Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/ebiederm...
[~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         void **slot;
339         unsigned int ret = 0;
340         struct radix_tree_iter iter;
341
342         if (!nr_pages)
343                 return 0;
344
345         rcu_read_lock();
346 restart:
347         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
348                 struct page *page;
349 repeat:
350                 page = radix_tree_deref_slot(slot);
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 != *slot)) {
368                         page_cache_release(page);
369                         goto repeat;
370                 }
371 export:
372                 indices[ret] = iter.index;
373                 pages[ret] = page;
374                 if (++ret == nr_pages)
375                         break;
376         }
377         rcu_read_unlock();
378         return ret;
379 }
380
381 /*
382  * Remove swap entry from radix tree, free the swap and its page cache.
383  */
384 static int shmem_free_swap(struct address_space *mapping,
385                            pgoff_t index, void *radswap)
386 {
387         int error;
388
389         spin_lock_irq(&mapping->tree_lock);
390         error = shmem_radix_tree_replace(mapping, index, radswap, NULL);
391         spin_unlock_irq(&mapping->tree_lock);
392         if (!error)
393                 free_swap_and_cache(radix_to_swp_entry(radswap));
394         return error;
395 }
396
397 /*
398  * Pagevec may contain swap entries, so shuffle up pages before releasing.
399  */
400 static void shmem_deswap_pagevec(struct pagevec *pvec)
401 {
402         int i, j;
403
404         for (i = 0, j = 0; i < pagevec_count(pvec); i++) {
405                 struct page *page = pvec->pages[i];
406                 if (!radix_tree_exceptional_entry(page))
407                         pvec->pages[j++] = page;
408         }
409         pvec->nr = j;
410 }
411
412 /*
413  * SysV IPC SHM_UNLOCK restore Unevictable pages to their evictable lists.
414  */
415 void shmem_unlock_mapping(struct address_space *mapping)
416 {
417         struct pagevec pvec;
418         pgoff_t indices[PAGEVEC_SIZE];
419         pgoff_t index = 0;
420
421         pagevec_init(&pvec, 0);
422         /*
423          * Minor point, but we might as well stop if someone else SHM_LOCKs it.
424          */
425         while (!mapping_unevictable(mapping)) {
426                 /*
427                  * Avoid pagevec_lookup(): find_get_pages() returns 0 as if it
428                  * has finished, if it hits a row of PAGEVEC_SIZE swap entries.
429                  */
430                 pvec.nr = shmem_find_get_pages_and_swap(mapping, index,
431                                         PAGEVEC_SIZE, pvec.pages, indices);
432                 if (!pvec.nr)
433                         break;
434                 index = indices[pvec.nr - 1] + 1;
435                 shmem_deswap_pagevec(&pvec);
436                 check_move_unevictable_pages(pvec.pages, pvec.nr);
437                 pagevec_release(&pvec);
438                 cond_resched();
439         }
440 }
441
442 /*
443  * Remove range of pages and swap entries from radix tree, and free them.
444  * If !unfalloc, truncate or punch hole; if unfalloc, undo failed fallocate.
445  */
446 static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
447                                                                  bool unfalloc)
448 {
449         struct address_space *mapping = inode->i_mapping;
450         struct shmem_inode_info *info = SHMEM_I(inode);
451         pgoff_t start = (lstart + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
452         pgoff_t end = (lend + 1) >> PAGE_CACHE_SHIFT;
453         unsigned int partial_start = lstart & (PAGE_CACHE_SIZE - 1);
454         unsigned int partial_end = (lend + 1) & (PAGE_CACHE_SIZE - 1);
455         struct pagevec pvec;
456         pgoff_t indices[PAGEVEC_SIZE];
457         long nr_swaps_freed = 0;
458         pgoff_t index;
459         int i;
460
461         if (lend == -1)
462                 end = -1;       /* unsigned, so actually very big */
463
464         pagevec_init(&pvec, 0);
465         index = start;
466         while (index < end) {
467                 pvec.nr = shmem_find_get_pages_and_swap(mapping, index,
468                                 min(end - index, (pgoff_t)PAGEVEC_SIZE),
469                                                         pvec.pages, indices);
470                 if (!pvec.nr)
471                         break;
472                 mem_cgroup_uncharge_start();
473                 for (i = 0; i < pagevec_count(&pvec); i++) {
474                         struct page *page = pvec.pages[i];
475
476                         index = indices[i];
477                         if (index >= end)
478                                 break;
479
480                         if (radix_tree_exceptional_entry(page)) {
481                                 if (unfalloc)
482                                         continue;
483                                 nr_swaps_freed += !shmem_free_swap(mapping,
484                                                                 index, page);
485                                 continue;
486                         }
487
488                         if (!trylock_page(page))
489                                 continue;
490                         if (!unfalloc || !PageUptodate(page)) {
491                                 if (page->mapping == mapping) {
492                                         VM_BUG_ON(PageWriteback(page));
493                                         truncate_inode_page(mapping, page);
494                                 }
495                         }
496                         unlock_page(page);
497                 }
498                 shmem_deswap_pagevec(&pvec);
499                 pagevec_release(&pvec);
500                 mem_cgroup_uncharge_end();
501                 cond_resched();
502                 index++;
503         }
504
505         if (partial_start) {
506                 struct page *page = NULL;
507                 shmem_getpage(inode, start - 1, &page, SGP_READ, NULL);
508                 if (page) {
509                         unsigned int top = PAGE_CACHE_SIZE;
510                         if (start > end) {
511                                 top = partial_end;
512                                 partial_end = 0;
513                         }
514                         zero_user_segment(page, partial_start, top);
515                         set_page_dirty(page);
516                         unlock_page(page);
517                         page_cache_release(page);
518                 }
519         }
520         if (partial_end) {
521                 struct page *page = NULL;
522                 shmem_getpage(inode, end, &page, SGP_READ, NULL);
523                 if (page) {
524                         zero_user_segment(page, 0, partial_end);
525                         set_page_dirty(page);
526                         unlock_page(page);
527                         page_cache_release(page);
528                 }
529         }
530         if (start >= end)
531                 return;
532
533         index = start;
534         for ( ; ; ) {
535                 cond_resched();
536                 pvec.nr = shmem_find_get_pages_and_swap(mapping, index,
537                                 min(end - index, (pgoff_t)PAGEVEC_SIZE),
538                                                         pvec.pages, indices);
539                 if (!pvec.nr) {
540                         if (index == start || unfalloc)
541                                 break;
542                         index = start;
543                         continue;
544                 }
545                 if ((index == start || unfalloc) && indices[0] >= end) {
546                         shmem_deswap_pagevec(&pvec);
547                         pagevec_release(&pvec);
548                         break;
549                 }
550                 mem_cgroup_uncharge_start();
551                 for (i = 0; i < pagevec_count(&pvec); i++) {
552                         struct page *page = pvec.pages[i];
553
554                         index = indices[i];
555                         if (index >= end)
556                                 break;
557
558                         if (radix_tree_exceptional_entry(page)) {
559                                 if (unfalloc)
560                                         continue;
561                                 nr_swaps_freed += !shmem_free_swap(mapping,
562                                                                 index, page);
563                                 continue;
564                         }
565
566                         lock_page(page);
567                         if (!unfalloc || !PageUptodate(page)) {
568                                 if (page->mapping == mapping) {
569                                         VM_BUG_ON(PageWriteback(page));
570                                         truncate_inode_page(mapping, page);
571                                 }
572                         }
573                         unlock_page(page);
574                 }
575                 shmem_deswap_pagevec(&pvec);
576                 pagevec_release(&pvec);
577                 mem_cgroup_uncharge_end();
578                 index++;
579         }
580
581         spin_lock(&info->lock);
582         info->swapped -= nr_swaps_freed;
583         shmem_recalc_inode(inode);
584         spin_unlock(&info->lock);
585 }
586
587 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
588 {
589         shmem_undo_range(inode, lstart, lend, false);
590         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
591 }
592 EXPORT_SYMBOL_GPL(shmem_truncate_range);
593
594 static int shmem_setattr(struct dentry *dentry, struct iattr *attr)
595 {
596         struct inode *inode = dentry->d_inode;
597         int error;
598
599         error = inode_change_ok(inode, attr);
600         if (error)
601                 return error;
602
603         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
604                 loff_t oldsize = inode->i_size;
605                 loff_t newsize = attr->ia_size;
606
607                 if (newsize != oldsize) {
608                         i_size_write(inode, newsize);
609                         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
610                 }
611                 if (newsize < oldsize) {
612                         loff_t holebegin = round_up(newsize, PAGE_SIZE);
613                         unmap_mapping_range(inode->i_mapping, holebegin, 0, 1);
614                         shmem_truncate_range(inode, newsize, (loff_t)-1);
615                         /* unmap again to remove racily COWed private pages */
616                         unmap_mapping_range(inode->i_mapping, holebegin, 0, 1);
617                 }
618         }
619
620         setattr_copy(inode, attr);
621 #ifdef CONFIG_TMPFS_POSIX_ACL
622         if (attr->ia_valid & ATTR_MODE)
623                 error = generic_acl_chmod(inode);
624 #endif
625         return error;
626 }
627
628 static void shmem_evict_inode(struct inode *inode)
629 {
630         struct shmem_inode_info *info = SHMEM_I(inode);
631
632         if (inode->i_mapping->a_ops == &shmem_aops) {
633                 shmem_unacct_size(info->flags, inode->i_size);
634                 inode->i_size = 0;
635                 shmem_truncate_range(inode, 0, (loff_t)-1);
636                 if (!list_empty(&info->swaplist)) {
637                         mutex_lock(&shmem_swaplist_mutex);
638                         list_del_init(&info->swaplist);
639                         mutex_unlock(&shmem_swaplist_mutex);
640                 }
641         } else
642                 kfree(info->symlink);
643
644         simple_xattrs_free(&info->xattrs);
645         WARN_ON(inode->i_blocks);
646         shmem_free_inode(inode->i_sb);
647         clear_inode(inode);
648 }
649
650 /*
651  * If swap found in inode, free it and move page from swapcache to filecache.
652  */
653 static int shmem_unuse_inode(struct shmem_inode_info *info,
654                              swp_entry_t swap, struct page **pagep)
655 {
656         struct address_space *mapping = info->vfs_inode.i_mapping;
657         void *radswap;
658         pgoff_t index;
659         gfp_t gfp;
660         int error = 0;
661
662         radswap = swp_to_radix_entry(swap);
663         index = radix_tree_locate_item(&mapping->page_tree, radswap);
664         if (index == -1)
665                 return 0;
666
667         /*
668          * Move _head_ to start search for next from here.
669          * But be careful: shmem_evict_inode checks list_empty without taking
670          * mutex, and there's an instant in list_move_tail when info->swaplist
671          * would appear empty, if it were the only one on shmem_swaplist.
672          */
673         if (shmem_swaplist.next != &info->swaplist)
674                 list_move_tail(&shmem_swaplist, &info->swaplist);
675
676         gfp = mapping_gfp_mask(mapping);
677         if (shmem_should_replace_page(*pagep, gfp)) {
678                 mutex_unlock(&shmem_swaplist_mutex);
679                 error = shmem_replace_page(pagep, gfp, info, index);
680                 mutex_lock(&shmem_swaplist_mutex);
681                 /*
682                  * We needed to drop mutex to make that restrictive page
683                  * allocation, but the inode might have been freed while we
684                  * dropped it: although a racing shmem_evict_inode() cannot
685                  * complete without emptying the radix_tree, our page lock
686                  * on this swapcache page is not enough to prevent that -
687                  * free_swap_and_cache() of our swap entry will only
688                  * trylock_page(), removing swap from radix_tree whatever.
