mm/shmem.c

   1 /*
   2  * Resizable virtual memory filesystem for Linux.
   3  *
   4  * Copyright (C) 2000 Linus Torvalds.
   5  *               2000 Transmeta Corp.
   6  *               2000-2001 Christoph Rohland
   7  *               2000-2001 SAP AG
   8  *               2002 Red Hat Inc.
   9  * Copyright (C) 2002-2005 Hugh Dickins.
  10  * Copyright (C) 2002-2005 VERITAS Software Corporation.
  11  * Copyright (C) 2004 Andi Kleen, SuSE Labs
  12  *
  13  * Extended attribute support for tmpfs:
  14  * Copyright (c) 2004, Luke Kenneth Casson Leighton <lkcl@lkcl.net>
  15  * Copyright (c) 2004 Red Hat, Inc., James Morris <jmorris@redhat.com>
  16  *
  17  * tiny-shmem:
  18  * Copyright (c) 2004, 2008 Matt Mackall <mpm@selenic.com>
  19  *
  20  * This file is released under the GPL.
  21  */
  22
  23 #include <linux/fs.h>
  24 #include <linux/init.h>
  25 #include <linux/vfs.h>
  26 #include <linux/mount.h>
  27 #include <linux/pagemap.h>
  28 #include <linux/file.h>
  29 #include <linux/mm.h>
  30 #include <linux/module.h>
  31 #include <linux/percpu_counter.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/security.h>
  55 #include <linux/swapops.h>
  56 #include <linux/mempolicy.h>
  57 #include <linux/namei.h>
  58 #include <linux/ctype.h>
  59 #include <linux/migrate.h>
  60 #include <linux/highmem.h>
  61 #include <linux/seq_file.h>
  62 #include <linux/magic.h>
  63
  64 #include <asm/uaccess.h>
  65 #include <asm/div64.h>
  66 #include <asm/pgtable.h>
  67
  68 /*
  69  * The maximum size of a shmem/tmpfs file is limited by the maximum size of
  70  * its triple-indirect swap vector - see illustration at shmem_swp_entry().
  71  *
  72  * With 4kB page size, maximum file size is just over 2TB on a 32-bit kernel,
  73  * but one eighth of that on a 64-bit kernel.  With 8kB page size, maximum
  74  * file size is just over 4TB on a 64-bit kernel, but 16TB on a 32-bit kernel,
  75  * MAX_LFS_FILESIZE being then more restrictive than swap vector layout.
  76  *
  77  * We use / and * instead of shifts in the definitions below, so that the swap
  78  * vector can be tested with small even values (e.g. 20) for ENTRIES_PER_PAGE.
  79  */
  80 #define ENTRIES_PER_PAGE (PAGE_CACHE_SIZE/sizeof(unsigned long))
  81 #define ENTRIES_PER_PAGEPAGE ((unsigned long long)ENTRIES_PER_PAGE*ENTRIES_PER_PAGE)
  82
  83 #define SHMSWP_MAX_INDEX (SHMEM_NR_DIRECT + (ENTRIES_PER_PAGEPAGE/2) * (ENTRIES_PER_PAGE+1))
  84 #define SHMSWP_MAX_BYTES (SHMSWP_MAX_INDEX << PAGE_CACHE_SHIFT)
  85
  86 #define SHMEM_MAX_BYTES  min_t(unsigned long long, SHMSWP_MAX_BYTES, MAX_LFS_FILESIZE)
  87 #define SHMEM_MAX_INDEX  ((unsigned long)((SHMEM_MAX_BYTES+1) >> PAGE_CACHE_SHIFT))
  88
  89 #define BLOCKS_PER_PAGE  (PAGE_CACHE_SIZE/512)
  90 #define VM_ACCT(size)    (PAGE_CACHE_ALIGN(size) >> PAGE_SHIFT)
  91
  92 /* info->flags needs VM_flags to handle pagein/truncate races efficiently */
  93 #define SHMEM_PAGEIN     VM_READ
  94 #define SHMEM_TRUNCATE   VM_WRITE
  95
  96 /* Definition to limit shmem_truncate's steps between cond_rescheds */
  97 #define LATENCY_LIMIT    64
  98
  99 /* Pretend that each entry is of this size in directory's i_size */
 100 #define BOGO_DIRENT_SIZE 20
 101
 102 /* Flag allocation requirements to shmem_getpage and shmem_swp_alloc */
 103 enum sgp_type {
 104         SGP_READ,       /* don't exceed i_size, don't allocate page */
 105         SGP_CACHE,      /* don't exceed i_size, may allocate page */
 106         SGP_DIRTY,      /* like SGP_CACHE, but set new page dirty */
 107         SGP_WRITE,      /* may exceed i_size, may allocate page */
 108 };
 109
 110 #ifdef CONFIG_TMPFS
 111 static unsigned long shmem_default_max_blocks(void)
 112 {
 113         return totalram_pages / 2;
 114 }
 115
 116 static unsigned long shmem_default_max_inodes(void)
 117 {
 118         return min(totalram_pages - totalhigh_pages, totalram_pages / 2);
 119 }
 120 #endif
 121
 122 static int shmem_getpage(struct inode *inode, unsigned long idx,
 123                          struct page **pagep, enum sgp_type sgp, int *type);
 124
 125 static inline struct page *shmem_dir_alloc(gfp_t gfp_mask)
 126 {
 127         /*
 128          * The above definition of ENTRIES_PER_PAGE, and the use of
 129          * BLOCKS_PER_PAGE on indirect pages, assume PAGE_CACHE_SIZE:
 130          * might be reconsidered if it ever diverges from PAGE_SIZE.
 131          *
 132          * Mobility flags are masked out as swap vectors cannot move
 133          */
 134         return alloc_pages((gfp_mask & ~GFP_MOVABLE_MASK) | __GFP_ZERO,
 135                                 PAGE_CACHE_SHIFT-PAGE_SHIFT);
 136 }
 137
 138 static inline void shmem_dir_free(struct page *page)
 139 {
 140         __free_pages(page, PAGE_CACHE_SHIFT-PAGE_SHIFT);
 141 }
 142
 143 static struct page **shmem_dir_map(struct page *page)
 144 {
 145         return (struct page **)kmap_atomic(page, KM_USER0);
 146 }
 147
 148 static inline void shmem_dir_unmap(struct page **dir)
 149 {
 150         kunmap_atomic(dir, KM_USER0);
 151 }
 152
 153 static swp_entry_t *shmem_swp_map(struct page *page)
 154 {
 155         return (swp_entry_t *)kmap_atomic(page, KM_USER1);
 156 }
 157
 158 static inline void shmem_swp_balance_unmap(void)
 159 {
 160         /*
 161          * When passing a pointer to an i_direct entry, to code which
 162          * also handles indirect entries and so will shmem_swp_unmap,
 163          * we must arrange for the preempt count to remain in balance.
 164          * What kmap_atomic of a lowmem page does depends on config
 165          * and architecture, so pretend to kmap_atomic some lowmem page.
 166          */
 167         (void) kmap_atomic(ZERO_PAGE(0), KM_USER1);
 168 }
 169
 170 static inline void shmem_swp_unmap(swp_entry_t *entry)
 171 {
 172         kunmap_atomic(entry, KM_USER1);
 173 }
 174
 175 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
 176 {
 177         return sb->s_fs_info;
 178 }
 179
 180 /*
 181  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
 182  * for shared memory and for shared anonymous (/dev/zero) mappings
 183  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
 184  * consistent with the pre-accounting of private mappings ...
 185  */
 186 static inline int shmem_acct_size(unsigned long flags, loff_t size)
 187 {
 188         return (flags & VM_NORESERVE) ?
 189                 0 : security_vm_enough_memory_kern(VM_ACCT(size));
 190 }
 191
 192 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
 193 {
 194         if (!(flags & VM_NORESERVE))
 195                 vm_unacct_memory(VM_ACCT(size));
 196 }
 197
 198 /*
 199  * ... whereas tmpfs objects are accounted incrementally as
 200  * pages are allocated, in order to allow huge sparse files.
 201  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
 202  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
 203  */
 204 static inline int shmem_acct_block(unsigned long flags)
 205 {
 206         return (flags & VM_NORESERVE) ?
 207                 security_vm_enough_memory_kern(VM_ACCT(PAGE_CACHE_SIZE)) : 0;
 208 }
 209
 210 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
 211 {
 212         if (flags & VM_NORESERVE)
 213                 vm_unacct_memory(pages * VM_ACCT(PAGE_CACHE_SIZE));
 214 }
 215
 216 static const struct super_operations shmem_ops;
 217 static const struct address_space_operations shmem_aops;
 218 static const struct file_operations shmem_file_operations;
 219 static const struct inode_operations shmem_inode_operations;
 220 static const struct inode_operations shmem_dir_inode_operations;
 221 static const struct inode_operations shmem_special_inode_operations;
 222 static const struct vm_operations_struct shmem_vm_ops;
 223
 224 static struct backing_dev_info shmem_backing_dev_info  __read_mostly = {
 225         .ra_pages       = 0,    /* No readahead */
 226         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_SWAP_BACKED,
 227 };
 228
 229 static LIST_HEAD(shmem_swaplist);
 230 static DEFINE_MUTEX(shmem_swaplist_mutex);
 231
 232 static void shmem_free_blocks(struct inode *inode, long pages)
 233 {
 234         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 235         if (sbinfo->max_blocks) {
 236                 percpu_counter_add(&sbinfo->used_blocks, -pages);
 237                 spin_lock(&inode->i_lock);
 238                 inode->i_blocks -= pages*BLOCKS_PER_PAGE;
 239                 spin_unlock(&inode->i_lock);
 240         }
 241 }
 242
 243 static int shmem_reserve_inode(struct super_block *sb)
 244 {
 245         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 246         if (sbinfo->max_inodes) {
 247                 spin_lock(&sbinfo->stat_lock);
 248                 if (!sbinfo->free_inodes) {
 249                         spin_unlock(&sbinfo->stat_lock);
 250                         return -ENOSPC;
 251                 }
 252                 sbinfo->free_inodes--;
 253                 spin_unlock(&sbinfo->stat_lock);
 254         }
 255         return 0;
 256 }
 257
 258 static void shmem_free_inode(struct super_block *sb)
 259 {
 260         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 261         if (sbinfo->max_inodes) {
 262                 spin_lock(&sbinfo->stat_lock);
 263                 sbinfo->free_inodes++;
 264                 spin_unlock(&sbinfo->stat_lock);
 265         }
 266 }
 267
 268 /**
 269  * shmem_recalc_inode - recalculate the size of an inode
 270  * @inode: inode to recalc
 271  *
 272  * We have to calculate the free blocks since the mm can drop
 273  * undirtied hole pages behind our back.
 274  *
 275  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
 276  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
 277  *
 278  * It has to be called with the spinlock held.
 279  */
 280 static void shmem_recalc_inode(struct inode *inode)
 281 {
 282         struct shmem_inode_info *info = SHMEM_I(inode);
 283         long freed;
 284
 285         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
 286         if (freed > 0) {
 287                 info->alloced -= freed;
 288                 shmem_unacct_blocks(info->flags, freed);
 289                 shmem_free_blocks(inode, freed);
 290         }
 291 }
 292
 293 /**
 294  * shmem_swp_entry - find the swap vector position in the info structure
 295  * @info:  info structure for the inode
 296  * @index: index of the page to find
 297  * @page:  optional page to add to the structure. Has to be preset to
 298  *         all zeros
 299  *
 300  * If there is no space allocated yet it will return NULL when
 301  * page is NULL, else it will use the page for the needed block,
 302  * setting it to NULL on return to indicate that it has been used.
 303  *
 304  * The swap vector is organized the following way:
 305  *
 306  * There are SHMEM_NR_DIRECT entries directly stored in the
 307  * shmem_inode_info structure. So small files do not need an addional
 308  * allocation.
