mm/shmem.c

   1 /*
   2  * Resizable virtual memory filesystem for Linux.
   3  *
   4  * Copyright (C) 2000 Linus Torvalds.
   5  *               2000 Transmeta Corp.
   6  *               2000-2001 Christoph Rohland
   7  *               2000-2001 SAP AG
   8  *               2002 Red Hat Inc.
   9  * Copyright (C) 2002-2011 Hugh Dickins.
  10  * Copyright (C) 2011 Google Inc.
  11  * Copyright (C) 2002-2005 VERITAS Software Corporation.
  12  * Copyright (C) 2004 Andi Kleen, SuSE Labs
  13  *
  14  * Extended attribute support for tmpfs:
  15  * Copyright (c) 2004, Luke Kenneth Casson Leighton <lkcl@lkcl.net>
  16  * Copyright (c) 2004 Red Hat, Inc., James Morris <jmorris@redhat.com>
  17  *
  18  * tiny-shmem:
  19  * Copyright (c) 2004, 2008 Matt Mackall <mpm@selenic.com>
  20  *
  21  * This file is released under the GPL.
  22  */
  23
  24 #include <linux/fs.h>
  25 #include <linux/init.h>
  26 #include <linux/vfs.h>
  27 #include <linux/mount.h>
  28 #include <linux/pagemap.h>
  29 #include <linux/file.h>
  30 #include <linux/mm.h>
  31 #include <linux/export.h>
  32 #include <linux/swap.h>
  33
  34 static struct vfsmount *shm_mnt;
  35
  36 #ifdef CONFIG_SHMEM
  37 /*
  38  * This virtual memory filesystem is heavily based on the ramfs. It
  39  * extends ramfs by the ability to use swap and honor resource limits
  40  * which makes it a completely usable filesystem.
  41  */
  42
  43 #include <linux/xattr.h>
  44 #include <linux/exportfs.h>
  45 #include <linux/posix_acl.h>
  46 #include <linux/generic_acl.h>
  47 #include <linux/mman.h>
  48 #include <linux/string.h>
  49 #include <linux/slab.h>
  50 #include <linux/backing-dev.h>
  51 #include <linux/shmem_fs.h>
  52 #include <linux/writeback.h>
  53 #include <linux/blkdev.h>
  54 #include <linux/pagevec.h>
  55 #include <linux/percpu_counter.h>
  56 #include <linux/falloc.h>
  57 #include <linux/splice.h>
  58 #include <linux/security.h>
  59 #include <linux/swapops.h>
  60 #include <linux/mempolicy.h>
  61 #include <linux/namei.h>
  62 #include <linux/ctype.h>
  63 #include <linux/migrate.h>
  64 #include <linux/highmem.h>
  65 #include <linux/seq_file.h>
  66 #include <linux/magic.h>
  67
  68 #include <asm/uaccess.h>
  69 #include <asm/pgtable.h>
  70
  71 #define BLOCKS_PER_PAGE  (PAGE_CACHE_SIZE/512)
  72 #define VM_ACCT(size)    (PAGE_CACHE_ALIGN(size) >> PAGE_SHIFT)
  73
  74 /* Pretend that each entry is of this size in directory's i_size */
  75 #define BOGO_DIRENT_SIZE 20
  76
  77 /* Symlink up to this size is kmalloc'ed instead of using a swappable page */
  78 #define SHORT_SYMLINK_LEN 128
  79
  80 /*
  81  * shmem_fallocate and shmem_writepage communicate via inode->i_private
  82  * (with i_mutex making sure that it has only one user at a time):
  83  * we would prefer not to enlarge the shmem inode just for that.
  84  */
  85 struct shmem_falloc {
  86         pgoff_t start;          /* start of range currently being fallocated */
  87         pgoff_t next;           /* the next page offset to be fallocated */
  88         pgoff_t nr_falloced;    /* how many new pages have been fallocated */
  89         pgoff_t nr_unswapped;   /* how often writepage refused to swap out */
  90 };
  91
  92 /* Flag allocation requirements to shmem_getpage */
  93 enum sgp_type {
  94         SGP_READ,       /* don't exceed i_size, don't allocate page */
  95         SGP_CACHE,      /* don't exceed i_size, may allocate page */
  96         SGP_DIRTY,      /* like SGP_CACHE, but set new page dirty */
  97         SGP_WRITE,      /* may exceed i_size, may allocate !Uptodate page */
  98         SGP_FALLOC,     /* like SGP_WRITE, but make existing page Uptodate */
  99 };
 100
 101 #ifdef CONFIG_TMPFS
 102 static unsigned long shmem_default_max_blocks(void)
 103 {
 104         return totalram_pages / 2;
 105 }
 106
 107 static unsigned long shmem_default_max_inodes(void)
 108 {
 109         return min(totalram_pages - totalhigh_pages, totalram_pages / 2);
 110 }
 111 #endif
 112
 113 static bool shmem_should_replace_page(struct page *page, gfp_t gfp);
 114 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
 115                                 struct shmem_inode_info *info, pgoff_t index);
 116 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
 117         struct page **pagep, enum sgp_type sgp, gfp_t gfp, int *fault_type);
 118
 119 static inline int shmem_getpage(struct inode *inode, pgoff_t index,
 120         struct page **pagep, enum sgp_type sgp, int *fault_type)
 121 {
 122         return shmem_getpage_gfp(inode, index, pagep, sgp,
 123                         mapping_gfp_mask(inode->i_mapping), fault_type);
 124 }
 125
 126 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
 127 {
 128         return sb->s_fs_info;
 129 }
 130
 131 /*
 132  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
 133  * for shared memory and for shared anonymous (/dev/zero) mappings
 134  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
 135  * consistent with the pre-accounting of private mappings ...
 136  */
 137 static inline int shmem_acct_size(unsigned long flags, loff_t size)
 138 {
 139         return (flags & VM_NORESERVE) ?
 140                 0 : security_vm_enough_memory_mm(current->mm, VM_ACCT(size));
 141 }
 142
 143 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
 144 {
 145         if (!(flags & VM_NORESERVE))
 146                 vm_unacct_memory(VM_ACCT(size));
 147 }
 148
 149 /*
 150  * ... whereas tmpfs objects are accounted incrementally as
 151  * pages are allocated, in order to allow huge sparse files.
 152  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
 153  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
 154  */
 155 static inline int shmem_acct_block(unsigned long flags)
 156 {
 157         return (flags & VM_NORESERVE) ?
 158                 security_vm_enough_memory_mm(current->mm, VM_ACCT(PAGE_CACHE_SIZE)) : 0;
 159 }
 160
 161 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
 162 {
 163         if (flags & VM_NORESERVE)
 164                 vm_unacct_memory(pages * VM_ACCT(PAGE_CACHE_SIZE));
 165 }
 166
 167 static const struct super_operations shmem_ops;
 168 static const struct address_space_operations shmem_aops;
 169 static const struct file_operations shmem_file_operations;
 170 static const struct inode_operations shmem_inode_operations;
 171 static const struct inode_operations shmem_dir_inode_operations;
 172 static const struct inode_operations shmem_special_inode_operations;
 173 static const struct vm_operations_struct shmem_vm_ops;
 174
 175 static struct backing_dev_info shmem_backing_dev_info  __read_mostly = {
 176         .ra_pages       = 0,    /* No readahead */
 177         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_SWAP_BACKED,
 178 };
 179
 180 static LIST_HEAD(shmem_swaplist);
 181 static DEFINE_MUTEX(shmem_swaplist_mutex);
 182
 183 static int shmem_reserve_inode(struct super_block *sb)
 184 {
 185         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 186         if (sbinfo->max_inodes) {
 187                 spin_lock(&sbinfo->stat_lock);
 188                 if (!sbinfo->free_inodes) {
 189                         spin_unlock(&sbinfo->stat_lock);
 190                         return -ENOSPC;
 191                 }
 192                 sbinfo->free_inodes--;
 193                 spin_unlock(&sbinfo->stat_lock);
 194         }
 195         return 0;
 196 }
 197
 198 static void shmem_free_inode(struct super_block *sb)
 199 {
 200         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 201         if (sbinfo->max_inodes) {
 202                 spin_lock(&sbinfo->stat_lock);
 203                 sbinfo->free_inodes++;
 204                 spin_unlock(&sbinfo->stat_lock);
 205         }
 206 }
 207
 208 /**
 209  * shmem_recalc_inode - recalculate the block usage of an inode
 210  * @inode: inode to recalc
 211  *
 212  * We have to calculate the free blocks since the mm can drop
 213  * undirtied hole pages behind our back.
 214  *
 215  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
 216  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
 217  *
 218  * It has to be called with the spinlock held.
 219  */
 220 static void shmem_recalc_inode(struct inode *inode)
 221 {
 222         struct shmem_inode_info *info = SHMEM_I(inode);
 223         long freed;
 224
 225         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
 226         if (freed > 0) {
 227                 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 228                 if (sbinfo->max_blocks)
 229                         percpu_counter_add(&sbinfo->used_blocks, -freed);
 230                 info->alloced -= freed;
 231                 inode->i_blocks -= freed * BLOCKS_PER_PAGE;
 232                 shmem_unacct_blocks(info->flags, freed);
 233         }
 234 }
 235
 236 /*
 237  * Replace item expected in radix tree by a new item, while holding tree lock.
 238  */
 239 static int shmem_radix_tree_replace(struct address_space *mapping,
 240                         pgoff_t index, void *expected, void *replacement)
 241 {
 242         void **pslot;
 243         void *item = NULL;
 244
 245         VM_BUG_ON(!expected);
 246         pslot = radix_tree_lookup_slot(&mapping->page_tree, index);
 247         if (pslot)
 248                 item = radix_tree_deref_slot_protected(pslot,
 249                                                         &mapping->tree_lock);
 250         if (item != expected)
 251                 return -ENOENT;
 252         if (replacement)
 253                 radix_tree_replace_slot(pslot, replacement);
 254         else
 255                 radix_tree_delete(&mapping->page_tree, index);
 256         return 0;
 257 }
 258
 259 /*
 260  * Sometimes, before we decide whether to proceed or to fail, we must check
 261  * that an entry was not already brought back from swap by a racing thread.
 262  *
 263  * Checking page is not enough: by the time a SwapCache page is locked, it
 264  * might be reused, and again be SwapCache, using the same swap as before.
 265  */
 266 static bool shmem_confirm_swap(struct address_space *mapping,
 267                                pgoff_t index, swp_entry_t swap)
 268 {
 269         void *item;
 270
 271         rcu_read_lock();
 272         item = radix_tree_lookup(&mapping->page_tree, index);
 273         rcu_read_unlock();
 274         return item == swp_to_radix_entry(swap);
 275 }
 276
 277 /*
 278  * Like add_to_page_cache_locked, but error if expected item has gone.
 279  */
 280 static int shmem_add_to_page_cache(struct page *page,
 281                                    struct address_space *mapping,
 282                                    pgoff_t index, gfp_t gfp, void *expected)
 283 {
 284         int error;
 285
 286         VM_BUG_ON(!PageLocked(page));
 287         VM_BUG_ON(!PageSwapBacked(page));
 288
 289         page_cache_get(page);
 290         page->mapping = mapping;
 291         page->index = index;
 292
 293         spin_lock_irq(&mapping->tree_lock);
 294         if (!expected)
 295                 error = radix_tree_insert(&mapping->page_tree, index, page);
 296         else
 297                 error = shmem_radix_tree_replace(mapping, index, expected,
 298                                                                  page);
 299         if (!error) {
 300                 mapping->nrpages++;
 301                 __inc_zone_page_state(page, NR_FILE_PAGES);
 302                 __inc_zone_page_state(page, NR_SHMEM);
 303                 spin_unlock_irq(&mapping->tree_lock);
 304         } else {
 305                 page->mapping = NULL;
 306                 spin_unlock_irq(&mapping->tree_lock);
 307                 page_cache_release(page);
 308         }
 309         return error;
 310 }
 311
 312 /*
 313  * Like delete_from_page_cache, but substitutes swap for page.
 314  */
 315 static void shmem_delete_from_page_cache(struct page *page, void *radswap)
 316 {
 317         struct address_space *mapping = page->mapping;
 318         int error;
 319
 320         spin_lock_irq(&mapping->tree_lock);
 321         error = shmem_radix_tree_replace(mapping, page->index, page, radswap);
 322         page->mapping = NULL;
 323         mapping->nrpages--;
 324         __dec_zone_page_state(page, NR_FILE_PAGES);
 325         __dec_zone_page_state(page, NR_SHMEM);
 326         spin_unlock_irq(&mapping->tree_lock);
 327         page_cache_release(page);
 328         BUG_ON(error);
 329 }
 330
 331 /*
 332  * Like find_get_pages, but collecting swap entries as well as pages.