689                  *
690                  * We must not proceed to shmem_add_to_page_cache() if the
691                  * inode has been freed, but of course we cannot rely on
692                  * inode or mapping or info to check that.  However, we can
693                  * safely check if our swap entry is still in use (and here
694                  * it can't have got reused for another page): if it's still
695                  * in use, then the inode cannot have been freed yet, and we
696                  * can safely proceed (if it's no longer in use, that tells
697                  * nothing about the inode, but we don't need to unuse swap).
698                  */
699                 if (!page_swapcount(*pagep))
700                         error = -ENOENT;
701         }
702
703         /*
704          * We rely on shmem_swaplist_mutex, not only to protect the swaplist,
705          * but also to hold up shmem_evict_inode(): so inode cannot be freed
706          * beneath us (pagelock doesn't help until the page is in pagecache).
707          */
708         if (!error)
709                 error = shmem_add_to_page_cache(*pagep, mapping, index,
710                                                 GFP_NOWAIT, radswap);
711         if (error != -ENOMEM) {
712                 /*
713                  * Truncation and eviction use free_swap_and_cache(), which
714                  * only does trylock page: if we raced, best clean up here.
715                  */
716                 delete_from_swap_cache(*pagep);
717                 set_page_dirty(*pagep);
718                 if (!error) {
719                         spin_lock(&info->lock);
720                         info->swapped--;
721                         spin_unlock(&info->lock);
722                         swap_free(swap);
723                 }
724                 error = 1;      /* not an error, but entry was found */
725         }
726         return error;
727 }
728
729 /*
730  * Search through swapped inodes to find and replace swap by page.
731  */
732 int shmem_unuse(swp_entry_t swap, struct page *page)
733 {
734         struct list_head *this, *next;
735         struct shmem_inode_info *info;
736         int found = 0;
737         int error = 0;
738
739         /*
740          * There's a faint possibility that swap page was replaced before
741          * caller locked it: caller will come back later with the right page.
742          */
743         if (unlikely(!PageSwapCache(page) || page_private(page) != swap.val))
744                 goto out;
745
746         /*
747          * Charge page using GFP_KERNEL while we can wait, before taking
748          * the shmem_swaplist_mutex which might hold up shmem_writepage().
749          * Charged back to the user (not to caller) when swap account is used.
750          */
751         error = mem_cgroup_cache_charge(page, current->mm, GFP_KERNEL);
752         if (error)
753                 goto out;
754         /* No radix_tree_preload: swap entry keeps a place for page in tree */
755
756         mutex_lock(&shmem_swaplist_mutex);
757         list_for_each_safe(this, next, &shmem_swaplist) {
758                 info = list_entry(this, struct shmem_inode_info, swaplist);
759                 if (info->swapped)
760                         found = shmem_unuse_inode(info, swap, &page);
761                 else
762                         list_del_init(&info->swaplist);
763                 cond_resched();
764                 if (found)
765                         break;
766         }
767         mutex_unlock(&shmem_swaplist_mutex);
768
769         if (found < 0)
770                 error = found;
771 out:
772         unlock_page(page);
773         page_cache_release(page);
774         return error;
775 }
776
777 /*
778  * Move the page from the page cache to the swap cache.
779  */
780 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
781 {
782         struct shmem_inode_info *info;
783         struct address_space *mapping;
784         struct inode *inode;
785         swp_entry_t swap;
786         pgoff_t index;
787
788         BUG_ON(!PageLocked(page));
789         mapping = page->mapping;
790         index = page->index;
791         inode = mapping->host;
792         info = SHMEM_I(inode);
793         if (info->flags & VM_LOCKED)
794                 goto redirty;
795         if (!total_swap_pages)
796                 goto redirty;
797
798         /*
799          * shmem_backing_dev_info's capabilities prevent regular writeback or
800          * sync from ever calling shmem_writepage; but a stacking filesystem
801          * might use ->writepage of its underlying filesystem, in which case
802          * tmpfs should write out to swap only in response to memory pressure,
803          * and not for the writeback threads or sync.
804          */
805         if (!wbc->for_reclaim) {
806                 WARN_ON_ONCE(1);        /* Still happens? Tell us about it! */
807                 goto redirty;
808         }
809
810         /*
811          * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
812          * value into swapfile.c, the only way we can correctly account for a
813          * fallocated page arriving here is now to initialize it and write it.
814          *
815          * That's okay for a page already fallocated earlier, but if we have
816          * not yet completed the fallocation, then (a) we want to keep track
817          * of this page in case we have to undo it, and (b) it may not be a
818          * good idea to continue anyway, once we're pushing into swap.  So
819          * reactivate the page, and let shmem_fallocate() quit when too many.
820          */
821         if (!PageUptodate(page)) {
822                 if (inode->i_private) {
823                         struct shmem_falloc *shmem_falloc;
824                         spin_lock(&inode->i_lock);
825                         shmem_falloc = inode->i_private;
826                         if (shmem_falloc &&
827                             index >= shmem_falloc->start &&
828                             index < shmem_falloc->next)
829                                 shmem_falloc->nr_unswapped++;
830                         else
831                                 shmem_falloc = NULL;
832                         spin_unlock(&inode->i_lock);
833                         if (shmem_falloc)
834                                 goto redirty;
835                 }
836                 clear_highpage(page);
837                 flush_dcache_page(page);
838                 SetPageUptodate(page);
839         }
840
841         swap = get_swap_page();
842         if (!swap.val)
843                 goto redirty;
844
845         /*
846          * Add inode to shmem_unuse()'s list of swapped-out inodes,
847          * if it's not already there.  Do it now before the page is
848          * moved to swap cache, when its pagelock no longer protects
849          * the inode from eviction.  But don't unlock the mutex until
850          * we've incremented swapped, because shmem_unuse_inode() will
851          * prune a !swapped inode from the swaplist under this mutex.
852          */
853         mutex_lock(&shmem_swaplist_mutex);
854         if (list_empty(&info->swaplist))
855                 list_add_tail(&info->swaplist, &shmem_swaplist);
856
857         if (add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
858                 swap_shmem_alloc(swap);
859                 shmem_delete_from_page_cache(page, swp_to_radix_entry(swap));
860
861                 spin_lock(&info->lock);
862                 info->swapped++;
863                 shmem_recalc_inode(inode);
864                 spin_unlock(&info->lock);
865
866                 mutex_unlock(&shmem_swaplist_mutex);
867                 BUG_ON(page_mapped(page));
868                 swap_writepage(page, wbc);
869                 return 0;
870         }
871
872         mutex_unlock(&shmem_swaplist_mutex);
873         swapcache_free(swap, NULL);
874 redirty:
875         set_page_dirty(page);
876         if (wbc->for_reclaim)
877                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
878         unlock_page(page);
879         return 0;
880 }
881
882 #ifdef CONFIG_NUMA
883 #ifdef CONFIG_TMPFS
884 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
885 {
886         char buffer[64];
887
888         if (!mpol || mpol->mode == MPOL_DEFAULT)
889                 return;         /* show nothing */
890
891         mpol_to_str(buffer, sizeof(buffer), mpol);
892
893         seq_printf(seq, ",mpol=%s", buffer);
894 }
895
896 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
897 {
898         struct mempolicy *mpol = NULL;
899         if (sbinfo->mpol) {
900                 spin_lock(&sbinfo->stat_lock);  /* prevent replace/use races */
901                 mpol = sbinfo->mpol;
902                 mpol_get(mpol);
903                 spin_unlock(&sbinfo->stat_lock);
904         }
905         return mpol;
906 }
907 #endif /* CONFIG_TMPFS */
908
909 static struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
910                         struct shmem_inode_info *info, pgoff_t index)
911 {
912         struct vm_area_struct pvma;
913         struct page *page;
914
915         /* Create a pseudo vma that just contains the policy */
916         pvma.vm_start = 0;
917         /* Bias interleave by inode number to distribute better across nodes */
918         pvma.vm_pgoff = index + info->vfs_inode.i_ino;
919         pvma.vm_ops = NULL;
920         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, index);
921
922         page = swapin_readahead(swap, gfp, &pvma, 0);
923
924         /* Drop reference taken by mpol_shared_policy_lookup() */
925         mpol_cond_put(pvma.vm_policy);
926
927         return page;
928 }
929
930 static struct page *shmem_alloc_page(gfp_t gfp,
931                         struct shmem_inode_info *info, pgoff_t index)
932 {
933         struct vm_area_struct pvma;
934         struct page *page;
935
936         /* Create a pseudo vma that just contains the policy */
937         pvma.vm_start = 0;
938         /* Bias interleave by inode number to distribute better across nodes */
939         pvma.vm_pgoff = index + info->vfs_inode.i_ino;
940         pvma.vm_ops = NULL;
941         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, index);
942
943         page = alloc_page_vma(gfp, &pvma, 0);
944
945         /* Drop reference taken by mpol_shared_policy_lookup() */
946         mpol_cond_put(pvma.vm_policy);
947
948         return page;
949 }
950 #else /* !CONFIG_NUMA */
951 #ifdef CONFIG_TMPFS
952 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
953 {
954 }
955 #endif /* CONFIG_TMPFS */
956
957 static inline struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
958                         struct shmem_inode_info *info, pgoff_t index)
959 {
960         return swapin_readahead(swap, gfp, NULL, 0);
961 }
962
963 static inline struct page *shmem_alloc_page(gfp_t gfp,
964                         struct shmem_inode_info *info, pgoff_t index)
965 {
966         return alloc_page(gfp);
967 }
968 #endif /* CONFIG_NUMA */
969
970 #if !defined(CONFIG_NUMA) || !defined(CONFIG_TMPFS)
971 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
972 {
973         return NULL;
974 }
975 #endif
976
977 /*
978  * When a page is moved from swapcache to shmem filecache (either by the
979  * usual swapin of shmem_getpage_gfp(), or by the less common swapoff of
980  * shmem_unuse_inode()), it may have been read in earlier from swap, in
981  * ignorance of the mapping it belongs to.  If that mapping has special
982  * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
983  * we may need to copy to a suitable page before moving to filecache.
984  *
985  * In a future release, this may well be extended to respect cpuset and
986  * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
987  * but for now it is a simple matter of zone.
988  */
989 static bool shmem_should_replace_page(struct page *page, gfp_t gfp)
990 {
991         return page_zonenum(page) > gfp_zone(gfp);
992 }
993
994 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
995                                 struct shmem_inode_info *info, pgoff_t index)
996 {
997         struct page *oldpage, *newpage;
998         struct address_space *swap_mapping;
999         pgoff_t swap_index;
1000         int error;
1001
1002         oldpage = *pagep;
1003         swap_index = page_private(oldpage);
1004         swap_mapping = page_mapping(oldpage);
1005
1006         /*
1007          * We have arrived here because our zones are constrained, so don't
1008          * limit chance of success by further cpuset and node constraints.
