct address_space *mapping = inode->i_mapping;
1066         struct shmem_inode_info *info;
1067         struct shmem_sb_info *sbinfo;
1068         struct page *page;
1069         swp_entry_t swap;
1070         int error;
1071         int once = 0;
1072         int alloced = 0;
1073
1074         if (index > (MAX_LFS_FILESIZE >> PAGE_CACHE_SHIFT))
1075                 return -EFBIG;
1076 repeat:
1077         swap.val = 0;
1078         page = find_lock_page(mapping, index);
1079         if (radix_tree_exceptional_entry(page)) {
1080                 swap = radix_to_swp_entry(page);
1081                 page = NULL;
1082         }
1083
1084         if (sgp != SGP_WRITE && sgp != SGP_FALLOC &&
1085             ((loff_t)index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
1086                 error = -EINVAL;
1087                 goto failed;
1088         }
1089
1090         /* fallocated page? */
1091         if (page && !PageUptodate(page)) {
1092                 if (sgp != SGP_READ)
1093                         goto clear;
1094                 unlock_page(page);
1095                 page_cache_release(page);
1096                 page = NULL;
1097         }
1098         if (page || (sgp == SGP_READ && !swap.val)) {
1099                 *pagep = page;
1100                 return 0;
1101         }
1102
1103         /*
1104          * Fast cache lookup did not find it:
1105          * bring it back from swap or allocate.
1106          */
1107         info = SHMEM_I(inode);
1108         sbinfo = SHMEM_SB(inode->i_sb);
1109
1110         if (swap.val) {
1111                 /* Look it up and read it in.. */
1112                 page = lookup_swap_cache(swap);
1113                 if (!page) {
1114                         /* here we actually do the io */
1115                         if (fault_type)
1116                                 *fault_type |= VM_FAULT_MAJOR;
1117                         page = shmem_swapin(swap, gfp, info, index);
1118                         if (!page) {
1119                                 error = -ENOMEM;
1120                                 goto failed;
1121                         }
1122                 }
1123
1124                 /* We have to do this with page locked to prevent races */
1125                 lock_page(page);
1126                 if (!PageSwapCache(page) || page_private(page) != swap.val ||
1127                     !shmem_confirm_swap(mapping, index, swap)) {
1128                         error = -EEXIST;        /* try again */
1129                         goto unlock;
1130                 }
1131                 if (!PageUptodate(page)) {
1132                         error = -EIO;
1133                         goto failed;
1134                 }
1135                 wait_on_page_writeback(page);
1136
1137                 if (shmem_should_replace_page(page, gfp)) {
1138                         error = shmem_replace_page(&page, gfp, info, index);
1139                         if (error)
1140                                 goto failed;
1141                 }
1142
1143                 error = mem_cgroup_cache_charge(page, current->mm,
1144                                                 gfp & GFP_RECLAIM_MASK);
1145                 if (!error) {
1146                         error = shmem_add_to_page_cache(page, mapping, index,
1147                                                 gfp, swp_to_radix_entry(swap));
1148                         /*
1149                          * We already confirmed swap under page lock, and make
1150                          * no memory allocation here, so usually no possibility
1151                          * of error; but free_swap_and_cache() only trylocks a
1152                          * page, so it is just possible that the entry has been
1153                          * truncated or holepunched since swap was confirmed.
1154                          * shmem_undo_range() will have done some of the
1155                          * unaccounting, now delete_from_swap_cache() will do
1156                          * the rest (including mem_cgroup_uncharge_swapcache).
1157                          * Reset swap.val? No, leave it so "failed" goes back to
1158                          * "repeat": reading a hole and writing should succeed.
1159                          */
1160                         if (error)
1161                                 delete_from_swap_cache(page);
1162                 }
1163                 if (error)
1164                         goto failed;
1165
1166                 spin_lock(&info->lock);
1167                 info->swapped--;
1168                 shmem_recalc_inode(inode);
1169                 spin_unlock(&info->lock);
1170
1171                 delete_from_swap_cache(page);
1172                 set_page_dirty(page);
1173                 swap_free(swap);
1174
1175         } else {
1176                 if (shmem_acct_block(info->flags)) {
1177                         error = -ENOSPC;
1178                         goto failed;
1179                 }
1180                 if (sbinfo->max_blocks) {
1181                         if (percpu_counter_compare(&sbinfo->used_blocks,
1182                                                 sbinfo->max_blocks) >= 0) {
1183                                 error = -ENOSPC;
1184                                 goto unacct;
1185                         }
1186                         percpu_counter_inc(&sbinfo->used_blocks);
1187                 }
1188
1189                 page = shmem_alloc_page(gfp, info, index);
1190                 if (!page) {
1191                         error = -ENOMEM;
1192                         goto decused;
1193                 }
1194
1195                 SetPageSwapBacked(page);
1196                 __set_page_locked(page);
1197                 error = mem_cgroup_cache_charge(page, current->mm,
1198                                                 gfp & GFP_RECLAIM_MASK);
1199                 if (error)
1200                         goto decused;
1201                 error = radix_tree_preload(gfp & GFP_RECLAIM_MASK);
1202                 if (!error) {
1203                         error = shmem_add_to_page_cache(page, mapping, index,
1204                                                         gfp, NULL);
1205                         radix_tree_preload_end();
1206                 }
1207                 if (error) {
1208                         mem_cgroup_uncharge_cache_page(page);
1209                         goto decused;
1210                 }
1211                 lru_cache_add_anon(page);
1212
1213                 spin_lock(&info->lock);
1214                 info->alloced++;
1215                 inode->i_blocks += BLOCKS_PER_PAGE;
1216                 shmem_recalc_inode(inode);
1217                 spin_unlock(&info->lock);
1218                 alloced = true;
1219
1220                 /*
1221                  * Let SGP_FALLOC use the SGP_WRITE optimization on a new page.
1222                  */
1223                 if (sgp == SGP_FALLOC)
1224                         sgp = SGP_WRITE;
1225 clear:
1226                 /*
1227                  * Let SGP_WRITE caller clear ends if write does not fill page;
1228                  * but SGP_FALLOC on a page fallocated earlier must initialize
1229                  * it now, lest undo on failure cancel our earlier guarantee.
1230                  */
1231                 if (sgp != SGP_WRITE) {
1232                         clear_highpage(page);
1233                         flush_dcache_page(page);
1234                         SetPageUptodate(page);
1235                 }
1236                 if (sgp == SGP_DIRTY)
1237                         set_page_dirty(page);
1238         }
1239
1240         /* Perhaps the file has been truncated since we checked */
1241         if (sgp != SGP_WRITE && sgp != SGP_FALLOC &&
1242             ((loff_t)index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
1243                 error = -EINVAL;
1244                 if (alloced)
1245                         goto trunc;
1246                 else
1247                         goto failed;
1248         }
1249         *pagep = page;
1250         return 0;
1251
1252         /*
1253          * Error recovery.
