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