 309  *
 310  * For pages with index > SHMEM_NR_DIRECT there is the pointer
 311  * i_indirect which points to a page which holds in the first half
 312  * doubly indirect blocks, in the second half triple indirect blocks:
 313  *
 314  * For an artificial ENTRIES_PER_PAGE = 4 this would lead to the
 315  * following layout (for SHMEM_NR_DIRECT == 16):
 316  *
 317  * i_indirect -> dir --> 16-19
 318  *            |      +-> 20-23
 319  *            |
 320  *            +-->dir2 --> 24-27
 321  *            |        +-> 28-31
 322  *            |        +-> 32-35
 323  *            |        +-> 36-39
 324  *            |
 325  *            +-->dir3 --> 40-43
 326  *                     +-> 44-47
 327  *                     +-> 48-51
 328  *                     +-> 52-55
 329  */
 330 static swp_entry_t *shmem_swp_entry(struct shmem_inode_info *info, unsigned long index, struct page **page)
 331 {
 332         unsigned long offset;
 333         struct page **dir;
 334         struct page *subdir;
 335
 336         if (index < SHMEM_NR_DIRECT) {
 337                 shmem_swp_balance_unmap();
 338                 return info->i_direct+index;
 339         }
 340         if (!info->i_indirect) {
 341                 if (page) {
 342                         info->i_indirect = *page;
 343                         *page = NULL;
 344                 }
 345                 return NULL;                    /* need another page */
 346         }
 347
 348         index -= SHMEM_NR_DIRECT;
 349         offset = index % ENTRIES_PER_PAGE;
 350         index /= ENTRIES_PER_PAGE;
 351         dir = shmem_dir_map(info->i_indirect);
 352
 353         if (index >= ENTRIES_PER_PAGE/2) {
 354                 index -= ENTRIES_PER_PAGE/2;
 355                 dir += ENTRIES_PER_PAGE/2 + index/ENTRIES_PER_PAGE;
 356                 index %= ENTRIES_PER_PAGE;
 357                 subdir = *dir;
 358                 if (!subdir) {
 359                         if (page) {
 360                                 *dir = *page;
 361                                 *page = NULL;
 362                         }
 363                         shmem_dir_unmap(dir);
 364                         return NULL;            /* need another page */
 365                 }
 366                 shmem_dir_unmap(dir);
 367                 dir = shmem_dir_map(subdir);
 368         }
 369
 370         dir += index;
 371         subdir = *dir;
 372         if (!subdir) {
 373                 if (!page || !(subdir = *page)) {
 374                         shmem_dir_unmap(dir);
 375                         return NULL;            /* need a page */
 376                 }
 377                 *dir = subdir;
 378                 *page = NULL;
 379         }
 380         shmem_dir_unmap(dir);
 381         return shmem_swp_map(subdir) + offset;
 382 }
 383
 384 static void shmem_swp_set(struct shmem_inode_info *info, swp_entry_t *entry, unsigned long value)
 385 {
 386         long incdec = value? 1: -1;
 387
 388         entry->val = value;
 389         info->swapped += incdec;
 390         if ((unsigned long)(entry - info->i_direct) >= SHMEM_NR_DIRECT) {
 391                 struct page *page = kmap_atomic_to_page(entry);
 392                 set_page_private(page, page_private(page) + incdec);
 393         }
 394 }
 395
 396 /**
 397  * shmem_swp_alloc - get the position of the swap entry for the page.
 398  * @info:       info structure for the inode
 399  * @index:      index of the page to find
 400  * @sgp:        check and recheck i_size? skip allocation?
 401  *
 402  * If the entry does not exist, allocate it.
 403  */
 404 static swp_entry_t *shmem_swp_alloc(struct shmem_inode_info *info, unsigned long index, enum sgp_type sgp)
 405 {
 406         struct inode *inode = &info->vfs_inode;
 407         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 408         struct page *page = NULL;
 409         swp_entry_t *entry;
 410
 411         if (sgp != SGP_WRITE &&
 412             ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode))
 413                 return ERR_PTR(-EINVAL);
 414
 415         while (!(entry = shmem_swp_entry(info, index, &page))) {
 416                 if (sgp == SGP_READ)
 417                         return shmem_swp_map(ZERO_PAGE(0));
 418                 /*
 419                  * Test used_blocks against 1 less max_blocks, since we have 1 data
 420                  * page (and perhaps indirect index pages) yet to allocate:
 421                  * a waste to allocate index if we cannot allocate data.
 422                  */
 423                 if (sbinfo->max_blocks) {
 424                         if (percpu_counter_compare(&sbinfo->used_blocks, (sbinfo->max_blocks - 1)) > 0)
 425                                 return ERR_PTR(-ENOSPC);
 426                         percpu_counter_inc(&sbinfo->used_blocks);
 427                         spin_lock(&inode->i_lock);
 428                         inode->i_blocks += BLOCKS_PER_PAGE;
 429                         spin_unlock(&inode->i_lock);
 430                 }
 431
 432                 spin_unlock(&info->lock);
 433                 page = shmem_dir_alloc(mapping_gfp_mask(inode->i_mapping));
 434                 spin_lock(&info->lock);
 435
 436                 if (!page) {
 437                         shmem_free_blocks(inode, 1);
 438                         return ERR_PTR(-ENOMEM);
 439                 }
 440                 if (sgp != SGP_WRITE &&
 441                     ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
 442                         entry = ERR_PTR(-EINVAL);
 443                         break;
 444                 }
 445                 if (info->next_index <= index)
 446                         info->next_index = index + 1;
 447         }
 448         if (page) {
 449                 /* another task gave its page, or truncated the file */
 450                 shmem_free_blocks(inode, 1);
 451                 shmem_dir_free(page);
 452         }
 453         if (info->next_index <= index && !IS_ERR(entry))
 454                 info->next_index = index + 1;
 455         return entry;
 456 }
 457
 458 /**
 459  * shmem_free_swp - free some swap entries in a directory
 460  * @dir:        pointer to the directory
 461  * @edir:       pointer after last entry of the directory
 462  * @punch_lock: pointer to spinlock when needed for the holepunch case
 463  */
 464 static int shmem_free_swp(swp_entry_t *dir, swp_entry_t *edir,
 465                                                 spinlock_t *punch_lock)
 466 {
 467         spinlock_t *punch_unlock = NULL;
 468         swp_entry_t *ptr;
 469         int freed = 0;
 470
 471         for (ptr = dir; ptr < edir; ptr++) {
 472                 if (ptr->val) {
 473                         if (unlikely(punch_lock)) {
 474                                 punch_unlock = punch_lock;
 475                                 punch_lock = NULL;
 476                                 spin_lock(punch_unlock);
 477                                 if (!ptr->val)
 478                                         continue;
 479                         }
 480                         free_swap_and_cache(*ptr);
 481                         *ptr = (swp_entry_t){0};
 482                         freed++;
 483                 }
 484         }
 485         if (punch_unlock)
 486                 spin_unlock(punch_unlock);
 487         return freed;
 488 }
 489
 490 static int shmem_map_and_free_swp(struct page *subdir, int offset,
 491                 int limit, struct page ***dir, spinlock_t *punch_lock)
 492 {
 493         swp_entry_t *ptr;
 494         int freed = 0;
 495
 496         ptr = shmem_swp_map(subdir);
 497         for (; offset < limit; offset += LATENCY_LIMIT) {
 498                 int size = limit - offset;
 499                 if (size > LATENCY_LIMIT)
 500                         size = LATENCY_LIMIT;
 501                 freed += shmem_free_swp(ptr+offset, ptr+offset+size,
 502                                                         punch_lock);
 503                 if (need_resched()) {
 504                         shmem_swp_unmap(ptr);
 505                         if (*dir) {
 506                                 shmem_dir_unmap(*dir);
 507                                 *dir = NULL;
 508                         }
 509                         cond_resched();
 510                         ptr = shmem_swp_map(subdir);
 511                 }
 512         }
 513         shmem_swp_unmap(ptr);
 514         return freed;
 515 }
 516
 517 static void shmem_free_pages(struct list_head *next)
 518 {
 519         struct page *page;
 520         int freed = 0;
 521
 522         do {
 523                 page = container_of(next, struct page, lru);
 524                 next = next->next;
 525                 shmem_dir_free(page);
 526                 freed++;
 527                 if (freed >= LATENCY_LIMIT) {
 528                         cond_resched();
 529                         freed = 0;
 530                 }
 531         } while (next);
 532 }
 533
 534 static void shmem_truncate_range(struct inode *inode, loff_t start, loff_t end)
 535 {
 536         struct shmem_inode_info *info = SHMEM_I(inode);
 537         unsigned long idx;
 538         unsigned long size;
 539         unsigned long limit;
 540         unsigned long stage;
 541         unsigned long diroff;
 542         struct page **dir;
 543         struct page *topdir;
 544         struct page *middir;
 545         struct page *subdir;
 546         swp_entry_t *ptr;
 547         LIST_HEAD(pages_to_free);
 548         long nr_pages_to_free = 0;
 549         long nr_swaps_freed = 0;
 550         int offset;
 551         int freed;
 552         int punch_hole;
 553         spinlock_t *needs_lock;
 554         spinlock_t *punch_lock;
 555         unsigned long upper_limit;
 556
 557         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 558         idx = (start + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 559         if (idx >= info->next_index)
 560                 return;
 561
 562         spin_lock(&info->lock);
 563         info->flags |= SHMEM_TRUNCATE;
 564         if (likely(end == (loff_t) -1)) {
 565                 limit = info->next_index;
 566                 upper_limit = SHMEM_MAX_INDEX;
 567                 info->next_index = idx;
 568                 needs_lock = NULL;
 569                 punch_hole = 0;
 570         } else {
 571                 if (end + 1 >= inode->i_size) { /* we may free a little more */
 572                         limit = (inode->i_size + PAGE_CACHE_SIZE - 1) >>
 573                                                         PAGE_CACHE_SHIFT;
 574                         upper_limit = SHMEM_MAX_INDEX;
 575                 } else {
 576                         limit = (end + 1) >> PAGE_CACHE_SHIFT;
 577                         upper_limit = limit;
 578                 }
 579                 needs_lock = &info->lock;
 580                 punch_hole = 1;
 581         }
 582
 583         topdir = info->i_indirect;
 584         if (topdir && idx <= SHMEM_NR_DIRECT && !punch_hole) {
 585                 info->i_indirect = NULL;
 586                 nr_pages_to_free++;
 587                 list_add(&topdir->lru, &pages_to_free);
 588         }
 589         spin_unlock(&info->lock);
 590
 591         if (info->swapped && idx < SHMEM_NR_DIRECT) {
 592                 ptr = info->i_direct;
 593                 size = limit;
 594                 if (size > SHMEM_NR_DIRECT)
 595                         size = SHMEM_NR_DIRECT;
 596                 nr_swaps_freed = shmem_free_swp(ptr+idx, ptr+size, needs_lock);
 597         }
 598
 599         /*
 600          * If there are no indirect blocks or we are punching a hole
 601          * below indirect blocks, nothing to be done.
 602          */
 603         if (!topdir || limit <= SHMEM_NR_DIRECT)
 604                 goto done2;
 605
 606         /*
 607          * The truncation case has already dropped info->lock, and we're safe
 608          * because i_size and next_index have already been lowered, preventing
 609          * access beyond.  But in the punch_hole case, we still need to take
 610          * the lock when updating the swap directory, because there might be
 611          * racing accesses by shmem_getpage(SGP_CACHE), shmem_unuse_inode or
 612          * shmem_writepage.  However, whenever we find we can remove a whole
 613          * directory page (not at the misaligned start or end of the range),
 614          * we first NULLify its pointer in the level above, and then have no
 615          * need to take the lock when updating its contents: needs_lock and
 616          * punch_lock (either pointing to info->lock or NULL) manage this.