1254          */
1255 trunc:
1256         info = SHMEM_I(inode);
1257         ClearPageDirty(page);
1258         delete_from_page_cache(page);
1259         spin_lock(&info->lock);
1260         info->alloced--;
1261         inode->i_blocks -= BLOCKS_PER_PAGE;
1262         spin_unlock(&info->lock);
1263 decused:
1264         sbinfo = SHMEM_SB(inode->i_sb);
1265         if (sbinfo->max_blocks)
1266                 percpu_counter_add(&sbinfo->used_blocks, -1);
1267 unacct:
1268         shmem_unacct_blocks(info->flags, 1);
1269 failed:
1270         if (swap.val && error != -EINVAL &&
1271             !shmem_confirm_swap(mapping, index, swap))
1272                 error = -EEXIST;
1273 unlock:
1274         if (page) {
1275                 unlock_page(page);
1276                 page_cache_release(page);
1277         }
1278         if (error == -ENOSPC && !once++) {
1279                 info = SHMEM_I(inode);
1280                 spin_lock(&info->lock);
1281                 shmem_recalc_inode(inode);
1282                 spin_unlock(&info->lock);
1283                 goto repeat;
1284         }
1285         if (error == -EEXIST)   /* from above or from radix_tree_insert */
1286                 goto repeat;
1287         return error;
1288 }
1289
1290 static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1291 {
1292         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1293         int error;
1294         int ret = VM_FAULT_LOCKED;
1295
1296         error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_CACHE, &ret);
1297         if (error)
1298                 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1299
1300         if (ret & VM_FAULT_MAJOR) {
1301                 count_vm_event(PGMAJFAULT);
1302                 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1303         }
1304         return ret;
1305 }
1306
1307 #ifdef CONFIG_NUMA
1308 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
1309 {
1310         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1311         return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
1312 }
1313
1314 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
1315                                           unsigned long addr)
1316 {
1317         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1318         pgoff_t index;
1319
1320         index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1321         return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
1322 }
1323 #endif
1324
1325 int shmem_lock(struct file *file, int lock, struct user_struct *user)
1326 {
1327         struct inode *inode = file->f_path.dentry->d_inode;
1328         struct shmem_inode_info *info = SHMEM_I(inode);
1329         int retval = -ENOMEM;
1330
1331         spin_lock(&info->lock);
1332         if (lock && !(info->flags & VM_LOCKED)) {
1333                 if (!user_shm_lock(inode->i_size, user))
1334                         goto out_nomem;
1335                 info->flags |= VM_LOCKED;
1336                 mapping_set_unevictable(file->f_mapping);
1337         }
1338         if (!lock && (info->flags & VM_LOCKED) && user) {
1339                 user_shm_unlock(inode->i_size, user);
1340                 info->flags &= ~VM_LOCKED;
1341                 mapping_clear_unevictable(file->f_mapping);
1342         }
1343         retval = 0;
1344
1345 out_nomem:
1346         spin_unlock(&info->lock);
1347         return retval;
1348 }
1349
1350 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1351 {
1352         file_accessed(file);
1353         vma->vm_ops = &shmem_vm_ops;
1354         return 0;
1355 }
1356
1357 static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
1358                                      umode_t mode, dev_t dev, unsigned long flags)
1359 {
1360         struct inode *inode;
1361         struct shmem_inode_info *info;
1362         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1363
1364         if (shmem_reserve_inode(sb))
1365                 return NULL;
1366
1367         inode = new_inode(sb);
1368         if (inode) {
1369                 inode->i_ino = get_next_ino();
1370                 inode_init_owner(inode, dir, mode);
1371                 inode->i_blocks = 0;
1372                 inode->i_mapping->backing_dev_info = &shmem_backing_dev_info;
1373                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1374                 inode->i_generation = get_seconds();
1375                 info = SHMEM_I(inode);
1376                 memset(info, 0, (char *)inode - (char *)info);
1377                 spin_lock_init(&info->lock);
1378                 info->flags = flags & VM_NORESERVE;
1379                 INIT_LIST_HEAD(&info->swaplist);
1380                 simple_xattrs_init(&info->xattrs);
1381                 cache_no_acl(inode);
1382
1383                 switch (mode & S_IFMT) {
1384                 default:
1385                         inode->i_op = &shmem_special_inode_operations;
1386                         init_special_inode(inode, mode, dev);
1387                         break;
1388                 case S_IFREG:
1389                         inode->i_mapping->a_ops = &shmem_aops;
1390                         inode->i_op = &shmem_inode_operations;
1391                         inode->i_fop = &shmem_file_operations;
1392                         mpol_shared_policy_init(&info->policy,
1393                                                  shmem_get_sbmpol(sbinfo));
1394                         break;
1395                 case S_IFDIR:
1396                         inc_nlink(inode);
1397                         /* Some things misbehave if size == 0 on a directory */
1398                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
1399                         inode->i_op = &shmem_dir_inode_operations;
1400                         inode->i_fop = &simple_dir_operations;
1401                         break;
1402                 case S_IFLNK:
1403                         /*
1404                          * Must not load anything in the rbtree,
1405                          * mpol_free_shared_policy will not be called.
1406                          */
1407                         mpol_shared_policy_init(&info->policy, NULL);
1408                         break;
1409                 }
1410         } else
1411                 shmem_free_inode(sb);
1412         return inode;
1413 }
1414
1415 #ifdef CONFIG_TMPFS
1416 static const struct inode_operations shmem_symlink_inode_operations;
1417 static const struct inode_operations shmem_short_symlink_operations;
1418
1419 #ifdef CONFIG_TMPFS_XATTR
1420 static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
1421 #else
1422 #define shmem_initxattrs NULL
1423 #endif
1424
1425 static int
1426 shmem_write_begin(struct file *file, struct address_space *mapping,
1427                         loff_t pos, unsigned len, unsigned flags,
1428                         struct page **pagep, void **fsdata)
1429 {
1430         struct inode *inode = mapping->host;
1431         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1432         return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1433 }
1434
1435 static int
1436 shmem_write_end(struct file *file, struct address_space *mapping,
1437                         loff_t pos, unsigned len, unsigned copied,
1438                         struct page *page, void *fsdata)
1439 {
1440         struct inode *inode = mapping->host;
1441
1442         if (pos + copied > inode->i_size)
1443                 i_size_write(inode, pos + copied);
1444
1445         if (!PageUptodate(page)) {
1446                 if (copied < PAGE_CACHE_SIZE) {
1447                         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1448                         zero_user_segments(page, 0, from,
1449                                         from + copied, PAGE_CACHE_SIZE);
1450                 }
1451                 SetPageUptodate(page);
1452         }
1453         set_page_dirty(page);
1454         unlock_page(page);
1455         page_cache_release(page);
1456
1457         return copied;
1458 }
1459
1460 static void do_shmem_file_read(struct file *filp, loff_t *ppos, read_descriptor_t *desc, read_actor_t actor)
1461 {
1462         struct inode *inode = filp->f_path.dentry->d_inode;
1463         struct address_space *mapping = inode->i_mapping;
1464         pgoff_t index;
1465         unsigned long offset;
1466         enum sgp_type sgp = SGP_READ;
1467
1468         /*
1469          * Might this read be for a stacking filesystem?  Then when reading
1470          * holes of a sparse file, we actually need to allocate those pages,
1471          * and even mark them dirty, so it cannot exceed the max_blocks limit.
1472          */
1473         if (segment_eq(get_fs(), KERNEL_DS))
1474                 sgp = SGP_DIRTY;
1475
1476         index = *ppos >> PAGE_CACHE_SHIFT;
1477         offset = *ppos & ~PAGE_CACHE_MASK;
1478
1479         for (;;) {
1480                 struct page *page = NULL;
1481                 pgoff_t end_index;
1482                 unsigned long nr, ret;
1483                 loff_t i_size = i_size_read(inode);
1484
1485                 end_index = i_size >> PAGE_CACHE_SHIFT;
1486                 if (index > end_index)
1487                         break;
1488                 if (index == end_index) {
1489                         nr = i_size & ~PAGE_CACHE_MASK;
1490                         if (nr <= offset)
1491                                 break;
1492                 }
1493
1494                 desc->error = shmem_getpage(inode, index, &page, sgp, NULL);
1495                 if (desc->error) {
1496                         if (desc->error == -EINVAL)
1497                                 desc->error = 0;
1498                         break;
1499                 }
1500                 if (page)
1501                         unlock_page(page);
1502
1503                 /*
1504                  * We must evaluate after, since reads (unlike writes)
1505                  * are called without i_mutex protection against truncate
1506                  */
1507                 nr = PAGE_CACHE_SIZE;
1508                 i_size = i_size_read(inode);
1509                 end_index = i_size >> PAGE_CACHE_SHIFT;
1510                 if (index == end_index) {
1511                         nr = i_size & ~PAGE_CACHE_MASK;
1512                         if (nr <= offset) {
1513                                 if (page)
1514                                         page_cache_release(page);
1515                                 break;
1516                         }
1517                 }
1518                 nr -= offset;
1519
1520                 if (page) {
1521                         /*
1522                          * If users can be writing to this page using arbitrary
1523                          * virtual addresses, take care about potential aliasing
1524                          * before reading the page on the kernel side.