 333  */
 334 static unsigned shmem_find_get_pages_and_swap(struct address_space *mapping,
 335                                         pgoff_t start, unsigned int nr_pages,
 336                                         struct page **pages, pgoff_t *indices)
 337 {
 338         unsigned int i;
 339         unsigned int ret;
 340         unsigned int nr_found;
 341
 342         rcu_read_lock();
 343 restart:
 344         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
 345                                 (void ***)pages, indices, start, nr_pages);
 346         ret = 0;
 347         for (i = 0; i < nr_found; i++) {
 348                 struct page *page;
 349 repeat:
 350                 page = radix_tree_deref_slot((void **)pages[i]);
 351                 if (unlikely(!page))
 352                         continue;
 353                 if (radix_tree_exception(page)) {
 354                         if (radix_tree_deref_retry(page))
 355                                 goto restart;
 356                         /*
 357                          * Otherwise, we must be storing a swap entry
 358                          * here as an exceptional entry: so return it
 359                          * without attempting to raise page count.
 360                          */
 361                         goto export;
 362                 }
 363                 if (!page_cache_get_speculative(page))
 364                         goto repeat;
 365
 366                 /* Has the page moved? */
 367                 if (unlikely(page != *((void **)pages[i]))) {
 368                         page_cache_release(page);
 369                         goto repeat;
 370                 }
 371 export:
 372                 indices[ret] = indices[i];
 373                 pages[ret] = page;
 374                 ret++;
 375         }
 376         if (unlikely(!ret && nr_found))
 377                 goto restart;
 378         rcu_read_unlock();
 379         return ret;
 380 }
 381
 382 /*
 383  * Remove swap entry from radix tree, free the swap and its page cache.
 384  */
 385 static int shmem_free_swap(struct address_space *mapping,
 386                            pgoff_t index, void *radswap)
 387 {
 388         int error;
 389
 390         spin_lock_irq(&mapping->tree_lock);
 391         error = shmem_radix_tree_replace(mapping, index, radswap, NULL);
 392         spin_unlock_irq(&mapping->tree_lock);
 393         if (!error)
 394                 free_swap_and_cache(radix_to_swp_entry(radswap));
 395         return error;
 396 }
 397
 398 /*
 399  * Pagevec may contain swap entries, so shuffle up pages before releasing.
 400  */
 401 static void shmem_deswap_pagevec(struct pagevec *pvec)
 402 {
 403         int i, j;
 404
 405         for (i = 0, j = 0; i < pagevec_count(pvec); i++) {
 406                 struct page *page = pvec->pages[i];
 407                 if (!radix_tree_exceptional_entry(page))
 408                         pvec->pages[j++] = page;
 409         }
 410         pvec->nr = j;
 411 }
 412
 413 /*
 414  * SysV IPC SHM_UNLOCK restore Unevictable pages to their evictable lists.
 415  */
 416 void shmem_unlock_mapping(struct address_space *mapping)
 417 {
 418         struct pagevec pvec;
 419         pgoff_t indices[PAGEVEC_SIZE];
 420         pgoff_t index = 0;
 421
 422         pagevec_init(&pvec, 0);
 423         /*
 424          * Minor point, but we might as well stop if someone else SHM_LOCKs it.
 425          */
 426         while (!mapping_unevictable(mapping)) {
 427                 /*
 428                  * Avoid pagevec_lookup(): find_get_pages() returns 0 as if it
 429                  * has finished, if it hits a row of PAGEVEC_SIZE swap entries.
 430                  */
 431                 pvec.nr = shmem_find_get_pages_and_swap(mapping, index,
 432                                         PAGEVEC_SIZE, pvec.pages, indices);
 433                 if (!pvec.nr)
 434                         break;
 435                 index = indices[pvec.nr - 1] + 1;
 436                 shmem_deswap_pagevec(&pvec);
 437                 check_move_unevictable_pages(pvec.pages, pvec.nr);
 438                 pagevec_release(&pvec);
 439                 cond_resched();
 440         }
 441 }
 442
 443 /*
 444  * Remove range of pages and swap entries from radix tree, and free them.
 445  * If !unfalloc, truncate or punch hole; if unfalloc, undo failed fallocate.
 446  */
 447 static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
 448                                                                  bool unfalloc)
 449 {
 450         struct address_space *mapping = inode->i_mapping;
 451         struct shmem_inode_info *info = SHMEM_I(inode);
 452         pgoff_t start = (lstart + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 453         pgoff_t end = (lend + 1) >> PAGE_CACHE_SHIFT;
 454         unsigned int partial_start = lstart & (PAGE_CACHE_SIZE - 1);
 455         unsigned int partial_end = (lend + 1) & (PAGE_CACHE_SIZE - 1);
 456         struct pagevec pvec;
 457         pgoff_t indices[PAGEVEC_SIZE];
 458         long nr_swaps_freed = 0;
 459         pgoff_t index;
 460         int i;
 461
 462         if (lend == -1)
 463                 end = -1;       /* unsigned, so actually very big */
 464
 465         pagevec_init(&pvec, 0);
 466         index = start;
 467         while (index < end) {
 468                 pvec.nr = shmem_find_get_pages_and_swap(mapping, index,
 469                                 min(end - index, (pgoff_t)PAGEVEC_SIZE),
 470                                                         pvec.pages, indices);
 471                 if (!pvec.nr)
 472                         break;
 473                 mem_cgroup_uncharge_start();
 474                 for (i = 0; i < pagevec_count(&pvec); i++) {
 475                         struct page *page = pvec.pages[i];
 476
 477                         index = indices[i];
 478                         if (index >= end)
 479                                 break;
 480
 481                         if (radix_tree_exceptional_entry(page)) {
 482                                 if (unfalloc)
 483                                         continue;
 484                                 nr_swaps_freed += !shmem_free_swap(mapping,
 485                                                                 index, page);
 486                                 continue;
 487                         }
 488
 489                         if (!trylock_page(page))
 490                                 continue;
 491                         if (!unfalloc || !PageUptodate(page)) {
 492                                 if (page->mapping == mapping) {
 493                                         VM_BUG_ON(PageWriteback(page));
 494                                         truncate_inode_page(mapping, page);
 495                                 }
 496                         }
 497                         unlock_page(page);
 498                 }
 499                 shmem_deswap_pagevec(&pvec);
 500                 pagevec_release(&pvec);
 501                 mem_cgroup_uncharge_end();
 502                 cond_resched();
 503                 index++;
 504         }
 505
 506         if (partial_start) {
 507                 struct page *page = NULL;
 508                 shmem_getpage(inode, start - 1, &page, SGP_READ, NULL);
 509                 if (page) {
 510                         unsigned int top = PAGE_CACHE_SIZE;
 511                         if (start > end) {
 512                                 top = partial_end;
 513                                 partial_end = 0;
 514                         }
 515                         zero_user_segment(page, partial_start, top);
 516                         set_page_dirty(page);
 517                         unlock_page(page);
 518                         page_cache_release(page);
 519                 }
 520         }
 521         if (partial_end) {
 522                 struct page *page = NULL;
 523                 shmem_getpage(inode, end, &page, SGP_READ, NULL);
 524                 if (page) {
 525                         zero_user_segment(page, 0, partial_end);
 526                         set_page_dirty(page);
 527                         unlock_page(page);
 528                         page_cache_release(page);
 529                 }
 530         }
 531         if (start >= end)
 532                 return;
 533
 534         index = start;
 535         for ( ; ; ) {
 536                 cond_resched();
 537                 pvec.nr = shmem_find_get_pages_and_swap(mapping, index,
 538                                 min(end - index, (pgoff_t)PAGEVEC_SIZE),
 539                                                         pvec.pages, indices);
 540                 if (!pvec.nr) {
 541                         if (index == start || unfalloc)
 542                                 break;
 543                         index = start;
 544                         continue;
 545                 }
 546                 if ((index == start || unfalloc) && indices[0] >= end) {
 547                         shmem_deswap_pagevec(&pvec);
 548                         pagevec_release(&pvec);
 549                         break;
 550                 }
 551                 mem_cgroup_uncharge_start();
 552                 for (i = 0; i < pagevec_count(&pvec); i++) {
 553                         struct page *page = pvec.pages[i];
 554
 555                         index = indices[i];
 556                         if (index >= end)
 557                                 break;
 558
 559                         if (radix_tree_exceptional_entry(page)) {
 560                                 if (unfalloc)
 561                                         continue;
 562                                 nr_swaps_freed += !shmem_free_swap(mapping,
 563                                                                 index, page);
 564                                 continue;
 565                         }
 566
 567                         lock_page(page);
 568                         if (!unfalloc || !PageUptodate(page)) {
 569                                 if (page->mapping == mapping) {
 570                                         VM_BUG_ON(PageWriteback(page));
 571                                         truncate_inode_page(mapping, page);
 572                                 }
 573                         }
 574                         unlock_page(page);
 575                 }
 576                 shmem_deswap_pagevec(&pvec);
 577                 pagevec_release(&pvec);
 578                 mem_cgroup_uncharge_end();
 579                 index++;
 580         }
 581
 582         spin_lock(&info->lock);
 583         info->swapped -= nr_swaps_freed;
 584         shmem_recalc_inode(inode);
 585         spin_unlock(&info->lock);
 586 }
 587
 588 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
 589 {
 590         shmem_undo_range(inode, lstart, lend, false);
 591         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 592 }
 593 EXPORT_SYMBOL_GPL(shmem_truncate_range);
 594
 595 static int shmem_setattr(struct dentry *dentry, struct iattr *attr)
 596 {
 597         struct inode *inode = dentry->d_inode;
 598         int error;
 599
 600         error = inode_change_ok(inode, attr);
 601         if (error)
 602                 return error;
 603
 604         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
 605                 loff_t oldsize = inode->i_size;
 606                 loff_t newsize = attr->ia_size;
 607
 608                 if (newsize != oldsize) {
 609                         i_size_write(inode, newsize);
 610                         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 611                 }
 612                 if (newsize < oldsize) {
 613                         loff_t holebegin = round_up(newsize, PAGE_SIZE);
 614                         unmap_mapping_range(inode->i_mapping, holebegin, 0, 1);
 615                         shmem_truncate_range(inode, newsize, (loff_t)-1);
 616                         /* unmap again to remove racily COWed private pages */
 617                         unmap_mapping_range(inode->i_mapping, holebegin, 0, 1);
 618                 }
 619         }
 620
 621         setattr_copy(inode, attr);
 622 #ifdef CONFIG_TMPFS_POSIX_ACL
 623         if (attr->ia_valid & ATTR_MODE)
 624                 error = generic_acl_chmod(inode);
 625 #endif
 626         return error;
 627 }
 628
 629 static void shmem_evict_inode(struct inode *inode)
 630 {
 631         struct shmem_inode_info *info = SHMEM_I(inode);
 632
 633         if (inode->i_mapping->a_ops == &shmem_aops) {
 634                 shmem_unacct_size(info->flags, inode->i_size);
 635                 inode->i_size = 0;
 636                 shmem_truncate_range(inode, 0, (loff_t)-1);
 637                 if (!list_empty(&info->swaplist)) {
 638                         mutex_lock(&shmem_swaplist_mutex);
 639                         list_del_init(&info->swaplist);
 640                         mutex_unlock(&shmem_swaplist_mutex);
 641                 }
 642         } else
 643                 kfree(info->symlink);
 644
 645         simple_xattrs_free(&info->xattrs);
 646         BUG_ON(inode->i_blocks);
 647         shmem_free_inode(inode->i_sb);
 648         clear_inode(inode);
 649 }
 650
 651 /*
 652  * If swap found in inode, free it and move page from swapcache to filecache.
 653  */
 654 static int shmem_unuse_inode(struct shmem_inode_info *info,
 655                              swp_entry_t swap, struct page **pagep)
 656 {
 657         struct address_space *mapping = info->vfs_inode.i_mapping;
 658         void *radswap;
 659         pgoff_t index;
 660         gfp_t gfp;
 661         int error = 0;
 662
 663         radswap = swp_to_radix_entry(swap);
 664         index = radix_tree_locate_item(&mapping->page_tree, radswap);
 665         if (index == -1)
 666                 return 0;
 667
 668         /*
 669          * Move _head_ to start search for next from here.
 670          * But be careful: shmem_evict_inode checks list_empty without taking
 671          * mutex, and there's an instant in list_move_tail when info->swaplist
 672          * would appear empty, if it were the only one on shmem_swaplist.
