mm/shmem.c

   1 /*
   2  * Resizable virtual memory filesystem for Linux.
   3  *
   4  * Copyright (C) 2000 Linus Torvalds.
   5  *               2000 Transmeta Corp.
   6  *               2000-2001 Christoph Rohland
   7  *               2000-2001 SAP AG
   8  *               2002 Red Hat Inc.
   9  * Copyright (C) 2002-2005 Hugh Dickins.
  10  * Copyright (C) 2002-2005 VERITAS Software Corporation.
  11  * Copyright (C) 2004 Andi Kleen, SuSE Labs
  12  *
  13  * Extended attribute support for tmpfs:
  14  * Copyright (c) 2004, Luke Kenneth Casson Leighton <lkcl@lkcl.net>
  15  * Copyright (c) 2004 Red Hat, Inc., James Morris <jmorris@redhat.com>
  16  *
  17  * tiny-shmem:
  18  * Copyright (c) 2004, 2008 Matt Mackall <mpm@selenic.com>
  19  *
  20  * This file is released under the GPL.
  21  */
  22
  23 #include <linux/fs.h>
  24 #include <linux/init.h>
  25 #include <linux/vfs.h>
  26 #include <linux/mount.h>
  27 #include <linux/pagemap.h>
  28 #include <linux/file.h>
  29 #include <linux/mm.h>
  30 #include <linux/module.h>
  31 #include <linux/percpu_counter.h>
  32 #include <linux/swap.h>
  33
  34 static struct vfsmount *shm_mnt;
  35
  36 #ifdef CONFIG_SHMEM
  37 /*
  38  * This virtual memory filesystem is heavily based on the ramfs. It
  39  * extends ramfs by the ability to use swap and honor resource limits
  40  * which makes it a completely usable filesystem.
  41  */
  42
  43 #include <linux/xattr.h>
  44 #include <linux/exportfs.h>
  45 #include <linux/posix_acl.h>
  46 #include <linux/generic_acl.h>
  47 #include <linux/mman.h>
  48 #include <linux/string.h>
  49 #include <linux/slab.h>
  50 #include <linux/backing-dev.h>
  51 #include <linux/shmem_fs.h>
  52 #include <linux/writeback.h>
  53 #include <linux/blkdev.h>
  54 #include <linux/security.h>
  55 #include <linux/swapops.h>
  56 #include <linux/mempolicy.h>
  57 #include <linux/namei.h>
  58 #include <linux/ctype.h>
  59 #include <linux/migrate.h>
  60 #include <linux/highmem.h>
  61 #include <linux/seq_file.h>
  62 #include <linux/magic.h>
  63
  64 #include <asm/uaccess.h>
  65 #include <asm/div64.h>
  66 #include <asm/pgtable.h>
  67
  68 /*
  69  * The maximum size of a shmem/tmpfs file is limited by the maximum size of
  70  * its triple-indirect swap vector - see illustration at shmem_swp_entry().
  71  *
  72  * With 4kB page size, maximum file size is just over 2TB on a 32-bit kernel,
  73  * but one eighth of that on a 64-bit kernel.  With 8kB page size, maximum
  74  * file size is just over 4TB on a 64-bit kernel, but 16TB on a 32-bit kernel,
  75  * MAX_LFS_FILESIZE being then more restrictive than swap vector layout.
  76  *
  77  * We use / and * instead of shifts in the definitions below, so that the swap
  78  * vector can be tested with small even values (e.g. 20) for ENTRIES_PER_PAGE.