 617          */
 618
 619         upper_limit -= SHMEM_NR_DIRECT;
 620         limit -= SHMEM_NR_DIRECT;
 621         idx = (idx > SHMEM_NR_DIRECT)? (idx - SHMEM_NR_DIRECT): 0;
 622         offset = idx % ENTRIES_PER_PAGE;
 623         idx -= offset;
 624
 625         dir = shmem_dir_map(topdir);
 626         stage = ENTRIES_PER_PAGEPAGE/2;
 627         if (idx < ENTRIES_PER_PAGEPAGE/2) {
 628                 middir = topdir;
 629                 diroff = idx/ENTRIES_PER_PAGE;
 630         } else {
 631                 dir += ENTRIES_PER_PAGE/2;
 632                 dir += (idx - ENTRIES_PER_PAGEPAGE/2)/ENTRIES_PER_PAGEPAGE;
 633                 while (stage <= idx)
 634                         stage += ENTRIES_PER_PAGEPAGE;
 635                 middir = *dir;
 636                 if (*dir) {
 637                         diroff = ((idx - ENTRIES_PER_PAGEPAGE/2) %
 638                                 ENTRIES_PER_PAGEPAGE) / ENTRIES_PER_PAGE;
 639                         if (!diroff && !offset && upper_limit >= stage) {
 640                                 if (needs_lock) {
 641                                         spin_lock(needs_lock);
 642                                         *dir = NULL;
 643                                         spin_unlock(needs_lock);
 644                                         needs_lock = NULL;
 645                                 } else
 646                                         *dir = NULL;
 647                                 nr_pages_to_free++;
 648                                 list_add(&middir->lru, &pages_to_free);
 649                         }
 650                         shmem_dir_unmap(dir);
 651                         dir = shmem_dir_map(middir);
 652                 } else {
 653                         diroff = 0;
 654                         offset = 0;
 655                         idx = stage;
 656                 }
 657         }
 658
 659         for (; idx < limit; idx += ENTRIES_PER_PAGE, diroff++) {
 660                 if (unlikely(idx == stage)) {
 661                         shmem_dir_unmap(dir);
 662                         dir = shmem_dir_map(topdir) +
 663                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
 664                         while (!*dir) {
 665                                 dir++;
 666                                 idx += ENTRIES_PER_PAGEPAGE;
 667                                 if (idx >= limit)
 668                                         goto done1;
 669                         }
 670                         stage = idx + ENTRIES_PER_PAGEPAGE;
 671                         middir = *dir;
 672                         if (punch_hole)
 673                                 needs_lock = &info->lock;
 674                         if (upper_limit >= stage) {
 675                                 if (needs_lock) {
 676                                         spin_lock(needs_lock);
 677                                         *dir = NULL;
 678                                         spin_unlock(needs_lock);
 679                                         needs_lock = NULL;
 680                                 } else
 681                                         *dir = NULL;
 682                                 nr_pages_to_free++;
 683                                 list_add(&middir->lru, &pages_to_free);
 684                         }
 685                         shmem_dir_unmap(dir);
 686                         cond_resched();
 687                         dir = shmem_dir_map(middir);
 688                         diroff = 0;
 689                 }
 690                 punch_lock = needs_lock;
 691                 subdir = dir[diroff];
 692                 if (subdir && !offset && upper_limit-idx >= ENTRIES_PER_PAGE) {
 693                         if (needs_lock) {
 694                                 spin_lock(needs_lock);
 695                                 dir[diroff] = NULL;
 696                                 spin_unlock(needs_lock);
 697                                 punch_lock = NULL;
 698                         } else
 699                                 dir[diroff] = NULL;
 700                         nr_pages_to_free++;
 701                         list_add(&subdir->lru, &pages_to_free);
 702                 }
 703                 if (subdir && page_private(subdir) /* has swap entries */) {
 704                         size = limit - idx;
 705                         if (size > ENTRIES_PER_PAGE)
 706                                 size = ENTRIES_PER_PAGE;
 707                         freed = shmem_map_and_free_swp(subdir,
 708                                         offset, size, &dir, punch_lock);
 709                         if (!dir)
 710                                 dir = shmem_dir_map(middir);
 711                         nr_swaps_freed += freed;
 712                         if (offset || punch_lock) {
 713                                 spin_lock(&info->lock);
 714                                 set_page_private(subdir,
 715                                         page_private(subdir) - freed);
 716                                 spin_unlock(&info->lock);
 717                         } else
 718                                 BUG_ON(page_private(subdir) != freed);
 719                 }
 720                 offset = 0;
 721         }
 722 done1:
 723         shmem_dir_unmap(dir);
 724 done2:
 725         if (inode->i_mapping->nrpages && (info->flags & SHMEM_PAGEIN)) {
 726                 /*
 727                  * Call truncate_inode_pages again: racing shmem_unuse_inode
 728                  * may have swizzled a page in from swap since
 729                  * truncate_pagecache or generic_delete_inode did it, before we
 730                  * lowered next_index.  Also, though shmem_getpage checks
 731                  * i_size before adding to cache, no recheck after: so fix the
 732                  * narrow window there too.
 733                  *
 734                  * Recalling truncate_inode_pages_range and unmap_mapping_range
 735                  * every time for punch_hole (which never got a chance to clear
 736                  * SHMEM_PAGEIN at the start of vmtruncate_range) is expensive,
 737                  * yet hardly ever necessary: try to optimize them out later.
 738                  */
 739                 truncate_inode_pages_range(inode->i_mapping, start, end);
 740                 if (punch_hole)
 741                         unmap_mapping_range(inode->i_mapping, start,
 742                                                         end - start, 1);
 743         }
 744
 745         spin_lock(&info->lock);
 746         info->flags &= ~SHMEM_TRUNCATE;
 747         info->swapped -= nr_swaps_freed;
 748         if (nr_pages_to_free)
 749                 shmem_free_blocks(inode, nr_pages_to_free);
 750         shmem_recalc_inode(inode);
 751         spin_unlock(&info->lock);
 752
 753         /*
 754          * Empty swap vector directory pages to be freed?
 755          */
 756         if (!list_empty(&pages_to_free)) {
 757                 pages_to_free.prev->next = NULL;
 758                 shmem_free_pages(pages_to_free.next);
 759         }
 760 }
 761
 762 static int shmem_notify_change(struct dentry *dentry, struct iattr *attr)
 763 {
 764         struct inode *inode = dentry->d_inode;
 765         loff_t newsize = attr->ia_size;
 766         int error;
 767
 768         error = inode_change_ok(inode, attr);
 769         if (error)
 770                 return error;
 771
 772         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)
 773                                         && newsize != inode->i_size) {
 774                 struct page *page = NULL;
 775
 776                 if (newsize < inode->i_size) {
 777                         /*
 778                          * If truncating down to a partial page, then
 779                          * if that page is already allocated, hold it
 780                          * in memory until the truncation is over, so
 781                          * truncate_partial_page cannot miss it were
 782                          * it assigned to swap.
 783                          */
 784                         if (newsize & (PAGE_CACHE_SIZE-1)) {
 785                                 (void) shmem_getpage(inode,
 786                                         newsize >> PAGE_CACHE_SHIFT,
 787                                                 &page, SGP_READ, NULL);
 788                                 if (page)
 789                                         unlock_page(page);
 790                         }
 791                         /*
 792                          * Reset SHMEM_PAGEIN flag so that shmem_truncate can
 793                          * detect if any pages might have been added to cache
 794                          * after truncate_inode_pages.  But we needn't bother
 795                          * if it's being fully truncated to zero-length: the
 796                          * nrpages check is efficient enough in that case.
 797                          */
 798                         if (newsize) {
 799                                 struct shmem_inode_info *info = SHMEM_I(inode);
 800                                 spin_lock(&info->lock);
 801                                 info->flags &= ~SHMEM_PAGEIN;
 802                                 spin_unlock(&info->lock);
 803                         }
 804                 }
 805
 806                 /* XXX(truncate): truncate_setsize should be called last */
 807                 truncate_setsize(inode, newsize);
 808                 if (page)
 809                         page_cache_release(page);
 810                 shmem_truncate_range(inode, newsize, (loff_t)-1);
 811         }
 812
 813         setattr_copy(inode, attr);
 814 #ifdef CONFIG_TMPFS_POSIX_ACL
 815         if (attr->ia_valid & ATTR_MODE)
 816                 error = generic_acl_chmod(inode);
 817 #endif
 818         return error;
 819 }
 820
 821 static void shmem_evict_inode(struct inode *inode)
 822 {
 823         struct shmem_inode_info *info = SHMEM_I(inode);
 824
 825         if (inode->i_mapping->a_ops == &shmem_aops) {
 826                 truncate_inode_pages(inode->i_mapping, 0);
 827                 shmem_unacct_size(info->flags, inode->i_size);
 828                 inode->i_size = 0;
 829                 shmem_truncate_range(inode, 0, (loff_t)-1);
 830                 if (!list_empty(&info->swaplist)) {
 831                         mutex_lock(&shmem_swaplist_mutex);
 832                         list_del_init(&info->swaplist);
 833                         mutex_unlock(&shmem_swaplist_mutex);
 834                 }
 835         }
 836         BUG_ON(inode->i_blocks);
 837         shmem_free_inode(inode->i_sb);
 838         end_writeback(inode);
 839 }
 840
 841 static inline int shmem_find_swp(swp_entry_t entry, swp_entry_t *dir, swp_entry_t *edir)
 842 {
 843         swp_entry_t *ptr;
 844
 845         for (ptr = dir; ptr < edir; ptr++) {
 846                 if (ptr->val == entry.val)
 847                         return ptr - dir;
 848         }
 849         return -1;
 850 }
 851
 852 static int shmem_unuse_inode(struct shmem_inode_info *info, swp_entry_t entry, struct page *page)
 853 {
 854         struct inode *inode;
 855         unsigned long idx;
 856         unsigned long size;
 857         unsigned long limit;
 858         unsigned long stage;
 859         struct page **dir;
 860         struct page *subdir;
 861         swp_entry_t *ptr;
 862         int offset;
 863         int error;
 864
 865         idx = 0;
 866         ptr = info->i_direct;
 867         spin_lock(&info->lock);
 868         if (!info->swapped) {
 869                 list_del_init(&info->swaplist);
 870                 goto lost2;
 871         }
 872         limit = info->next_index;
 873         size = limit;
 874         if (size > SHMEM_NR_DIRECT)
 875                 size = SHMEM_NR_DIRECT;
 876         offset = shmem_find_swp(entry, ptr, ptr+size);
 877         if (offset >= 0)
 878                 goto found;
 879         if (!info->i_indirect)
 880                 goto lost2;
 881
 882         dir = shmem_dir_map(info->i_indirect);
 883         stage = SHMEM_NR_DIRECT + ENTRIES_PER_PAGEPAGE/2;
 884
 885         for (idx = SHMEM_NR_DIRECT; idx < limit; idx += ENTRIES_PER_PAGE, dir++) {
 886                 if (unlikely(idx == stage)) {
 887                         shmem_dir_unmap(dir-1);
 888                         if (cond_resched_lock(&info->lock)) {
 889                                 /* check it has not been truncated */
 890                                 if (limit > info->next_index) {
 891                                         limit = info->next_index;
 892                                         if (idx >= limit)
 893                                                 goto lost2;
 894                                 }
 895                         }
 896                         dir = shmem_dir_map(info->i_indirect) +
 897                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
 898                         while (!*dir) {
 899                                 dir++;
 900                                 idx += ENTRIES_PER_PAGEPAGE;
 901                                 if (idx >= limit)
 902                                         goto lost1;
 903                         }
 904                         stage = idx + ENTRIES_PER_PAGEPAGE;
 905                         subdir = *dir;
 906                         shmem_dir_unmap(dir);
 907                         dir = shmem_dir_map(subdir);
 908                 }
 909                 subdir = *dir;
 910                 if (subdir && page_private(subdir)) {
 911                         ptr = shmem_swp_map(subdir);
 912                         size = limit - idx;
 913                         if (size > ENTRIES_PER_PAGE)
 914                                 size = ENTRIES_PER_PAGE;
 915                         offset = shmem_find_swp(entry, ptr, ptr+size);
 916                         shmem_swp_unmap(ptr);
 917                         if (offset >= 0) {
 918                                 shmem_dir_unmap(dir);
 919                                 goto found;
 920                         }
 921                 }
 922         }
 923 lost1:
 924         shmem_dir_unmap(dir-1);
 925 lost2:
 926         spin_unlock(&info->lock);
 927         return 0;
 928 found:
 929         idx += offset;
 930         inode = igrab(&info->vfs_inode);
 931         spin_unlock(&info->lock);
 932
 933         /*
 934          * Move _head_ to start search for next from here.
 935          * But be careful: shmem_evict_inode checks list_empty without taking
 936          * mutex, and there's an instant in list_move_tail when info->swaplist
 937          * would appear empty, if it were the only one on shmem_swaplist.  We
 938          * could avoid doing it if inode NULL; or use this minor optimization.
 939          */
 940         if (shmem_swaplist.next != &info->swaplist)
 941                 list_move_tail(&shmem_swaplist, &info->swaplist);
 942         mutex_unlock(&shmem_swaplist_mutex);
 943
 944         error = 1;
 945         if (!inode)
 946                 goto out;
 947         /*
 948          * Charge page using GFP_KERNEL while we can wait.
 949          * Charged back to the user(not to caller) when swap account is used.
 950          * add_to_page_cache() will be called with GFP_NOWAIT.
 951          */
 952         error = mem_cgroup_cache_charge(page, current->mm, GFP_KERNEL);
 953         if (error)
 954                 goto out;
 955         error = radix_tree_preload(GFP_KERNEL);
 956         if (error) {
 957                 mem_cgroup_uncharge_cache_page(page);
 958                 goto out;
 959         }
 960         error = 1;
 961
 962         spin_lock(&info->lock);
 963         ptr = shmem_swp_entry(info, idx, NULL);
 964         if (ptr && ptr->val == entry.val) {
 965                 error = add_to_page_cache_locked(page, inode->i_mapping,
 966                                                 idx, GFP_NOWAIT);
 967                 /* does mem_cgroup_uncharge_cache_page on error */
 968         } else  /* we must compensate for our precharge above */
 969                 mem_cgroup_uncharge_cache_page(page);
 970
 971         if (error == -EEXIST) {
 972                 struct page *filepage = find_get_page(inode->i_mapping, idx);
 973                 error = 1;
 974                 if (filepage) {
 975                         /*
 976                          * There might be a more uptodate page coming down
 977                          * from a stacked writepage: forget our swappage if so.