 673          */
 674         if (shmem_swaplist.next != &info->swaplist)
 675                 list_move_tail(&shmem_swaplist, &info->swaplist);
 676
 677         gfp = mapping_gfp_mask(mapping);
 678         if (shmem_should_replace_page(*pagep, gfp)) {
 679                 mutex_unlock(&shmem_swaplist_mutex);
 680                 error = shmem_replace_page(pagep, gfp, info, index);
 681                 mutex_lock(&shmem_swaplist_mutex);
 682                 /*
 683                  * We needed to drop mutex to make that restrictive page
 684                  * allocation, but the inode might have been freed while we
 685                  * dropped it: although a racing shmem_evict_inode() cannot
 686                  * complete without emptying the radix_tree, our page lock
 687                  * on this swapcache page is not enough to prevent that -
 688                  * free_swap_and_cache() of our swap entry will only
 689                  * trylock_page(), removing swap from radix_tree whatever.
 690                  *
 691                  * We must not proceed to shmem_add_to_page_cache() if the
 692                  * inode has been freed, but of course we cannot rely on
 693                  * inode or mapping or info to check that.  However, we can
 694                  * safely check if our swap entry is still in use (and here
 695                  * it can't have got reused for another page): if it's still
 696                  * in use, then the inode cannot have been freed yet, and we
 697                  * can safely proceed (if it's no longer in use, that tells
 698                  * nothing about the inode, but we don't need to unuse swap).
 699                  */
 700                 if (!page_swapcount(*pagep))
 701                         error = -ENOENT;
 702         }
 703
 704         /*
 705          * We rely on shmem_swaplist_mutex, not only to protect the swaplist,
 706          * but also to hold up shmem_evict_inode(): so inode cannot be freed
 707          * beneath us (pagelock doesn't help until the page is in pagecache).
 708          */
 709         if (!error)
 710                 error = shmem_add_to_page_cache(*pagep, mapping, index,
 711                                                 GFP_NOWAIT, radswap);
 712         if (error != -ENOMEM) {
 713                 /*
 714                  * Truncation and eviction use free_swap_and_cache(), which
 715                  * only does trylock page: if we raced, best clean up here.
 716                  */
 717                 delete_from_swap_cache(*pagep);
 718                 set_page_dirty(*pagep);
 719                 if (!error) {
 720                         spin_lock(&info->lock);
 721                         info->swapped--;
 722                         spin_unlock(&info->lock);
 723                         swap_free(swap);
 724                 }
 725                 error = 1;      /* not an error, but entry was found */
 726         }
 727         return error;
 728 }
 729
 730 /*
 731  * Search through swapped inodes to find and replace swap by page.
 732  */
 733 int shmem_unuse(swp_entry_t swap, struct page *page)
 734 {
 735         struct list_head *this, *next;
 736         struct shmem_inode_info *info;
 737         int found = 0;
 738         int error = 0;
 739
 740         /*
 741          * There's a faint possibility that swap page was replaced before
 742          * caller locked it: caller will come back later with the right page.
 743          */
 744         if (unlikely(!PageSwapCache(page) || page_private(page) != swap.val))
 745                 goto out;
 746
 747         /*
 748          * Charge page using GFP_KERNEL while we can wait, before taking
 749          * the shmem_swaplist_mutex which might hold up shmem_writepage().
 750          * Charged back to the user (not to caller) when swap account is used.
 751          */
 752         error = mem_cgroup_cache_charge(page, current->mm, GFP_KERNEL);
 753         if (error)
 754                 goto out;
 755         /* No radix_tree_preload: swap entry keeps a place for page in tree */
 756
 757         mutex_lock(&shmem_swaplist_mutex);
 758         list_for_each_safe(this, next, &shmem_swaplist) {
 759                 info = list_entry(this, struct shmem_inode_info, swaplist);
 760                 if (info->swapped)
 761                         found = shmem_unuse_inode(info, swap, &page);
 762                 else
 763                         list_del_init(&info->swaplist);
 764                 cond_resched();
 765                 if (found)
 766                         break;
 767         }
 768         mutex_unlock(&shmem_swaplist_mutex);
 769
 770         if (found < 0)
 771                 error = found;
 772 out:
 773         unlock_page(page);
 774         page_cache_release(page);
 775         return error;
 776 }
 777
 778 /*
 779  * Move the page from the page cache to the swap cache.
 780  */
 781 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
 782 {
 783         struct shmem_inode_info *info;
 784         struct address_space *mapping;
 785         struct inode *inode;
 786         swp_entry_t swap;
 787         pgoff_t index;
 788
 789         BUG_ON(!PageLocked(page));
 790         mapping = page->mapping;
 791         index = page->index;
 792         inode = mapping->host;
 793         info = SHMEM_I(inode);
 794         if (info->flags & VM_LOCKED)
 795                 goto redirty;
 796         if (!total_swap_pages)
 797                 goto redirty;
 798
 799         /*
 800          * shmem_backing_dev_info's capabilities prevent regular writeback or
 801          * sync from ever calling shmem_writepage; but a stacking filesystem
 802          * might use ->writepage of its underlying filesystem, in which case
 803          * tmpfs should write out to swap only in response to memory pressure,
 804          * and not for the writeback threads or sync.
 805          */
 806         if (!wbc->for_reclaim) {
 807                 WARN_ON_ONCE(1);        /* Still happens? Tell us about it! */
 808                 goto redirty;
 809         }
 810
 811         /*
 812          * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
 813          * value into swapfile.c, the only way we can correctly account for a
 814          * fallocated page arriving here is now to initialize it and write it.
 815          *
 816          * That's okay for a page already fallocated earlier, but if we have
 817          * not yet completed the fallocation, then (a) we want to keep track
 818          * of this page in case we have to undo it, and (b) it may not be a
 819          * good idea to continue anyway, once we're pushing into swap.  So
 820          * reactivate the page, and let shmem_fallocate() quit when too many.
 821          */
 822         if (!PageUptodate(page)) {
 823                 if (inode->i_private) {
 824                         struct shmem_falloc *shmem_falloc;
 825                         spin_lock(&inode->i_lock);
 826                         shmem_falloc = inode->i_private;
 827                         if (shmem_falloc &&
 828                             index >= shmem_falloc->start &&
 829                             index < shmem_falloc->next)
 830                                 shmem_falloc->nr_unswapped++;
 831                         else
 832                                 shmem_falloc = NULL;
 833                         spin_unlock(&inode->i_lock);
 834                         if (shmem_falloc)
 835                                 goto redirty;
 836                 }
 837                 clear_highpage(page);
 838                 flush_dcache_page(page);
 839                 SetPageUptodate(page);
 840         }
 841
 842         swap = get_swap_page();
 843         if (!swap.val)
 844                 goto redirty;
 845
 846         /*
 847          * Add inode to shmem_unuse()'s list of swapped-out inodes,
 848          * if it's not already there.  Do it now before the page is
 849          * moved to swap cache, when its pagelock no longer protects
 850          * the inode from eviction.  But don't unlock the mutex until
 851          * we've incremented swapped, because shmem_unuse_inode() will
 852          * prune a !swapped inode from the swaplist under this mutex.
 853          */
 854         mutex_lock(&shmem_swaplist_mutex);
 855         if (list_empty(&info->swaplist))
 856                 list_add_tail(&info->swaplist, &shmem_swaplist);
 857
 858         if (add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
 859                 swap_shmem_alloc(swap);
 860                 shmem_delete_from_page_cache(page, swp_to_radix_entry(swap));
 861
 862                 spin_lock(&info->lock);
 863                 info->swapped++;
 864                 shmem_recalc_inode(inode);
 865                 spin_unlock(&info->lock);
 866
 867                 mutex_unlock(&shmem_swaplist_mutex);
 868                 BUG_ON(page_mapped(page));
 869                 swap_writepage(page, wbc);
 870                 return 0;
 871         }
 872
 873         mutex_unlock(&shmem_swaplist_mutex);
 874         swapcache_free(swap, NULL);
 875 redirty:
 876         set_page_dirty(page);
 877         if (wbc->for_reclaim)
 878                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
 879         unlock_page(page);
 880         return 0;
 881 }
 882
 883 #ifdef CONFIG_NUMA
 884 #ifdef CONFIG_TMPFS
 885 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
 886 {
 887         char buffer[64];
 888
 889         if (!mpol || mpol->mode == MPOL_DEFAULT)
 890                 return;         /* show nothing */
 891
 892         mpol_to_str(buffer, sizeof(buffer), mpol, 1);
 893
 894         seq_printf(seq, ",mpol=%s", buffer);
 895 }
 896
 897 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
 898 {
 899         struct mempolicy *mpol = NULL;
 900         if (sbinfo->mpol) {
 901                 spin_lock(&sbinfo->stat_lock);  /* prevent replace/use races */
 902                 mpol = sbinfo->mpol;
 903                 mpol_get(mpol);
 904                 spin_unlock(&sbinfo->stat_lock);
 905         }
 906         return mpol;
 907 }
 908 #endif /* CONFIG_TMPFS */
 909
 910 static struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
 911                         struct shmem_inode_info *info, pgoff_t index)
 912 {
 913         struct mempolicy mpol, *spol;
 914         struct vm_area_struct pvma;
 915
 916         spol = mpol_cond_copy(&mpol,
 917                         mpol_shared_policy_lookup(&info->policy, index));
 918
 919         /* Create a pseudo vma that just contains the policy */
 920         pvma.vm_start = 0;
 921         /* Bias interleave by inode number to distribute better across nodes */
 922         pvma.vm_pgoff = index + info->vfs_inode.i_ino;
 923         pvma.vm_ops = NULL;
 924         pvma.vm_policy = spol;
 925         return swapin_readahead(swap, gfp, &pvma, 0);
 926 }
 927
 928 static struct page *shmem_alloc_page(gfp_t gfp,
 929                         struct shmem_inode_info *info, pgoff_t index)
 930 {
 931         struct vm_area_struct pvma;
 932
 933         /* Create a pseudo vma that just contains the policy */
 934         pvma.vm_start = 0;
 935         /* Bias interleave by inode number to distribute better across nodes */
 936         pvma.vm_pgoff = index + info->vfs_inode.i_ino;
 937         pvma.vm_ops = NULL;
 938         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, index);
 939
 940         /*
 941          * alloc_page_vma() will drop the shared policy reference
 942          */
 943         return alloc_page_vma(gfp, &pvma, 0);
 944 }
 945 #else /* !CONFIG_NUMA */
 946 #ifdef CONFIG_TMPFS
 947 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
 948 {
 949 }
 950 #endif /* CONFIG_TMPFS */
 951
 952 static inline struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
 953                         struct shmem_inode_info *info, pgoff_t index)
 954 {
 955         return swapin_readahead(swap, gfp, NULL, 0);
 956 }
 957
 958 static inline struct page *shmem_alloc_page(gfp_t gfp,
 959                         struct shmem_inode_info *info, pgoff_t index)
 960 {
 961         return alloc_page(gfp);
 962 }
 963 #endif /* CONFIG_NUMA */
 964
 965 #if !defined(CONFIG_NUMA) || !defined(CONFIG_TMPFS)
 966 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
 967 {
 968         return NULL;
 969 }
 970 #endif
 971
 972 /*
 973  * When a page is moved from swapcache to shmem filecache (either by the
 974  * usual swapin of shmem_getpage_gfp(), or by the less common swapoff of
 975  * shmem_unuse_inode()), it may have been read in earlier from swap, in
 976  * ignorance of the mapping it belongs to.  If that mapping has special
 977  * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
 978  * we may need to copy to a suitable page before moving to filecache.
 979  *
 980  * In a future release, this may well be extended to respect cpuset and
 981  * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
 982  * but for now it is a simple matter of zone.
 983  */
 984 static bool shmem_should_replace_page(struct page *page, gfp_t gfp)
 985 {
 986         return page_zonenum(page) > gfp_zone(gfp);
 987 }
 988
 989 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
 990                                 struct shmem_inode_info *info, pgoff_t index)
 991 {
 992         struct page *oldpage, *newpage;
 993         struct address_space *swap_mapping;
 994         pgoff_t swap_index;
 995         int error;
 996
 997         oldpage = *pagep;
 998         swap_index = page_private(oldpage);
 999         swap_mapping = page_mapping(oldpage);
1000
1001         /*
1002          * We have arrived here because our zones are constrained, so don't
1003          * limit chance of success by further cpuset and node constraints.
1004          */
1005         gfp &= ~GFP_CONSTRAINT_MASK;
1006         newpage = shmem_alloc_page(gfp, info, index);
1007         if (!newpage)
1008                 return -ENOMEM;
1009
1010         page_cache_get(newpage);
1011         copy_highpage(newpage, oldpage);
1012         flush_dcache_page(newpage);
1013
1014         __set_page_locked(newpage);
1015         SetPageUptodate(newpage);
1016         SetPageSwapBacked(newpage);
1017         set_page_private(newpage, swap_index);
1018         SetPageSwapCache(newpage);
1019
1020         /*
1021          * Our caller will very soon move newpage out of swapcache, but it's
1022          * a nice clean interface for us to replace oldpage by newpage there.