  79  */
  80 #define ENTRIES_PER_PAGE (PAGE_CACHE_SIZE/sizeof(unsigned long))
  81 #define ENTRIES_PER_PAGEPAGE ((unsigned long long)ENTRIES_PER_PAGE*ENTRIES_PER_PAGE)
  82
  83 #define SHMSWP_MAX_INDEX (SHMEM_NR_DIRECT + (ENTRIES_PER_PAGEPAGE/2) * (ENTRIES_PER_PAGE+1))
  84 #define SHMSWP_MAX_BYTES (SHMSWP_MAX_INDEX << PAGE_CACHE_SHIFT)
  85
  86 #define SHMEM_MAX_BYTES  min_t(unsigned long long, SHMSWP_MAX_BYTES, MAX_LFS_FILESIZE)
  87 #define SHMEM_MAX_INDEX  ((unsigned long)((SHMEM_MAX_BYTES+1) >> PAGE_CACHE_SHIFT))
  88
  89 #define BLOCKS_PER_PAGE  (PAGE_CACHE_SIZE/512)
  90 #define VM_ACCT(size)    (PAGE_CACHE_ALIGN(size) >> PAGE_SHIFT)
  91
  92 /* info->flags needs VM_flags to handle pagein/truncate races efficiently */
  93 #define SHMEM_PAGEIN     VM_READ
  94 #define SHMEM_TRUNCATE   VM_WRITE
  95
  96 /* Definition to limit shmem_truncate's steps between cond_rescheds */
  97 #define LATENCY_LIMIT    64
  98
  99 /* Pretend that each entry is of this size in directory's i_size */
 100 #define BOGO_DIRENT_SIZE 20
 101
 102 /* Flag allocation requirements to shmem_getpage and shmem_swp_alloc */
 103 enum sgp_type {
 104         SGP_READ,       /* don't exceed i_size, don't allocate page */
 105         SGP_CACHE,      /* don't exceed i_size, may allocate page */
 106         SGP_DIRTY,      /* like SGP_CACHE, but set new page dirty */
 107         SGP_WRITE,      /* may exceed i_size, may allocate page */
 108 };
 109
 110 #ifdef CONFIG_TMPFS
 111 static unsigned long shmem_default_max_blocks(void)
 112 {
 113         return totalram_pages / 2;
 114 }
 115
 116 static unsigned long shmem_default_max_inodes(void)
 117 {
 118         return min(totalram_pages - totalhigh_pages, totalram_pages / 2);
 119 }
 120 #endif
 121
 122 static int shmem_getpage(struct inode *inode, unsigned long idx,
 123                          struct page **pagep, enum sgp_type sgp, int *type);
 124
 125 static inline struct page *shmem_dir_alloc(gfp_t gfp_mask)
 126 {
 127         /*
 128          * The above definition of ENTRIES_PER_PAGE, and the use of
 129          * BLOCKS_PER_PAGE on indirect pages, assume PAGE_CACHE_SIZE:
 130          * might be reconsidered if it ever diverges from PAGE_SIZE.
 131          *
 132          * Mobility flags are masked out as swap vectors cannot move
 133          */
 134         return alloc_pages((gfp_mask & ~GFP_MOVABLE_MASK) | __GFP_ZERO,
 135                                 PAGE_CACHE_SHIFT-PAGE_SHIFT);
 136 }
 137
 138 static inline void shmem_dir_free(struct page *page)
 139 {
 140         __free_pages(page, PAGE_CACHE_SHIFT-PAGE_SHIFT);
 141 }
 142
 143 static struct page **shmem_dir_map(struct page *page)
 144 {
 145         return (struct page **)kmap_atomic(page, KM_USER0);
 146 }
 147
 148 static inline void shmem_dir_unmap(struct page **dir)
 149 {
 150         kunmap_atomic(dir, KM_USER0);
 151 }
 152
 153 static swp_entry_t *shmem_swp_map(struct page *page)
 154 {
 155         return (swp_entry_t *)kmap_atomic(page, KM_USER1);
 156 }
 157
 158 static inline void shmem_swp_balance_unmap(void)
 159 {
 160         /*
 161          * When passing a pointer to an i_direct entry, to code which
 162          * also handles indirect entries and so will shmem_swp_unmap,
 163          * we must arrange for the preempt count to remain in balance.
 164          * What kmap_atomic of a lowmem page does depends on config
 165          * and architecture, so pretend to kmap_atomic some lowmem page.
 166          */
 167         (void) kmap_atomic(ZERO_PAGE(0), KM_USER1);
 168 }
 169
 170 static inline void shmem_swp_unmap(swp_entry_t *entry)
 171 {
 172         kunmap_atomic(entry, KM_USER1);
 173 }
 174
 175 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
 176 {
 177         return sb->s_fs_info;
 178 }
 179
 180 /*
 181  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
 182  * for shared memory and for shared anonymous (/dev/zero) mappings
 183  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
 184  * consistent with the pre-accounting of private mappings ...
 185  */
 186 static inline int shmem_acct_size(unsigned long flags, loff_t size)
 187 {
 188         return (flags & VM_NORESERVE) ?
 189                 0 : security_vm_enough_memory_kern(VM_ACCT(size));
 190 }
 191
 192 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
 193 {
 194         if (!(flags & VM_NORESERVE))
 195                 vm_unacct_memory(VM_ACCT(size));
 196 }
 197
 198 /*
 199  * ... whereas tmpfs objects are accounted incrementally as
 200  * pages are allocated, in order to allow huge sparse files.
 201  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
 202  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
 203  */
 204 static inline int shmem_acct_block(unsigned long flags)
 205 {
 206         return (flags & VM_NORESERVE) ?