1525                          */
1526                         if (mapping_writably_mapped(mapping))
1527                                 flush_dcache_page(page);
1528                         /*
1529                          * Mark the page accessed if we read the beginning.
1530                          */
1531                         if (!offset)
1532                                 mark_page_accessed(page);
1533                 } else {
1534                         page = ZERO_PAGE(0);
1535                         page_cache_get(page);
1536                 }
1537
1538                 /*
1539                  * Ok, we have the page, and it's up-to-date, so
1540                  * now we can copy it to user space...
1541                  *
1542                  * The actor routine returns how many bytes were actually used..
1543                  * NOTE! This may not be the same as how much of a user buffer
1544                  * we filled up (we may be padding etc), so we can only update
1545                  * "pos" here (the actor routine has to update the user buffer
1546                  * pointers and the remaining count).
1547                  */
1548                 ret = actor(desc, page, offset, nr);
1549                 offset += ret;
1550                 index += offset >> PAGE_CACHE_SHIFT;
1551                 offset &= ~PAGE_CACHE_MASK;
1552
1553                 page_cache_release(page);
1554                 if (ret != nr || !desc->count)
1555                         break;
1556
1557                 cond_resched();
1558         }
1559
1560         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1561         file_accessed(filp);
1562 }
1563
1564 static ssize_t shmem_file_aio_read(struct kiocb *iocb,
1565                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
1566 {
1567         struct file *filp = iocb->ki_filp;
1568         ssize_t retval;
1569         unsigned long seg;
1570         size_t count;
1571         loff_t *ppos = &iocb->ki_pos;
1572
1573         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1574         if (retval)
1575                 return retval;
1576
1577         for (seg = 0; seg < nr_segs; seg++) {
1578                 read_descriptor_t desc;
1579
1580                 desc.written = 0;
1581                 desc.arg.buf = iov[seg].iov_base;
1582                 desc.count = iov[seg].iov_len;
1583                 if (desc.count == 0)
1584                         continue;
1585                 desc.error = 0;
1586                 do_shmem_file_read(filp, ppos, &desc, file_read_actor);
1587                 retval += desc.written;
1588                 if (desc.error) {
1589                         retval = retval ?: desc.error;
1590                         break;
1591                 }
1592                 if (desc.count > 0)
1593                         break;
1594         }
1595         return retval;
1596 }
1597
1598 static ssize_t shmem_file_splice_read(struct file *in, loff_t *ppos,
1599                                 struct pipe_inode_info *pipe, size_t len,
1600                                 unsigned int flags)
1601 {
1602         struct address_space *mapping = in->f_mapping;
1603         struct inode *inode = mapping->host;
1604         unsigned int loff, nr_pages, req_pages;
1605         struct page *pages[PIPE_DEF_BUFFERS];
1606         struct partial_page partial[PIPE_DEF_BUFFERS];
1607         struct page *page;
1608         pgoff_t index, end_index;
1609         loff_t isize, left;
1610         int error, page_nr;
1611         struct splice_pipe_desc spd = {
1612                 .pages = pages,
1613                 .partial = partial,
1614                 .nr_pages_max = PIPE_DEF_BUFFERS,
1615                 .flags = flags,
1616                 .ops = &page_cache_pipe_buf_ops,
1617                 .spd_release = spd_release_page,
1618         };
1619
1620         isize = i_size_read(inode);
1621         if (unlikely(*ppos >= isize))
1622                 return 0;
1623
1624         left = isize - *ppos;
1625         if (unlikely(left < len))
1626                 len = left;
1627
1628         if (splice_grow_spd(pipe, &spd))
1629                 return -ENOMEM;
1630
1631         index = *ppos >> PAGE_CACHE_SHIFT;
1632         loff = *ppos & ~PAGE_CACHE_MASK;
1633         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1634         nr_pages = min(req_pages, pipe->buffers);
1635
1636         spd.nr_pages = find_get_pages_contig(mapping, index,
1637                                                 nr_pages, spd.pages);
1638         index += spd.nr_pages;
1639         error = 0;
1640
1641         while (spd.nr_pages < nr_pages) {
1642                 error = shmem_getpage(inode, index, &page, SGP_CACHE, NULL);
1643                 if (error)
1644                         break;
1645                 unlock_page(page);
1646                 spd.pages[spd.nr_pages++] = page;
1647                 index++;
1648         }
1649
1650         index = *ppos >> PAGE_CACHE_SHIFT;
1651         nr_pages = spd.nr_pages;
1652         spd.nr_pages = 0;
1653
1654         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
1655                 unsigned int this_len;
1656
1657                 if (!len)
1658                         break;
1659
1660                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
1661                 page = spd.pages[page_nr];
1662
1663                 if (!PageUptodate(page) || page->mapping != mapping) {
1664                         error = shmem_getpage(inode, index, &page,
1665                                                         SGP_CACHE, NULL);
1666                         if (error)
1667                                 break;
1668                         unlock_page(page);
1669                         page_cache_release(spd.pages[page_nr]);
1670                         spd.pages[page_nr] = page;
1671                 }
1672
1673                 isize = i_size_read(inode);
1674                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1675                 if (unlikely(!isize || index > end_index))
1676                         break;
1677
1678                 if (end_index == index) {
1679                         unsigned int plen;
1680
1681                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1682                         if (plen <= loff)
1683                                 break;
1684
1685                         this_len = min(this_len, plen - loff);
1686                         len = this_len;
1687                 }
1688
1689                 spd.partial[page_nr].offset = loff;
1690                 spd.partial[page_nr].len = this_len;
1691                 len -= this_len;
1692                 loff = 0;
1693                 spd.nr_pages++;
1694                 index++;
1695         }
1696
1697         while (page_nr < nr_pages)
1698                 page_cache_release(spd.pages[page_nr++]);
1699
1700         if (spd.nr_pages)
1701                 error = splice_to_pipe(pipe, &spd);
1702
1703         splice_shrink_spd(&spd);
1704
1705         if (error > 0) {
1706                 *ppos += error;
1707                 file_accessed(in);
1708         }
1709         return error;
1710 }
1711
1712 static long shmem_fallocate(struct file *file, int mode, loff_t offset,
1713                                                          loff_t len)
1714 {
1715         struct inode *inode = file->f_path.dentry->d_inode;
1716         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1717         struct shmem_falloc shmem_falloc;
1718         pgoff_t start, index, end;
1719         int error;
1720
1721         mutex_lock(&inode->i_mutex);
1722
1723         if (mode & FALLOC_FL_PUNCH_HOLE) {
1724                 struct address_space *mapping = file->f_mapping;
1725                 loff_t unmap_start = round_up(offset, PAGE_SIZE);
1726                 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
1727
1728                 if ((u64)unmap_end > (u64)unmap_start)
1729                         unmap_mapping_range(mapping, unmap_start,
1730                                             1 + unmap_end - unmap_start, 0);
1731                 shmem_truncate_range(inode, offset, offset + len - 1);
1732                 /* No need to unmap again: hole-punching leaves COWed pages */
1733                 error = 0;
1734                 goto out;
1735         }
1736
1737         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
1738         error = inode_newsize_ok(inode, offset + len);
1739         if (error)
1740                 goto out;
1741
1742         start = offset >> PAGE_CACHE_SHIFT;
1743         end = (offset + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1744         /* Try to avoid a swapstorm if len is impossible to satisfy */
1745         if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
1746                 error = -ENOSPC;
1747                 goto out;
1748         }
1749
1750         shmem_falloc.start = start;
1751         shmem_falloc.next  = start;
1752         shmem_falloc.nr_falloced = 0;
1753         shmem_falloc.nr_unswapped = 0;
1754         spin_lock(&inode->i_lock);
1755         inode->i_private = &shmem_falloc;
1756         spin_unlock(&inode->i_lock);
1757
1758         for (index = start; index < end; index++) {
1759                 struct page *page;
1760
1761                 /*
1762                  * Good, the fallocate(2) manpage permits EINTR: we may have
1763                  * been interrupted because we are using up too much memory.