1009          */
1010         gfp &= ~GFP_CONSTRAINT_MASK;
1011         newpage = shmem_alloc_page(gfp, info, index);
1012         if (!newpage)
1013                 return -ENOMEM;
1014
1015         page_cache_get(newpage);
1016         copy_highpage(newpage, oldpage);
1017         flush_dcache_page(newpage);
1018
1019         __set_page_locked(newpage);
1020         SetPageUptodate(newpage);
1021         SetPageSwapBacked(newpage);
1022         set_page_private(newpage, swap_index);
1023         SetPageSwapCache(newpage);
1024
1025         /*
1026          * Our caller will very soon move newpage out of swapcache, but it's
1027          * a nice clean interface for us to replace oldpage by newpage there.
1028          */
1029         spin_lock_irq(&swap_mapping->tree_lock);
1030         error = shmem_radix_tree_replace(swap_mapping, swap_index, oldpage,
1031                                                                    newpage);
1032         if (!error) {
1033                 __inc_zone_page_state(newpage, NR_FILE_PAGES);
1034                 __dec_zone_page_state(oldpage, NR_FILE_PAGES);
1035         }
1036         spin_unlock_irq(&swap_mapping->tree_lock);
1037
1038         if (unlikely(error)) {
1039                 /*
1040                  * Is this possible?  I think not, now that our callers check
1041                  * both PageSwapCache and page_private after getting page lock;
1042                  * but be defensive.  Reverse old to newpage for clear and free.
1043                  */
1044                 oldpage = newpage;
1045         } else {
1046                 mem_cgroup_replace_page_cache(oldpage, newpage);
1047                 lru_cache_add_anon(newpage);
1048                 *pagep = newpage;
1049         }
1050
1051         ClearPageSwapCache(oldpage);
1052         set_page_private(oldpage, 0);
1053
1054         unlock_page(oldpage);
1055         page_cache_release(oldpage);
1056         page_cache_release(oldpage);
1057         return error;
1058 }
1059
1060 /*
1061  * shmem_getpage_gfp - find page in cache, or get from swap, or allocate
1062  *
1063  * If we allocate a new one we do not mark it dirty. That's up to the
1064  * vm. If we swap it in we mark it dirty since we also free the swap
1065  * entry since a page cannot live in both the swap and page cache
1066  */
1067 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
1068         struct page **pagep, enum sgp_type sgp, gfp_t gfp, int *fault_type)
1069 {
1070         struct address_space *mapping = inode->i_mapping;
1071         struct shmem_inode_info *info;
1072         struct shmem_sb_info *sbinfo;
1073         struct page *page;
1074         swp_entry_t swap;
1075         int error;
1076         int once = 0;
1077         int alloced = 0;
1078
1079         if (index > (MAX_LFS_FILESIZE >> PAGE_CACHE_SHIFT))
1080                 return -EFBIG;
1081 repeat:
1082         swap.val = 0;
1083         page = find_lock_page(mapping, index);
1084         if (radix_tree_exceptional_entry(page)) {
1085                 swap = radix_to_swp_entry(page);
1086                 page = NULL;
1087         }
1088
1089         if (sgp != SGP_WRITE && sgp != SGP_FALLOC &&
1090             ((loff_t)index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
1091                 error = -EINVAL;
1092                 goto failed;
1093         }
1094
1095         /* fallocated page? */
1096         if (page && !PageUptodate(page)) {
1097                 if (sgp != SGP_READ)
1098                         goto clear;
1099                 unlock_page(page);
1100                 page_cache_release(page);
1101                 page = NULL;
1102         }
1103         if (page || (sgp == SGP_READ && !swap.val)) {
1104                 *pagep = page;
1105                 return 0;
1106         }
1107
1108         /*
1109          * Fast cache lookup did not find it:
1110          * bring it back from swap or allocate.
1111          */
1112         info = SHMEM_I(inode);
1113         sbinfo = SHMEM_SB(inode->i_sb);
1114
1115         if (swap.val) {
1116                 /* Look it up and read it in.. */
1117                 page = lookup_swap_cache(swap);
1118                 if (!page) {
1119                         /* here we actually do the io */
1120                         if (fault_type)
1121                                 *fault_type |= VM_FAULT_MAJOR;
1122                         page = shmem_swapin(swap, gfp, info, index);
1123                         if (!page) {
1124                                 error = -ENOMEM;
1125                                 goto failed;
1126                         }
1127                 }
1128
1129                 /* We have to do this with page locked to prevent races */
1130                 lock_page(page);
1131                 if (!PageSwapCache(page) || page_private(page) != swap.val ||
1132                     !shmem_confirm_swap(mapping, index, swap)) {
1133                         error = -EEXIST;        /* try again */
1134                         goto unlock;
1135                 }
1136                 if (!PageUptodate(page)) {
1137                         error = -EIO;
1138                         goto failed;
1139                 }
1140                 wait_on_page_writeback(page);
1141
1142                 if (shmem_should_replace_page(page, gfp)) {
1143                         error = shmem_replace_page(&page, gfp, info, index);
1144                         if (error)
1145                                 goto failed;
1146                 }
1147
1148                 error = mem_cgroup_cache_charge(page, current->mm,
1149                                                 gfp & GFP_RECLAIM_MASK);
1150                 if (!error) {
1151                         error = shmem_add_to_page_cache(page, mapping, index,
1152                                                 gfp, swp_to_radix_entry(swap));
1153                         /*
1154                          * We already confirmed swap under page lock, and make
1155                          * no memory allocation here, so usually no possibility
1156                          * of error; but free_swap_and_cache() only trylocks a
1157                          * page, so it is just possible that the entry has been
1158                          * truncated or holepunched since swap was confirmed.
1159                          * shmem_undo_range() will have done some of the
1160                          * unaccounting, now delete_from_swap_cache() will do
1161                          * the rest (including mem_cgroup_uncharge_swapcache).
1162                          * Reset swap.val? No, leave it so "failed" goes back to
1163                          * "repeat": reading a hole and writing should succeed.
1164                          */
1165                         if (error)
1166                                 delete_from_swap_cache(page);
1167                 }
1168                 if (error)
1169                         goto failed;
1170
1171                 spin_lock(&info->lock);
1172                 info->swapped--;
1173                 shmem_recalc_inode(inode);
1174                 spin_unlock(&info->lock);
1175
1176                 delete_from_swap_cache(page);
1177                 set_page_dirty(page);
1178                 swap_free(swap);
1179
1180         } else {
1181                 if (shmem_acct_block(info->flags)) {
1182                         error = -ENOSPC;
1183                         goto failed;
1184                 }
1185                 if (sbinfo->max_blocks) {
1186                         if (percpu_counter_compare(&sbinfo->used_blocks,
1187                                                 sbinfo->max_blocks) >= 0) {
1188                                 error = -ENOSPC;
1189                                 goto unacct;
1190                         }
1191                         percpu_counter_inc(&sbinfo->used_blocks);
1192                 }
1193
1194                 page = shmem_alloc_page(gfp, info, index);
1195                 if (!page) {
1196                         error = -ENOMEM;
1197                         goto decused;
1198                 }
1199
1200                 SetPageSwapBacked(page);
1201                 __set_page_locked(page);
1202                 error = mem_cgroup_cache_charge(page, current->mm,
1203                                                 gfp & GFP_RECLAIM_MASK);
1204                 if (error)
1205                         goto decused;
1206                 error = radix_tree_preload(gfp & GFP_RECLAIM_MASK);
1207                 if (!error) {
1208                         error = shmem_add_to_page_cache(page, mapping, index,
1209                                                         gfp, NULL);
1210                         radix_tree_preload_end();
1211                 }
1212                 if (error) {
1213                         mem_cgroup_uncharge_cache_page(page);
1214                         goto decused;
1215                 }
1216                 lru_cache_add_anon(page);
1217
1218                 spin_lock(&info->lock);
1219                 info->alloced++;
1220                 inode->i_blocks += BLOCKS_PER_PAGE;
1221                 shmem_recalc_inode(inode);
1222                 spin_unlock(&info->lock);
1223                 alloced = true;
1224
1225                 /*
1226                  * Let SGP_FALLOC use the SGP_WRITE optimization on a new page.
1227                  */
1228                 if (sgp == SGP_FALLOC)
1229                         sgp = SGP_WRITE;
1230 clear:
1231                 /*
1232                  * Let SGP_WRITE caller clear ends if write does not fill page;
1233                  * but SGP_FALLOC on a page fallocated earlier must initialize
1234                  * it now, lest undo on failure cancel our earlier guarantee.
1235                  */
1236                 if (sgp != SGP_WRITE) {
1237                         clear_highpage(page);
1238                         flush_dcache_page(page);
1239                         SetPageUptodate(page);
1240                 }
1241                 if (sgp == SGP_DIRTY)
1242                         set_page_dirty(page);
1243         }
1244
1245         /* Perhaps the file has been truncated since we checked */
1246         if (sgp != SGP_WRITE && sgp != SGP_FALLOC &&
1247             ((loff_t)index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
1248                 error = -EINVAL;
1249                 if (alloced)
1250                         goto trunc;
1251                 else
1252                         goto failed;
1253         }
1254         *pagep = page;
1255         return 0;
1256
1257         /*
1258          * Error recovery.
1259          */
1260 trunc:
1261         info = SHMEM_I(inode);
1262         ClearPageDirty(page);
1263         delete_from_page_cache(page);
1264         spin_lock(&info->lock);
1265         info->alloced--;
1266         inode->i_blocks -= BLOCKS_PER_PAGE;
1267         spin_unlock(&info->lock);
1268 decused:
1269         sbinfo = SHMEM_SB(inode->i_sb);
1270         if (sbinfo->max_blocks)
1271                 percpu_counter_add(&sbinfo->used_blocks, -1);
1272 unacct:
1273         shmem_unacct_blocks(info->flags, 1);
1274 failed:
1275         if (swap.val && error != -EINVAL &&
1276             !shmem_confirm_swap(mapping, index, swap))
1277                 error = -EEXIST;
1278 unlock:
1279         if (page) {
1280                 unlock_page(page);
1281                 page_cache_release(page);
1282         }
1283         if (error == -ENOSPC && !once++) {
1284                 info = SHMEM_I(inode);
1285                 spin_lock(&info->lock);
1286                 shmem_recalc_inode(inode);
1287                 spin_unlock(&info->lock);
1288                 goto repeat;
1289         }
1290         if (error == -EEXIST)   /* from above or from radix_tree_insert */
1291                 goto repeat;
1292         return error;
1293 }
1294
1295 static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1296 {
1297         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1298         int error;
1299         int ret = VM_FAULT_LOCKED;
1300
1301         error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_CACHE, &ret);
1302         if (error)
1303                 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1304
1305         if (ret & VM_FAULT_MAJOR) {
1306                 count_vm_event(PGMAJFAULT);
1307                 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1308         }
1309         return ret;
1310 }
1311
1312 #ifdef CONFIG_NUMA
1313 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
1314 {
1315         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1316         return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
1317 }
1318
1319 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
1320                                           unsigned long addr)
1321 {
1322         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1323         pgoff_t index;
1324
1325         index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1326         return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
1327 }
1328 #endif
1329
1330 int shmem_lock(struct file *file, int lock, struct user_struct *user)
1331 {
1332         struct inode *inode = file->f_path.dentry->d_inode;
1333         struct shmem_inode_info *info = SHMEM_I(inode);
1334         int retval = -ENOMEM;
1335
1336         spin_lock(&info->lock);
1337         if (lock && !(info->flags & VM_LOCKED)) {
1338                 if (!user_shm_lock(inode->i_size, user))
1339                         goto out_nomem;
1340                 info->flags |= VM_LOCKED;
1341                 mapping_set_unevictable(file->f_mapping);
1342         }
1343         if (!lock && (info->flags & VM_LOCKED) && user) {
1344                 user_shm_unlock(inode->i_size, user);
1345                 info->flags &= ~VM_LOCKED;
1346                 mapping_clear_unevictable(file->f_mapping);
1347         }
1348         retval = 0;
1349
1350 out_nomem:
1351         spin_unlock(&info->lock);
1352         return retval;
1353 }
1354
1355 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1356 {
1357         file_accessed(file);
1358         vma->vm_ops = &shmem_vm_ops;
1359         return 0;
1360 }
1361
1362 static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
1363                                      umode_t mode, dev_t dev, unsigned long flags)
1364 {
1365         struct inode *inode;
1366         struct shmem_inode_info *info;
1367         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1368
1369         if (shmem_reserve_inode(sb))
1370                 return NULL;
1371
1372         inode = new_inode(sb);
1373         if (inode) {
1374                 inode->i_ino = get_next_ino();
1375                 inode_init_owner(inode, dir, mode);
1376                 inode->i_blocks = 0;
1377                 inode->i_mapping->backing_dev_info = &shmem_backing_dev_info;
1378                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1379                 inode->i_generation = get_seconds();
1380                 info = SHMEM_I(inode);
1381                 memset(info, 0, (char *)inode - (char *)info);
1382                 spin_lock_init(&info->lock);
1383                 info->flags = flags & VM_NORESERVE;
1384                 INIT_LIST_HEAD(&info->swaplist);
1385                 simple_xattrs_init(&info->xattrs);
1386                 cache_no_acl(inode);
1387
1388                 switch (mode & S_IFMT) {
1389                 default:
1390                         inode->i_op = &shmem_special_inode_operations;
1391                         init_special_inode(inode, mode, dev);
1392                         break;
1393                 case S_IFREG:
1394                         inode->i_mapping->a_ops = &shmem_aops;
1395                         inode->i_op = &shmem_inode_operations;
1396                         inode->i_fop = &shmem_file_operations;
1397                         mpol_shared_policy_init(&info->policy,
1398                                                  shmem_get_sbmpol(sbinfo));
1399                         break;
1400                 case S_IFDIR:
1401                         inc_nlink(inode);
1402                         /* Some things misbehave if size == 0 on a directory */
1403                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
1404                         inode->i_op = &shmem_dir_inode_operations;
1405                         inode->i_fop = &simple_dir_operations;
1406                         break;
1407                 case S_IFLNK:
1408                         /*
1409                          * Must not load anything in the rbtree,
1410                          * mpol_free_shared_policy will not be called.