1023          */
1024         spin_lock_irq(&swap_mapping->tree_lock);
1025         error = shmem_radix_tree_replace(swap_mapping, swap_index, oldpage,
1026                                                                    newpage);
1027         if (!error) {
1028                 __inc_zone_page_state(newpage, NR_FILE_PAGES);
1029                 __dec_zone_page_state(oldpage, NR_FILE_PAGES);
1030         }
1031         spin_unlock_irq(&swap_mapping->tree_lock);
1032
1033         if (unlikely(error)) {
1034                 /*
1035                  * Is this possible?  I think not, now that our callers check
1036                  * both PageSwapCache and page_private after getting page lock;
1037                  * but be defensive.  Reverse old to newpage for clear and free.
1038                  */
1039                 oldpage = newpage;
1040         } else {
1041                 mem_cgroup_replace_page_cache(oldpage, newpage);
1042                 lru_cache_add_anon(newpage);
1043                 *pagep = newpage;
1044         }
1045
1046         ClearPageSwapCache(oldpage);
1047         set_page_private(oldpage, 0);
1048
1049         unlock_page(oldpage);
1050         page_cache_release(oldpage);
1051         page_cache_release(oldpage);
1052         return error;
1053 }
1054
1055 /*
1056  * shmem_getpage_gfp - find page in cache, or get from swap, or allocate
1057  *
1058  * If we allocate a new one we do not mark it dirty. That's up to the
1059  * vm. If we swap it in we mark it dirty since we also free the swap
1060  * entry since a page cannot live in both the swap and page cache
1061  */
1062 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
1063         struct page **pagep, enum sgp_type sgp, gfp_t gfp, int *fault_type)
1064 {
1065         struct address_space *mapping = inode->i_mapping;
1066         struct shmem_inode_info *info;
1067         struct shmem_sb_info *sbinfo;
1068         struct page *page;
1069         swp_entry_t swap;
1070         int error;
1071         int once = 0;
1072         int alloced = 0;
1073
1074         if (index > (MAX_LFS_FILESIZE >> PAGE_CACHE_SHIFT))
1075                 return -EFBIG;
1076 repeat:
1077         swap.val = 0;
1078         page = find_lock_page(mapping, index);
1079         if (radix_tree_exceptional_entry(page)) {
1080                 swap = radix_to_swp_entry(page);
1081                 page = NULL;
1082         }
1083
1084         if (sgp != SGP_WRITE && sgp != SGP_FALLOC &&
1085             ((loff_t)index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
1086                 error = -EINVAL;
1087                 goto failed;
1088         }
1089
1090         /* fallocated page? */
1091         if (page && !PageUptodate(page)) {
1092                 if (sgp != SGP_READ)
1093                         goto clear;
1094                 unlock_page(page);
1095                 page_cache_release(page);
1096                 page = NULL;
1097         }
1098         if (page || (sgp == SGP_READ && !swap.val)) {
1099                 *pagep = page;
1100                 return 0;
1101         }
1102
1103         /*
1104          * Fast cache lookup did not find it:
1105          * bring it back from swap or allocate.
1106          */
1107         info = SHMEM_I(inode);
1108         sbinfo = SHMEM_SB(inode->i_sb);
1109
1110         if (swap.val) {
1111                 /* Look it up and read it in.. */
1112                 page = lookup_swap_cache(swap);
1113                 if (!page) {
1114                         /* here we actually do the io */
1115                         if (fault_type)
1116                                 *fault_type |= VM_FAULT_MAJOR;
1117                         page = shmem_swapin(swap, gfp, info, index);
1118                         if (!page) {
1119                                 error = -ENOMEM;
1120                                 goto failed;
1121                         }
1122                 }
1123
1124                 /* We have to do this with page locked to prevent races */
1125                 lock_page(page);
1126                 if (!PageSwapCache(page) || page_private(page) != swap.val ||
1127                     !shmem_confirm_swap(mapping, index, swap)) {
1128                         error = -EEXIST;        /* try again */
1129                         goto unlock;
1130                 }
1131                 if (!PageUptodate(page)) {
1132                         error = -EIO;
1133                         goto failed;
1134                 }
1135                 wait_on_page_writeback(page);
1136
1137                 if (shmem_should_replace_page(page, gfp)) {
1138                         error = shmem_replace_page(&page, gfp, info, index);
1139                         if (error)
1140                                 goto failed;
1141                 }
1142
1143                 error = mem_cgroup_cache_charge(page, current->mm,
1144                                                 gfp & GFP_RECLAIM_MASK);
1145                 if (!error) {
1146                         error = shmem_add_to_page_cache(page, mapping, index,
1147                                                 gfp, swp_to_radix_entry(swap));
1148                         /* We already confirmed swap, and make no allocation */
1149                         VM_BUG_ON(error);
1150                 }
1151                 if (error)
1152                         goto failed;
1153
1154                 spin_lock(&info->lock);
1155                 info->swapped--;
1156                 shmem_recalc_inode(inode);
1157                 spin_unlock(&info->lock);
1158
1159                 delete_from_swap_cache(page);
1160                 set_page_dirty(page);
1161                 swap_free(swap);
1162
1163         } else {
1164                 if (shmem_acct_block(info->flags)) {
1165                         error = -ENOSPC;
1166                         goto failed;
1167                 }
1168                 if (sbinfo->max_blocks) {
1169                         if (percpu_counter_compare(&sbinfo->used_blocks,
1170                                                 sbinfo->max_blocks) >= 0) {
1171                                 error = -ENOSPC;
1172                                 goto unacct;
1173                         }
1174                         percpu_counter_inc(&sbinfo->used_blocks);
1175                 }
1176
1177                 page = shmem_alloc_page(gfp, info, index);
1178                 if (!page) {
1179                         error = -ENOMEM;
1180                         goto decused;
1181                 }
1182
1183                 SetPageSwapBacked(page);
1184                 __set_page_locked(page);
1185                 error = mem_cgroup_cache_charge(page, current->mm,
1186                                                 gfp & GFP_RECLAIM_MASK);
1187                 if (error)
1188                         goto decused;
1189                 error = radix_tree_preload(gfp & GFP_RECLAIM_MASK);
1190                 if (!error) {
1191                         error = shmem_add_to_page_cache(page, mapping, index,
1192                                                         gfp, NULL);
1193                         radix_tree_preload_end();
1194                 }
1195                 if (error) {
1196                         mem_cgroup_uncharge_cache_page(page);
1197                         goto decused;
1198                 }
1199                 lru_cache_add_anon(page);
1200
1201                 spin_lock(&info->lock);
1202                 info->alloced++;
1203                 inode->i_blocks += BLOCKS_PER_PAGE;
1204                 shmem_recalc_inode(inode);
1205                 spin_unlock(&info->lock);
1206                 alloced = true;
1207
1208                 /*
1209                  * Let SGP_FALLOC use the SGP_WRITE optimization on a new page.
1210                  */
1211                 if (sgp == SGP_FALLOC)
1212                         sgp = SGP_WRITE;
1213 clear:
1214                 /*
1215                  * Let SGP_WRITE caller clear ends if write does not fill page;
1216                  * but SGP_FALLOC on a page fallocated earlier must initialize
1217                  * it now, lest undo on failure cancel our earlier guarantee.
1218                  */
1219                 if (sgp != SGP_WRITE) {
1220                         clear_highpage(page);
1221                         flush_dcache_page(page);
1222                         SetPageUptodate(page);
1223                 }
1224                 if (sgp == SGP_DIRTY)
1225                         set_page_dirty(page);
1226         }
1227
1228         /* Perhaps the file has been truncated since we checked */
1229         if (sgp != SGP_WRITE && sgp != SGP_FALLOC &&
1230             ((loff_t)index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
1231                 error = -EINVAL;
1232                 if (alloced)
1233                         goto trunc;
1234                 else
1235                         goto failed;
1236         }
1237         *pagep = page;
1238         return 0;
1239
1240         /*
1241          * Error recovery.
1242          */
1243 trunc:
1244         info = SHMEM_I(inode);
1245         ClearPageDirty(page);
1246         delete_from_page_cache(page);
1247         spin_lock(&info->lock);
1248         info->alloced--;
1249         inode->i_blocks -= BLOCKS_PER_PAGE;
1250         spin_unlock(&info->lock);
1251 decused:
1252         sbinfo = SHMEM_SB(inode->i_sb);
1253         if (sbinfo->max_blocks)
1254                 percpu_counter_add(&sbinfo->used_blocks, -1);
1255 unacct:
1256         shmem_unacct_blocks(info->flags, 1);
1257 failed:
1258         if (swap.val && error != -EINVAL &&
1259             !shmem_confirm_swap(mapping, index, swap))
1260                 error = -EEXIST;
1261 unlock:
1262         if (page) {
1263                 unlock_page(page);
1264                 page_cache_release(page);
1265         }
1266         if (error == -ENOSPC && !once++) {
1267                 info = SHMEM_I(inode);
1268                 spin_lock(&info->lock);
1269                 shmem_recalc_inode(inode);
1270                 spin_unlock(&info->lock);
1271                 goto repeat;
1272         }
1273         if (error == -EEXIST)   /* from above or from radix_tree_insert */
1274                 goto repeat;
1275         return error;
1276 }
1277
1278 static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1279 {
1280         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1281         int error;
1282         int ret = VM_FAULT_LOCKED;
1283
1284         error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_CACHE, &ret);
1285         if (error)
1286                 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1287
1288         if (ret & VM_FAULT_MAJOR) {
1289                 count_vm_event(PGMAJFAULT);
1290                 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1291         }
1292         return ret;
1293 }
1294
1295 #ifdef CONFIG_NUMA
1296 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
1297 {
1298         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1299         return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
1300 }
1301
1302 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
1303                                           unsigned long addr)
1304 {
1305         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1306         pgoff_t index;
1307
1308         index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1309         return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
1310 }
1311 #endif
1312
1313 int shmem_lock(struct file *file, int lock, struct user_struct *user)
1314 {
1315         struct inode *inode = file->f_path.dentry->d_inode;
1316         struct shmem_inode_info *info = SHMEM_I(inode);
1317         int retval = -ENOMEM;
1318
1319         spin_lock(&info->lock);
1320         if (lock && !(info->flags & VM_LOCKED)) {
1321                 if (!user_shm_lock(inode->i_size, user))
1322                         goto out_nomem;
1323                 info->flags |= VM_LOCKED;
1324                 mapping_set_unevictable(file->f_mapping);
1325         }
1326         if (!lock && (info->flags & VM_LOCKED) && user) {
1327                 user_shm_unlock(inode->i_size, user);
1328                 info->flags &= ~VM_LOCKED;
1329                 mapping_clear_unevictable(file->f_mapping);
1330         }
1331         retval = 0;
1332
1333 out_nomem:
1334         spin_unlock(&info->lock);
1335         return retval;
1336 }
1337
1338 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1339 {
1340         file_accessed(file);
1341         vma->vm_ops = &shmem_vm_ops;
1342         vma->vm_flags |= VM_CAN_NONLINEAR;
1343         return 0;
1344 }
1345
1346 static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
1347                                      umode_t mode, dev_t dev, unsigned long flags)
1348 {
1349         struct inode *inode;
1350         struct shmem_inode_info *info;
1351         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1352
1353         if (shmem_reserve_inode(sb))
1354                 return NULL;
1355
1356         inode = new_inode(sb);
1357         if (inode) {
1358                 inode->i_ino = get_next_ino();
1359                 inode_init_owner(inode, dir, mode);
1360                 inode->i_blocks = 0;
1361                 inode->i_mapping->backing_dev_info = &shmem_backing_dev_info;
1362                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1363                 inode->i_generation = get_seconds();
1364                 info = SHMEM_I(inode);
1365                 memset(info, 0, (char *)inode - (char *)info);
1366                 spin_lock_init(&info->lock);
1367                 info->flags = flags & VM_NORESERVE;
1368                 INIT_LIST_HEAD(&info->swaplist);
1369                 simple_xattrs_init(&info->xattrs);
1370                 cache_no_acl(inode);
1371
1372                 switch (mode & S_IFMT) {
1373                 default:
1374                         inode->i_op = &shmem_special_inode_operations;
1375                         init_special_inode(inode, mode, dev);
1376                         break;
1377                 case S_IFREG:
1378                         inode->i_mapping->a_ops = &shmem_aops;
1379                         inode->i_op = &shmem_inode_operations;
1380                         inode->i_fop = &shmem_file_operations;
1381                         mpol_shared_policy_init(&info->policy,
1382                                                  shmem_get_sbmpol(sbinfo));
1383                         break;
1384                 case S_IFDIR:
1385                         inc_nlink(inode);
1386                         /* Some things misbehave if size == 0 on a directory */
1387                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
1388                         inode->i_op = &shmem_dir_inode_operations;
1389                         inode->i_fop = &simple_dir_operations;
1390                         break;
1391                 case S_IFLNK:
1392                         /*
1393                          * Must not load anything in the rbtree,
1394                          * mpol_free_shared_policy will not be called.