1764                  */
1765                 if (signal_pending(current))
1766                         error = -EINTR;
1767                 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
1768                         error = -ENOMEM;
1769                 else
1770                         error = shmem_getpage(inode, index, &page, SGP_FALLOC,
1771                                                                         NULL);
1772                 if (error) {
1773                         /* Remove the !PageUptodate pages we added */
1774                         shmem_undo_range(inode,
1775                                 (loff_t)start << PAGE_CACHE_SHIFT,
1776                                 (loff_t)index << PAGE_CACHE_SHIFT, true);
1777                         goto undone;
1778                 }
1779
1780                 /*
1781                  * Inform shmem_writepage() how far we have reached.
1782                  * No need for lock or barrier: we have the page lock.
1783                  */
1784                 shmem_falloc.next++;
1785                 if (!PageUptodate(page))
1786                         shmem_falloc.nr_falloced++;
1787
1788                 /*
1789                  * If !PageUptodate, leave it that way so that freeable pages
1790                  * can be recognized if we need to rollback on error later.
1791                  * But set_page_dirty so that memory pressure will swap rather
1792                  * than free the pages we are allocating (and SGP_CACHE pages
1793                  * might still be clean: we now need to mark those dirty too).
1794                  */
1795                 set_page_dirty(page);
1796                 unlock_page(page);
1797                 page_cache_release(page);
1798                 cond_resched();
1799         }
1800
1801         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
1802                 i_size_write(inode, offset + len);
1803         inode->i_ctime = CURRENT_TIME;
1804 undone:
1805         spin_lock(&inode->i_lock);
1806         inode->i_private = NULL;
1807         spin_unlock(&inode->i_lock);
1808 out:
1809         mutex_unlock(&inode->i_mutex);
1810         return error;
1811 }
1812
1813 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
1814 {
1815         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
1816
1817         buf->f_type = TMPFS_MAGIC;
1818         buf->f_bsize = PAGE_CACHE_SIZE;
1819         buf->f_namelen = NAME_MAX;
1820         if (sbinfo->max_blocks) {
1821                 buf->f_blocks = sbinfo->max_blocks;
1822                 buf->f_bavail =
1823                 buf->f_bfree  = sbinfo->max_blocks -
1824                                 percpu_counter_sum(&sbinfo->used_blocks);
1825         }
1826         if (sbinfo->max_inodes) {
1827                 buf->f_files = sbinfo->max_inodes;
1828                 buf->f_ffree = sbinfo->free_inodes;
1829         }
1830         /* else leave those fields 0 like simple_statfs */
1831         return 0;
1832 }
1833
1834 /*
1835  * File creation. Allocate an inode, and we're done..
1836  */
1837 static int
1838 shmem_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
1839 {
1840         struct inode *inode;
1841         int error = -ENOSPC;
1842
1843         inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
1844         if (inode) {
1845                 error = security_inode_init_security(inode, dir,
1846                                                      &dentry->d_name,
1847                                                      shmem_initxattrs, NULL);
1848                 if (error) {
1849                         if (error != -EOPNOTSUPP) {
1850                                 iput(inode);
1851                                 return error;
1852                         }
1853                 }
1854 #ifdef CONFIG_TMPFS_POSIX_ACL
1855                 error = generic_acl_init(inode, dir);
1856                 if (error) {
1857                         iput(inode);
1858                         return error;
1859                 }
1860 #else
1861                 error = 0;
1862 #endif
1863                 dir->i_size += BOGO_DIRENT_SIZE;
1864                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1865                 d_instantiate(dentry, inode);
1866                 dget(dentry); /* Extra count - pin the dentry in core */
1867         }
1868         return error;
1869 }
1870
1871 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
1872 {
1873         int error;
1874
1875         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
1876                 return error;
1877         inc_nlink(dir);
1878         return 0;
1879 }
1880
1881 static int shmem_create(struct inode *dir, struct dentry *dentry, umode_t mode,
1882                 bool excl)
1883 {
1884         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
1885 }
1886
1887 /*
1888  * Link a file..
1889  */
1890 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
1891 {
1892         struct inode *inode = old_dentry->d_inode;
1893         int ret;
1894
1895         /*
1896          * No ordinary (disk based) filesystem counts links as inodes;
1897          * but each new link needs a new dentry, pinning lowmem, and
1898          * tmpfs dentries cannot be pruned until they are unlinked.
1899          */
1900         ret = shmem_reserve_inode(inode->i_sb);
1901         if (ret)
1902                 goto out;
1903
1904         dir->i_size += BOGO_DIRENT_SIZE;
1905         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1906         inc_nlink(inode);
1907         ihold(inode);   /* New dentry reference */
1908         dget(dentry);           /* Extra pinning count for the created dentry */
1909         d_instantiate(dentry, inode);
1910 out:
1911         return ret;
1912 }
1913
1914 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
1915 {
1916         struct inode *inode = dentry->d_inode;
1917
1918         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
1919                 shmem_free_inode(inode->i_sb);
1920
1921         dir->i_size -= BOGO_DIRENT_SIZE;
1922         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1923         drop_nlink(inode);
1924         dput(dentry);   /* Undo the count from "create" - this does all the work */
1925         return 0;
1926 }
1927
1928 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
1929 {
1930         if (!simple_empty(dentry))
1931                 return -ENOTEMPTY;
1932
1933         drop_nlink(dentry->d_inode);
1934         drop_nlink(dir);
1935         return shmem_unlink(dir, dentry);
1936 }
1937
1938 /*
1939  * The VFS layer already does all the dentry stuff for rename,
1940  * we just have to decrement the usage count for the target if
1941  * it exists so that the VFS layer correctly free's it when it
1942  * gets overwritten.