1411                          */
1412                         mpol_shared_policy_init(&info->policy, NULL);
1413                         break;
1414                 }
1415         } else
1416                 shmem_free_inode(sb);
1417         return inode;
1418 }
1419
1420 #ifdef CONFIG_TMPFS
1421 static const struct inode_operations shmem_symlink_inode_operations;
1422 static const struct inode_operations shmem_short_symlink_operations;
1423
1424 #ifdef CONFIG_TMPFS_XATTR
1425 static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
1426 #else
1427 #define shmem_initxattrs NULL
1428 #endif
1429
1430 static int
1431 shmem_write_begin(struct file *file, struct address_space *mapping,
1432                         loff_t pos, unsigned len, unsigned flags,
1433                         struct page **pagep, void **fsdata)
1434 {
1435         struct inode *inode = mapping->host;
1436         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1437         return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1438 }
1439
1440 static int
1441 shmem_write_end(struct file *file, struct address_space *mapping,
1442                         loff_t pos, unsigned len, unsigned copied,
1443                         struct page *page, void *fsdata)
1444 {
1445         struct inode *inode = mapping->host;
1446
1447         if (pos + copied > inode->i_size)
1448                 i_size_write(inode, pos + copied);
1449
1450         if (!PageUptodate(page)) {
1451                 if (copied < PAGE_CACHE_SIZE) {
1452                         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1453                         zero_user_segments(page, 0, from,
1454                                         from + copied, PAGE_CACHE_SIZE);
1455                 }
1456                 SetPageUptodate(page);
1457         }
1458         set_page_dirty(page);
1459         unlock_page(page);
1460         page_cache_release(page);
1461
1462         return copied;
1463 }
1464
1465 static void do_shmem_file_read(struct file *filp, loff_t *ppos, read_descriptor_t *desc, read_actor_t actor)
1466 {
1467         struct inode *inode = filp->f_path.dentry->d_inode;
1468         struct address_space *mapping = inode->i_mapping;
1469         pgoff_t index;
1470         unsigned long offset;
1471         enum sgp_type sgp = SGP_READ;
1472
1473         /*
1474          * Might this read be for a stacking filesystem?  Then when reading
1475          * holes of a sparse file, we actually need to allocate those pages,
1476          * and even mark them dirty, so it cannot exceed the max_blocks limit.
1477          */
1478         if (segment_eq(get_fs(), KERNEL_DS))
1479                 sgp = SGP_DIRTY;
1480
1481         index = *ppos >> PAGE_CACHE_SHIFT;
1482         offset = *ppos & ~PAGE_CACHE_MASK;
1483
1484         for (;;) {
1485                 struct page *page = NULL;
1486                 pgoff_t end_index;
1487                 unsigned long nr, ret;
1488                 loff_t i_size = i_size_read(inode);
1489
1490                 end_index = i_size >> PAGE_CACHE_SHIFT;
1491                 if (index > end_index)
1492                         break;
1493                 if (index == end_index) {
1494                         nr = i_size & ~PAGE_CACHE_MASK;
1495                         if (nr <= offset)
1496                                 break;
1497                 }
1498
1499                 desc->error = shmem_getpage(inode, index, &page, sgp, NULL);
1500                 if (desc->error) {
1501                         if (desc->error == -EINVAL)
1502                                 desc->error = 0;
1503                         break;
1504                 }
1505                 if (page)
1506                         unlock_page(page);
1507
1508                 /*
1509                  * We must evaluate after, since reads (unlike writes)
1510                  * are called without i_mutex protection against truncate
1511                  */
1512                 nr = PAGE_CACHE_SIZE;
1513                 i_size = i_size_read(inode);
1514                 end_index = i_size >> PAGE_CACHE_SHIFT;
1515                 if (index == end_index) {
1516                         nr = i_size & ~PAGE_CACHE_MASK;
1517                         if (nr <= offset) {
1518                                 if (page)
1519                                         page_cache_release(page);
1520                                 break;
1521                         }
1522                 }
1523                 nr -= offset;
1524
1525                 if (page) {
1526                         /*
1527                          * If users can be writing to this page using arbitrary
1528                          * virtual addresses, take care about potential aliasing
1529                          * before reading the page on the kernel side.
1530                          */
1531                         if (mapping_writably_mapped(mapping))
1532                                 flush_dcache_page(page);
1533                         /*
1534                          * Mark the page accessed if we read the beginning.
1535                          */
1536                         if (!offset)
1537                                 mark_page_accessed(page);
1538                 } else {
1539                         page = ZERO_PAGE(0);
1540                         page_cache_get(page);
1541                 }
1542
1543                 /*
1544                  * Ok, we have the page, and it's up-to-date, so
1545                  * now we can copy it to user space...
1546                  *
1547                  * The actor routine returns how many bytes were actually used..
1548                  * NOTE! This may not be the same as how much of a user buffer
1549                  * we filled up (we may be padding etc), so we can only update
1550                  * "pos" here (the actor routine has to update the user buffer
1551                  * pointers and the remaining count).
1552                  */
1553                 ret = actor(desc, page, offset, nr);
1554                 offset += ret;
1555                 index += offset >> PAGE_CACHE_SHIFT;
1556                 offset &= ~PAGE_CACHE_MASK;
1557
1558                 page_cache_release(page);
1559                 if (ret != nr || !desc->count)
1560                         break;
1561
1562                 cond_resched();
1563         }
1564
1565         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1566         file_accessed(filp);
1567 }
1568
1569 static ssize_t shmem_file_aio_read(struct kiocb *iocb,
1570                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
1571 {
1572         struct file *filp = iocb->ki_filp;
1573         ssize_t retval;
1574         unsigned long seg;
1575         size_t count;
1576         loff_t *ppos = &iocb->ki_pos;
1577
1578         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1579         if (retval)
1580                 return retval;
1581
1582         for (seg = 0; seg < nr_segs; seg++) {
1583                 read_descriptor_t desc;
1584
1585                 desc.written = 0;
1586                 desc.arg.buf = iov[seg].iov_base;
1587                 desc.count = iov[seg].iov_len;
1588                 if (desc.count == 0)
1589                         continue;
1590                 desc.error = 0;
1591                 do_shmem_file_read(filp, ppos, &desc, file_read_actor);
1592                 retval += desc.written;
1593                 if (desc.error) {
1594                         retval = retval ?: desc.error;
1595                         break;
1596                 }
1597                 if (desc.count > 0)
1598                         break;
1599         }
1600         return retval;
1601 }
1602
1603 static ssize_t shmem_file_splice_read(struct file *in, loff_t *ppos,
1604                                 struct pipe_inode_info *pipe, size_t len,
1605                                 unsigned int flags)
1606 {
1607         struct address_space *mapping = in->f_mapping;
1608         struct inode *inode = mapping->host;
1609         unsigned int loff, nr_pages, req_pages;
1610         struct page *pages[PIPE_DEF_BUFFERS];
1611         struct partial_page partial[PIPE_DEF_BUFFERS];
1612         struct page *page;
1613         pgoff_t index, end_index;
1614         loff_t isize, left;
1615         int error, page_nr;
1616         struct splice_pipe_desc spd = {
1617                 .pages = pages,
1618                 .partial = partial,
1619                 .nr_pages_max = PIPE_DEF_BUFFERS,
1620                 .flags = flags,
1621                 .ops = &page_cache_pipe_buf_ops,
1622                 .spd_release = spd_release_page,
1623         };
1624
1625         isize = i_size_read(inode);
1626         if (unlikely(*ppos >= isize))
1627                 return 0;
1628
1629         left = isize - *ppos;
1630         if (unlikely(left < len))
1631                 len = left;
1632
1633         if (splice_grow_spd(pipe, &spd))
1634                 return -ENOMEM;
1635
1636         index = *ppos >> PAGE_CACHE_SHIFT;
1637         loff = *ppos & ~PAGE_CACHE_MASK;
1638         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1639         nr_pages = min(req_pages, pipe->buffers);
1640
1641         spd.nr_pages = find_get_pages_contig(mapping, index,
1642                                                 nr_pages, spd.pages);
1643         index += spd.nr_pages;
1644         error = 0;
1645
1646         while (spd.nr_pages < nr_pages) {
1647                 error = shmem_getpage(inode, index, &page, SGP_CACHE, NULL);
1648                 if (error)
1649                         break;
1650                 unlock_page(page);
1651                 spd.pages[spd.nr_pages++] = page;
1652                 index++;
1653         }
1654
1655         index = *ppos >> PAGE_CACHE_SHIFT;
1656         nr_pages = spd.nr_pages;
1657         spd.nr_pages = 0;
1658
1659         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
1660                 unsigned int this_len;
1661
1662                 if (!len)
1663                         break;
1664
1665                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
1666                 page = spd.pages[page_nr];
1667
1668                 if (!PageUptodate(page) || page->mapping != mapping) {
1669                         error = shmem_getpage(inode, index, &page,
1670                                                         SGP_CACHE, NULL);
1671                         if (error)
1672                                 break;
1673                         unlock_page(page);
1674                         page_cache_release(spd.pages[page_nr]);
1675                         spd.pages[page_nr] = page;
1676                 }
1677
1678                 isize = i_size_read(inode);
1679                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1680                 if (unlikely(!isize || index > end_index))
1681                         break;
1682
1683                 if (end_index == index) {
1684                         unsigned int plen;
1685
1686                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1687                         if (plen <= loff)
1688                                 break;
1689
1690                         this_len = min(this_len, plen - loff);
1691                         len = this_len;
1692                 }
1693
1694                 spd.partial[page_nr].offset = loff;
1695                 spd.partial[page_nr].len = this_len;
1696                 len -= this_len;
1697                 loff = 0;
1698                 spd.nr_pages++;
1699                 index++;
1700         }
1701
1702         while (page_nr < nr_pages)
1703                 page_cache_release(spd.pages[page_nr++]);
1704
1705         if (spd.nr_pages)
1706                 error = splice_to_pipe(pipe, &spd);
1707
1708         splice_shrink_spd(&spd);
1709
1710         if (error > 0) {
1711                 *ppos += error;
1712                 file_accessed(in);
1713         }
1714         return error;
1715 }
1716
1717 /*
1718  * llseek SEEK_DATA or SEEK_HOLE through the radix_tree.