1395                          */
1396                         mpol_shared_policy_init(&info->policy, NULL);
1397                         break;
1398                 }
1399         } else
1400                 shmem_free_inode(sb);
1401         return inode;
1402 }
1403
1404 #ifdef CONFIG_TMPFS
1405 static const struct inode_operations shmem_symlink_inode_operations;
1406 static const struct inode_operations shmem_short_symlink_operations;
1407
1408 #ifdef CONFIG_TMPFS_XATTR
1409 static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
1410 #else
1411 #define shmem_initxattrs NULL
1412 #endif
1413
1414 static int
1415 shmem_write_begin(struct file *file, struct address_space *mapping,
1416                         loff_t pos, unsigned len, unsigned flags,
1417                         struct page **pagep, void **fsdata)
1418 {
1419         struct inode *inode = mapping->host;
1420         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1421         return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1422 }
1423
1424 static int
1425 shmem_write_end(struct file *file, struct address_space *mapping,
1426                         loff_t pos, unsigned len, unsigned copied,
1427                         struct page *page, void *fsdata)
1428 {
1429         struct inode *inode = mapping->host;
1430
1431         if (pos + copied > inode->i_size)
1432                 i_size_write(inode, pos + copied);
1433
1434         if (!PageUptodate(page)) {
1435                 if (copied < PAGE_CACHE_SIZE) {
1436                         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1437                         zero_user_segments(page, 0, from,
1438                                         from + copied, PAGE_CACHE_SIZE);
1439                 }
1440                 SetPageUptodate(page);
1441         }
1442         set_page_dirty(page);
1443         unlock_page(page);
1444         page_cache_release(page);
1445
1446         return copied;
1447 }
1448
1449 static void do_shmem_file_read(struct file *filp, loff_t *ppos, read_descriptor_t *desc, read_actor_t actor)
1450 {
1451         struct inode *inode = filp->f_path.dentry->d_inode;
1452         struct address_space *mapping = inode->i_mapping;
1453         pgoff_t index;
1454         unsigned long offset;
1455         enum sgp_type sgp = SGP_READ;
1456
1457         /*
1458          * Might this read be for a stacking filesystem?  Then when reading
1459          * holes of a sparse file, we actually need to allocate those pages,
1460          * and even mark them dirty, so it cannot exceed the max_blocks limit.
1461          */
1462         if (segment_eq(get_fs(), KERNEL_DS))
1463                 sgp = SGP_DIRTY;
1464
1465         index = *ppos >> PAGE_CACHE_SHIFT;
1466         offset = *ppos & ~PAGE_CACHE_MASK;
1467
1468         for (;;) {
1469                 struct page *page = NULL;
1470                 pgoff_t end_index;
1471                 unsigned long nr, ret;
1472                 loff_t i_size = i_size_read(inode);
1473
1474                 end_index = i_size >> PAGE_CACHE_SHIFT;
1475                 if (index > end_index)
1476                         break;
1477                 if (index == end_index) {
1478                         nr = i_size & ~PAGE_CACHE_MASK;
1479                         if (nr <= offset)
1480                                 break;
1481                 }
1482
1483                 desc->error = shmem_getpage(inode, index, &page, sgp, NULL);
1484                 if (desc->error) {
1485                         if (desc->error == -EINVAL)
1486                                 desc->error = 0;
1487                         break;
1488                 }
1489                 if (page)
1490                         unlock_page(page);
1491
1492                 /*
1493                  * We must evaluate after, since reads (unlike writes)
1494                  * are called without i_mutex protection against truncate
1495                  */
1496                 nr = PAGE_CACHE_SIZE;
1497                 i_size = i_size_read(inode);
1498                 end_index = i_size >> PAGE_CACHE_SHIFT;
1499                 if (index == end_index) {
1500                         nr = i_size & ~PAGE_CACHE_MASK;
1501                         if (nr <= offset) {
1502                                 if (page)
1503                                         page_cache_release(page);
1504                                 break;
1505                         }
1506                 }
1507                 nr -= offset;
1508
1509                 if (page) {
1510                         /*
1511                          * If users can be writing to this page using arbitrary
1512                          * virtual addresses, take care about potential aliasing
1513                          * before reading the page on the kernel side.
1514                          */
1515                         if (mapping_writably_mapped(mapping))
1516                                 flush_dcache_page(page);
1517                         /*
1518                          * Mark the page accessed if we read the beginning.
1519                          */
1520                         if (!offset)
1521                                 mark_page_accessed(page);
1522                 } else {
1523                         page = ZERO_PAGE(0);
1524                         page_cache_get(page);
1525                 }
1526
1527                 /*
1528                  * Ok, we have the page, and it's up-to-date, so
1529                  * now we can copy it to user space...
1530                  *
1531                  * The actor routine returns how many bytes were actually used..
1532                  * NOTE! This may not be the same as how much of a user buffer
1533                  * we filled up (we may be padding etc), so we can only update
1534                  * "pos" here (the actor routine has to update the user buffer
1535                  * pointers and the remaining count).
1536                  */
1537                 ret = actor(desc, page, offset, nr);
1538                 offset += ret;
1539                 index += offset >> PAGE_CACHE_SHIFT;
1540                 offset &= ~PAGE_CACHE_MASK;
1541
1542                 page_cache_release(page);
1543                 if (ret != nr || !desc->count)
1544                         break;
1545
1546                 cond_resched();
1547         }
1548
1549         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1550         file_accessed(filp);
1551 }
1552
1553 static ssize_t shmem_file_aio_read(struct kiocb *iocb,
1554                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
1555 {
1556         struct file *filp = iocb->ki_filp;
1557         ssize_t retval;
1558         unsigned long seg;
1559         size_t count;
1560         loff_t *ppos = &iocb->ki_pos;
1561
1562         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1563         if (retval)
1564                 return retval;
1565
1566         for (seg = 0; seg < nr_segs; seg++) {
1567                 read_descriptor_t desc;
1568
1569                 desc.written = 0;
1570                 desc.arg.buf = iov[seg].iov_base;
1571                 desc.count = iov[seg].iov_len;
1572                 if (desc.count == 0)
1573                         continue;
1574                 desc.error = 0;
1575                 do_shmem_file_read(filp, ppos, &desc, file_read_actor);
1576                 retval += desc.written;
1577                 if (desc.error) {
1578                         retval = retval ?: desc.error;
1579                         break;
1580                 }
1581                 if (desc.count > 0)
1582                         break;
1583         }
1584         return retval;
1585 }
1586
1587 static ssize_t shmem_file_splice_read(struct file *in, loff_t *ppos,
1588                                 struct pipe_inode_info *pipe, size_t len,
1589                                 unsigned int flags)
1590 {
1591         struct address_space *mapping = in->f_mapping;
1592         struct inode *inode = mapping->host;
1593         unsigned int loff, nr_pages, req_pages;
1594         struct page *pages[PIPE_DEF_BUFFERS];
1595         struct partial_page partial[PIPE_DEF_BUFFERS];
1596         struct page *page;
1597         pgoff_t index, end_index;
1598         loff_t isize, left;
1599         int error, page_nr;
1600         struct splice_pipe_desc spd = {
1601                 .pages = pages,
1602                 .partial = partial,
1603                 .nr_pages_max = PIPE_DEF_BUFFERS,
1604                 .flags = flags,
1605                 .ops = &page_cache_pipe_buf_ops,
1606                 .spd_release = spd_release_page,
1607         };
1608
1609         isize = i_size_read(inode);
1610         if (unlikely(*ppos >= isize))
1611                 return 0;
1612
1613         left = isize - *ppos;
1614         if (unlikely(left < len))
1615                 len = left;
1616
1617         if (splice_grow_spd(pipe, &spd))
1618                 return -ENOMEM;
1619
1620         index = *ppos >> PAGE_CACHE_SHIFT;
1621         loff = *ppos & ~PAGE_CACHE_MASK;
1622         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1623         nr_pages = min(req_pages, pipe->buffers);
1624
1625         spd.nr_pages = find_get_pages_contig(mapping, index,
1626                                                 nr_pages, spd.pages);
1627         index += spd.nr_pages;
1628         error = 0;
1629
1630         while (spd.nr_pages < nr_pages) {
1631                 error = shmem_getpage(inode, index, &page, SGP_CACHE, NULL);
1632                 if (error)
1633                         break;
1634                 unlock_page(page);
1635                 spd.pages[spd.nr_pages++] = page;
1636                 index++;
1637         }
1638
1639         index = *ppos >> PAGE_CACHE_SHIFT;
1640         nr_pages = spd.nr_pages;
1641         spd.nr_pages = 0;
1642
1643         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
1644                 unsigned int this_len;
1645
1646                 if (!len)
1647                         break;
1648
1649                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
1650                 page = spd.pages[page_nr];
1651
1652                 if (!PageUptodate(page) || page->mapping != mapping) {
1653                         error = shmem_getpage(inode, index, &page,
1654                                                         SGP_CACHE, NULL);
1655                         if (error)
1656                                 break;
1657                         unlock_page(page);
1658                         page_cache_release(spd.pages[page_nr]);
1659                         spd.pages[page_nr] = page;
1660                 }
1661
1662                 isize = i_size_read(inode);
1663                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1664                 if (unlikely(!isize || index > end_index))
1665                         break;
1666
1667                 if (end_index == index) {
1668                         unsigned int plen;
1669
1670                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1671                         if (plen <= loff)
1672                                 break;
1673
1674                         this_len = min(this_len, plen - loff);
1675                         len = this_len;
1676                 }
1677
1678                 spd.partial[page_nr].offset = loff;
1679                 spd.partial[page_nr].len = this_len;
1680                 len -= this_len;
1681                 loff = 0;
1682                 spd.nr_pages++;
1683                 index++;
1684         }
1685
1686         while (page_nr < nr_pages)
1687                 page_cache_release(spd.pages[page_nr++]);
1688
1689         if (spd.nr_pages)
1690                 error = splice_to_pipe(pipe, &spd);
1691
1692         splice_shrink_spd(&spd);
1693
1694         if (error > 0) {
1695                 *ppos += error;
1696                 file_accessed(in);
1697         }
1698         return error;
1699 }
1700
1701 static long shmem_fallocate(struct file *file, int mode, loff_t offset,
1702                                                          loff_t len)
1703 {
1704         struct inode *inode = file->f_path.dentry->d_inode;
1705         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1706         struct shmem_falloc shmem_falloc;
1707         pgoff_t start, index, end;
1708         int error;
1709
1710         mutex_lock(&inode->i_mutex);
1711
1712         if (mode & FALLOC_FL_PUNCH_HOLE) {
1713                 struct address_space *mapping = file->f_mapping;
1714                 loff_t unmap_start = round_up(offset, PAGE_SIZE);
1715                 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
1716
1717                 if ((u64)unmap_end > (u64)unmap_start)
1718                         unmap_mapping_range(mapping, unmap_start,
1719                                             1 + unmap_end - unmap_start, 0);
1720                 shmem_truncate_range(inode, offset, offset + len - 1);
1721                 /* No need to unmap again: hole-punching leaves COWed pages */
1722                 error = 0;
1723                 goto out;
1724         }
1725
1726         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
1727         error = inode_newsize_ok(inode, offset + len);
1728         if (error)
1729                 goto out;
1730
1731         start = offset >> PAGE_CACHE_SHIFT;
1732         end = (offset + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1733         /* Try to avoid a swapstorm if len is impossible to satisfy */
1734         if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
1735                 error = -ENOSPC;
1736                 goto out;
1737         }
1738
1739         shmem_falloc.start = start;
1740         shmem_falloc.next  = start;
1741         shmem_falloc.nr_falloced = 0;
1742         shmem_falloc.nr_unswapped = 0;
1743         spin_lock(&inode->i_lock);
1744         inode->i_private = &shmem_falloc;
1745         spin_unlock(&inode->i_lock);
1746
1747         for (index = start; index < end; index++) {
1748                 struct page *page;
1749
1750                 /*
1751                  * Good, the fallocate(2) manpage permits EINTR: we may have
1752                  * been interrupted because we are using up too much memory.
1753                  */
1754                 if (signal_pending(current))
1755                         error = -EINTR;
1756                 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
1757                         error = -ENOMEM;
1758                 else
1759                         error = shmem_getpage(inode, index, &page, SGP_FALLOC,
1760                                                                         NULL);
1761                 if (error) {
1762                         /* Remove the !PageUptodate pages we added */
1763                         shmem_undo_range(inode,
1764                                 (loff_t)start << PAGE_CACHE_SHIFT,
1765                                 (loff_t)index << PAGE_CACHE_SHIFT, true);
1766                         goto undone;
1767                 }
1768
1769                 /*
1770                  * Inform shmem_writepage() how far we have reached.