1943  */
1944 static int shmem_rename(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
1945 {
1946         struct inode *inode = old_dentry->d_inode;
1947         int they_are_dirs = S_ISDIR(inode->i_mode);
1948
1949         if (!simple_empty(new_dentry))
1950                 return -ENOTEMPTY;
1951
1952         if (new_dentry->d_inode) {
1953                 (void) shmem_unlink(new_dir, new_dentry);
1954                 if (they_are_dirs)
1955                         drop_nlink(old_dir);
1956         } else if (they_are_dirs) {
1957                 drop_nlink(old_dir);
1958                 inc_nlink(new_dir);
1959         }
1960
1961         old_dir->i_size -= BOGO_DIRENT_SIZE;
1962         new_dir->i_size += BOGO_DIRENT_SIZE;
1963         old_dir->i_ctime = old_dir->i_mtime =
1964         new_dir->i_ctime = new_dir->i_mtime =
1965         inode->i_ctime = CURRENT_TIME;
1966         return 0;
1967 }
1968
1969 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
1970 {
1971         int error;
1972         int len;
1973         struct inode *inode;
1974         struct page *page;
1975         char *kaddr;
1976         struct shmem_inode_info *info;
1977
1978         len = strlen(symname) + 1;
1979         if (len > PAGE_CACHE_SIZE)
1980                 return -ENAMETOOLONG;
1981
1982         inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0, VM_NORESERVE);
1983         if (!inode)
1984                 return -ENOSPC;
1985
1986         error = security_inode_init_security(inode, dir, &dentry->d_name,
1987                                              shmem_initxattrs, NULL);
1988         if (error) {
1989                 if (error != -EOPNOTSUPP) {
1990                         iput(inode);
1991                         return error;
1992                 }
1993                 error = 0;
1994         }
1995
1996         info = SHMEM_I(inode);
1997         inode->i_size = len-1;
1998         if (len <= SHORT_SYMLINK_LEN) {
1999                 info->symlink = kmemdup(symname, len, GFP_KERNEL);
2000                 if (!info->symlink) {
2001                         iput(inode);
2002                         return -ENOMEM;
2003                 }
2004                 inode->i_op = &shmem_short_symlink_operations;
2005         } else {
2006                 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
2007                 if (error) {
2008                         iput(inode);
2009                         return error;
2010                 }
2011                 inode->i_mapping->a_ops = &shmem_aops;
2012                 inode->i_op = &shmem_symlink_inode_operations;
2013                 kaddr = kmap_atomic(page);
2014                 memcpy(kaddr, symname, len);
2015                 kunmap_atomic(kaddr);
2016                 SetPageUptodate(page);
2017                 set_page_dirty(page);
2018                 unlock_page(page);
2019                 page_cache_release(page);
2020         }
2021         dir->i_size += BOGO_DIRENT_SIZE;
2022         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2023         d_instantiate(dentry, inode);
2024         dget(dentry);
2025         return 0;
2026 }
2027
2028 static void *shmem_follow_short_symlink(struct dentry *dentry, struct nameidata *nd)
2029 {
2030         nd_set_link(nd, SHMEM_I(dentry->d_inode)->symlink);
2031         return NULL;
2032 }
2033
2034 static void *shmem_follow_link(struct dentry *dentry, struct nameidata *nd)
2035 {
2036         struct page *page = NULL;
2037         int error = shmem_getpage(dentry->d_inode, 0, &page, SGP_READ, NULL);
2038         nd_set_link(nd, error ? ERR_PTR(error) : kmap(page));
2039         if (page)
2040                 unlock_page(page);
2041         return page;
2042 }
2043
2044 static void shmem_put_link(struct dentry *dentry, struct nameidata *nd, void *cookie)
2045 {
2046         if (!IS_ERR(nd_get_link(nd))) {
2047                 struct page *page = cookie;
2048                 kunmap(page);
2049                 mark_page_accessed(page);
2050                 page_cache_release(page);
2051         }
2052 }
2053
2054 #ifdef CONFIG_TMPFS_XATTR
2055 /*
2056  * Superblocks without xattr inode operations may get some security.* xattr
2057  * support from the LSM "for free". As soon as we have any other xattrs
2058  * like ACLs, we also need to implement the security.* handlers at
2059  * filesystem level, though.
2060  */
2061
2062 /*
2063  * Callback for security_inode_init_security() for acquiring xattrs.
2064  */
2065 static int shmem_initxattrs(struct inode *inode,
2066                             const struct xattr *xattr_array,
2067                             void *fs_info)
2068 {
2069         struct shmem_inode_info *info = SHMEM_I(inode);
2070         const struct xattr *xattr;
2071         struct simple_xattr *new_xattr;
2072         size_t len;
2073
2074         for (xattr = xattr_array; xattr->name != NULL; xattr++) {
2075                 new_xattr = simple_xattr_alloc(xattr->value, xattr->value_len);
2076                 if (!new_xattr)
2077                         return -ENOMEM;
2078
2079                 len = strlen(xattr->name) + 1;
2080                 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
2081                                           GFP_KERNEL);
2082                 if (!new_xattr->name) {
2083                         kfree(new_xattr);
2084                         return -ENOMEM;
2085                 }
2086
2087                 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
2088                        XATTR_SECURITY_PREFIX_LEN);
2089                 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
2090                        xattr->name, len);
2091
2092                 simple_xattr_list_add(&info->xattrs, new_xattr);
2093         }
2094
2095         return 0;
2096 }
2097
2098 static const struct xattr_handler *shmem_xattr_handlers[] = {
2099 #ifdef CONFIG_TMPFS_POSIX_ACL
2100         &generic_acl_access_handler,
2101         &generic_acl_default_handler,
2102 #endif
2103         NULL
2104 };
2105
2106 static int shmem_xattr_validate(const char *name)
2107 {
2108         struct { const char *prefix; size_t len; } arr[] = {
2109                 { XATTR_SECURITY_PREFIX, XATTR_SECURITY_PREFIX_LEN },
2110                 { XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN }
2111         };
2112         int i;
2113
2114         for (i = 0; i < ARRAY_SIZE(arr); i++) {
2115                 size_t preflen = arr[i].len;
2116                 if (strncmp(name, arr[i].prefix, preflen) == 0) {
2117                         if (!name[preflen])
2118                                 return -EINVAL;
2119                         return 0;
2120                 }
2121         }
2122         return -EOPNOTSUPP;
2123 }
2124
2125 static ssize_t shmem_getxattr(struct dentry *dentry, const char *name,
2126                               void *buffer, size_t size)
2127 {
2128         struct shmem_inode_info *info = SHMEM_I(dentry->d_inode);
2129         int err;
2130
2131         /*
2132          * If this is a request for a synthetic attribute in the system.*
2133          * namespace use the generic infrastructure to resolve a handler
2134          * for it via sb->s_xattr.
2135          */
2136         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2137                 return generic_getxattr(dentry, name, buffer, size);
2138
2139         err = shmem_xattr_validate(name);
2140         if (err)
2141                 return err;
2142
2143         return simple_xattr_get(&info->xattrs, name, buffer, size);
2144 }
2145
2146 static int shmem_setxattr(struct dentry *dentry, const char *name,
2147                           const void *value, size_t size, int flags)
2148 {
2149         struct shmem_inode_info *info = SHMEM_I(dentry->d_inode);
2150         int err;
2151
2152         /*
2153          * If this is a request for a synthetic attribute in the system.*
2154          * namespace use the generic infrastructure to resolve a handler
2155          * for it via sb->s_xattr.
2156          */
2157         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2158                 return generic_setxattr(dentry, name, value, size, flags);
2159
2160         err = shmem_xattr_validate(name);
2161         if (err)
2162                 return err;
2163
2164         return simple_xattr_set(&info->xattrs, name, value, size, flags);
2165 }
2166
2167 static int shmem_removexattr(struct dentry *dentry, const char *name)
2168 {
2169         struct shmem_inode_info *info = SHMEM_I(dentry->d_inode);
2170         int err;
2171
2172         /*
2173          * If this is a request for a synthetic attribute in the system.*
2174          * namespace use the generic infrastructure to resolve a handler
2175          * for it via sb->s_xattr.