1719  */
1720 static pgoff_t shmem_seek_hole_data(struct address_space *mapping,
1721                                     pgoff_t index, pgoff_t end, int whence)
1722 {
1723         struct page *page;
1724         struct pagevec pvec;
1725         pgoff_t indices[PAGEVEC_SIZE];
1726         bool done = false;
1727         int i;
1728
1729         pagevec_init(&pvec, 0);
1730         pvec.nr = 1;            /* start small: we may be there already */
1731         while (!done) {
1732                 pvec.nr = shmem_find_get_pages_and_swap(mapping, index,
1733                                         pvec.nr, pvec.pages, indices);
1734                 if (!pvec.nr) {
1735                         if (whence == SEEK_DATA)
1736                                 index = end;
1737                         break;
1738                 }
1739                 for (i = 0; i < pvec.nr; i++, index++) {
1740                         if (index < indices[i]) {
1741                                 if (whence == SEEK_HOLE) {
1742                                         done = true;
1743                                         break;
1744                                 }
1745                                 index = indices[i];
1746                         }
1747                         page = pvec.pages[i];
1748                         if (page && !radix_tree_exceptional_entry(page)) {
1749                                 if (!PageUptodate(page))
1750                                         page = NULL;
1751                         }
1752                         if (index >= end ||
1753                             (page && whence == SEEK_DATA) ||
1754                             (!page && whence == SEEK_HOLE)) {
1755                                 done = true;
1756                                 break;
1757                         }
1758                 }
1759                 shmem_deswap_pagevec(&pvec);
1760                 pagevec_release(&pvec);
1761                 pvec.nr = PAGEVEC_SIZE;
1762                 cond_resched();
1763         }
1764         return index;
1765 }
1766
1767 static loff_t shmem_file_llseek(struct file *file, loff_t offset, int whence)
1768 {
1769         struct address_space *mapping = file->f_mapping;
1770         struct inode *inode = mapping->host;
1771         pgoff_t start, end;
1772         loff_t new_offset;
1773
1774         if (whence != SEEK_DATA && whence != SEEK_HOLE)
1775                 return generic_file_llseek_size(file, offset, whence,
1776                                         MAX_LFS_FILESIZE, i_size_read(inode));
1777         mutex_lock(&inode->i_mutex);
1778         /* We're holding i_mutex so we can access i_size directly */
1779
1780         if (offset < 0)
1781                 offset = -EINVAL;
1782         else if (offset >= inode->i_size)
1783                 offset = -ENXIO;
1784         else {
1785                 start = offset >> PAGE_CACHE_SHIFT;
1786                 end = (inode->i_size + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1787                 new_offset = shmem_seek_hole_data(mapping, start, end, whence);
1788                 new_offset <<= PAGE_CACHE_SHIFT;
1789                 if (new_offset > offset) {
1790                         if (new_offset < inode->i_size)
1791                                 offset = new_offset;
1792                         else if (whence == SEEK_DATA)
1793                                 offset = -ENXIO;
1794                         else
1795                                 offset = inode->i_size;
1796                 }
1797         }
1798
1799         if (offset >= 0 && offset != file->f_pos) {
1800                 file->f_pos = offset;
1801                 file->f_version = 0;
1802         }
1803         mutex_unlock(&inode->i_mutex);
1804         return offset;
1805 }
1806
1807 static long shmem_fallocate(struct file *file, int mode, loff_t offset,
1808                                                          loff_t len)
1809 {
1810         struct inode *inode = file->f_path.dentry->d_inode;
1811         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1812         struct shmem_falloc shmem_falloc;
1813         pgoff_t start, index, end;
1814         int error;
1815
1816         mutex_lock(&inode->i_mutex);
1817
1818         if (mode & FALLOC_FL_PUNCH_HOLE) {
1819                 struct address_space *mapping = file->f_mapping;
1820                 loff_t unmap_start = round_up(offset, PAGE_SIZE);
1821                 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
1822
1823                 if ((u64)unmap_end > (u64)unmap_start)
1824                         unmap_mapping_range(mapping, unmap_start,
1825                                             1 + unmap_end - unmap_start, 0);
1826                 shmem_truncate_range(inode, offset, offset + len - 1);
1827                 /* No need to unmap again: hole-punching leaves COWed pages */
1828                 error = 0;
1829                 goto out;
1830         }
1831
1832         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
1833         error = inode_newsize_ok(inode, offset + len);
1834         if (error)
1835                 goto out;
1836
1837         start = offset >> PAGE_CACHE_SHIFT;
1838         end = (offset + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1839         /* Try to avoid a swapstorm if len is impossible to satisfy */
1840         if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
1841                 error = -ENOSPC;
1842                 goto out;
1843         }
1844
1845         shmem_falloc.start = start;
1846         shmem_falloc.next  = start;
1847         shmem_falloc.nr_falloced = 0;
1848         shmem_falloc.nr_unswapped = 0;
1849         spin_lock(&inode->i_lock);
1850         inode->i_private = &shmem_falloc;
1851         spin_unlock(&inode->i_lock);
1852
1853         for (index = start; index < end; index++) {
1854                 struct page *page;
1855
1856                 /*
1857                  * Good, the fallocate(2) manpage permits EINTR: we may have
1858                  * been interrupted because we are using up too much memory.
1859                  */
1860                 if (signal_pending(current))
1861                         error = -EINTR;
1862                 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
1863                         error = -ENOMEM;
1864                 else
1865                         error = shmem_getpage(inode, index, &page, SGP_FALLOC,
1866                                                                         NULL);
1867                 if (error) {
1868                         /* Remove the !PageUptodate pages we added */
1869                         shmem_undo_range(inode,
1870                                 (loff_t)start << PAGE_CACHE_SHIFT,
1871                                 (loff_t)index << PAGE_CACHE_SHIFT, true);
1872                         goto undone;
1873                 }
1874
1875                 /*
1876                  * Inform shmem_writepage() how far we have reached.
1877                  * No need for lock or barrier: we have the page lock.
1878                  */
1879                 shmem_falloc.next++;
1880                 if (!PageUptodate(page))
1881                         shmem_falloc.nr_falloced++;
1882
1883                 /*
1884                  * If !PageUptodate, leave it that way so that freeable pages
1885                  * can be recognized if we need to rollback on error later.
1886                  * But set_page_dirty so that memory pressure will swap rather
1887                  * than free the pages we are allocating (and SGP_CACHE pages
1888                  * might still be clean: we now need to mark those dirty too).
1889                  */
1890                 set_page_dirty(page);
1891                 unlock_page(page);
1892                 page_cache_release(page);
1893                 cond_resched();
1894         }
1895
1896         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
1897                 i_size_write(inode, offset + len);
1898         inode->i_ctime = CURRENT_TIME;
1899 undone:
1900         spin_lock(&inode->i_lock);
1901         inode->i_private = NULL;
1902         spin_unlock(&inode->i_lock);
1903 out:
1904         mutex_unlock(&inode->i_mutex);
1905         return error;
1906 }
1907
1908 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
1909 {
1910         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
1911
1912         buf->f_type = TMPFS_MAGIC;
1913         buf->f_bsize = PAGE_CACHE_SIZE;
1914         buf->f_namelen = NAME_MAX;
1915         if (sbinfo->max_blocks) {
1916                 buf->f_blocks = sbinfo->max_blocks;
1917                 buf->f_bavail =
1918                 buf->f_bfree  = sbinfo->max_blocks -
1919                                 percpu_counter_sum(&sbinfo->used_blocks);
1920         }
1921         if (sbinfo->max_inodes) {
1922                 buf->f_files = sbinfo->max_inodes;
1923                 buf->f_ffree = sbinfo->free_inodes;
1924         }
1925         /* else leave those fields 0 like simple_statfs */
1926         return 0;
1927 }
1928
1929 /*
1930  * File creation. Allocate an inode, and we're done..
1931  */
1932 static int
1933 shmem_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
1934 {
1935         struct inode *inode;
1936         int error = -ENOSPC;
1937
1938         inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
1939         if (inode) {
1940                 error = security_inode_init_security(inode, dir,
1941                                                      &dentry->d_name,
1942                                                      shmem_initxattrs, NULL);
1943                 if (error) {
1944                         if (error != -EOPNOTSUPP) {
1945                                 iput(inode);
1946                                 return error;
1947                         }
1948                 }
1949 #ifdef CONFIG_TMPFS_POSIX_ACL
1950                 error = generic_acl_init(inode, dir);
1951                 if (error) {
1952                         iput(inode);
1953                         return error;
1954                 }
1955 #else
1956                 error = 0;
1957 #endif
1958                 dir->i_size += BOGO_DIRENT_SIZE;
1959                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1960                 d_instantiate(dentry, inode);
1961                 dget(dentry); /* Extra count - pin the dentry in core */
1962         }
1963         return error;
1964 }
1965
1966 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
1967 {
1968         int error;
1969
1970         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
1971                 return error;
1972         inc_nlink(dir);
1973         return 0;
1974 }
1975
1976 static int shmem_create(struct inode *dir, struct dentry *dentry, umode_t mode,
1977                 bool excl)
1978 {
1979         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
1980 }
1981
1982 /*
1983  * Link a file..
1984  */
1985 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
1986 {
1987         struct inode *inode = old_dentry->d_inode;
1988         int ret;
1989
1990         /*
1991          * No ordinary (disk based) filesystem counts links as inodes;
1992          * but each new link needs a new dentry, pinning lowmem, and
1993          * tmpfs dentries cannot be pruned until they are unlinked.