 978                          */
 979                         if (PageUptodate(filepage))
 980                                 error = 0;
 981                         page_cache_release(filepage);
 982                 }
 983         }
 984         if (!error) {
 985                 delete_from_swap_cache(page);
 986                 set_page_dirty(page);
 987                 info->flags |= SHMEM_PAGEIN;
 988                 shmem_swp_set(info, ptr, 0);
 989                 swap_free(entry);
 990                 error = 1;      /* not an error, but entry was found */
 991         }
 992         if (ptr)
 993                 shmem_swp_unmap(ptr);
 994         spin_unlock(&info->lock);
 995         radix_tree_preload_end();
 996 out:
 997         unlock_page(page);
 998         page_cache_release(page);
 999         iput(inode);            /* allows for NULL */
1000         return error;
1001 }
1002
1003 /*
1004  * shmem_unuse() search for an eventually swapped out shmem page.
1005  */
1006 int shmem_unuse(swp_entry_t entry, struct page *page)
1007 {
1008         struct list_head *p, *next;
1009         struct shmem_inode_info *info;
1010         int found = 0;
1011
1012         mutex_lock(&shmem_swaplist_mutex);
1013         list_for_each_safe(p, next, &shmem_swaplist) {
1014                 info = list_entry(p, struct shmem_inode_info, swaplist);
1015                 found = shmem_unuse_inode(info, entry, page);
1016                 cond_resched();
1017                 if (found)
1018                         goto out;
1019         }
1020         mutex_unlock(&shmem_swaplist_mutex);
1021         /*
1022          * Can some race bring us here?  We've been holding page lock,
1023          * so I think not; but would rather try again later than BUG()
1024          */
1025         unlock_page(page);
1026         page_cache_release(page);
1027 out:
1028         return (found < 0) ? found : 0;
1029 }
1030
1031 /*
1032  * Move the page from the page cache to the swap cache.
1033  */
1034 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1035 {
1036         struct shmem_inode_info *info;
1037         swp_entry_t *entry, swap;
1038         struct address_space *mapping;
1039         unsigned long index;
1040         struct inode *inode;
1041
1042         BUG_ON(!PageLocked(page));
1043         mapping = page->mapping;
1044         index = page->index;
1045         inode = mapping->host;
1046         info = SHMEM_I(inode);
1047         if (info->flags & VM_LOCKED)
1048                 goto redirty;
1049         if (!total_swap_pages)
1050                 goto redirty;
1051
1052         /*
1053          * shmem_backing_dev_info's capabilities prevent regular writeback or
1054          * sync from ever calling shmem_writepage; but a stacking filesystem
1055          * may use the ->writepage of its underlying filesystem, in which case
1056          * tmpfs should write out to swap only in response to memory pressure,
1057          * and not for the writeback threads or sync.  However, in those cases,
1058          * we do still want to check if there's a redundant swappage to be
1059          * discarded.
1060          */
1061         if (wbc->for_reclaim)
1062                 swap = get_swap_page();
1063         else
1064                 swap.val = 0;
1065
1066         spin_lock(&info->lock);
1067         if (index >= info->next_index) {
1068                 BUG_ON(!(info->flags & SHMEM_TRUNCATE));
1069                 goto unlock;
1070         }
1071         entry = shmem_swp_entry(info, index, NULL);
1072         if (entry->val) {
1073                 /*
1074                  * The more uptodate page coming down from a stacked
1075                  * writepage should replace our old swappage.
1076                  */
1077                 free_swap_and_cache(*entry);
1078                 shmem_swp_set(info, entry, 0);
1079         }
1080         shmem_recalc_inode(inode);
1081
1082         if (swap.val && add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
1083                 delete_from_page_cache(page);
1084                 shmem_swp_set(info, entry, swap.val);
1085                 shmem_swp_unmap(entry);
1086                 if (list_empty(&info->swaplist))
1087                         inode = igrab(inode);
1088                 else
1089                         inode = NULL;
1090                 spin_unlock(&info->lock);
1091                 swap_shmem_alloc(swap);
1092                 BUG_ON(page_mapped(page));
1093                 swap_writepage(page, wbc);
1094                 if (inode) {
1095                         mutex_lock(&shmem_swaplist_mutex);
1096                         /* move instead of add in case we're racing */
1097                         list_move_tail(&info->swaplist, &shmem_swaplist);
1098                         mutex_unlock(&shmem_swaplist_mutex);
1099                         iput(inode);
1100                 }
1101                 return 0;
1102         }
1103
1104         shmem_swp_unmap(entry);
1105 unlock:
1106         spin_unlock(&info->lock);
1107         /*
1108          * add_to_swap_cache() doesn't return -EEXIST, so we can safely
1109          * clear SWAP_HAS_CACHE flag.
1110          */
1111         swapcache_free(swap, NULL);
1112 redirty:
1113         set_page_dirty(page);
1114         if (wbc->for_reclaim)
1115                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
1116         unlock_page(page);
1117         return 0;
1118 }
1119
1120 #ifdef CONFIG_NUMA
1121 #ifdef CONFIG_TMPFS
1122 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1123 {
1124         char buffer[64];
1125
1126         if (!mpol || mpol->mode == MPOL_DEFAULT)
1127                 return;         /* show nothing */
1128
1129         mpol_to_str(buffer, sizeof(buffer), mpol, 1);
1130
1131         seq_printf(seq, ",mpol=%s", buffer);
1132 }
1133
1134 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1135 {
1136         struct mempolicy *mpol = NULL;
1137         if (sbinfo->mpol) {
1138                 spin_lock(&sbinfo->stat_lock);  /* prevent replace/use races */
1139                 mpol = sbinfo->mpol;
1140                 mpol_get(mpol);
1141                 spin_unlock(&sbinfo->stat_lock);
1142         }
1143         return mpol;
1144 }
1145 #endif /* CONFIG_TMPFS */
1146
1147 static struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1148                         struct shmem_inode_info *info, unsigned long idx)
1149 {
1150         struct mempolicy mpol, *spol;
1151         struct vm_area_struct pvma;
1152         struct page *page;
1153
1154         spol = mpol_cond_copy(&mpol,
1155                                 mpol_shared_policy_lookup(&info->policy, idx));
1156
1157         /* Create a pseudo vma that just contains the policy */
1158         pvma.vm_start = 0;
1159         pvma.vm_pgoff = idx;
1160         pvma.vm_ops = NULL;
1161         pvma.vm_policy = spol;
1162         page = swapin_readahead(entry, gfp, &pvma, 0);
1163         return page;
1164 }
1165
1166 static struct page *shmem_alloc_page(gfp_t gfp,
1167                         struct shmem_inode_info *info, unsigned long idx)
1168 {
1169         struct vm_area_struct pvma;
1170
1171         /* Create a pseudo vma that just contains the policy */
1172         pvma.vm_start = 0;
1173         pvma.vm_pgoff = idx;
1174         pvma.vm_ops = NULL;
1175         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, idx);
1176
1177         /*
1178          * alloc_page_vma() will drop the shared policy reference
1179          */
1180         return alloc_page_vma(gfp, &pvma, 0);
1181 }
1182 #else /* !CONFIG_NUMA */
1183 #ifdef CONFIG_TMPFS
1184 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *p)
1185 {
1186 }
1187 #endif /* CONFIG_TMPFS */
1188
1189 static inline struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1190                         struct shmem_inode_info *info, unsigned long idx)
1191 {
1192         return swapin_readahead(entry, gfp, NULL, 0);
1193 }
1194
1195 static inline struct page *shmem_alloc_page(gfp_t gfp,
1196                         struct shmem_inode_info *info, unsigned long idx)
1197 {
1198         return alloc_page(gfp);
1199 }
1200 #endif /* CONFIG_NUMA */
1201
1202 #if !defined(CONFIG_NUMA) || !defined(CONFIG_TMPFS)
1203 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1204 {
1205         return NULL;
1206 }
1207 #endif
1208
1209 /*
1210  * shmem_getpage - either get the page from swap or allocate a new one
1211  *
1212  * If we allocate a new one we do not mark it dirty. That's up to the
1213  * vm. If we swap it in we mark it dirty since we also free the swap
1214  * entry since a page cannot live in both the swap and page cache
1215  */
1216 static int shmem_getpage(struct inode *inode, unsigned long idx,
1217                         struct page **pagep, enum sgp_type sgp, int *type)
1218 {
1219         struct address_space *mapping = inode->i_mapping;
1220         struct shmem_inode_info *info = SHMEM_I(inode);
1221         struct shmem_sb_info *sbinfo;
1222         struct page *filepage = *pagep;
1223         struct page *swappage;
1224         struct page *prealloc_page = NULL;
1225         swp_entry_t *entry;
1226         swp_entry_t swap;
1227         gfp_t gfp;
1228         int error;
1229
1230         if (idx >= SHMEM_MAX_INDEX)
1231                 return -EFBIG;
1232
1233         if (type)
1234                 *type = 0;
1235
1236         /*
1237          * Normally, filepage is NULL on entry, and either found
1238          * uptodate immediately, or allocated and zeroed, or read
1239          * in under swappage, which is then assigned to filepage.
1240          * But shmem_readpage (required for splice) passes in a locked
1241          * filepage, which may be found not uptodate by other callers
1242          * too, and may need to be copied from the swappage read in.
1243          */
1244 repeat:
1245         if (!filepage)
1246                 filepage = find_lock_page(mapping, idx);
1247         if (filepage && PageUptodate(filepage))
1248                 goto done;
1249         gfp = mapping_gfp_mask(mapping);
1250         if (!filepage) {
1251                 /*
1252                  * Try to preload while we can wait, to not make a habit of
1253                  * draining atomic reserves; but don't latch on to this cpu.
1254                  */
1255                 error = radix_tree_preload(gfp & ~__GFP_HIGHMEM);
1256                 if (error)
1257                         goto failed;
1258                 radix_tree_preload_end();
1259                 if (sgp != SGP_READ && !prealloc_page) {
1260                         /* We don't care if this fails */
1261                         prealloc_page = shmem_alloc_page(gfp, info, idx);
1262                         if (prealloc_page) {
1263                                 if (mem_cgroup_cache_charge(prealloc_page,
1264                                                 current->mm, GFP_KERNEL)) {
1265                                         page_cache_release(prealloc_page);
1266                                         prealloc_page = NULL;
1267                                 }
1268                         }
1269                 }
1270         }
1271         error = 0;
1272
1273         spin_lock(&info->lock);
1274         shmem_recalc_inode(inode);
1275         entry = shmem_swp_alloc(info, idx, sgp);
1276         if (IS_ERR(entry)) {
1277                 spin_unlock(&info->lock);
1278                 error = PTR_ERR(entry);
1279                 goto failed;
1280         }
1281         swap = *entry;
1282
1283         if (swap.val) {
1284                 /* Look it up and read it in.. */
1285                 swappage = lookup_swap_cache(swap);
1286                 if (!swappage) {
1287                         shmem_swp_unmap(entry);
1288                         /* here we actually do the io */
1289                         if (type && !(*type & VM_FAULT_MAJOR)) {
1290                                 __count_vm_event(PGMAJFAULT);
1291                                 *type |= VM_FAULT_MAJOR;
1292                         }
1293                         spin_unlock(&info->lock);
1294                         swappage = shmem_swapin(swap, gfp, info, idx);
1295                         if (!swappage) {
1296                                 spin_lock(&info->lock);
1297                                 entry = shmem_swp_alloc(info, idx, sgp);
1298                                 if (IS_ERR(entry))
1299                                         error = PTR_ERR(entry);
1300                                 else {
1301                                         if (entry->val == swap.val)
1302                                                 error = -ENOMEM;
1303                                         shmem_swp_unmap(entry);
1304                                 }
1305                                 spin_unlock(&info->lock);
1306                                 if (error)
1307                                         goto failed;
1308                                 goto repeat;
1309                         }
1310                         wait_on_page_locked(swappage);
1311                         page_cache_release(swappage);
1312                         goto repeat;
1313                 }
1314
1315                 /* We have to do this with page locked to prevent races */
1316                 if (!trylock_page(swappage)) {
1317                         shmem_swp_unmap(entry);
1318                         spin_unlock(&info->lock);
1319                         wait_on_page_locked(swappage);
1320                         page_cache_release(swappage);
1321                         goto repeat;
1322                 }
1323                 if (PageWriteback(swappage)) {
1324                         shmem_swp_unmap(entry);
1325                         spin_unlock(&info->lock);
1326                         wait_on_page_writeback(swappage);
1327                         unlock_page(swappage);
1328                         page_cache_release(swappage);
1329                         goto repeat;
1330                 }
1331                 if (!PageUptodate(swappage)) {
1332                         shmem_swp_unmap(entry);
1333                         spin_unlock(&info->lock);
1334                         unlock_page(swappage);
1335                         page_cache_release(swappage);
1336                         error = -EIO;
1337                         goto failed;
1338                 }
1339
1340                 if (filepage) {
1341                         shmem_swp_set(info, entry, 0);
1342                         shmem_swp_unmap(entry);
1343                         delete_from_swap_cache(swappage);
1344                         spin_unlock(&info->lock);
1345                         copy_highpage(filepage, swappage);
1346                         unlock_page(swappage);
1347                         page_cache_release(swappage);
1348                         flush_dcache_page(filepage);
1349                         SetPageUptodate(filepage);
1350                         set_page_dirty(filepage);
1351                         swap_free(swap);
1352                 } else if (!(error = add_to_page_cache_locked(swappage, mapping,
1353                                         idx, GFP_NOWAIT))) {
1354                         info->flags |= SHMEM_PAGEIN;
1355                         shmem_swp_set(info, entry, 0);
1356                         shmem_swp_unmap(entry);
1357                         delete_from_swap_cache(swappage);
1358                         spin_unlock(&info->lock);
1359                         filepage = swappage;
1360                         set_page_dirty(filepage);
1361                         swap_free(swap);
1362                 } else {
1363                         shmem_swp_unmap(entry);
1364                         spin_unlock(&info->lock);
1365                         if (error == -ENOMEM) {
1366                                 /*
1367                                  * reclaim from proper memory cgroup and
1368                                  * call memcg's OOM if needed.