1771                  * No need for lock or barrier: we have the page lock.
1772                  */
1773                 shmem_falloc.next++;
1774                 if (!PageUptodate(page))
1775                         shmem_falloc.nr_falloced++;
1776
1777                 /*
1778                  * If !PageUptodate, leave it that way so that freeable pages
1779                  * can be recognized if we need to rollback on error later.
1780                  * But set_page_dirty so that memory pressure will swap rather
1781                  * than free the pages we are allocating (and SGP_CACHE pages
1782                  * might still be clean: we now need to mark those dirty too).
1783                  */
1784                 set_page_dirty(page);
1785                 unlock_page(page);
1786                 page_cache_release(page);
1787                 cond_resched();
1788         }
1789
1790         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
1791                 i_size_write(inode, offset + len);
1792         inode->i_ctime = CURRENT_TIME;
1793 undone:
1794         spin_lock(&inode->i_lock);
1795         inode->i_private = NULL;
1796         spin_unlock(&inode->i_lock);
1797 out:
1798         mutex_unlock(&inode->i_mutex);
1799         return error;
1800 }
1801
1802 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
1803 {
1804         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
1805
1806         buf->f_type = TMPFS_MAGIC;
1807         buf->f_bsize = PAGE_CACHE_SIZE;
1808         buf->f_namelen = NAME_MAX;
1809         if (sbinfo->max_blocks) {
1810                 buf->f_blocks = sbinfo->max_blocks;
1811                 buf->f_bavail =
1812                 buf->f_bfree  = sbinfo->max_blocks -
1813                                 percpu_counter_sum(&sbinfo->used_blocks);
1814         }
1815         if (sbinfo->max_inodes) {
1816                 buf->f_files = sbinfo->max_inodes;
1817                 buf->f_ffree = sbinfo->free_inodes;
1818         }
1819         /* else leave those fields 0 like simple_statfs */
1820         return 0;
1821 }
1822
1823 /*
1824  * File creation. Allocate an inode, and we're done..
1825  */
1826 static int
1827 shmem_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
1828 {
1829         struct inode *inode;
1830         int error = -ENOSPC;
1831
1832         inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
1833         if (inode) {
1834                 error = security_inode_init_security(inode, dir,
1835                                                      &dentry->d_name,
1836                                                      shmem_initxattrs, NULL);
1837                 if (error) {
1838                         if (error != -EOPNOTSUPP) {
1839                                 iput(inode);
1840                                 return error;
1841                         }
1842                 }
1843 #ifdef CONFIG_TMPFS_POSIX_ACL
1844                 error = generic_acl_init(inode, dir);
1845                 if (error) {
1846                         iput(inode);
1847                         return error;
1848                 }
1849 #else
1850                 error = 0;
1851 #endif
1852                 dir->i_size += BOGO_DIRENT_SIZE;
1853                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1854                 d_instantiate(dentry, inode);
1855                 dget(dentry); /* Extra count - pin the dentry in core */
1856         }
1857         return error;
1858 }
1859
1860 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
1861 {
1862         int error;
1863
1864         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
1865                 return error;
1866         inc_nlink(dir);
1867         return 0;
1868 }
1869
1870 static int shmem_create(struct inode *dir, struct dentry *dentry, umode_t mode,
1871                 bool excl)
1872 {
1873         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
1874 }
1875
1876 /*
1877  * Link a file..
1878  */
1879 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
1880 {
1881         struct inode *inode = old_dentry->d_inode;
1882         int ret;
1883
1884         /*
1885          * No ordinary (disk based) filesystem counts links as inodes;
1886          * but each new link needs a new dentry, pinning lowmem, and
1887          * tmpfs dentries cannot be pruned until they are unlinked.
1888          */
1889         ret = shmem_reserve_inode(inode->i_sb);
1890         if (ret)
1891                 goto out;
1892
1893         dir->i_size += BOGO_DIRENT_SIZE;
1894         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1895         inc_nlink(inode);
1896         ihold(inode);   /* New dentry reference */
1897         dget(dentry);           /* Extra pinning count for the created dentry */
1898         d_instantiate(dentry, inode);
1899 out:
1900         return ret;
1901 }
1902
1903 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
1904 {
1905         struct inode *inode = dentry->d_inode;
1906
1907         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
1908                 shmem_free_inode(inode->i_sb);
1909
1910         dir->i_size -= BOGO_DIRENT_SIZE;
1911         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1912         drop_nlink(inode);
1913         dput(dentry);   /* Undo the count from "create" - this does all the work */
1914         return 0;
1915 }
1916
1917 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
1918 {
1919         if (!simple_empty(dentry))
1920                 return -ENOTEMPTY;
1921
1922         drop_nlink(dentry->d_inode);
1923         drop_nlink(dir);
1924         return shmem_unlink(dir, dentry);
1925 }
1926
1927 /*
1928  * The VFS layer already does all the dentry stuff for rename,
1929  * we just have to decrement the usage count for the target if
1930  * it exists so that the VFS layer correctly free's it when it
1931  * gets overwritten.
1932  */
1933 static int shmem_rename(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
1934 {
1935         struct inode *inode = old_dentry->d_inode;
1936         int they_are_dirs = S_ISDIR(inode->i_mode);
1937
1938         if (!simple_empty(new_dentry))
1939                 return -ENOTEMPTY;
1940
1941         if (new_dentry->d_inode) {
1942                 (void) shmem_unlink(new_dir, new_dentry);
1943                 if (they_are_dirs)
1944                         drop_nlink(old_dir);
1945         } else if (they_are_dirs) {
1946                 drop_nlink(old_dir);
1947                 inc_nlink(new_dir);
1948         }
1949
1950         old_dir->i_size -= BOGO_DIRENT_SIZE;
1951         new_dir->i_size += BOGO_DIRENT_SIZE;
1952         old_dir->i_ctime = old_dir->i_mtime =
1953         new_dir->i_ctime = new_dir->i_mtime =
1954         inode->i_ctime = CURRENT_TIME;
1955         return 0;
1956 }
1957
1958 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
1959 {
1960         int error;
1961         int len;
1962         struct inode *inode;
1963         struct page *page;
1964         char *kaddr;
1965         struct shmem_inode_info *info;
1966
1967         len = strlen(symname) + 1;
1968         if (len > PAGE_CACHE_SIZE)
1969                 return -ENAMETOOLONG;
1970
1971         inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0, VM_NORESERVE);
1972         if (!inode)
1973                 return -ENOSPC;
1974
1975         error = security_inode_init_security(inode, dir, &dentry->d_name,
1976                                              shmem_initxattrs, NULL);
1977         if (error) {
1978                 if (error != -EOPNOTSUPP) {
1979                         iput(inode);
1980                         return error;
1981                 }
1982                 error = 0;
1983         }
1984
1985         info = SHMEM_I(inode);
1986         inode->i_size = len-1;
1987         if (len <= SHORT_SYMLINK_LEN) {
1988                 info->symlink = kmemdup(symname, len, GFP_KERNEL);
1989                 if (!info->symlink) {
1990                         iput(inode);
1991                         return -ENOMEM;
1992                 }
1993                 inode->i_op = &shmem_short_symlink_operations;
1994         } else {
1995                 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
1996                 if (error) {
1997                         iput(inode);
1998                         return error;
1999                 }
2000                 inode->i_mapping->a_ops = &shmem_aops;
2001                 inode->i_op = &shmem_symlink_inode_operations;
2002                 kaddr = kmap_atomic(page);
2003                 memcpy(kaddr, symname, len);
2004                 kunmap_atomic(kaddr);
2005                 SetPageUptodate(page);
2006                 set_page_dirty(page);
2007                 unlock_page(page);
2008                 page_cache_release(page);
2009         }
2010         dir->i_size += BOGO_DIRENT_SIZE;
2011         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2012         d_instantiate(dentry, inode);
2013         dget(dentry);
2014         return 0;
2015 }
2016
2017 static void *shmem_follow_short_symlink(struct dentry *dentry, struct nameidata *nd)
2018 {
2019         nd_set_link(nd, SHMEM_I(dentry->d_inode)->symlink);
2020         return NULL;
2021 }
2022
2023 static void *shmem_follow_link(struct dentry *dentry, struct nameidata *nd)
2024 {
2025         struct page *page = NULL;
2026         int error = shmem_getpage(dentry->d_inode, 0, &page, SGP_READ, NULL);
2027         nd_set_link(nd, error ? ERR_PTR(error) : kmap(page));
2028         if (page)
2029                 unlock_page(page);
2030         return page;
2031 }
2032
2033 static void shmem_put_link(struct dentry *dentry, struct nameidata *nd, void *cookie)
2034 {
2035         if (!IS_ERR(nd_get_link(nd))) {
2036                 struct page *page = cookie;
2037                 kunmap(page);
2038                 mark_page_accessed(page);
2039                 page_cache_release(page);
2040         }
2041 }
2042
2043 #ifdef CONFIG_TMPFS_XATTR
2044 /*
2045  * Superblocks without xattr inode operations may get some security.* xattr
2046  * support from the LSM "for free". As soon as we have any other xattrs
2047  * like ACLs, we also need to implement the security.* handlers at
2048  * filesystem level, though.
2049  */
2050
2051 /*
2052  * Callback for security_inode_init_security() for acquiring xattrs.
2053  */
2054 static int shmem_initxattrs(struct inode *inode,
2055                             const struct xattr *xattr_array,
2056                             void *fs_info)
2057 {
2058         struct shmem_inode_info *info = SHMEM_I(inode);
2059         const struct xattr *xattr;
2060         struct simple_xattr *new_xattr;
2061         size_t len;
2062
2063         for (xattr = xattr_array; xattr->name != NULL; xattr++) {
2064                 new_xattr = simple_xattr_alloc(xattr->value, xattr->value_len);
2065                 if (!new_xattr)
2066                         return -ENOMEM;
2067
2068                 len = strlen(xattr->name) + 1;
2069                 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
2070                                           GFP_KERNEL);
2071                 if (!new_xattr->name) {
2072                         kfree(new_xattr);
2073                         return -ENOMEM;
2074                 }
2075
2076                 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
2077                        XATTR_SECURITY_PREFIX_LEN);
2078                 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
2079                        xattr->name, len);
2080
2081                 simple_xattr_list_add(&info->xattrs, new_xattr);
2082         }
2083
2084         return 0;
2085 }
2086
2087 static const struct xattr_handler *shmem_xattr_handlers[] = {
2088 #ifdef CONFIG_TMPFS_POSIX_ACL
2089         &generic_acl_access_handler,
2090         &generic_acl_default_handler,
2091 #endif
2092         NULL
2093 };
2094
2095 static int shmem_xattr_validate(const char *name)
2096 {
2097         struct { const char *prefix; size_t len; } arr[] = {
2098                 { XATTR_SECURITY_PREFIX, XATTR_SECURITY_PREFIX_LEN },
2099                 { XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN }
2100         };
2101         int i;
2102
2103         for (i = 0; i < ARRAY_SIZE(arr); i++) {
2104                 size_t preflen = arr[i].len;
2105                 if (strncmp(name, arr[i].prefix, preflen) == 0) {
2106                         if (!name[preflen])
2107                                 return -EINVAL;
2108                         return 0;
2109                 }
2110         }
2111         return -EOPNOTSUPP;
2112 }
2113
2114 static ssize_t shmem_getxattr(struct dentry *dentry, const char *name,
2115                               void *buffer, size_t size)
2116 {
2117         struct shmem_inode_info *info = SHMEM_I(dentry->d_inode);
2118         int err;
2119
2120         /*
2121          * If this is a request for a synthetic attribute in the system.*
2122          * namespace use the generic infrastructure to resolve a handler
2123          * for it via sb->s_xattr.
2124          */
2125         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2126                 return generic_getxattr(dentry, name, buffer, size);
2127
2128         err = shmem_xattr_validate(name);
2129         if (err)
2130                 return err;
2131
2132         return simple_xattr_get(&info->xattrs, name, buffer, size);
2133 }
2134
2135 static int shmem_setxattr(struct dentry *dentry, const char *name,
2136                           const void *value, size_t size, int flags)
2137 {
2138         struct shmem_inode_info *info = SHMEM_I(dentry->d_inode);
2139         int err;
2140
2141         /*
2142          * If this is a request for a synthetic attribute in the system.*
2143          * namespace use the generic infrastructure to resolve a handler
2144          * for it via sb->s_xattr.