1994          */
1995         ret = shmem_reserve_inode(inode->i_sb);
1996         if (ret)
1997                 goto out;
1998
1999         dir->i_size += BOGO_DIRENT_SIZE;
2000         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2001         inc_nlink(inode);
2002         ihold(inode);   /* New dentry reference */
2003         dget(dentry);           /* Extra pinning count for the created dentry */
2004         d_instantiate(dentry, inode);
2005 out:
2006         return ret;
2007 }
2008
2009 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
2010 {
2011         struct inode *inode = dentry->d_inode;
2012
2013         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
2014                 shmem_free_inode(inode->i_sb);
2015
2016         dir->i_size -= BOGO_DIRENT_SIZE;
2017         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2018         drop_nlink(inode);
2019         dput(dentry);   /* Undo the count from "create" - this does all the work */
2020         return 0;
2021 }
2022
2023 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
2024 {
2025         if (!simple_empty(dentry))
2026                 return -ENOTEMPTY;
2027
2028         drop_nlink(dentry->d_inode);
2029         drop_nlink(dir);
2030         return shmem_unlink(dir, dentry);
2031 }
2032
2033 /*
2034  * The VFS layer already does all the dentry stuff for rename,
2035  * we just have to decrement the usage count for the target if
2036  * it exists so that the VFS layer correctly free's it when it
2037  * gets overwritten.
2038  */
2039 static int shmem_rename(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
2040 {
2041         struct inode *inode = old_dentry->d_inode;
2042         int they_are_dirs = S_ISDIR(inode->i_mode);
2043
2044         if (!simple_empty(new_dentry))
2045                 return -ENOTEMPTY;
2046
2047         if (new_dentry->d_inode) {
2048                 (void) shmem_unlink(new_dir, new_dentry);
2049                 if (they_are_dirs)
2050                         drop_nlink(old_dir);
2051         } else if (they_are_dirs) {
2052                 drop_nlink(old_dir);
2053                 inc_nlink(new_dir);
2054         }
2055
2056         old_dir->i_size -= BOGO_DIRENT_SIZE;
2057         new_dir->i_size += BOGO_DIRENT_SIZE;
2058         old_dir->i_ctime = old_dir->i_mtime =
2059         new_dir->i_ctime = new_dir->i_mtime =
2060         inode->i_ctime = CURRENT_TIME;
2061         return 0;
2062 }
2063
2064 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
2065 {
2066         int error;
2067         int len;
2068         struct inode *inode;
2069         struct page *page;
2070         char *kaddr;
2071         struct shmem_inode_info *info;
2072
2073         len = strlen(symname) + 1;
2074         if (len > PAGE_CACHE_SIZE)
2075                 return -ENAMETOOLONG;
2076
2077         inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0, VM_NORESERVE);
2078         if (!inode)
2079                 return -ENOSPC;
2080
2081         error = security_inode_init_security(inode, dir, &dentry->d_name,
2082                                              shmem_initxattrs, NULL);
2083         if (error) {
2084                 if (error != -EOPNOTSUPP) {
2085                         iput(inode);
2086                         return error;
2087                 }
2088                 error = 0;
2089         }
2090
2091         info = SHMEM_I(inode);
2092         inode->i_size = len-1;
2093         if (len <= SHORT_SYMLINK_LEN) {
2094                 info->symlink = kmemdup(symname, len, GFP_KERNEL);
2095                 if (!info->symlink) {
2096                         iput(inode);
2097                         return -ENOMEM;
2098                 }
2099                 inode->i_op = &shmem_short_symlink_operations;
2100         } else {
2101                 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
2102                 if (error) {
2103                         iput(inode);
2104                         return error;
2105                 }
2106                 inode->i_mapping->a_ops = &shmem_aops;
2107                 inode->i_op = &shmem_symlink_inode_operations;
2108                 kaddr = kmap_atomic(page);
2109                 memcpy(kaddr, symname, len);
2110                 kunmap_atomic(kaddr);
2111                 SetPageUptodate(page);
2112                 set_page_dirty(page);
2113                 unlock_page(page);
2114                 page_cache_release(page);
2115         }
2116         dir->i_size += BOGO_DIRENT_SIZE;
2117         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2118         d_instantiate(dentry, inode);
2119         dget(dentry);
2120         return 0;
2121 }
2122
2123 static void *shmem_follow_short_symlink(struct dentry *dentry, struct nameidata *nd)
2124 {
2125         nd_set_link(nd, SHMEM_I(dentry->d_inode)->symlink);
2126         return NULL;
2127 }
2128
2129 static void *shmem_follow_link(struct dentry *dentry, struct nameidata *nd)
2130 {
2131         struct page *page = NULL;
2132         int error = shmem_getpage(dentry->d_inode, 0, &page, SGP_READ, NULL);
2133         nd_set_link(nd, error ? ERR_PTR(error) : kmap(page));
2134         if (page)
2135                 unlock_page(page);
2136         return page;
2137 }
2138
2139 static void shmem_put_link(struct dentry *dentry, struct nameidata *nd, void *cookie)
2140 {
2141         if (!IS_ERR(nd_get_link(nd))) {
2142                 struct page *page = cookie;
2143                 kunmap(page);
2144                 mark_page_accessed(page);
2145                 page_cache_release(page);
2146         }
2147 }
2148
2149 #ifdef CONFIG_TMPFS_XATTR
2150 /*
2151  * Superblocks without xattr inode operations may get some security.* xattr
2152  * support from the LSM "for free". As soon as we have any other xattrs
2153  * like ACLs, we also need to implement the security.* handlers at
2154  * filesystem level, though.
2155  */
2156
2157 /*
2158  * Callback for security_inode_init_security() for acquiring xattrs.
2159  */
2160 static int shmem_initxattrs(struct inode *inode,
2161                             const struct xattr *xattr_array,
2162                             void *fs_info)
2163 {
2164         struct shmem_inode_info *info = SHMEM_I(inode);
2165         const struct xattr *xattr;
2166         struct simple_xattr *new_xattr;
2167         size_t len;
2168
2169         for (xattr = xattr_array; xattr->name != NULL; xattr++) {
2170                 new_xattr = simple_xattr_alloc(xattr->value, xattr->value_len);
2171                 if (!new_xattr)
2172                         return -ENOMEM;
2173
2174                 len = strlen(xattr->name) + 1;
2175                 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
2176                                           GFP_KERNEL);
2177                 if (!new_xattr->name) {
2178                         kfree(new_xattr);
2179                         return -ENOMEM;
2180                 }
2181
2182                 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
2183                        XATTR_SECURITY_PREFIX_LEN);
2184                 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
2185                        xattr->name, len);
2186
2187                 simple_xattr_list_add(&info->xattrs, new_xattr);
2188         }
2189
2190         return 0;
2191 }
2192
2193 static const struct xattr_handler *shmem_xattr_handlers[] = {
2194 #ifdef CONFIG_TMPFS_POSIX_ACL
2195         &generic_acl_access_handler,
2196         &generic_acl_default_handler,
2197 #endif
2198         NULL
2199 };
2200
2201 static int shmem_xattr_validate(const char *name)
2202 {
2203         struct { const char *prefix; size_t len; } arr[] = {
2204                 { XATTR_SECURITY_PREFIX, XATTR_SECURITY_PREFIX_LEN },
2205                 { XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN }
2206         };
2207         int i;
2208
2209         for (i = 0; i < ARRAY_SIZE(arr); i++) {
2210                 size_t preflen = arr[i].len;
2211                 if (strncmp(name, arr[i].prefix, preflen) == 0) {
2212                         if (!name[preflen])
2213                                 return -EINVAL;
2214                         return 0;
2215                 }
2216         }
2217         return -EOPNOTSUPP;
2218 }
2219
2220 static ssize_t shmem_getxattr(struct dentry *dentry, const char *name,
2221                               void *buffer, size_t size)
2222 {
2223         struct shmem_inode_info *info = SHMEM_I(dentry->d_inode);
2224         int err;
2225
2226         /*
2227          * If this is a request for a synthetic attribute in the system.*
2228          * namespace use the generic infrastructure to resolve a handler
2229          * for it via sb->s_xattr.
2230          */
2231         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2232                 return generic_getxattr(dentry, name, buffer, size);
2233
2234         err = shmem_xattr_validate(name);
2235         if (err)
2236                 return err;
2237
2238         return simple_xattr_get(&info->xattrs, name, buffer, size);
2239 }
2240
2241 static int shmem_setxattr(struct dentry *dentry, const char *name,
2242                           const void *value, size_t size, int flags)
2243 {
2244         struct shmem_inode_info *info = SHMEM_I(dentry->d_inode);
2245         int err;
2246
2247         /*
2248          * If this is a request for a synthetic attribute in the system.*
2249          * namespace use the generic infrastructure to resolve a handler
2250          * for it via sb->s_xattr.
2251          */
2252         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2253                 return generic_setxattr(dentry, name, value, size, flags);
2254
2255         err = shmem_xattr_validate(name);
2256         if (err)
2257                 return err;
2258
2259         return simple_xattr_set(&info->xattrs, name, value, size, flags);
2260 }
2261
2262 static int shmem_removexattr(struct dentry *dentry, const char *name)
2263 {
2264         struct shmem_inode_info *info = SHMEM_I(dentry->d_inode);
2265         int err;
2266
2267         /*
2268          * If this is a request for a synthetic attribute in the system.*
2269          * namespace use the generic infrastructure to resolve a handler
2270          * for it via sb->s_xattr.
2271          */
2272         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2273                 return generic_removexattr(dentry, name);
2274
2275         err = shmem_xattr_validate(name);
2276         if (err)
2277                 return err;
2278
2279         return simple_xattr_remove(&info->xattrs, name);
2280 }
2281
2282 static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
2283 {
2284         struct shmem_inode_info *info = SHMEM_I(dentry->d_inode);
2285         return simple_xattr_list(&info->xattrs, buffer, size);
2286 }
2287 #endif /* CONFIG_TMPFS_XATTR */
2288
2289 static const struct inode_operations shmem_short_symlink_operations = {
2290         .readlink       = generic_readlink,
2291         .follow_link    = shmem_follow_short_symlink,
2292 #ifdef CONFIG_TMPFS_XATTR
2293         .setxattr       = shmem_setxattr,
2294         .getxattr       = shmem_getxattr,
2295         .listxattr      = shmem_listxattr,
2296         .removexattr    = shmem_removexattr,
2297 #endif
2298 };
2299
2300 static const struct inode_operations shmem_symlink_inode_operations = {
2301         .readlink       = generic_readlink,
2302         .follow_link    = shmem_follow_link,
2303         .put_link       = shmem_put_link,
2304 #ifdef CONFIG_TMPFS_XATTR
2305         .setxattr       = shmem_setxattr,
2306         .getxattr       = shmem_getxattr,
2307         .listxattr      = shmem_listxattr,
2308         .removexattr    = shmem_removexattr,
2309 #endif
2310 };
2311
2312 static struct dentry *shmem_get_parent(struct dentry *child)
2313 {
2314         return ERR_PTR(-ESTALE);
2315 }
2316
2317 static int shmem_match(struct inode *ino, void *vfh)
2318 {
2319         __u32 *fh = vfh;
2320         __u64 inum = fh[2];
2321         inum = (inum << 32) | fh[1];
2322         return ino->i_ino == inum && fh[0] == ino->i_generation;
2323 }
2324
2325 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
2326                 struct fid *fid, int fh_len, int fh_type)
2327 {
2328         struct inode *inode;
2329         struct dentry *dentry = NULL;
2330         u64 inum;
2331
2332         if (fh_len < 3)
2333                 return NULL;
2334
2335         inum = fid->raw[2];
2336         inum = (inum << 32) | fid->raw[1];
2337
2338         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
2339                         shmem_match, fid->raw);
2340         if (inode) {
2341                 dentry = d_find_alias(inode);
2342                 iput(inode);
2343         }
2344
2345         return dentry;
2346 }
2347
2348 static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
2349                                 struct inode *parent)
2350 {
2351         if (*len < 3) {
2352                 *len = 3;
2353                 return 255;
2354         }
2355
2356         if (inode_unhashed(inode)) {
2357                 /* Unfortunately insert_inode_hash is not idempotent,
2358                  * so as we hash inodes here rather than at creation
2359                  * time, we need a lock to ensure we only try
2360                  * to do it once
2361                  */
2362                 static DEFINE_SPINLOCK(lock);
2363                 spin_lock(&lock);
2364                 if (inode_unhashed(inode))
2365                         __insert_inode_hash(inode,
2366                                             inode->i_ino + inode->i_generation);
2367                 spin_unlock(&lock);
2368         }
2369
2370         fh[0] = inode->i_generation;
2371         fh[1] = inode->i_ino;
2372         fh[2] = ((__u64)inode->i_ino) >> 32;
2373
2374         *len = 3;
2375         return 1;
2376 }
2377
2378 static const struct export_operations shmem_export_ops = {
2379         .get_parent     = shmem_get_parent,
2380         .encode_fh      = shmem_encode_fh,
2381         .fh_to_dentry   = shmem_fh_to_dentry,
2382 };
2383
2384 static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
2385                                bool remount)
2386 {
2387         char *this_char, *value, *rest;
2388         struct mempolicy *mpol = NULL;
2389         uid_t uid;
2390         gid_t gid;
2391
2392         while (options != NULL) {
2393                 this_char = options;
2394                 for (;;) {
2395                         /*
2396                          * NUL-terminate this option: unfortunately,
2397                          * mount options form a comma-separated list,
2398                          * but mpol's nodelist may also contain commas.