1369                                  */
1370                                 error = mem_cgroup_shmem_charge_fallback(
1371                                                                 swappage,
1372                                                                 current->mm,
1373                                                                 gfp);
1374                                 if (error) {
1375                                         unlock_page(swappage);
1376                                         page_cache_release(swappage);
1377                                         goto failed;
1378                                 }
1379                         }
1380                         unlock_page(swappage);
1381                         page_cache_release(swappage);
1382                         goto repeat;
1383                 }
1384         } else if (sgp == SGP_READ && !filepage) {
1385                 shmem_swp_unmap(entry);
1386                 filepage = find_get_page(mapping, idx);
1387                 if (filepage &&
1388                     (!PageUptodate(filepage) || !trylock_page(filepage))) {
1389                         spin_unlock(&info->lock);
1390                         wait_on_page_locked(filepage);
1391                         page_cache_release(filepage);
1392                         filepage = NULL;
1393                         goto repeat;
1394                 }
1395                 spin_unlock(&info->lock);
1396         } else {
1397                 shmem_swp_unmap(entry);
1398                 sbinfo = SHMEM_SB(inode->i_sb);
1399                 if (sbinfo->max_blocks) {
1400                         if ((percpu_counter_compare(&sbinfo->used_blocks, sbinfo->max_blocks) > 0) ||
1401                             shmem_acct_block(info->flags)) {
1402                                 spin_unlock(&info->lock);
1403                                 error = -ENOSPC;
1404                                 goto failed;
1405                         }
1406                         percpu_counter_inc(&sbinfo->used_blocks);
1407                         spin_lock(&inode->i_lock);
1408                         inode->i_blocks += BLOCKS_PER_PAGE;
1409                         spin_unlock(&inode->i_lock);
1410                 } else if (shmem_acct_block(info->flags)) {
1411                         spin_unlock(&info->lock);
1412                         error = -ENOSPC;
1413                         goto failed;
1414                 }
1415
1416                 if (!filepage) {
1417                         int ret;
1418
1419                         if (!prealloc_page) {
1420                                 spin_unlock(&info->lock);
1421                                 filepage = shmem_alloc_page(gfp, info, idx);
1422                                 if (!filepage) {
1423                                         shmem_unacct_blocks(info->flags, 1);
1424                                         shmem_free_blocks(inode, 1);
1425                                         error = -ENOMEM;
1426                                         goto failed;
1427                                 }
1428                                 SetPageSwapBacked(filepage);
1429
1430                                 /*
1431                                  * Precharge page while we can wait, compensate
1432                                  * after
1433                                  */
1434                                 error = mem_cgroup_cache_charge(filepage,
1435                                         current->mm, GFP_KERNEL);
1436                                 if (error) {
1437                                         page_cache_release(filepage);
1438                                         shmem_unacct_blocks(info->flags, 1);
1439                                         shmem_free_blocks(inode, 1);
1440                                         filepage = NULL;
1441                                         goto failed;
1442                                 }
1443
1444                                 spin_lock(&info->lock);
1445                         } else {
1446                                 filepage = prealloc_page;
1447                                 prealloc_page = NULL;
1448                                 SetPageSwapBacked(filepage);
1449                         }
1450
1451                         entry = shmem_swp_alloc(info, idx, sgp);
1452                         if (IS_ERR(entry))
1453                                 error = PTR_ERR(entry);
1454                         else {
1455                                 swap = *entry;
1456                                 shmem_swp_unmap(entry);
1457                         }
1458                         ret = error || swap.val;
1459                         if (ret)
1460                                 mem_cgroup_uncharge_cache_page(filepage);
1461                         else
1462                                 ret = add_to_page_cache_lru(filepage, mapping,
1463                                                 idx, GFP_NOWAIT);
1464                         /*
1465                          * At add_to_page_cache_lru() failure, uncharge will
1466                          * be done automatically.
1467                          */
1468                         if (ret) {
1469                                 spin_unlock(&info->lock);
1470                                 page_cache_release(filepage);
1471                                 shmem_unacct_blocks(info->flags, 1);
1472                                 shmem_free_blocks(inode, 1);
1473                                 filepage = NULL;
1474                                 if (error)
1475                                         goto failed;
1476                                 goto repeat;
1477                         }
1478                         info->flags |= SHMEM_PAGEIN;
1479                 }
1480
1481                 info->alloced++;
1482                 spin_unlock(&info->lock);
1483                 clear_highpage(filepage);
1484                 flush_dcache_page(filepage);
1485                 SetPageUptodate(filepage);
1486                 if (sgp == SGP_DIRTY)
1487                         set_page_dirty(filepage);
1488         }
1489 done:
1490         *pagep = filepage;
1491         error = 0;
1492         goto out;
1493
1494 failed:
1495         if (*pagep != filepage) {
1496                 unlock_page(filepage);
1497                 page_cache_release(filepage);
1498         }
1499 out:
1500         if (prealloc_page) {
1501                 mem_cgroup_uncharge_cache_page(prealloc_page);
1502                 page_cache_release(prealloc_page);
1503         }
1504         return error;
1505 }
1506
1507 static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1508 {
1509         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1510         int error;
1511         int ret;
1512
1513         if (((loff_t)vmf->pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
1514                 return VM_FAULT_SIGBUS;
1515
1516         error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_CACHE, &ret);
1517         if (error)
1518                 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1519
1520         return ret | VM_FAULT_LOCKED;
1521 }
1522
1523 #ifdef CONFIG_NUMA
1524 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *new)
1525 {
1526         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1527         return mpol_set_shared_policy(&SHMEM_I(i)->policy, vma, new);
1528 }
1529
1530 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
1531                                           unsigned long addr)
1532 {
1533         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1534         unsigned long idx;
1535
1536         idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1537         return mpol_shared_policy_lookup(&SHMEM_I(i)->policy, idx);
1538 }
1539 #endif
1540
1541 int shmem_lock(struct file *file, int lock, struct user_struct *user)
1542 {
1543         struct inode *inode = file->f_path.dentry->d_inode;
1544         struct shmem_inode_info *info = SHMEM_I(inode);
1545         int retval = -ENOMEM;
1546
1547         spin_lock(&info->lock);
1548         if (lock && !(info->flags & VM_LOCKED)) {
1549                 if (!user_shm_lock(inode->i_size, user))
1550                         goto out_nomem;
1551                 info->flags |= VM_LOCKED;
1552                 mapping_set_unevictable(file->f_mapping);
1553         }
1554         if (!lock && (info->flags & VM_LOCKED) && user) {
1555                 user_shm_unlock(inode->i_size, user);
1556                 info->flags &= ~VM_LOCKED;
1557                 mapping_clear_unevictable(file->f_mapping);
1558                 scan_mapping_unevictable_pages(file->f_mapping);
1559         }
1560         retval = 0;
1561
1562 out_nomem:
1563         spin_unlock(&info->lock);
1564         return retval;
1565 }
1566
1567 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1568 {
1569         file_accessed(file);
1570         vma->vm_ops = &shmem_vm_ops;
1571         vma->vm_flags |= VM_CAN_NONLINEAR;
1572         return 0;
1573 }
1574
1575 static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
1576                                      int mode, dev_t dev, unsigned long flags)
1577 {
1578         struct inode *inode;
1579         struct shmem_inode_info *info;
1580         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1581
1582         if (shmem_reserve_inode(sb))
1583                 return NULL;
1584
1585         inode = new_inode(sb);
1586         if (inode) {
1587                 inode->i_ino = get_next_ino();
1588                 inode_init_owner(inode, dir, mode);
1589                 inode->i_blocks = 0;
1590                 inode->i_mapping->backing_dev_info = &shmem_backing_dev_info;
1591                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1592                 inode->i_generation = get_seconds();
1593                 info = SHMEM_I(inode);
1594                 memset(info, 0, (char *)inode - (char *)info);
1595                 spin_lock_init(&info->lock);
1596                 info->flags = flags & VM_NORESERVE;
1597                 INIT_LIST_HEAD(&info->swaplist);
1598                 cache_no_acl(inode);
1599
1600                 switch (mode & S_IFMT) {
1601                 default:
1602                         inode->i_op = &shmem_special_inode_operations;
1603                         init_special_inode(inode, mode, dev);
1604                         break;
1605                 case S_IFREG:
1606                         inode->i_mapping->a_ops = &shmem_aops;
1607                         inode->i_op = &shmem_inode_operations;
1608                         inode->i_fop = &shmem_file_operations;
1609                         mpol_shared_policy_init(&info->policy,
1610                                                  shmem_get_sbmpol(sbinfo));
1611                         break;
1612                 case S_IFDIR:
1613                         inc_nlink(inode);
1614                         /* Some things misbehave if size == 0 on a directory */
1615                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
1616                         inode->i_op = &shmem_dir_inode_operations;
1617                         inode->i_fop = &simple_dir_operations;
1618                         break;
1619                 case S_IFLNK:
1620                         /*
1621                          * Must not load anything in the rbtree,
1622                          * mpol_free_shared_policy will not be called.
1623                          */
1624                         mpol_shared_policy_init(&info->policy, NULL);
1625                         break;
1626                 }
1627         } else
1628                 shmem_free_inode(sb);
1629         return inode;
1630 }
1631
1632 #ifdef CONFIG_TMPFS
1633 static const struct inode_operations shmem_symlink_inode_operations;
1634 static const struct inode_operations shmem_symlink_inline_operations;
1635
1636 /*
1637  * Normally tmpfs avoids the use of shmem_readpage and shmem_write_begin;
1638  * but providing them allows a tmpfs file to be used for splice, sendfile, and
1639  * below the loop driver, in the generic fashion that many filesystems support.
1640  */
1641 static int shmem_readpage(struct file *file, struct page *page)
1642 {
1643         struct inode *inode = page->mapping->host;
1644         int error = shmem_getpage(inode, page->index, &page, SGP_CACHE, NULL);
1645         unlock_page(page);
1646         return error;
1647 }
1648
1649 static int
1650 shmem_write_begin(struct file *file, struct address_space *mapping,
1651                         loff_t pos, unsigned len, unsigned flags,
1652                         struct page **pagep, void **fsdata)
1653 {
1654         struct inode *inode = mapping->host;
1655         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1656         *pagep = NULL;
1657         return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1658 }
1659
1660 static int
1661 shmem_write_end(struct file *file, struct address_space *mapping,
1662                         loff_t pos, unsigned len, unsigned copied,
1663                         struct page *page, void *fsdata)
1664 {
1665         struct inode *inode = mapping->host;
1666
1667         if (pos + copied > inode->i_size)
1668                 i_size_write(inode, pos + copied);
1669
1670         set_page_dirty(page);
1671         unlock_page(page);
1672         page_cache_release(page);
1673
1674         return copied;
1675 }
1676
1677 static void do_shmem_file_read(struct file *filp, loff_t *ppos, read_descriptor_t *desc, read_actor_t actor)
1678 {
1679         struct inode *inode = filp->f_path.dentry->d_inode;
1680         struct address_space *mapping = inode->i_mapping;
1681         unsigned long index, offset;
1682         enum sgp_type sgp = SGP_READ;
1683
1684         /*
1685          * Might this read be for a stacking filesystem?  Then when reading
1686          * holes of a sparse file, we actually need to allocate those pages,
1687          * and even mark them dirty, so it cannot exceed the max_blocks limit.