2145          */
2146         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2147                 return generic_setxattr(dentry, name, value, size, flags);
2148
2149         err = shmem_xattr_validate(name);
2150         if (err)
2151                 return err;
2152
2153         return simple_xattr_set(&info->xattrs, name, value, size, flags);
2154 }
2155
2156 static int shmem_removexattr(struct dentry *dentry, const char *name)
2157 {
2158         struct shmem_inode_info *info = SHMEM_I(dentry->d_inode);
2159         int err;
2160
2161         /*
2162          * If this is a request for a synthetic attribute in the system.*
2163          * namespace use the generic infrastructure to resolve a handler
2164          * for it via sb->s_xattr.
2165          */
2166         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2167                 return generic_removexattr(dentry, name);
2168
2169         err = shmem_xattr_validate(name);
2170         if (err)
2171                 return err;
2172
2173         return simple_xattr_remove(&info->xattrs, name);
2174 }
2175
2176 static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
2177 {
2178         struct shmem_inode_info *info = SHMEM_I(dentry->d_inode);
2179         return simple_xattr_list(&info->xattrs, buffer, size);
2180 }
2181 #endif /* CONFIG_TMPFS_XATTR */
2182
2183 static const struct inode_operations shmem_short_symlink_operations = {
2184         .readlink       = generic_readlink,
2185         .follow_link    = shmem_follow_short_symlink,
2186 #ifdef CONFIG_TMPFS_XATTR
2187         .setxattr       = shmem_setxattr,
2188         .getxattr       = shmem_getxattr,
2189         .listxattr      = shmem_listxattr,
2190         .removexattr    = shmem_removexattr,
2191 #endif
2192 };
2193
2194 static const struct inode_operations shmem_symlink_inode_operations = {
2195         .readlink       = generic_readlink,
2196         .follow_link    = shmem_follow_link,
2197         .put_link       = shmem_put_link,
2198 #ifdef CONFIG_TMPFS_XATTR
2199         .setxattr       = shmem_setxattr,
2200         .getxattr       = shmem_getxattr,
2201         .listxattr      = shmem_listxattr,
2202         .removexattr    = shmem_removexattr,
2203 #endif
2204 };
2205
2206 static struct dentry *shmem_get_parent(struct dentry *child)
2207 {
2208         return ERR_PTR(-ESTALE);
2209 }
2210
2211 static int shmem_match(struct inode *ino, void *vfh)
2212 {
2213         __u32 *fh = vfh;
2214         __u64 inum = fh[2];
2215         inum = (inum << 32) | fh[1];
2216         return ino->i_ino == inum && fh[0] == ino->i_generation;
2217 }
2218
2219 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
2220                 struct fid *fid, int fh_len, int fh_type)
2221 {
2222         struct inode *inode;
2223         struct dentry *dentry = NULL;
2224         u64 inum = fid->raw[2];
2225         inum = (inum << 32) | fid->raw[1];
2226
2227         if (fh_len < 3)
2228                 return NULL;
2229
2230         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
2231                         shmem_match, fid->raw);
2232         if (inode) {
2233                 dentry = d_find_alias(inode);
2234                 iput(inode);
2235         }
2236
2237         return dentry;
2238 }
2239
2240 static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
2241                                 struct inode *parent)
2242 {
2243         if (*len < 3) {
2244                 *len = 3;
2245                 return 255;
2246         }
2247
2248         if (inode_unhashed(inode)) {
2249                 /* Unfortunately insert_inode_hash is not idempotent,
2250                  * so as we hash inodes here rather than at creation
2251                  * time, we need a lock to ensure we only try
2252                  * to do it once
2253                  */
2254                 static DEFINE_SPINLOCK(lock);
2255                 spin_lock(&lock);
2256                 if (inode_unhashed(inode))
2257                         __insert_inode_hash(inode,
2258                                             inode->i_ino + inode->i_generation);
2259                 spin_unlock(&lock);
2260         }
2261
2262         fh[0] = inode->i_generation;
2263         fh[1] = inode->i_ino;
2264         fh[2] = ((__u64)inode->i_ino) >> 32;
2265
2266         *len = 3;
2267         return 1;
2268 }
2269
2270 static const struct export_operations shmem_export_ops = {
2271         .get_parent     = shmem_get_parent,
2272         .encode_fh      = shmem_encode_fh,
2273         .fh_to_dentry   = shmem_fh_to_dentry,
2274 };
2275
2276 static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
2277                                bool remount)
2278 {
2279         char *this_char, *value, *rest;
2280         uid_t uid;
2281         gid_t gid;
2282
2283         while (options != NULL) {
2284                 this_char = options;
2285                 for (;;) {
2286                         /*
2287                          * NUL-terminate this option: unfortunately,
2288                          * mount options form a comma-separated list,
2289                          * but mpol's nodelist may also contain commas.
2290                          */
2291                         options = strchr(options, ',');
2292                         if (options == NULL)
2293                                 break;
2294                         options++;
2295                         if (!isdigit(*options)) {
2296                                 options[-1] = '\0';
2297                                 break;
2298                         }
2299                 }
2300                 if (!*this_char)
2301                         continue;
2302                 if ((value = strchr(this_char,'=')) != NULL) {
2303                         *value++ = 0;
2304                 } else {
2305                         printk(KERN_ERR
2306                             "tmpfs: No value for mount option '%s'\n",
2307                             this_char);
2308                         return 1;
2309                 }
2310
2311                 if (!strcmp(this_char,"size")) {
2312                         unsigned long long size;
2313                         size = memparse(value,&rest);
2314                         if (*rest == '%') {
2315                                 size <<= PAGE_SHIFT;
2316                                 size *= totalram_pages;
2317                                 do_div(size, 100);
2318                                 rest++;
2319                         }
2320                         if (*rest)
2321                                 goto bad_val;
2322                         sbinfo->max_blocks =
2323                                 DIV_ROUND_UP(size, PAGE_CACHE_SIZE);
2324                 } else if (!strcmp(this_char,"nr_blocks")) {
2325                         sbinfo->max_blocks = memparse(value, &rest);
2326                         if (*rest)
2327                                 goto bad_val;
2328                 } else if (!strcmp(this_char,"nr_inodes")) {
2329                         sbinfo->max_inodes = memparse(value, &rest);
2330                         if (*rest)
2331                                 goto bad_val;
2332                 } else if (!strcmp(this_char,"mode")) {
2333                         if (remount)
2334                                 continue;
2335                         sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
2336                         if (*rest)
2337                                 goto bad_val;
2338                 } else if (!strcmp(this_char,"uid")) {
2339                         if (remount)
2340                                 continue;
2341                         uid = simple_strtoul(value, &rest, 0);
2342                         if (*rest)
2343                                 goto bad_val;
2344                         sbinfo->uid = make_kuid(current_user_ns(), uid);
2345                         if (!uid_valid(sbinfo->uid))
2346                                 goto bad_val;
2347                 } else if (!strcmp(this_char,"gid")) {
2348                         if (remount)
2349                                 continue;
2350                         gid = simple_strtoul(value, &rest, 0);
2351                         if (*rest)
2352                                 goto bad_val;
2353                         sbinfo->gid = make_kgid(current_user_ns(), gid);
2354                         if (!gid_valid(sbinfo->gid))
2355                                 goto bad_val;
2356                 } else if (!strcmp(this_char,"mpol")) {
2357                         if (mpol_parse_str(value, &sbinfo->mpol, 1))
2358                                 goto bad_val;
2359                 } else {
2360                         printk(KERN_ERR "tmpfs: Bad mount option %s\n",
2361                                this_char);
2362                         return 1;
2363                 }
2364         }
2365         return 0;
2366
2367 bad_val:
2368         printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
2369                value, this_char);
2370         return 1;
2371
2372 }
2373
2374 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2375 {
2376         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2377         struct shmem_sb_info config = *sbinfo;
2378         unsigned long inodes;
2379         int error = -EINVAL;
2380
2381         if (shmem_parse_options(data, &config, true))
2382                 return error;
2383
2384         spin_lock(&sbinfo->stat_lock);
2385         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2386         if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
2387                 goto out;
2388         if (config.max_inodes < inodes)
2389                 goto out;
2390         /*
2391          * Those tests disallow limited->unlimited while any are in use;
2392          * but we must separately disallow unlimited->limited, because
2393          * in that case we have no record of how much is already in use.
2394          */
2395         if (config.max_blocks && !sbinfo->max_blocks)
2396                 goto out;
2397         if (config.max_inodes && !sbinfo->max_inodes)
2398                 goto out;
2399
2400         error = 0;
2401         sbinfo->max_blocks  = config.max_blocks;
2402         sbinfo->max_inodes  = config.max_inodes;
2403         sbinfo->free_inodes = config.max_inodes - inodes;
2404
2405         mpol_put(sbinfo->mpol);
2406         sbinfo->mpol        = config.mpol;      /* transfers initial ref */
2407 out:
2408         spin_unlock(&sbinfo->stat_lock);
2409         return error;
2410 }
2411
2412 static int shmem_show_options(struct seq_file *seq, struct dentry *root)
2413 {
2414         struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
2415
2416         if (sbinfo->max_blocks != shmem_default_max_blocks())
2417                 seq_printf(seq, ",size=%luk",
2418                         sbinfo->max_blocks << (PAGE_CACHE_SHIFT - 10));
2419         if (sbinfo->max_inodes != shmem_default_max_inodes())
2420                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
2421         if (sbinfo->mode != (S_IRWXUGO | S_ISVTX))
2422                 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
2423         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
2424                 seq_printf(seq, ",uid=%u",
2425                                 from_kuid_munged(&init_user_ns, sbinfo->uid));
2426         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
2427                 seq_printf(seq, ",gid=%u",
2428                                 from_kgid_munged(&init_user_ns, sbinfo->gid));
2429         shmem_show_mpol(seq, sbinfo->mpol);
2430         return 0;
2431 }
2432 #endif /* CONFIG_TMPFS */
2433
2434 static void shmem_put_super(struct super_block *sb)
2435 {
2436         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2437
2438         percpu_counter_destroy(&sbinfo->used_blocks);
2439         kfree(sbinfo);
2440         sb->s_fs_info = NULL;
2441 }
2442
2443 int shmem_fill_super(struct super_block *sb, void *data, int silent)
2444 {
2445         struct inode *inode;
2446         struct shmem_sb_info *sbinfo;
2447         int err = -ENOMEM;
2448
2449         /* Round up to L1_CACHE_BYTES to resist false sharing */
2450         sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
2451                                 L1_CACHE_BYTES), GFP_KERNEL);
2452         if (!sbinfo)
2453                 return -ENOMEM;
2454
2455         sbinfo->mode = S_IRWXUGO | S_ISVTX;
2456         sbinfo->uid = current_fsuid();
2457         sbinfo->gid = current_fsgid();
2458         sb->s_fs_info = sbinfo;
2459
2460 #ifdef CONFIG_TMPFS
2461         /*
2462          * Per default we only allow half of the physical ram per
2463          * tmpfs instance, limiting inodes to one per page of lowmem;
2464          * but the internal instance is left unlimited.