2176          */
2177         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2178                 return generic_removexattr(dentry, name);
2179
2180         err = shmem_xattr_validate(name);
2181         if (err)
2182                 return err;
2183
2184         return simple_xattr_remove(&info->xattrs, name);
2185 }
2186
2187 static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
2188 {
2189         struct shmem_inode_info *info = SHMEM_I(dentry->d_inode);
2190         return simple_xattr_list(&info->xattrs, buffer, size);
2191 }
2192 #endif /* CONFIG_TMPFS_XATTR */
2193
2194 static const struct inode_operations shmem_short_symlink_operations = {
2195         .readlink       = generic_readlink,
2196         .follow_link    = shmem_follow_short_symlink,
2197 #ifdef CONFIG_TMPFS_XATTR
2198         .setxattr       = shmem_setxattr,
2199         .getxattr       = shmem_getxattr,
2200         .listxattr      = shmem_listxattr,
2201         .removexattr    = shmem_removexattr,
2202 #endif
2203 };
2204
2205 static const struct inode_operations shmem_symlink_inode_operations = {
2206         .readlink       = generic_readlink,
2207         .follow_link    = shmem_follow_link,
2208         .put_link       = shmem_put_link,
2209 #ifdef CONFIG_TMPFS_XATTR
2210         .setxattr       = shmem_setxattr,
2211         .getxattr       = shmem_getxattr,
2212         .listxattr      = shmem_listxattr,
2213         .removexattr    = shmem_removexattr,
2214 #endif
2215 };
2216
2217 static struct dentry *shmem_get_parent(struct dentry *child)
2218 {
2219         return ERR_PTR(-ESTALE);
2220 }
2221
2222 static int shmem_match(struct inode *ino, void *vfh)
2223 {
2224         __u32 *fh = vfh;
2225         __u64 inum = fh[2];
2226         inum = (inum << 32) | fh[1];
2227         return ino->i_ino == inum && fh[0] == ino->i_generation;
2228 }
2229
2230 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
2231                 struct fid *fid, int fh_len, int fh_type)
2232 {
2233         struct inode *inode;
2234         struct dentry *dentry = NULL;
2235         u64 inum;
2236
2237         if (fh_len < 3)
2238                 return NULL;
2239
2240         inum = fid->raw[2];
2241         inum = (inum << 32) | fid->raw[1];
2242
2243         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
2244                         shmem_match, fid->raw);
2245         if (inode) {
2246                 dentry = d_find_alias(inode);
2247                 iput(inode);
2248         }
2249
2250         return dentry;
2251 }
2252
2253 static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
2254                                 struct inode *parent)
2255 {
2256         if (*len < 3) {
2257                 *len = 3;
2258                 return 255;
2259         }
2260
2261         if (inode_unhashed(inode)) {
2262                 /* Unfortunately insert_inode_hash is not idempotent,
2263                  * so as we hash inodes here rather than at creation
2264                  * time, we need a lock to ensure we only try
2265                  * to do it once
2266                  */
2267                 static DEFINE_SPINLOCK(lock);
2268                 spin_lock(&lock);
2269                 if (inode_unhashed(inode))
2270                         __insert_inode_hash(inode,
2271                                             inode->i_ino + inode->i_generation);
2272                 spin_unlock(&lock);
2273         }
2274
2275         fh[0] = inode->i_generation;
2276         fh[1] = inode->i_ino;
2277         fh[2] = ((__u64)inode->i_ino) >> 32;
2278
2279         *len = 3;
2280         return 1;
2281 }
2282
2283 static const struct export_operations shmem_export_ops = {
2284         .get_parent     = shmem_get_parent,
2285         .encode_fh      = shmem_encode_fh,
2286         .fh_to_dentry   = shmem_fh_to_dentry,
2287 };
2288
2289 static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
2290                                bool remount)
2291 {
2292         char *this_char, *value, *rest;
2293         uid_t uid;
2294         gid_t gid;
2295
2296         while (options != NULL) {
2297                 this_char = options;
2298                 for (;;) {
2299                         /*
2300                          * NUL-terminate this option: unfortunately,
2301                          * mount options form a comma-separated list,
2302                          * but mpol's nodelist may also contain commas.
2303                          */
2304                         options = strchr(options, ',');
2305                         if (options == NULL)
2306                                 break;
2307                         options++;
2308                         if (!isdigit(*options)) {
2309                                 options[-1] = '\0';
2310                                 break;
2311                         }
2312                 }
2313                 if (!*this_char)
2314                         continue;
2315                 if ((value = strchr(this_char,'=')) != NULL) {
2316                         *value++ = 0;
2317                 } else {
2318                         printk(KERN_ERR
2319                             "tmpfs: No value for mount option '%s'\n",
2320                             this_char);
2321                         return 1;
2322                 }
2323
2324                 if (!strcmp(this_char,"size")) {
2325                         unsigned long long size;
2326                         size = memparse(value,&rest);
2327                         if (*rest == '%') {
2328                                 size <<= PAGE_SHIFT;
2329                                 size *= totalram_pages;
2330                                 do_div(size, 100);
2331                                 rest++;
2332                         }
2333                         if (*rest)
2334                                 goto bad_val;
2335                         sbinfo->max_blocks =
2336                                 DIV_ROUND_UP(size, PAGE_CACHE_SIZE);
2337                 } else if (!strcmp(this_char,"nr_blocks")) {
2338                         sbinfo->max_blocks = memparse(value, &rest);
2339                         if (*rest)
2340                                 goto bad_val;
2341                 } else if (!strcmp(this_char,"nr_inodes")) {
2342                         sbinfo->max_inodes = memparse(value, &rest);
2343                         if (*rest)
2344                                 goto bad_val;
2345                 } else if (!strcmp(this_char,"mode")) {
2346                         if (remount)
2347                                 continue;
2348                         sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
2349                         if (*rest)
2350                                 goto bad_val;
2351                 } else if (!strcmp(this_char,"uid")) {
2352                         if (remount)
2353                                 continue;
2354                         uid = simple_strtoul(value, &rest, 0);
2355                         if (*rest)
2356                                 goto bad_val;
2357                         sbinfo->uid = make_kuid(current_user_ns(), uid);
2358                         if (!uid_valid(sbinfo->uid))
2359                                 goto bad_val;
2360                 } else if (!strcmp(this_char,"gid")) {
2361                         if (remount)
2362                                 continue;
2363                         gid = simple_strtoul(value, &rest, 0);
2364                         if (*rest)
2365                                 goto bad_val;
2366                         sbinfo->gid = make_kgid(current_user_ns(), gid);
2367                         if (!gid_valid(sbinfo->gid))
2368                                 goto bad_val;
2369                 } else if (!strcmp(this_char,"mpol")) {
2370                         if (mpol_parse_str(value, &sbinfo->mpol, 1))
2371                                 goto bad_val;
2372                 } else {
2373                         printk(KERN_ERR "tmpfs: Bad mount option %s\n",
2374                                this_char);
2375                         return 1;
2376                 }
2377         }
2378         return 0;
2379
2380 bad_val:
2381         printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
2382                value, this_char);
2383         return 1;
2384
2385 }
2386
2387 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2388 {
2389         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2390         struct shmem_sb_info config = *sbinfo;
2391         unsigned long inodes;
2392         int error = -EINVAL;
2393
2394         if (shmem_parse_options(data, &config, true))
2395                 return error;
2396
2397         spin_lock(&sbinfo->stat_lock);
2398         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2399         if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
2400                 goto out;
2401         if (config.max_inodes < inodes)
2402                 goto out;
2403         /*
2404          * Those tests disallow limited->unlimited while any are in use;
2405          * but we must separately disallow unlimited->limited, because
2406          * in that case we have no record of how much is already in use.
2407          */
2408         if (config.max_blocks && !sbinfo->max_blocks)
2409                 goto out;
2410         if (config.max_inodes && !sbinfo->max_inodes)
2411                 goto out;
2412
2413         error = 0;
2414         sbinfo->max_blocks  = config.max_blocks;
2415         sbinfo->max_inodes  = config.max_inodes;
2416         sbinfo->free_inodes = config.max_inodes - inodes;
2417
2418         mpol_put(sbinfo->mpol);
2419         sbinfo->mpol        = config.mpol;      /* transfers initial ref */
2420 out:
2421         spin_unlock(&sbinfo->stat_lock);
2422         return error;
2423 }
2424
2425 static int shmem_show_options(struct seq_file *seq, struct dentry *root)
2426 {
2427         struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
2428
2429         if (sbinfo->max_blocks != shmem_default_max_blocks())
2430                 seq_printf(seq, ",size=%luk",
2431                         sbinfo->max_blocks << (PAGE_CACHE_SHIFT - 10));
2432         if (sbinfo->max_inodes != shmem_default_max_inodes())
2433                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
2434         if (sbinfo->mode != (S_IRWXUGO | S_ISVTX))
2435                 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
2436         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
2437                 seq_printf(seq, ",uid=%u",
2438                                 from_kuid_munged(&init_user_ns, sbinfo->uid));
2439         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
2440                 seq_printf(seq, ",gid=%u",
2441                                 from_kgid_munged(&init_user_ns, sbinfo->gid));
2442         shmem_show_mpol(seq, sbinfo->mpol);
2443         return 0;
2444 }
2445 #endif /* CONFIG_TMPFS */
2446
2447 static void shmem_put_super(struct super_block *sb)
2448 {
2449         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2450
2451         percpu_counter_destroy(&sbinfo->used_blocks);
2452         kfree(sbinfo);
2453         sb->s_fs_info = NULL;
2454 }
2455
2456 int shmem_fill_super(struct super_block *sb, void *data, int silent)
2457 {
2458         struct inode *inode;
2459         struct shmem_sb_info *sbinfo;
2460         int err = -ENOMEM;
2461
2462         /* Round up to L1_CACHE_BYTES to resist false sharing */
2463         sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
2464                                 L1_CACHE_BYTES), GFP_KERNEL);
2465         if (!sbinfo)
2466                 return -ENOMEM;
2467
2468         sbinfo->mode = S_IRWXUGO | S_ISVTX;
2469         sbinfo->uid = current_fsuid();
2470         sbinfo->gid = current_fsgid();
2471         sb->s_fs_info = sbinfo;
2472
2473 #ifdef CONFIG_TMPFS
2474         /*
2475          * Per default we only allow half of the physical ram per
2476          * tmpfs instance, limiting inodes to one per page of lowmem;
2477          * but the internal instance is left unlimited.