2399                          */
2400                         options = strchr(options, ',');
2401                         if (options == NULL)
2402                                 break;
2403                         options++;
2404                         if (!isdigit(*options)) {
2405                                 options[-1] = '\0';
2406                                 break;
2407                         }
2408                 }
2409                 if (!*this_char)
2410                         continue;
2411                 if ((value = strchr(this_char,'=')) != NULL) {
2412                         *value++ = 0;
2413                 } else {
2414                         printk(KERN_ERR
2415                             "tmpfs: No value for mount option '%s'\n",
2416                             this_char);
2417                         goto error;
2418                 }
2419
2420                 if (!strcmp(this_char,"size")) {
2421                         unsigned long long size;
2422                         size = memparse(value,&rest);
2423                         if (*rest == '%') {
2424                                 size <<= PAGE_SHIFT;
2425                                 size *= totalram_pages;
2426                                 do_div(size, 100);
2427                                 rest++;
2428                         }
2429                         if (*rest)
2430                                 goto bad_val;
2431                         sbinfo->max_blocks =
2432                                 DIV_ROUND_UP(size, PAGE_CACHE_SIZE);
2433                 } else if (!strcmp(this_char,"nr_blocks")) {
2434                         sbinfo->max_blocks = memparse(value, &rest);
2435                         if (*rest)
2436                                 goto bad_val;
2437                 } else if (!strcmp(this_char,"nr_inodes")) {
2438                         sbinfo->max_inodes = memparse(value, &rest);
2439                         if (*rest)
2440                                 goto bad_val;
2441                 } else if (!strcmp(this_char,"mode")) {
2442                         if (remount)
2443                                 continue;
2444                         sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
2445                         if (*rest)
2446                                 goto bad_val;
2447                 } else if (!strcmp(this_char,"uid")) {
2448                         if (remount)
2449                                 continue;
2450                         uid = simple_strtoul(value, &rest, 0);
2451                         if (*rest)
2452                                 goto bad_val;
2453                         sbinfo->uid = make_kuid(current_user_ns(), uid);
2454                         if (!uid_valid(sbinfo->uid))
2455                                 goto bad_val;
2456                 } else if (!strcmp(this_char,"gid")) {
2457                         if (remount)
2458                                 continue;
2459                         gid = simple_strtoul(value, &rest, 0);
2460                         if (*rest)
2461                                 goto bad_val;
2462                         sbinfo->gid = make_kgid(current_user_ns(), gid);
2463                         if (!gid_valid(sbinfo->gid))
2464                                 goto bad_val;
2465                 } else if (!strcmp(this_char,"mpol")) {
2466                         mpol_put(mpol);
2467                         mpol = NULL;
2468                         if (mpol_parse_str(value, &mpol))
2469                                 goto bad_val;
2470                 } else {
2471                         printk(KERN_ERR "tmpfs: Bad mount option %s\n",
2472                                this_char);
2473                         goto error;
2474                 }
2475         }
2476         sbinfo->mpol = mpol;
2477         return 0;
2478
2479 bad_val:
2480         printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
2481                value, this_char);
2482 error:
2483         mpol_put(mpol);
2484         return 1;
2485
2486 }
2487
2488 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2489 {
2490         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2491         struct shmem_sb_info config = *sbinfo;
2492         unsigned long inodes;
2493         int error = -EINVAL;
2494
2495         config.mpol = NULL;
2496         if (shmem_parse_options(data, &config, true))
2497                 return error;
2498
2499         spin_lock(&sbinfo->stat_lock);
2500         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2501         if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
2502                 goto out;
2503         if (config.max_inodes < inodes)
2504                 goto out;
2505         /*
2506          * Those tests disallow limited->unlimited while any are in use;
2507          * but we must separately disallow unlimited->limited, because
2508          * in that case we have no record of how much is already in use.
2509          */
2510         if (config.max_blocks && !sbinfo->max_blocks)
2511                 goto out;
2512         if (config.max_inodes && !sbinfo->max_inodes)
2513                 goto out;
2514
2515         error = 0;
2516         sbinfo->max_blocks  = config.max_blocks;
2517         sbinfo->max_inodes  = config.max_inodes;
2518         sbinfo->free_inodes = config.max_inodes - inodes;
2519
2520         /*
2521          * Preserve previous mempolicy unless mpol remount option was specified.
2522          */
2523         if (config.mpol) {
2524                 mpol_put(sbinfo->mpol);
2525                 sbinfo->mpol = config.mpol;     /* transfers initial ref */
2526         }
2527 out:
2528         spin_unlock(&sbinfo->stat_lock);
2529         return error;
2530 }
2531
2532 static int shmem_show_options(struct seq_file *seq, struct dentry *root)
2533 {
2534         struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
2535
2536         if (sbinfo->max_blocks != shmem_default_max_blocks())
2537                 seq_printf(seq, ",size=%luk",
2538                         sbinfo->max_blocks << (PAGE_CACHE_SHIFT - 10));
2539         if (sbinfo->max_inodes != shmem_default_max_inodes())
2540                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
2541         if (sbinfo->mode != (S_IRWXUGO | S_ISVTX))
2542                 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
2543         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
2544                 seq_printf(seq, ",uid=%u",
2545                                 from_kuid_munged(&init_user_ns, sbinfo->uid));
2546         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
2547                 seq_printf(seq, ",gid=%u",
2548                                 from_kgid_munged(&init_user_ns, sbinfo->gid));
2549         shmem_show_mpol(seq, sbinfo->mpol);
2550         return 0;
2551 }
2552 #endif /* CONFIG_TMPFS */
2553
2554 static void shmem_put_super(struct super_block *sb)
2555 {
2556         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2557
2558         percpu_counter_destroy(&sbinfo->used_blocks);
2559         mpol_put(sbinfo->mpol);
2560         kfree(sbinfo);
2561         sb->s_fs_info = NULL;
2562 }
2563
2564 int shmem_fill_super(struct super_block *sb, void *data, int silent)
2565 {
2566         struct inode *inode;
2567         struct shmem_sb_info *sbinfo;
2568         int err = -ENOMEM;
2569
2570         /* Round up to L1_CACHE_BYTES to resist false sharing */
2571         sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
2572                                 L1_CACHE_BYTES), GFP_KERNEL);
2573         if (!sbinfo)
2574                 return -ENOMEM;
2575
2576         sbinfo->mode = S_IRWXUGO | S_ISVTX;
2577         sbinfo->uid = current_fsuid();
2578         sbinfo->gid = current_fsgid();
2579         sb->s_fs_info = sbinfo;
2580
2581 #ifdef CONFIG_TMPFS
2582         /*
2583          * Per default we only allow half of the physical ram per
2584          * tmpfs instance, limiting inodes to one per page of lowmem;
2585          * but the internal instance is left unlimited.