2465          */
2466         if (!(sb->s_flags & MS_NOUSER)) {
2467                 sbinfo->max_blocks = shmem_default_max_blocks();
2468                 sbinfo->max_inodes = shmem_default_max_inodes();
2469                 if (shmem_parse_options(data, sbinfo, false)) {
2470                         err = -EINVAL;
2471                         goto failed;
2472                 }
2473         }
2474         sb->s_export_op = &shmem_export_ops;
2475         sb->s_flags |= MS_NOSEC;
2476 #else
2477         sb->s_flags |= MS_NOUSER;
2478 #endif
2479
2480         spin_lock_init(&sbinfo->stat_lock);
2481         if (percpu_counter_init(&sbinfo->used_blocks, 0))
2482                 goto failed;
2483         sbinfo->free_inodes = sbinfo->max_inodes;
2484
2485         sb->s_maxbytes = MAX_LFS_FILESIZE;
2486         sb->s_blocksize = PAGE_CACHE_SIZE;
2487         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
2488         sb->s_magic = TMPFS_MAGIC;
2489         sb->s_op = &shmem_ops;
2490         sb->s_time_gran = 1;
2491 #ifdef CONFIG_TMPFS_XATTR
2492         sb->s_xattr = shmem_xattr_handlers;
2493 #endif
2494 #ifdef CONFIG_TMPFS_POSIX_ACL
2495         sb->s_flags |= MS_POSIXACL;
2496 #endif
2497
2498         inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
2499         if (!inode)
2500                 goto failed;
2501         inode->i_uid = sbinfo->uid;
2502         inode->i_gid = sbinfo->gid;
2503         sb->s_root = d_make_root(inode);
2504         if (!sb->s_root)
2505                 goto failed;
2506         return 0;
2507
2508 failed:
2509         shmem_put_super(sb);
2510         return err;
2511 }
2512
2513 static struct kmem_cache *shmem_inode_cachep;
2514
2515 static struct inode *shmem_alloc_inode(struct super_block *sb)
2516 {
2517         struct shmem_inode_info *info;
2518         info = kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
2519         if (!info)
2520                 return NULL;
2521         return &info->vfs_inode;
2522 }
2523
2524 static void shmem_destroy_callback(struct rcu_head *head)
2525 {
2526         struct inode *inode = container_of(head, struct inode, i_rcu);
2527         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
2528 }
2529
2530 static void shmem_destroy_inode(struct inode *inode)
2531 {
2532         if (S_ISREG(inode->i_mode))
2533                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
2534         call_rcu(&inode->i_rcu, shmem_destroy_callback);
2535 }
2536
2537 static void shmem_init_inode(void *foo)
2538 {
2539         struct shmem_inode_info *info = foo;
2540         inode_init_once(&info->vfs_inode);
2541 }
2542
2543 static int shmem_init_inodecache(void)
2544 {
2545         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
2546                                 sizeof(struct shmem_inode_info),
2547                                 0, SLAB_PANIC, shmem_init_inode);
2548         return 0;
2549 }
2550
2551 static void shmem_destroy_inodecache(void)
2552 {
2553         kmem_cache_destroy(shmem_inode_cachep);
2554 }
2555
2556 static const struct address_space_operations shmem_aops = {
2557         .writepage      = shmem_writepage,
2558         .set_page_dirty = __set_page_dirty_no_writeback,
2559 #ifdef CONFIG_TMPFS
2560         .write_begin    = shmem_write_begin,
2561         .write_end      = shmem_write_end,
2562 #endif
2563         .migratepage    = migrate_page,
2564         .error_remove_page = generic_error_remove_page,
2565 };
2566
2567 static const struct file_operations shmem_file_operations = {
2568         .mmap           = shmem_mmap,
2569 #ifdef CONFIG_TMPFS
2570         .llseek         = generic_file_llseek,
2571         .read           = do_sync_read,
2572         .write          = do_sync_write,
2573         .aio_read       = shmem_file_aio_read,
2574         .aio_write      = generic_file_aio_write,
2575         .fsync          = noop_fsync,
2576         .splice_read    = shmem_file_splice_read,
2577         .splice_write   = generic_file_splice_write,
2578         .fallocate      = shmem_fallocate,
2579 #endif
2580 };
2581
2582 static const struct inode_operations shmem_inode_operations = {
2583         .setattr        = shmem_setattr,
2584 #ifdef CONFIG_TMPFS_XATTR
2585         .setxattr       = shmem_setxattr,
2586         .getxattr       = shmem_getxattr,
2587         .listxattr      = shmem_listxattr,
2588         .removexattr    = shmem_removexattr,
2589 #endif
2590 };
2591
2592 static const struct inode_operations shmem_dir_inode_operations = {
2593 #ifdef CONFIG_TMPFS
2594         .create         = shmem_create,
2595         .lookup         = simple_lookup,
2596         .link           = shmem_link,
2597         .unlink         = shmem_unlink,
2598         .symlink        = shmem_symlink,
2599         .mkdir          = shmem_mkdir,
2600         .rmdir          = shmem_rmdir,
2601         .mknod          = shmem_mknod,
2602         .rename         = shmem_rename,
2603 #endif
2604 #ifdef CONFIG_TMPFS_XATTR
2605         .setxattr       = shmem_setxattr,
2606         .getxattr       = shmem_getxattr,
2607         .listxattr      = shmem_listxattr,
2608         .removexattr    = shmem_removexattr,
2609 #endif
2610 #ifdef CONFIG_TMPFS_POSIX_ACL
2611         .setattr        = shmem_setattr,
2612 #endif
2613 };
2614
2615 static const struct inode_operations shmem_special_inode_operations = {
2616 #ifdef CONFIG_TMPFS_XATTR
2617         .setxattr       = shmem_setxattr,
2618         .getxattr       = shmem_getxattr,
2619         .listxattr      = shmem_listxattr,
2620         .removexattr    = shmem_removexattr,
2621 #endif
2622 #ifdef CONFIG_TMPFS_POSIX_ACL
2623         .setattr        = shmem_setattr,
2624 #endif
2625 };
2626
2627 static const struct super_operations shmem_ops = {
2628         .alloc_inode    = shmem_alloc_inode,
2629         .destroy_inode  = shmem_destroy_inode,
2630 #ifdef CONFIG_TMPFS
2631         .statfs         = shmem_statfs,
2632         .remount_fs     = shmem_remount_fs,
2633         .show_options   = shmem_show_options,
2634 #endif
2635         .evict_inode    = shmem_evict_inode,
2636         .drop_inode     = generic_delete_inode,
2637         .put_super      = shmem_put_super,
2638 };
2639
2640 static const struct vm_operations_struct shmem_vm_ops = {
2641         .fault          = shmem_fault,
2642 #ifdef CONFIG_NUMA
2643         .set_policy     = shmem_set_policy,
2644         .get_policy     = shmem_get_policy,
2645 #endif
2646 };
2647
2648 static struct dentry *shmem_mount(struct file_system_type *fs_type,
2649         int flags, const char *dev_name, void *data)
2650 {
2651         return mount_nodev(fs_type, flags, data, shmem_fill_super);
2652 }
2653
2654 static struct file_system_type shmem_fs_type = {
2655         .owner          = THIS_MODULE,
2656         .name           = "tmpfs",
2657         .mount          = shmem_mount,
2658         .kill_sb        = kill_litter_super,
2659 };
2660
2661 int __init shmem_init(void)
2662 {
2663         int error;
2664
2665         error = bdi_init(&shmem_backing_dev_info);
2666         if (error)
2667                 goto out4;
2668
2669         error = shmem_init_inodecache();
2670         if (error)
2671                 goto out3;
2672
2673         error = register_filesystem(&shmem_fs_type);
2674         if (error) {
2675                 printk(KERN_ERR "Could not register tmpfs\n");
2676                 goto out2;
2677         }
2678
2679         shm_mnt = vfs_kern_mount(&shmem_fs_type, MS_NOUSER,
2680                                  shmem_fs_type.name, NULL);
2681         if (IS_ERR(shm_mnt)) {
2682                 error = PTR_ERR(shm_mnt);
2683                 printk(KERN_ERR "Could not kern_mount tmpfs\n");
2684                 goto out1;
2685         }
2686         return 0;
2687
2688 out1:
2689         unregister_filesystem(&shmem_fs_type);
2690 out2:
2691         shmem_destroy_inodecache();
2692 out3:
2693         bdi_destroy(&shmem_backing_dev_info);
2694 out4:
2695         shm_mnt = ERR_PTR(error);
2696         return error;
2697 }
2698
2699 #else /* !CONFIG_SHMEM */
2700
2701 /*
2702  * tiny-shmem: simple shmemfs and tmpfs using ramfs code
2703  *
2704  * This is intended for small system where the benefits of the full
2705  * shmem code (swap-backed and resource-limited) are outweighed by
2706  * their complexity. On systems without swap this code should be
2707  * effectively equivalent, but much lighter weight.
2708  */
2709
2710 #include <linux/ramfs.h>
2711
2712 static struct file_system_type shmem_fs_type = {
2713         .name           = "tmpfs",
2714         .mount          = ramfs_mount,
2715         .kill_sb        = kill_litter_super,
2716 };
2717
2718 int __init shmem_init(void)
2719 {
2720         BUG_ON(register_filesystem(&shmem_fs_type) != 0);
2721
2722         shm_mnt = kern_mount(&shmem_fs_type);
2723         BUG_ON(IS_ERR(shm_mnt));
2724
2725         return 0;
2726 }
2727
2728 int shmem_unuse(swp_entry_t swap, struct page *page)
2729 {
2730         return 0;
2731 }
2732
2733 int shmem_lock(struct file *file, int lock, struct user_struct *user)
2734 {
2735         return 0;
2736 }
2737
2738 void shmem_unlock_mapping(struct address_space *mapping)
2739 {
2740 }
2741
2742 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
2743 {
2744         truncate_inode_pages_range(inode->i_mapping, lstart, lend);
2745 }
2746 EXPORT_SYMBOL_GPL(shmem_truncate_range);
2747
2748 #define shmem_vm_ops                            generic_file_vm_ops
2749 #define shmem_file_operations                   ramfs_file_operations
2750 #define shmem_get_inode(sb, dir, mode, dev, flags)      ramfs_get_inode(sb, dir, mode, dev)
2751 #define shmem_acct_size(flags, size)            0
2752 #define shmem_unacct_size(flags, size)          do {} while (0)
2753
2754 #endif /* CONFIG_SHMEM */
2755
2756 /* common code */
2757
2758 /**
2759  * shmem_file_setup - get an unlinked file living in tmpfs
2760  * @name: name for dentry (to be seen in /proc/<pid>/maps
2761  * @size: size to be set for the file
2762  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
2763  */
2764 struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
2765 {
2766         int error;
2767         struct file *file;
2768         struct inode *inode;
2769         struct path path;
2770         struct dentry *root;
2771         struct qstr this;
2772
2773         if (IS_ERR(shm_mnt))
2774                 return (void *)shm_mnt;
2775
2776         if (size < 0 || size > MAX_LFS_FILESIZE)
2777                 return ERR_PTR(-EINVAL);
2778
2779         if (shmem_acct_size(flags, size))
2780                 return ERR_PTR(-ENOMEM);
2781
2782         error = -ENOMEM;
2783         this.name = name;
2784         this.len = strlen(name);
2785         this.hash = 0; /* will go */
2786         root = shm_mnt->mnt_root;
2787         path.dentry = d_alloc(root, &this);
2788         if (!path.dentry)
2789                 goto put_memory;
2790         path.mnt = mntget(shm_mnt);
2791
2792         error = -ENOSPC;
2793         inode = shmem_get_inode(root->d_sb, NULL, S_IFREG | S_IRWXUGO, 0, flags);
2794         if (!inode)
2795                 goto put_dentry;
2796
2797         d_instantiate(path.dentry, inode);
2798         inode->i_size = size;
2799         clear_nlink(inode);     /* It is unlinked */
2800 #ifndef CONFIG_MMU
2801         error = ramfs_nommu_expand_for_mapping(inode, size);
2802         if (error)
2803                 goto put_dentry;
2804 #endif
2805
2806         error = -ENFILE;
2807         file = alloc_file(&path, FMODE_WRITE | FMODE_READ,
2808                   &shmem_file_operations);
2809         if (!file)
2810                 goto put_dentry;
2811
2812         return file;
2813
2814 put_dentry:
2815         path_put(&path);
2816 put_memory:
2817         shmem_unacct_size(flags, size);
2818         return ERR_PTR(error);
2819 }
2820 EXPORT_SYMBOL_GPL(shmem_file_setup);
2821
2822 /**
2823  * shmem_zero_setup - setup a shared anonymous mapping
2824  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
2825  */
2826 int shmem_zero_setup(struct vm_area_struct *vma)
2827 {
2828         struct file *file;
2829         loff_t size = vma->vm_end - vma->vm_start;
2830
2831         file = shmem_file_setup("dev/zero", size, vma->vm_flags);
2832         if (IS_ERR(file))
2833                 return PTR_ERR(file);
2834
2835         if (vma->vm_file)
2836                 fput(vma->vm_file);
2837         vma->vm_file = file;
2838         vma->vm_ops = &shmem_vm_ops;
2839         vma->vm_flags |= VM_CAN_NONLINEAR;
2840         return 0;
2841 }
2842
2843 /**
2844  * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
2845  * @mapping:    the page's address_space
2846  * @index:      the page index
2847  * @gfp:        the page allocator flags to use if allocating
2848  *
2849  * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
2850  * with any new page allocations done using the specified allocation flags.
2851  * But read_cache_page_gfp() uses the ->readpage() method: which does not
2852  * suit tmpfs, since it may have pages in swapcache, and needs to find those
2853  * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
2854  *
2855  * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
2856  * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
2857  */
2858 struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
2859                                          pgoff_t index, gfp_t gfp)
2860 {
2861 #ifdef CONFIG_SHMEM
2862         struct inode *inode = mapping->host;
2863         struct page *page;
2864         int error;
2865
2866         BUG_ON(mapping->a_ops != &shmem_aops);
2867         error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE, gfp, NULL);
2868         if (error)
2869                 page = ERR_PTR(error);
2870         else
2871                 unlock_page(page);
2872         return page;
2873 #else
2874         /*
2875          * The tiny !SHMEM case uses ramfs without swap
2876          */
2877         return read_cache_page_gfp(mapping, index, gfp);
2878 #endif
2879 }
2880 EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);