2478          */
2479         if (!(sb->s_flags & MS_NOUSER)) {
2480                 sbinfo->max_blocks = shmem_default_max_blocks();
2481                 sbinfo->max_inodes = shmem_default_max_inodes();
2482                 if (shmem_parse_options(data, sbinfo, false)) {
2483                         err = -EINVAL;
2484                         goto failed;
2485                 }
2486         }
2487         sb->s_export_op = &shmem_export_ops;
2488         sb->s_flags |= MS_NOSEC;
2489 #else
2490         sb->s_flags |= MS_NOUSER;
2491 #endif
2492
2493         spin_lock_init(&sbinfo->stat_lock);
2494         if (percpu_counter_init(&sbinfo->used_blocks, 0))
2495                 goto failed;
2496         sbinfo->free_inodes = sbinfo->max_inodes;
2497
2498         sb->s_maxbytes = MAX_LFS_FILESIZE;
2499         sb->s_blocksize = PAGE_CACHE_SIZE;
2500         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
2501         sb->s_magic = TMPFS_MAGIC;
2502         sb->s_op = &shmem_ops;
2503         sb->s_time_gran = 1;
2504 #ifdef CONFIG_TMPFS_XATTR
2505         sb->s_xattr = shmem_xattr_handlers;
2506 #endif
2507 #ifdef CONFIG_TMPFS_POSIX_ACL
2508         sb->s_flags |= MS_POSIXACL;
2509 #endif
2510
2511         inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
2512         if (!inode)
2513                 goto failed;
2514         inode->i_uid = sbinfo->uid;
2515         inode->i_gid = sbinfo->gid;
2516         sb->s_root = d_make_root(inode);
2517         if (!sb->s_root)
2518                 goto failed;
2519         return 0;
2520
2521 failed:
2522         shmem_put_super(sb);
2523         return err;
2524 }
2525
2526 static struct kmem_cache *shmem_inode_cachep;
2527
2528 static struct inode *shmem_alloc_inode(struct super_block *sb)
2529 {
2530         struct shmem_inode_info *info;
2531         info = kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
2532         if (!info)
2533                 return NULL;
2534         return &info->vfs_inode;
2535 }
2536
2537 static void shmem_destroy_callback(struct rcu_head *head)
2538 {
2539         struct inode *inode = container_of(head, struct inode, i_rcu);
2540         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
2541 }
2542
2543 static void shmem_destroy_inode(struct inode *inode)
2544 {
2545         if (S_ISREG(inode->i_mode))
2546                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
2547         call_rcu(&inode->i_rcu, shmem_destroy_callback);
2548 }
2549
2550 static void shmem_init_inode(void *foo)
2551 {
2552         struct shmem_inode_info *info = foo;
2553         inode_init_once(&info->vfs_inode);
2554 }
2555
2556 static int shmem_init_inodecache(void)
2557 {
2558         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
2559                                 sizeof(struct shmem_inode_info),
2560                                 0, SLAB_PANIC, shmem_init_inode);
2561         return 0;
2562 }
2563
2564 static void shmem_destroy_inodecache(void)
2565 {
2566         kmem_cache_destroy(shmem_inode_cachep);
2567 }
2568
2569 static const struct address_space_operations shmem_aops = {
2570         .writepage      = shmem_writepage,
2571         .set_page_dirty = __set_page_dirty_no_writeback,
2572 #ifdef CONFIG_TMPFS
2573         .write_begin    = shmem_write_begin,
2574         .write_end      = shmem_write_end,
2575 #endif
2576         .migratepage    = migrate_page,
2577         .error_remove_page = generic_error_remove_page,
2578 };
2579
2580 static const struct file_operations shmem_file_operations = {
2581         .mmap           = shmem_mmap,
2582 #ifdef CONFIG_TMPFS
2583         .llseek         = generic_file_llseek,
2584         .read           = do_sync_read,
2585         .write          = do_sync_write,
2586         .aio_read       = shmem_file_aio_read,
2587         .aio_write      = generic_file_aio_write,
2588         .fsync          = noop_fsync,
2589         .splice_read    = shmem_file_splice_read,
2590         .splice_write   = generic_file_splice_write,
2591         .fallocate      = shmem_fallocate,
2592 #endif
2593 };
2594
2595 static const struct inode_operations shmem_inode_operations = {
2596         .setattr        = shmem_setattr,
2597 #ifdef CONFIG_TMPFS_XATTR
2598         .setxattr       = shmem_setxattr,
2599         .getxattr       = shmem_getxattr,
2600         .listxattr      = shmem_listxattr,
2601         .removexattr    = shmem_removexattr,
2602 #endif
2603 };
2604
2605 static const struct inode_operations shmem_dir_inode_operations = {
2606 #ifdef CONFIG_TMPFS
2607         .create         = shmem_create,
2608         .lookup         = simple_lookup,
2609         .link           = shmem_link,
2610         .unlink         = shmem_unlink,
2611         .symlink        = shmem_symlink,
2612         .mkdir          = shmem_mkdir,
2613         .rmdir          = shmem_rmdir,
2614         .mknod          = shmem_mknod,
2615         .rename         = shmem_rename,
2616 #endif
2617 #ifdef CONFIG_TMPFS_XATTR
2618         .setxattr       = shmem_setxattr,
2619         .getxattr       = shmem_getxattr,
2620         .listxattr      = shmem_listxattr,
2621         .removexattr    = shmem_removexattr,
2622 #endif
2623 #ifdef CONFIG_TMPFS_POSIX_ACL
2624         .setattr        = shmem_setattr,
2625 #endif
2626 };
2627
2628 static const struct inode_operations shmem_special_inode_operations = {
2629 #ifdef CONFIG_TMPFS_XATTR
2630         .setxattr       = shmem_setxattr,
2631         .getxattr       = shmem_getxattr,
2632         .listxattr      = shmem_listxattr,
2633         .removexattr    = shmem_removexattr,
2634 #endif
2635 #ifdef CONFIG_TMPFS_POSIX_ACL
2636         .setattr        = shmem_setattr,
2637 #endif
2638 };
2639
2640 static const struct super_operations shmem_ops = {
2641         .alloc_inode    = shmem_alloc_inode,
2642         .destroy_inode  = shmem_destroy_inode,
2643 #ifdef CONFIG_TMPFS
2644         .statfs         = shmem_statfs,
2645         .remount_fs     = shmem_remount_fs,
2646         .show_options   = shmem_show_options,
2647 #endif
2648         .evict_inode    = shmem_evict_inode,
2649         .drop_inode     = generic_delete_inode,
2650         .put_super      = shmem_put_super,
2651 };
2652
2653 static const struct vm_operations_struct shmem_vm_ops = {
2654         .fault          = shmem_fault,
2655 #ifdef CONFIG_NUMA
2656         .set_policy     = shmem_set_policy,
2657         .get_policy     = shmem_get_policy,
2658 #endif
2659         .remap_pages    = generic_file_remap_pages,
2660 };
2661
2662 static struct dentry *shmem_mount(struct file_system_type *fs_type,
2663         int flags, const char *dev_name, void *data)
2664 {
2665         return mount_nodev(fs_type, flags, data, shmem_fill_super);
2666 }
2667
2668 static struct file_system_type shmem_fs_type = {
2669         .owner          = THIS_MODULE,
2670         .name           = "tmpfs",
2671         .mount          = shmem_mount,
2672         .kill_sb        = kill_litter_super,
2673 };
2674
2675 int __init shmem_init(void)
2676 {
2677         int error;
2678
2679         error = bdi_init(&shmem_backing_dev_info);
2680         if (error)
2681                 goto out4;
2682
2683         error = shmem_init_inodecache();
2684         if (error)
2685                 goto out3;
2686
2687         error = register_filesystem(&shmem_fs_type);
2688         if (error) {
2689                 printk(KERN_ERR "Could not register tmpfs\n");
2690                 goto out2;
2691         }
2692
2693         shm_mnt = vfs_kern_mount(&shmem_fs_type, MS_NOUSER,
2694                                  shmem_fs_type.name, NULL);
2695         if (IS_ERR(shm_mnt)) {
2696                 error = PTR_ERR(shm_mnt);
2697                 printk(KERN_ERR "Could not kern_mount tmpfs\n");
2698                 goto out1;
2699         }
2700         return 0;
2701
2702 out1:
2703         unregister_filesystem(&shmem_fs_type);
2704 out2:
2705         shmem_destroy_inodecache();
2706 out3:
2707         bdi_destroy(&shmem_backing_dev_info);
2708 out4:
2709         shm_mnt = ERR_PTR(error);
2710         return error;
2711 }
2712
2713 #else /* !CONFIG_SHMEM */
2714
2715 /*
2716  * tiny-shmem: simple shmemfs and tmpfs using ramfs code
2717  *
2718  * This is intended for small system where the benefits of the full
2719  * shmem code (swap-backed and resource-limited) are outweighed by
2720  * their complexity. On systems without swap this code should be
2721  * effectively equivalent, but much lighter weight.