1688          */
1689         if (segment_eq(get_fs(), KERNEL_DS))
1690                 sgp = SGP_DIRTY;
1691
1692         index = *ppos >> PAGE_CACHE_SHIFT;
1693         offset = *ppos & ~PAGE_CACHE_MASK;
1694
1695         for (;;) {
1696                 struct page *page = NULL;
1697                 unsigned long end_index, nr, ret;
1698                 loff_t i_size = i_size_read(inode);
1699
1700                 end_index = i_size >> PAGE_CACHE_SHIFT;
1701                 if (index > end_index)
1702                         break;
1703                 if (index == end_index) {
1704                         nr = i_size & ~PAGE_CACHE_MASK;
1705                         if (nr <= offset)
1706                                 break;
1707                 }
1708
1709                 desc->error = shmem_getpage(inode, index, &page, sgp, NULL);
1710                 if (desc->error) {
1711                         if (desc->error == -EINVAL)
1712                                 desc->error = 0;
1713                         break;
1714                 }
1715                 if (page)
1716                         unlock_page(page);
1717
1718                 /*
1719                  * We must evaluate after, since reads (unlike writes)
1720                  * are called without i_mutex protection against truncate
1721                  */
1722                 nr = PAGE_CACHE_SIZE;
1723                 i_size = i_size_read(inode);
1724                 end_index = i_size >> PAGE_CACHE_SHIFT;
1725                 if (index == end_index) {
1726                         nr = i_size & ~PAGE_CACHE_MASK;
1727                         if (nr <= offset) {
1728                                 if (page)
1729                                         page_cache_release(page);
1730                                 break;
1731                         }
1732                 }
1733                 nr -= offset;
1734
1735                 if (page) {
1736                         /*
1737                          * If users can be writing to this page using arbitrary
1738                          * virtual addresses, take care about potential aliasing
1739                          * before reading the page on the kernel side.
1740                          */
1741                         if (mapping_writably_mapped(mapping))
1742                                 flush_dcache_page(page);
1743                         /*
1744                          * Mark the page accessed if we read the beginning.
1745                          */
1746                         if (!offset)
1747                                 mark_page_accessed(page);
1748                 } else {
1749                         page = ZERO_PAGE(0);
1750                         page_cache_get(page);
1751                 }
1752
1753                 /*
1754                  * Ok, we have the page, and it's up-to-date, so
1755                  * now we can copy it to user space...
1756                  *
1757                  * The actor routine returns how many bytes were actually used..
1758                  * NOTE! This may not be the same as how much of a user buffer
1759                  * we filled up (we may be padding etc), so we can only update
1760                  * "pos" here (the actor routine has to update the user buffer
1761                  * pointers and the remaining count).
1762                  */
1763                 ret = actor(desc, page, offset, nr);
1764                 offset += ret;
1765                 index += offset >> PAGE_CACHE_SHIFT;
1766                 offset &= ~PAGE_CACHE_MASK;
1767
1768                 page_cache_release(page);
1769                 if (ret != nr || !desc->count)
1770                         break;
1771
1772                 cond_resched();
1773         }
1774
1775         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1776         file_accessed(filp);
1777 }
1778
1779 static ssize_t shmem_file_aio_read(struct kiocb *iocb,
1780                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
1781 {
1782         struct file *filp = iocb->ki_filp;
1783         ssize_t retval;
1784         unsigned long seg;
1785         size_t count;
1786         loff_t *ppos = &iocb->ki_pos;
1787
1788         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1789         if (retval)
1790                 return retval;
1791
1792         for (seg = 0; seg < nr_segs; seg++) {
1793                 read_descriptor_t desc;
1794
1795                 desc.written = 0;
1796                 desc.arg.buf = iov[seg].iov_base;
1797                 desc.count = iov[seg].iov_len;
1798                 if (desc.count == 0)
1799                         continue;
1800                 desc.error = 0;
1801                 do_shmem_file_read(filp, ppos, &desc, file_read_actor);
1802                 retval += desc.written;
1803                 if (desc.error) {
1804                         retval = retval ?: desc.error;
1805                         break;
1806                 }
1807                 if (desc.count > 0)
1808                         break;
1809         }
1810         return retval;
1811 }
1812
1813 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
1814 {
1815         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
1816
1817         buf->f_type = TMPFS_MAGIC;
1818         buf->f_bsize = PAGE_CACHE_SIZE;
1819         buf->f_namelen = NAME_MAX;
1820         if (sbinfo->max_blocks) {
1821                 buf->f_blocks = sbinfo->max_blocks;
1822                 buf->f_bavail = buf->f_bfree =
1823                                 sbinfo->max_blocks - percpu_counter_sum(&sbinfo->used_blocks);
1824         }
1825         if (sbinfo->max_inodes) {
1826                 buf->f_files = sbinfo->max_inodes;
1827                 buf->f_ffree = sbinfo->free_inodes;
1828         }
1829         /* else leave those fields 0 like simple_statfs */
1830         return 0;
1831 }
1832
1833 /*
1834  * File creation. Allocate an inode, and we're done..
1835  */
1836 static int
1837 shmem_mknod(struct inode *dir, struct dentry *dentry, int mode, dev_t dev)
1838 {
1839         struct inode *inode;
1840         int error = -ENOSPC;
1841
1842         inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
1843         if (inode) {
1844                 error = security_inode_init_security(inode, dir,
1845                                                      &dentry->d_name, NULL,
1846                                                      NULL, NULL);
1847                 if (error) {
1848                         if (error != -EOPNOTSUPP) {
1849                                 iput(inode);
1850                                 return error;
1851                         }
1852                 }
1853 #ifdef CONFIG_TMPFS_POSIX_ACL
1854                 error = generic_acl_init(inode, dir);
1855                 if (error) {
1856                         iput(inode);
1857                         return error;
1858                 }
1859 #else
1860                 error = 0;
1861 #endif
1862                 dir->i_size += BOGO_DIRENT_SIZE;
1863                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1864                 d_instantiate(dentry, inode);
1865                 dget(dentry); /* Extra count - pin the dentry in core */
1866         }
1867         return error;
1868 }
1869
1870 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, int mode)
1871 {
1872         int error;
1873
1874         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
1875                 return error;
1876         inc_nlink(dir);
1877         return 0;
1878 }
1879
1880 static int shmem_create(struct inode *dir, struct dentry *dentry, int mode,
1881                 struct nameidata *nd)
1882 {
1883         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
1884 }
1885
1886 /*
1887  * Link a file..
1888  */
1889 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
1890 {
1891         struct inode *inode = old_dentry->d_inode;
1892         int ret;
1893
1894         /*
1895          * No ordinary (disk based) filesystem counts links as inodes;
1896          * but each new link needs a new dentry, pinning lowmem, and
1897          * tmpfs dentries cannot be pruned until they are unlinked.
1898          */
1899         ret = shmem_reserve_inode(inode->i_sb);
1900         if (ret)
1901                 goto out;
1902
1903         dir->i_size += BOGO_DIRENT_SIZE;
1904         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1905         inc_nlink(inode);
1906         ihold(inode);   /* New dentry reference */
1907         dget(dentry);           /* Extra pinning count for the created dentry */
1908         d_instantiate(dentry, inode);
1909 out:
1910         return ret;
1911 }
1912
1913 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
1914 {
1915         struct inode *inode = dentry->d_inode;
1916
1917         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
1918                 shmem_free_inode(inode->i_sb);
1919
1920         dir->i_size -= BOGO_DIRENT_SIZE;
1921         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1922         drop_nlink(inode);
1923         dput(dentry);   /* Undo the count from "create" - this does all the work */
1924         return 0;
1925 }
1926
1927 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
1928 {
1929         if (!simple_empty(dentry))
1930                 return -ENOTEMPTY;
1931
1932         drop_nlink(dentry->d_inode);
1933         drop_nlink(dir);
1934         return shmem_unlink(dir, dentry);
1935 }
1936
1937 /*
1938  * The VFS layer already does all the dentry stuff for rename,
1939  * we just have to decrement the usage count for the target if
1940  * it exists so that the VFS layer correctly free's it when it
1941  * gets overwritten.
1942  */
1943 static int shmem_rename(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
1944 {
1945         struct inode *inode = old_dentry->d_inode;
1946         int they_are_dirs = S_ISDIR(inode->i_mode);
1947
1948         if (!simple_empty(new_dentry))
1949                 return -ENOTEMPTY;
1950
1951         if (new_dentry->d_inode) {
1952                 (void) shmem_unlink(new_dir, new_dentry);
1953                 if (they_are_dirs)
1954                         drop_nlink(old_dir);
1955         } else if (they_are_dirs) {
1956                 drop_nlink(old_dir);
1957                 inc_nlink(new_dir);
1958         }
1959
1960         old_dir->i_size -= BOGO_DIRENT_SIZE;
1961         new_dir->i_size += BOGO_DIRENT_SIZE;
1962         old_dir->i_ctime = old_dir->i_mtime =
1963         new_dir->i_ctime = new_dir->i_mtime =
1964         inode->i_ctime = CURRENT_TIME;
1965         return 0;
1966 }
1967
1968 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
1969 {
1970         int error;
1971         int len;
1972         struct inode *inode;
1973         struct page *page = NULL;
1974         char *kaddr;
1975         struct shmem_inode_info *info;
1976
1977         len = strlen(symname) + 1;
1978         if (len > PAGE_CACHE_SIZE)
1979                 return -ENAMETOOLONG;
1980
1981         inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0, VM_NORESERVE);
1982         if (!inode)
1983                 return -ENOSPC;
1984
1985         error = security_inode_init_security(inode, dir, &dentry->d_name, NULL,
1986                                              NULL, NULL);
1987         if (error) {
1988                 if (error != -EOPNOTSUPP) {
1989                         iput(inode);
1990                         return error;
1991                 }
1992                 error = 0;
1993         }
1994
1995         info = SHMEM_I(inode);
1996         inode->i_size = len-1;
1997         if (len <= (char *)inode - (char *)info) {
1998                 /* do it inline */
1999                 memcpy(info, symname, len);
2000                 inode->i_op = &shmem_symlink_inline_operations;
2001         } else {
2002                 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
2003                 if (error) {
2004                         iput(inode);
2005                         return error;
2006                 }
2007                 inode->i_mapping->a_ops = &shmem_aops;
2008                 inode->i_op = &shmem_symlink_inode_operations;
2009                 kaddr = kmap_atomic(page, KM_USER0);
2010                 memcpy(kaddr, symname, len);
2011                 kunmap_atomic(kaddr, KM_USER0);
2012                 set_page_dirty(page);
2013                 unlock_page(page);
2014                 page_cache_release(page);
2015         }
2016         dir->i_size += BOGO_DIRENT_SIZE;
2017         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2018         d_instantiate(dentry, inode);
2019         dget(dentry);
2020         return 0;
2021 }
2022
2023 static void *shmem_follow_link_inline(struct dentry *dentry, struct nameidata *nd)
2024 {
2025         nd_set_link(nd, (char *)SHMEM_I(dentry->d_inode));
2026         return NULL;
2027 }
2028
2029 static void *shmem_follow_link(struct dentry *dentry, struct nameidata *nd)
2030 {
2031         struct page *page = NULL;
2032         int res = shmem_getpage(dentry->d_inode, 0, &page, SGP_READ, NULL);
2033         nd_set_link(nd, res ? ERR_PTR(res) : kmap(page));
2034         if (page)
2035                 unlock_page(page);
2036         return page;
2037 }
2038
2039 static void shmem_put_link(struct dentry *dentry, struct nameidata *nd, void *cookie)
2040 {
2041         if (!IS_ERR(nd_get_link(nd))) {
2042                 struct page *page = cookie;
2043                 kunmap(page);
2044                 mark_page_accessed(page);
2045                 page_cache_release(page);
2046         }
2047 }
2048
2049 static const struct inode_operations shmem_symlink_inline_operations = {
2050         .readlink       = generic_readlink,
2051         .follow_link    = shmem_follow_link_inline,
2052 };
2053
2054 static const struct inode_operations shmem_symlink_inode_operations = {
2055         .readlink       = generic_readlink,
2056         .follow_link    = shmem_follow_link,
2057         .put_link       = shmem_put_link,
2058 };
2059
2060 #ifdef CONFIG_TMPFS_POSIX_ACL
2061 /*
2062  * Superblocks without xattr inode operations will get security.* xattr
2063  * support from the VFS "for free". As soon as we have any other xattrs
2064  * like ACLs, we also need to implement the security.* handlers at
2065  * filesystem level, though.