2586          */
2587         if (!(sb->s_flags & MS_NOUSER)) {
2588                 sbinfo->max_blocks = shmem_default_max_blocks();
2589                 sbinfo->max_inodes = shmem_default_max_inodes();
2590                 if (shmem_parse_options(data, sbinfo, false)) {
2591                         err = -EINVAL;
2592                         goto failed;
2593                 }
2594         }
2595         sb->s_export_op = &shmem_export_ops;
2596         sb->s_flags |= MS_NOSEC;
2597 #else
2598         sb->s_flags |= MS_NOUSER;
2599 #endif
2600
2601         spin_lock_init(&sbinfo->stat_lock);
2602         if (percpu_counter_init(&sbinfo->used_blocks, 0))
2603                 goto failed;
2604         sbinfo->free_inodes = sbinfo->max_inodes;
2605
2606         sb->s_maxbytes = MAX_LFS_FILESIZE;
2607         sb->s_blocksize = PAGE_CACHE_SIZE;
2608         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
2609         sb->s_magic = TMPFS_MAGIC;
2610         sb->s_op = &shmem_ops;
2611         sb->s_time_gran = 1;
2612 #ifdef CONFIG_TMPFS_XATTR
2613         sb->s_xattr = shmem_xattr_handlers;
2614 #endif
2615 #ifdef CONFIG_TMPFS_POSIX_ACL
2616         sb->s_flags |= MS_POSIXACL;
2617 #endif
2618
2619         inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
2620         if (!inode)
2621                 goto failed;
2622         inode->i_uid = sbinfo->uid;
2623         inode->i_gid = sbinfo->gid;
2624         sb->s_root = d_make_root(inode);
2625         if (!sb->s_root)
2626                 goto failed;
2627         return 0;
2628
2629 failed:
2630         shmem_put_super(sb);
2631         return err;
2632 }
2633
2634 static struct kmem_cache *shmem_inode_cachep;
2635
2636 static struct inode *shmem_alloc_inode(struct super_block *sb)
2637 {
2638         struct shmem_inode_info *info;
2639         info = kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
2640         if (!info)
2641                 return NULL;
2642         return &info->vfs_inode;
2643 }
2644
2645 static void shmem_destroy_callback(struct rcu_head *head)
2646 {
2647         struct inode *inode = container_of(head, struct inode, i_rcu);
2648         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
2649 }
2650
2651 static void shmem_destroy_inode(struct inode *inode)
2652 {
2653         if (S_ISREG(inode->i_mode))
2654                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
2655         call_rcu(&inode->i_rcu, shmem_destroy_callback);
2656 }
2657
2658 static void shmem_init_inode(void *foo)
2659 {
2660         struct shmem_inode_info *info = foo;
2661         inode_init_once(&info->vfs_inode);
2662 }
2663
2664 static int shmem_init_inodecache(void)
2665 {
2666         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
2667                                 sizeof(struct shmem_inode_info),
2668                                 0, SLAB_PANIC, shmem_init_inode);
2669         return 0;
2670 }
2671
2672 static void shmem_destroy_inodecache(void)
2673 {
2674         kmem_cache_destroy(shmem_inode_cachep);
2675 }
2676
2677 static const struct address_space_operations shmem_aops = {
2678         .writepage      = shmem_writepage,
2679         .set_page_dirty = __set_page_dirty_no_writeback,
2680 #ifdef CONFIG_TMPFS
2681         .write_begin    = shmem_write_begin,
2682         .write_end      = shmem_write_end,
2683 #endif
2684         .migratepage    = migrate_page,
2685         .error_remove_page = generic_error_remove_page,
2686 };
2687
2688 static const struct file_operations shmem_file_operations = {
2689         .mmap           = shmem_mmap,
2690 #ifdef CONFIG_TMPFS
2691         .llseek         = shmem_file_llseek,
2692         .read           = do_sync_read,
2693         .write          = do_sync_write,
2694         .aio_read       = shmem_file_aio_read,
2695         .aio_write      = generic_file_aio_write,
2696         .fsync          = noop_fsync,
2697         .splice_read    = shmem_file_splice_read,
2698         .splice_write   = generic_file_splice_write,
2699         .fallocate      = shmem_fallocate,
2700 #endif
2701 };
2702
2703 static const struct inode_operations shmem_inode_operations = {
2704         .setattr        = shmem_setattr,
2705 #ifdef CONFIG_TMPFS_XATTR
2706         .setxattr       = shmem_setxattr,
2707         .getxattr       = shmem_getxattr,
2708         .listxattr      = shmem_listxattr,
2709         .removexattr    = shmem_removexattr,
2710 #endif
2711 };
2712
2713 static const struct inode_operations shmem_dir_inode_operations = {
2714 #ifdef CONFIG_TMPFS
2715         .create         = shmem_create,
2716         .lookup         = simple_lookup,
2717         .link           = shmem_link,
2718         .unlink         = shmem_unlink,
2719         .symlink        = shmem_symlink,
2720         .mkdir          = shmem_mkdir,
2721         .rmdir          = shmem_rmdir,
2722         .mknod          = shmem_mknod,
2723         .rename         = shmem_rename,
2724 #endif
2725 #ifdef CONFIG_TMPFS_XATTR
2726         .setxattr       = shmem_setxattr,
2727         .getxattr       = shmem_getxattr,
2728         .listxattr      = shmem_listxattr,
2729         .removexattr    = shmem_removexattr,
2730 #endif
2731 #ifdef CONFIG_TMPFS_POSIX_ACL
2732         .setattr        = shmem_setattr,
2733 #endif
2734 };
2735
2736 static const struct inode_operations shmem_special_inode_operations = {
2737 #ifdef CONFIG_TMPFS_XATTR
2738         .setxattr       = shmem_setxattr,
2739         .getxattr       = shmem_getxattr,
2740         .listxattr      = shmem_listxattr,
2741         .removexattr    = shmem_removexattr,
2742 #endif
2743 #ifdef CONFIG_TMPFS_POSIX_ACL
2744         .setattr        = shmem_setattr,
2745 #endif
2746 };
2747
2748 static const struct super_operations shmem_ops = {
2749         .alloc_inode    = shmem_alloc_inode,
2750         .destroy_inode  = shmem_destroy_inode,
2751 #ifdef CONFIG_TMPFS
2752         .statfs         = shmem_statfs,
2753         .remount_fs     = shmem_remount_fs,
2754         .show_options   = shmem_show_options,
2755 #endif
2756         .evict_inode    = shmem_evict_inode,
2757         .drop_inode     = generic_delete_inode,
2758         .put_super      = shmem_put_super,
2759 };
2760
2761 static const struct vm_operations_struct shmem_vm_ops = {
2762         .fault          = shmem_fault,
2763 #ifdef CONFIG_NUMA
2764         .set_policy     = shmem_set_policy,
2765         .get_policy     = shmem_get_policy,
2766 #endif
2767         .remap_pages    = generic_file_remap_pages,
2768 };
2769
2770 static struct dentry *shmem_mount(struct file_system_type *fs_type,
2771         int flags, const char *dev_name, void *data)
2772 {
2773         return mount_nodev(fs_type, flags, data, shmem_fill_super);
2774 }
2775
2776 static struct file_system_type shmem_fs_type = {
2777         .owner          = THIS_MODULE,
2778         .name           = "tmpfs",
2779         .mount          = shmem_mount,
2780         .kill_sb        = kill_litter_super,
2781         .fs_flags       = FS_USERNS_MOUNT,
2782 };
2783
2784 int __init shmem_init(void)
2785 {
2786         int error;
2787
2788         error = bdi_init(&shmem_backing_dev_info);
2789         if (error)
2790                 goto out4;
2791
2792         error = shmem_init_inodecache();
2793         if (error)
2794                 goto out3;
2795
2796         error = register_filesystem(&shmem_fs_type);
2797         if (error) {
2798                 printk(KERN_ERR "Could not register tmpfs\n");
2799                 goto out2;
2800         }
2801
2802         shm_mnt = vfs_kern_mount(&shmem_fs_type, MS_NOUSER,
2803                                  shmem_fs_type.name, NULL);
2804         if (IS_ERR(shm_mnt)) {
2805                 error = PTR_ERR(shm_mnt);
2806                 printk(KERN_ERR "Could not kern_mount tmpfs\n");
2807                 goto out1;
2808         }
2809         return 0;
2810
2811 out1:
2812         unregister_filesystem(&shmem_fs_type);
2813 out2:
2814         shmem_destroy_inodecache();
2815 out3:
2816         bdi_destroy(&shmem_backing_dev_info);
2817 out4:
2818         shm_mnt = ERR_PTR(error);
2819         return error;
2820 }
2821
2822 #else /* !CONFIG_SHMEM */
2823
2824 /*
2825  * tiny-shmem: simple shmemfs and tmpfs using ramfs code
2826  *
2827  * This is intended for small system where the benefits of the full
2828  * shmem code (swap-backed and resource-limited) are outweighed by
2829  * their complexity. On systems without swap this code should be
2830  * effectively equivalent, but much lighter weight.
2831  */
2832
2833 #include <linux/ramfs.h>
2834
2835 static struct file_system_type shmem_fs_type = {
2836         .name           = "tmpfs",
2837         .mount          = ramfs_mount,
2838         .kill_sb        = kill_litter_super,
2839         .fs_flags       = FS_USERNS_MOUNT,
2840 };
2841
2842 int __init shmem_init(void)
2843 {
2844         BUG_ON(register_filesystem(&shmem_fs_type) != 0);
2845
2846         shm_mnt = kern_mount(&shmem_fs_type);
2847         BUG_ON(IS_ERR(shm_mnt));
2848
2849         return 0;
2850 }
2851
2852 int shmem_unuse(swp_entry_t swap, struct page *page)
2853 {
2854         return 0;
2855 }
2856
2857 int shmem_lock(struct file *file, int lock, struct user_struct *user)
2858 {
2859         return 0;
2860 }
2861
2862 void shmem_unlock_mapping(struct address_space *mapping)
2863 {
2864 }
2865
2866 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
2867 {
2868         truncate_inode_pages_range(inode->i_mapping, lstart, lend);
2869 }
2870 EXPORT_SYMBOL_GPL(shmem_truncate_range);
2871
2872 #define shmem_vm_ops                            generic_file_vm_ops
2873 #define shmem_file_operations                   ramfs_file_operations
2874 #define shmem_get_inode(sb, dir, mode, dev, flags)      ramfs_get_inode(sb, dir, mode, dev)
2875 #define shmem_acct_size(flags, size)            0
2876 #define shmem_unacct_size(flags, size)          do {} while (0)
2877
2878 #endif /* CONFIG_SHMEM */
2879
2880 /* common code */
2881
2882 /**
2883  * shmem_file_setup - get an unlinked file living in tmpfs
2884  * @name: name for dentry (to be seen in /proc/<pid>/maps
2885  * @size: size to be set for the file
2886  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
2887  */
2888 struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
2889 {
2890         int error;
2891         struct file *file;
2892         struct inode *inode;
2893         struct path path;
2894         struct dentry *root;
2895         struct qstr this;
2896
2897         if (IS_ERR(shm_mnt))
2898                 return (void *)shm_mnt;
2899
2900         if (size < 0 || size > MAX_LFS_FILESIZE)
2901                 return ERR_PTR(-EINVAL);
2902
2903         if (shmem_acct_size(flags, size))
2904                 return ERR_PTR(-ENOMEM);
2905
2906         error = -ENOMEM;
2907         this.name = name;
2908         this.len = strlen(name);
2909         this.hash = 0; /* will go */
2910         root = shm_mnt->mnt_root;
2911         path.dentry = d_alloc(root, &this);
2912         if (!path.dentry)
2913                 goto put_memory;
2914         path.mnt = mntget(shm_mnt);
2915
2916         error = -ENOSPC;
2917         inode = shmem_get_inode(root->d_sb, NULL, S_IFREG | S_IRWXUGO, 0, flags);
2918         if (!inode)
2919                 goto put_dentry;
2920
2921         d_instantiate(path.dentry, inode);
2922         inode->i_size = size;
2923         clear_nlink(inode);     /* It is unlinked */
2924 #ifndef CONFIG_MMU
2925         error = ramfs_nommu_expand_for_mapping(inode, size);
2926         if (error)
2927                 goto put_dentry;
2928 #endif
2929
2930         error = -ENFILE;
2931         file = alloc_file(&path, FMODE_WRITE | FMODE_READ,
2932                   &shmem_file_operations);
2933         if (!file)
2934                 goto put_dentry;
2935
2936         return file;
2937
2938 put_dentry:
2939         path_put(&path);
2940 put_memory:
2941         shmem_unacct_size(flags, size);
2942         return ERR_PTR(error);
2943 }
2944 EXPORT_SYMBOL_GPL(shmem_file_setup);
2945
2946 /**
2947  * shmem_zero_setup - setup a shared anonymous mapping
2948  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
2949  */
2950 int shmem_zero_setup(struct vm_area_struct *vma)
2951 {
2952         struct file *file;
2953         loff_t size = vma->vm_end - vma->vm_start;
2954
2955         file = shmem_file_setup("dev/zero", size, vma->vm_flags);
2956         if (IS_ERR(file))
2957                 return PTR_ERR(file);
2958
2959         if (vma->vm_file)
2960                 fput(vma->vm_file);
2961         vma->vm_file = file;
2962         vma->vm_ops = &shmem_vm_ops;
2963         return 0;
2964 }
2965
2966 /**
2967  * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
2968  * @mapping:    the page's address_space
2969  * @index:      the page index
2970  * @gfp:        the page allocator flags to use if allocating
2971  *
2972  * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
2973  * with any new page allocations done using the specified allocation flags.
2974  * But read_cache_page_gfp() uses the ->readpage() method: which does not
2975  * suit tmpfs, since it may have pages in swapcache, and needs to find those
2976  * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
2977  *
2978  * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
2979  * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
2980  */
2981 struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
2982                                          pgoff_t index, gfp_t gfp)
2983 {
2984 #ifdef CONFIG_SHMEM
2985         struct inode *inode = mapping->host;
2986         struct page *page;
2987         int error;
2988
2989         BUG_ON(mapping->a_ops != &shmem_aops);
2990         error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE, gfp, NULL);
2991         if (error)
2992                 page = ERR_PTR(error);
2993         else
2994                 unlock_page(page);
2995         return page;
2996 #else
2997         /*
2998          * The tiny !SHMEM case uses ramfs without swap
2999          */
3000         return read_cache_page_gfp(mapping, index, gfp);
3001 #endif
3002 }
3003 EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);