2722  */
2723
2724 #include <linux/ramfs.h>
2725
2726 static struct file_system_type shmem_fs_type = {
2727         .name           = "tmpfs",
2728         .mount          = ramfs_mount,
2729         .kill_sb        = kill_litter_super,
2730 };
2731
2732 int __init shmem_init(void)
2733 {
2734         BUG_ON(register_filesystem(&shmem_fs_type) != 0);
2735
2736         shm_mnt = kern_mount(&shmem_fs_type);
2737         BUG_ON(IS_ERR(shm_mnt));
2738
2739         return 0;
2740 }
2741
2742 int shmem_unuse(swp_entry_t swap, struct page *page)
2743 {
2744         return 0;
2745 }
2746
2747 int shmem_lock(struct file *file, int lock, struct user_struct *user)
2748 {
2749         return 0;
2750 }
2751
2752 void shmem_unlock_mapping(struct address_space *mapping)
2753 {
2754 }
2755
2756 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
2757 {
2758         truncate_inode_pages_range(inode->i_mapping, lstart, lend);
2759 }
2760 EXPORT_SYMBOL_GPL(shmem_truncate_range);
2761
2762 #define shmem_vm_ops                            generic_file_vm_ops
2763 #define shmem_file_operations                   ramfs_file_operations
2764 #define shmem_get_inode(sb, dir, mode, dev, flags)      ramfs_get_inode(sb, dir, mode, dev)
2765 #define shmem_acct_size(flags, size)            0
2766 #define shmem_unacct_size(flags, size)          do {} while (0)
2767
2768 #endif /* CONFIG_SHMEM */
2769
2770 /* common code */
2771
2772 /**
2773  * shmem_file_setup - get an unlinked file living in tmpfs
2774  * @name: name for dentry (to be seen in /proc/<pid>/maps
2775  * @size: size to be set for the file
2776  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
2777  */
2778 struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
2779 {
2780         int error;
2781         struct file *file;
2782         struct inode *inode;
2783         struct path path;
2784         struct dentry *root;
2785         struct qstr this;
2786
2787         if (IS_ERR(shm_mnt))
2788                 return (void *)shm_mnt;
2789
2790         if (size < 0 || size > MAX_LFS_FILESIZE)
2791                 return ERR_PTR(-EINVAL);
2792
2793         if (shmem_acct_size(flags, size))
2794                 return ERR_PTR(-ENOMEM);
2795
2796         error = -ENOMEM;
2797         this.name = name;
2798         this.len = strlen(name);
2799         this.hash = 0; /* will go */
2800         root = shm_mnt->mnt_root;
2801         path.dentry = d_alloc(root, &this);
2802         if (!path.dentry)
2803                 goto put_memory;
2804         path.mnt = mntget(shm_mnt);
2805
2806         error = -ENOSPC;
2807         inode = shmem_get_inode(root->d_sb, NULL, S_IFREG | S_IRWXUGO, 0, flags);
2808         if (!inode)
2809                 goto put_dentry;
2810
2811         d_instantiate(path.dentry, inode);
2812         inode->i_size = size;
2813         clear_nlink(inode);     /* It is unlinked */
2814 #ifndef CONFIG_MMU
2815         error = ramfs_nommu_expand_for_mapping(inode, size);
2816         if (error)
2817                 goto put_dentry;
2818 #endif
2819
2820         error = -ENFILE;
2821         file = alloc_file(&path, FMODE_WRITE | FMODE_READ,
2822                   &shmem_file_operations);
2823         if (!file)
2824                 goto put_dentry;
2825
2826         return file;
2827
2828 put_dentry:
2829         path_put(&path);
2830 put_memory:
2831         shmem_unacct_size(flags, size);
2832         return ERR_PTR(error);
2833 }
2834 EXPORT_SYMBOL_GPL(shmem_file_setup);
2835
2836 /**
2837  * shmem_zero_setup - setup a shared anonymous mapping
2838  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
2839  */
2840 int shmem_zero_setup(struct vm_area_struct *vma)
2841 {
2842         struct file *file;
2843         loff_t size = vma->vm_end - vma->vm_start;
2844
2845         file = shmem_file_setup("dev/zero", size, vma->vm_flags);
2846         if (IS_ERR(file))
2847                 return PTR_ERR(file);
2848
2849         if (vma->vm_file)
2850                 fput(vma->vm_file);
2851         vma->vm_file = file;
2852         vma->vm_ops = &shmem_vm_ops;
2853         return 0;
2854 }
2855
2856 /**
2857  * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
2858  * @mapping:    the page's address_space
2859  * @index:      the page index
2860  * @gfp:        the page allocator flags to use if allocating
2861  *
2862  * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
2863  * with any new page allocations done using the specified allocation flags.
2864  * But read_cache_page_gfp() uses the ->readpage() method: which does not
2865  * suit tmpfs, since it may have pages in swapcache, and needs to find those
2866  * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
2867  *
2868  * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
2869  * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
2870  */
2871 struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
2872                                          pgoff_t index, gfp_t gfp)
2873 {
2874 #ifdef CONFIG_SHMEM
2875         struct inode *inode = mapping->host;
2876         struct page *page;
2877         int error;
2878
2879         BUG_ON(mapping->a_ops != &shmem_aops);
2880         error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE, gfp, NULL);
2881         if (error)
2882                 page = ERR_PTR(error);
2883         else
2884                 unlock_page(page);
2885         return page;
2886 #else
2887         /*
2888          * The tiny !SHMEM case uses ramfs without swap
2889          */
2890         return read_cache_page_gfp(mapping, index, gfp);
2891 #endif
2892 }
2893 EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);