2066  */
2067
2068 static size_t shmem_xattr_security_list(struct dentry *dentry, char *list,
2069                                         size_t list_len, const char *name,
2070                                         size_t name_len, int handler_flags)
2071 {
2072         return security_inode_listsecurity(dentry->d_inode, list, list_len);
2073 }
2074
2075 static int shmem_xattr_security_get(struct dentry *dentry, const char *name,
2076                 void *buffer, size_t size, int handler_flags)
2077 {
2078         if (strcmp(name, "") == 0)
2079                 return -EINVAL;
2080         return xattr_getsecurity(dentry->d_inode, name, buffer, size);
2081 }
2082
2083 static int shmem_xattr_security_set(struct dentry *dentry, const char *name,
2084                 const void *value, size_t size, int flags, int handler_flags)
2085 {
2086         if (strcmp(name, "") == 0)
2087                 return -EINVAL;
2088         return security_inode_setsecurity(dentry->d_inode, name, value,
2089                                           size, flags);
2090 }
2091
2092 static const struct xattr_handler shmem_xattr_security_handler = {
2093         .prefix = XATTR_SECURITY_PREFIX,
2094         .list   = shmem_xattr_security_list,
2095         .get    = shmem_xattr_security_get,
2096         .set    = shmem_xattr_security_set,
2097 };
2098
2099 static const struct xattr_handler *shmem_xattr_handlers[] = {
2100         &generic_acl_access_handler,
2101         &generic_acl_default_handler,
2102         &shmem_xattr_security_handler,
2103         NULL
2104 };
2105 #endif
2106
2107 static struct dentry *shmem_get_parent(struct dentry *child)
2108 {
2109         return ERR_PTR(-ESTALE);
2110 }
2111
2112 static int shmem_match(struct inode *ino, void *vfh)
2113 {
2114         __u32 *fh = vfh;
2115         __u64 inum = fh[2];
2116         inum = (inum << 32) | fh[1];
2117         return ino->i_ino == inum && fh[0] == ino->i_generation;
2118 }
2119
2120 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
2121                 struct fid *fid, int fh_len, int fh_type)
2122 {
2123         struct inode *inode;
2124         struct dentry *dentry = NULL;
2125         u64 inum = fid->raw[2];
2126         inum = (inum << 32) | fid->raw[1];
2127
2128         if (fh_len < 3)
2129                 return NULL;
2130
2131         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
2132                         shmem_match, fid->raw);
2133         if (inode) {
2134                 dentry = d_find_alias(inode);
2135                 iput(inode);
2136         }
2137
2138         return dentry;
2139 }
2140
2141 static int shmem_encode_fh(struct dentry *dentry, __u32 *fh, int *len,
2142                                 int connectable)
2143 {
2144         struct inode *inode = dentry->d_inode;
2145
2146         if (*len < 3) {
2147                 *len = 3;
2148                 return 255;
2149         }
2150
2151         if (inode_unhashed(inode)) {
2152                 /* Unfortunately insert_inode_hash is not idempotent,
2153                  * so as we hash inodes here rather than at creation
2154                  * time, we need a lock to ensure we only try
2155                  * to do it once
2156                  */
2157                 static DEFINE_SPINLOCK(lock);
2158                 spin_lock(&lock);
2159                 if (inode_unhashed(inode))
2160                         __insert_inode_hash(inode,
2161                                             inode->i_ino + inode->i_generation);
2162                 spin_unlock(&lock);
2163         }
2164
2165         fh[0] = inode->i_generation;
2166         fh[1] = inode->i_ino;
2167         fh[2] = ((__u64)inode->i_ino) >> 32;
2168
2169         *len = 3;
2170         return 1;
2171 }
2172
2173 static const struct export_operations shmem_export_ops = {
2174         .get_parent     = shmem_get_parent,
2175         .encode_fh      = shmem_encode_fh,
2176         .fh_to_dentry   = shmem_fh_to_dentry,
2177 };
2178
2179 static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
2180                                bool remount)
2181 {
2182         char *this_char, *value, *rest;
2183
2184         while (options != NULL) {
2185                 this_char = options;
2186                 for (;;) {
2187                         /*
2188                          * NUL-terminate this option: unfortunately,
2189                          * mount options form a comma-separated list,
2190                          * but mpol's nodelist may also contain commas.
2191                          */
2192                         options = strchr(options, ',');
2193                         if (options == NULL)
2194                                 break;
2195                         options++;
2196                         if (!isdigit(*options)) {
2197                                 options[-1] = '\0';
2198                                 break;
2199                         }
2200                 }
2201                 if (!*this_char)
2202                         continue;
2203                 if ((value = strchr(this_char,'=')) != NULL) {
2204                         *value++ = 0;
2205                 } else {
2206                         printk(KERN_ERR
2207                             "tmpfs: No value for mount option '%s'\n",
2208                             this_char);
2209                         return 1;
2210                 }
2211
2212                 if (!strcmp(this_char,"size")) {
2213                         unsigned long long size;
2214                         size = memparse(value,&rest);
2215                         if (*rest == '%') {
2216                                 size <<= PAGE_SHIFT;
2217                                 size *= totalram_pages;
2218                                 do_div(size, 100);
2219                                 rest++;
2220                         }
2221                         if (*rest)
2222                                 goto bad_val;
2223                         sbinfo->max_blocks =
2224                                 DIV_ROUND_UP(size, PAGE_CACHE_SIZE);
2225                 } else if (!strcmp(this_char,"nr_blocks")) {
2226                         sbinfo->max_blocks = memparse(value, &rest);
2227                         if (*rest)
2228                                 goto bad_val;
2229                 } else if (!strcmp(this_char,"nr_inodes")) {
2230                         sbinfo->max_inodes = memparse(value, &rest);
2231                         if (*rest)
2232                                 goto bad_val;
2233                 } else if (!strcmp(this_char,"mode")) {
2234                         if (remount)
2235                                 continue;
2236                         sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
2237                         if (*rest)
2238                                 goto bad_val;
2239                 } else if (!strcmp(this_char,"uid")) {
2240                         if (remount)
2241                                 continue;
2242                         sbinfo->uid = simple_strtoul(value, &rest, 0);
2243                         if (*rest)
2244                                 goto bad_val;
2245                 } else if (!strcmp(this_char,"gid")) {
2246                         if (remount)
2247                                 continue;
2248                         sbinfo->gid = simple_strtoul(value, &rest, 0);
2249                         if (*rest)
2250                                 goto bad_val;
2251                 } else if (!strcmp(this_char,"mpol")) {
2252                         if (mpol_parse_str(value, &sbinfo->mpol, 1))
2253                                 goto bad_val;
2254                 } else {
2255                         printk(KERN_ERR "tmpfs: Bad mount option %s\n",
2256                                this_char);
2257                         return 1;
2258                 }
2259         }
2260         return 0;
2261
2262 bad_val:
2263         printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
2264                value, this_char);
2265         return 1;
2266
2267 }
2268
2269 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2270 {
2271         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2272         struct shmem_sb_info config = *sbinfo;
2273         unsigned long inodes;
2274         int error = -EINVAL;
2275
2276         if (shmem_parse_options(data, &config, true))
2277                 return error;
2278
2279         spin_lock(&sbinfo->stat_lock);
2280         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2281         if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
2282                 goto out;
2283         if (config.max_inodes < inodes)
2284                 goto out;
2285         /*
2286          * Those tests also disallow limited->unlimited while any are in
2287          * use, so i_blocks will always be zero when max_blocks is zero;
2288          * but we must separately disallow unlimited->limited, because
2289          * in that case we have no record of how much is already in use.
2290          */
2291         if (config.max_blocks && !sbinfo->max_blocks)
2292                 goto out;
2293         if (config.max_inodes && !sbinfo->max_inodes)
2294                 goto out;
2295
2296         error = 0;
2297         sbinfo->max_blocks  = config.max_blocks;
2298         sbinfo->max_inodes  = config.max_inodes;
2299         sbinfo->free_inodes = config.max_inodes - inodes;
2300
2301         mpol_put(sbinfo->mpol);
2302         sbinfo->mpol        = config.mpol;      /* transfers initial ref */
2303 out:
2304         spin_unlock(&sbinfo->stat_lock);
2305         return error;
2306 }
2307
2308 static int shmem_show_options(struct seq_file *seq, struct vfsmount *vfs)
2309 {
2310         struct shmem_sb_info *sbinfo = SHMEM_SB(vfs->mnt_sb);
2311
2312         if (sbinfo->max_blocks != shmem_default_max_blocks())
2313                 seq_printf(seq, ",size=%luk",
2314                         sbinfo->max_blocks << (PAGE_CACHE_SHIFT - 10));
2315         if (sbinfo->max_inodes != shmem_default_max_inodes())
2316                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
2317         if (sbinfo->mode != (S_IRWXUGO | S_ISVTX))
2318                 seq_printf(seq, ",mode=%03o", sbinfo->mode);
2319         if (sbinfo->uid != 0)
2320                 seq_printf(seq, ",uid=%u", sbinfo->uid);
2321         if (sbinfo->gid != 0)
2322                 seq_printf(seq, ",gid=%u", sbinfo->gid);
2323         shmem_show_mpol(seq, sbinfo->mpol);
2324         return 0;
2325 }
2326 #endif /* CONFIG_TMPFS */
2327
2328 static void shmem_put_super(struct super_block *sb)
2329 {
2330         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2331
2332         percpu_counter_destroy(&sbinfo->used_blocks);
2333         kfree(sbinfo);
2334         sb->s_fs_info = NULL;
2335 }
2336
2337 int shmem_fill_super(struct super_block *sb, void *data, int silent)
2338 {
2339         struct inode *inode;
2340         struct dentry *root;
2341         struct shmem_sb_info *sbinfo;
2342         int err = -ENOMEM;
2343
2344         /* Round up to L1_CACHE_BYTES to resist false sharing */
2345         sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
2346                                 L1_CACHE_BYTES), GFP_KERNEL);
2347         if (!sbinfo)
2348                 return -ENOMEM;
2349
2350         sbinfo->mode = S_IRWXUGO | S_ISVTX;
2351         sbinfo->uid = current_fsuid();
2352         sbinfo->gid = current_fsgid();
2353         sb->s_fs_info = sbinfo;
2354
2355 #ifdef CONFIG_TMPFS
2356         /*
2357          * Per default we only allow half of the physical ram per
2358          * tmpfs instance, limiting inodes to one per page of lowmem;
2359          * but the internal instance is left unlimited.
2360          */
2361         if (!(sb->s_flags & MS_NOUSER)) {
2362                 sbinfo->max_blocks = shmem_default_max_blocks();
2363                 sbinfo->max_inodes = shmem_default_max_inodes();
2364                 if (shmem_parse_options(data, sbinfo, false)) {
2365                         err = -EINVAL;
2366                         goto failed;
2367                 }
2368         }
2369         sb->s_export_op = &shmem_export_ops;
2370 #else
2371         sb->s_flags |= MS_NOUSER;
2372 #endif
2373
2374         spin_lock_init(&sbinfo->stat_lock);
2375         if (percpu_counter_init(&sbinfo->used_blocks, 0))
2376                 goto failed;
2377         sbinfo->free_inodes = sbinfo->max_inodes;
2378
2379         sb->s_maxbytes = SHMEM_MAX_BYTES;
2380         sb->s_blocksize = PAGE_CACHE_SIZE;
2381         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
2382         sb->s_magic = TMPFS_MAGIC;
2383         sb->s_op = &shmem_ops;
2384         sb->s_time_gran = 1;
2385 #ifdef CONFIG_TMPFS_POSIX_ACL
2386         sb->s_xattr = shmem_xattr_handlers;
2387         sb->s_flags |= MS_POSIXACL;
2388 #endif
2389
2390         inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
2391         if (!inode)
2392                 goto failed;
2393         inode->i_uid = sbinfo->uid;
2394         inode->i_gid = sbinfo->gid;
2395         root = d_alloc_root(inode);
2396         if (!root)
2397                 goto failed_iput;
2398         sb->s_root = root;
2399         return 0;
2400
2401 failed_iput:
2402         iput(inode);
2403 failed:
2404         shmem_put_super(sb);
2405         return err;
2406 }
2407
2408 static struct kmem_cache *shmem_inode_cachep;
2409
2410 static struct inode *shmem_alloc_inode(struct super_block *sb)
2411 {
2412         struct shmem_inode_info *p;
2413         p = (struct shmem_inode_info *)kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
2414         if (!p)
2415                 return NULL;
2416         return &p->vfs_inode;
2417 }
2418
2419 static void shmem_i_callback(struct rcu_head *head)
2420 {
2421         struct inode *inode = container_of(head, struct inode, i_rcu);
2422         INIT_LIST_HEAD(&inode->i_dentry);
2423         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
2424 }
2425
2426 static void shmem_destroy_inode(struct inode *inode)
2427 {
2428         if ((inode->i_mode & S_IFMT) == S_IFREG) {
2429                 /* only struct inode is valid if it's an inline symlink */
2430                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
2431         }
2432         call_rcu(&inode->i_rcu, shmem_i_callback);
2433 }
2434
2435 static void init_once(void *foo)
2436 {
2437         struct shmem_inode_info *p = (struct shmem_inode_info *) foo;
2438
2439         inode_init_once(&p->vfs_inode);
2440 }
2441
2442 static int init_inodecache(void)
2443 {
2444         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
2445                                 sizeof(struct shmem_inode_info),
2446                                 0, SLAB_PANIC, init_once);
2447         return 0;
2448 }
2449
2450 static void destroy_inodecache(void)
2451 {
2452         kmem_cache_destroy(shmem_inode_cachep);
2453 }
2454
2455 static const struct address_space_operations shmem_aops = {
2456         .writepage      = shmem_writepage,
2457         .set_page_dirty = __set_page_dirty_no_writeback,
2458 #ifdef CONFIG_TMPFS
2459         .readpage       = shmem_readpage,
2460         .write_begin    = shmem_write_begin,
2461         .write_end      = shmem_write_end,
2462 #endif
2463         .migratepage    = migrate_page,
2464         .error_remove_page = generic_error_remove_page,
2465 };
2466
2467 static const struct file_operations shmem_file_operations = {
2468         .mmap           = shmem_mmap,
2469 #ifdef CONFIG_TMPFS
2470         .llseek         = generic_file_llseek,
2471         .read           = do_sync_read,
2472         .write          = do_sync_write,
2473         .aio_read       = shmem_file_aio_read,
2474         .aio_write      = generic_file_aio_write,
2475         .fsync          = noop_fsync,
2476         .splice_read    = generic_file_splice_read,
2477         .splice_write   = generic_file_splice_write,
2478 #endif
2479 };
2480
2481 static const struct inode_operations shmem_inode_operations = {
2482         .setattr        = shmem_notify_change,
2483         .truncate_range = shmem_truncate_range,
2484 #ifdef CONFIG_TMPFS_POSIX_ACL
2485         .setxattr       = generic_setxattr,
2486         .getxattr       = generic_getxattr,
2487         .listxattr      = generic_listxattr,
2488         .removexattr    = generic_removexattr,
2489         .check_acl      = generic_check_acl,
2490 #endif
2491
2492 };
2493
2494 static const struct inode_operations shmem_dir_inode_operations = {
2495 #ifdef CONFIG_TMPFS
2496         .create         = shmem_create,
2497         .lookup         = simple_lookup,
2498         .link           = shmem_link,
2499         .unlink         = shmem_unlink,
2500         .symlink        = shmem_symlink,
2501         .mkdir          = shmem_mkdir,
2502         .rmdir          = shmem_rmdir,
2503         .mknod          = shmem_mknod,
2504         .rename         = shmem_rename,
2505 #endif
2506 #ifdef CONFIG_TMPFS_POSIX_ACL
2507         .setattr        = shmem_notify_change,
2508         .setxattr       = generic_setxattr,
2509         .getxattr       = generic_getxattr,
2510         .listxattr      = generic_listxattr,
2511         .removexattr    = generic_removexattr,
2512         .check_acl      = generic_check_acl,
2513 #endif
2514 };
2515
2516 static const struct inode_operations shmem_special_inode_operations = {
2517 #ifdef CONFIG_TMPFS_POSIX_ACL
2518         .setattr        = shmem_notify_change,
2519         .setxattr       = generic_setxattr,
2520         .getxattr       = generic_getxattr,
2521         .listxattr      = generic_listxattr,
2522         .removexattr    = generic_removexattr,
2523         .check_acl      = generic_check_acl,
2524 #endif
2525 };
2526
2527 static const struct super_operations shmem_ops = {
2528         .alloc_inode    = shmem_alloc_inode,
2529         .destroy_inode  = shmem_destroy_inode,
2530 #ifdef CONFIG_TMPFS
2531         .statfs         = shmem_statfs,
2532         .remount_fs     = shmem_remount_fs,
2533         .show_options   = shmem_show_options,
2534 #endif
2535         .evict_inode    = shmem_evict_inode,
2536         .drop_inode     = generic_delete_inode,
2537         .put_super      = shmem_put_super,
2538 };
2539
2540 static const struct vm_operations_struct shmem_vm_ops = {
2541         .fault          = shmem_fault,
2542 #ifdef CONFIG_NUMA
2543         .set_policy     = shmem_set_policy,
2544         .get_policy     = shmem_get_policy,
2545 #endif
2546 };
2547
2548
2549 static struct dentry *shmem_mount(struct file_system_type *fs_type,
2550         int flags, const char *dev_name, void *data)
2551 {
2552         return mount_nodev(fs_type, flags, data, shmem_fill_super);
2553 }
2554
2555 static struct file_system_type tmpfs_fs_type = {
2556         .owner          = THIS_MODULE,
2557         .name           = "tmpfs",
2558         .mount          = shmem_mount,
2559         .kill_sb        = kill_litter_super,
2560 };
2561
2562 int __init init_tmpfs(void)
2563 {
2564         int error;
2565
2566         error = bdi_init(&shmem_backing_dev_info);
2567         if (error)
2568                 goto out4;
2569
2570         error = init_inodecache();
2571         if (error)
2572                 goto out3;
2573
2574         error = register_filesystem(&tmpfs_fs_type);
2575         if (error) {
2576                 printk(KERN_ERR "Could not register tmpfs\n");
2577                 goto out2;
2578         }
2579
2580         shm_mnt = vfs_kern_mount(&tmpfs_fs_type, MS_NOUSER,
2581                                 tmpfs_fs_type.name, NULL);
2582         if (IS_ERR(shm_mnt)) {
2583                 error = PTR_ERR(shm_mnt);
2584                 printk(KERN_ERR "Could not kern_mount tmpfs\n");
2585                 goto out1;
2586         }
2587         return 0;
2588
2589 out1:
2590         unregister_filesystem(&tmpfs_fs_type);
2591 out2:
2592         destroy_inodecache();
2593 out3:
2594         bdi_destroy(&shmem_backing_dev_info);
2595 out4:
2596         shm_mnt = ERR_PTR(error);
2597         return error;
2598 }
2599
2600 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2601 /**
2602  * mem_cgroup_get_shmem_target - find a page or entry assigned to the shmem file
2603  * @inode: the inode to be searched
2604  * @pgoff: the offset to be searched
2605  * @pagep: the pointer for the found page to be stored
2606  * @ent: the pointer for the found swap entry to be stored
2607  *
2608  * If a page is found, refcount of it is incremented. Callers should handle
2609  * these refcount.
2610  */
2611 void mem_cgroup_get_shmem_target(struct inode *inode, pgoff_t pgoff,
2612                                         struct page **pagep, swp_entry_t *ent)
2613 {
2614         swp_entry_t entry = { .val = 0 }, *ptr;
2615         struct page *page = NULL;
2616         struct shmem_inode_info *info = SHMEM_I(inode);
2617
2618         if ((pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
2619                 goto out;
2620
2621         spin_lock(&info->lock);
2622         ptr = shmem_swp_entry(info, pgoff, NULL);
2623 #ifdef CONFIG_SWAP
2624         if (ptr && ptr->val) {
2625                 entry.val = ptr->val;
2626                 page = find_get_page(&swapper_space, entry.val);
2627         } else
2628 #endif
2629                 page = find_get_page(inode->i_mapping, pgoff);
2630         if (ptr)
2631                 shmem_swp_unmap(ptr);
2632         spin_unlock(&info->lock);
2633 out:
2634         *pagep = page;
2635         *ent = entry;
2636 }
2637 #endif
2638
2639 #else /* !CONFIG_SHMEM */
2640
2641 /*
2642  * tiny-shmem: simple shmemfs and tmpfs using ramfs code
2643  *
2644  * This is intended for small system where the benefits of the full
2645  * shmem code (swap-backed and resource-limited) are outweighed by
2646  * their complexity. On systems without swap this code should be
2647  * effectively equivalent, but much lighter weight.
2648  */
2649
2650 #include <linux/ramfs.h>
2651
2652 static struct file_system_type tmpfs_fs_type = {
2653         .name           = "tmpfs",
2654         .mount          = ramfs_mount,
2655         .kill_sb        = kill_litter_super,
2656 };
2657
2658 int __init init_tmpfs(void)
2659 {
2660         BUG_ON(register_filesystem(&tmpfs_fs_type) != 0);
2661
2662         shm_mnt = kern_mount(&tmpfs_fs_type);
2663         BUG_ON(IS_ERR(shm_mnt));
2664
2665         return 0;
2666 }
2667
2668 int shmem_unuse(swp_entry_t entry, struct page *page)
2669 {
2670         return 0;
2671 }
2672
2673 int shmem_lock(struct file *file, int lock, struct user_struct *user)
2674 {
2675         return 0;
2676 }
2677
2678 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2679 /**
2680  * mem_cgroup_get_shmem_target - find a page or entry assigned to the shmem file
2681  * @inode: the inode to be searched
2682  * @pgoff: the offset to be searched
2683  * @pagep: the pointer for the found page to be stored
2684  * @ent: the pointer for the found swap entry to be stored
2685  *
2686  * If a page is found, refcount of it is incremented. Callers should handle
2687  * these refcount.
2688  */
2689 void mem_cgroup_get_shmem_target(struct inode *inode, pgoff_t pgoff,
2690                                         struct page **pagep, swp_entry_t *ent)
2691 {
2692         struct page *page = NULL;
2693
2694         if ((pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
2695                 goto out;
2696         page = find_get_page(inode->i_mapping, pgoff);
2697 out:
2698         *pagep = page;
2699         *ent = (swp_entry_t){ .val = 0 };
2700 }
2701 #endif
2702
2703 #define shmem_vm_ops                            generic_file_vm_ops
2704 #define shmem_file_operations                   ramfs_file_operations
2705 #define shmem_get_inode(sb, dir, mode, dev, flags)      ramfs_get_inode(sb, dir, mode, dev)
2706 #define shmem_acct_size(flags, size)            0
2707 #define shmem_unacct_size(flags, size)          do {} while (0)
2708 #define SHMEM_MAX_BYTES                         MAX_LFS_FILESIZE
2709
2710 #endif /* CONFIG_SHMEM */
2711
2712 /* common code */
2713
2714 /**
2715  * shmem_file_setup - get an unlinked file living in tmpfs
2716  * @name: name for dentry (to be seen in /proc/<pid>/maps
2717  * @size: size to be set for the file
2718  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
2719  */
2720 struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
2721 {
2722         int error;
2723         struct file *file;
2724         struct inode *inode;
2725         struct path path;
2726         struct dentry *root;
2727         struct qstr this;
2728
2729         if (IS_ERR(shm_mnt))
2730                 return (void *)shm_mnt;
2731
2732         if (size < 0 || size > SHMEM_MAX_BYTES)
2733                 return ERR_PTR(-EINVAL);
2734
2735         if (shmem_acct_size(flags, size))
2736                 return ERR_PTR(-ENOMEM);
2737
2738         error = -ENOMEM;
2739         this.name = name;
2740         this.len = strlen(name);
2741         this.hash = 0; /* will go */
2742         root = shm_mnt->mnt_root;
2743         path.dentry = d_alloc(root, &this);
2744         if (!path.dentry)
2745                 goto put_memory;
2746         path.mnt = mntget(shm_mnt);
2747
2748         error = -ENOSPC;
2749         inode = shmem_get_inode(root->d_sb, NULL, S_IFREG | S_IRWXUGO, 0, flags);
2750         if (!inode)
2751                 goto put_dentry;
2752
2753         d_instantiate(path.dentry, inode);
2754         inode->i_size = size;
2755         inode->i_nlink = 0;     /* It is unlinked */
2756 #ifndef CONFIG_MMU
2757         error = ramfs_nommu_expand_for_mapping(inode, size);
2758         if (error)
2759                 goto put_dentry;
2760 #endif
2761
2762         error = -ENFILE;
2763         file = alloc_file(&path, FMODE_WRITE | FMODE_READ,
2764                   &shmem_file_operations);
2765         if (!file)
2766                 goto put_dentry;
2767
2768         return file;
2769
2770 put_dentry:
2771         path_put(&path);
2772 put_memory:
2773         shmem_unacct_size(flags, size);
2774         return ERR_PTR(error);
2775 }
2776 EXPORT_SYMBOL_GPL(shmem_file_setup);
2777
2778 /**
2779  * shmem_zero_setup - setup a shared anonymous mapping
2780  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
2781  */
2782 int shmem_zero_setup(struct vm_area_struct *vma)
2783 {
2784         struct file *file;
2785         loff_t size = vma->vm_end - vma->vm_start;
2786
2787         file = shmem_file_setup("dev/zero", size, vma->vm_flags);
2788         if (IS_ERR(file))
2789                 return PTR_ERR(file);
2790
2791         if (vma->vm_file)
2792                 fput(vma->vm_file);
2793         vma->vm_file = file;
2794         vma->vm_ops = &shmem_vm_ops;
2795         vma->vm_flags |= VM_CAN_NONLINEAR;
2796         return 0;
2797 }