 207                 security_vm_enough_memory_kern(VM_ACCT(PAGE_CACHE_SIZE)) : 0;
 208 }
 209
 210 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
 211 {
 212         if (flags & VM_NORESERVE)
 213                 vm_unacct_memory(pages * VM_ACCT(PAGE_CACHE_SIZE));
 214 }
 215
 216 static const struct super_operations shmem_ops;
 217 static const struct address_space_operations shmem_aops;
 218 static const struct file_operations shmem_file_operations;
 219 static const struct inode_operations shmem_inode_operations;
 220 static const struct inode_operations shmem_dir_inode_operations;
 221 static const struct inode_operations shmem_special_inode_operations;
 222 static const struct vm_operations_struct shmem_vm_ops;
 223
 224 static struct backing_dev_info shmem_backing_dev_info  __read_mostly = {
 225         .ra_pages       = 0,    /* No readahead */
 226         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_SWAP_BACKED,
 227 };
 228
 229 static LIST_HEAD(shmem_swaplist);
 230 static DEFINE_MUTEX(shmem_swaplist_mutex);
 231
 232 static void shmem_free_blocks(struct inode *inode, long pages)
 233 {
 234         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 235         if (sbinfo->max_blocks) {
 236                 percpu_counter_add(&sbinfo->used_blocks, -pages);
 237                 spin_lock(&inode->i_lock);
 238                 inode->i_blocks -= pages*BLOCKS_PER_PAGE;
 239                 spin_unlock(&inode->i_lock);
 240         }
 241 }
 242
 243 static int shmem_reserve_inode(struct super_block *sb)
 244 {
 245         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 246         if (sbinfo->max_inodes) {
 247                 spin_lock(&sbinfo->stat_lock);
 248                 if (!sbinfo->free_inodes) {
 249                         spin_unlock(&sbinfo->stat_lock);
 250                         return -ENOSPC;
 251                 }
 252                 sbinfo->free_inodes--;
 253                 spin_unlock(&sbinfo->stat_lock);
 254         }
 255         return 0;
 256 }
 257
 258 static void shmem_free_inode(struct super_block *sb)
 259 {
 260         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 261         if (sbinfo->max_inodes) {
 262                 spin_lock(&sbinfo->stat_lock);
 263                 sbinfo->free_inodes++;
 264                 spin_unlock(&sbinfo->stat_lock);
 265         }
 266 }
 267
 268 /**
 269  * shmem_recalc_inode - recalculate the size of an inode
 270  * @inode: inode to recalc
 271  *
 272  * We have to calculate the free blocks since the mm can drop
 273  * undirtied hole pages behind our back.
 274  *
 275  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
 276  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
 277  *
 278  * It has to be called with the spinlock held.
 279  */
 280 static void shmem_recalc_inode(struct inode *inode)
 281 {
 282         struct shmem_inode_info *info = SHMEM_I(inode);
 283         long freed;
 284
 285         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
 286         if (freed > 0) {
 287                 info->alloced -= freed;
 288                 shmem_unacct_blocks(info->flags, freed);
 289                 shmem_free_blocks(inode, freed);
 290         }
 291 }
 292
 293 /**
 294  * shmem_swp_entry - find the swap vector position in the info structure
 295  * @info:  info structure for the inode
 296  * @index: index of the page to find
 297  * @page:  optional page to add to the structure. Has to be preset to
 298  *         all zeros
 299  *
 300  * If there is no space allocated yet it will return NULL when
 301  * page is NULL, else it will use the page for the needed block,
 302  * setting it to NULL on return to indicate that it has been used.
 303  *
 304  * The swap vector is organized the following way:
 305  *
 306  * There are SHMEM_NR_DIRECT entries directly stored in the
 307  * shmem_inode_info structure. So small files do not need an addional
 308  * allocation.
 309  *
 310  * For pages with index > SHMEM_NR_DIRECT there is the pointer
 311  * i_indirect which points to a page which holds in the first half
 312  * doubly indirect blocks, in the second half triple indirect blocks:
 313  *
 314  * For an artificial ENTRIES_PER_PAGE = 4 this would lead to the
 315  * following layout (for SHMEM_NR_DIRECT == 16):
 316  *
 317  * i_indirect -> dir --> 16-19
 318  *            |      +-> 20-23
 319  *            |
 320  *            +-->dir2 --> 24-27
 321  *            |        +-> 28-31
 322  *            |        +-> 32-35
 323  *            |        +-> 36-39
 324  *            |
 325  *            +-->dir3 --> 40-43
 326  *                     +-> 44-47
 327  *                     +-> 48-51
 328  *                     +-> 52-55
 329  */
 330 static swp_entry_t *shmem_swp_entry(struct shmem_inode_info *info, unsigned long index, struct page **page)
 331 {
 332         unsigned long offset;
 333         struct page **dir;
 334         struct page *subdir;
 335
 336         if (index < SHMEM_NR_DIRECT) {
 337                 shmem_swp_balance_unmap();
 338                 return info->i_direct+index;
 339         }
 340         if (!info->i_indirect) {
 341                 if (page) {
 342                         info->i_indirect = *page;
 343                         *page = NULL;
 344                 }
 345                 return NULL;                    /* need another page */
 346         }
 347
 348         index -= SHMEM_NR_DIRECT;
 349         offset = index % ENTRIES_PER_PAGE;
 350         index /= ENTRIES_PER_PAGE;
 351         dir = shmem_dir_map(info->i_indirect);
 352
 353         if (index >= ENTRIES_PER_PAGE/2) {
 354                 index -= ENTRIES_PER_PAGE/2;
 355                 dir += ENTRIES_PER_PAGE/2 + index/ENTRIES_PER_PAGE;
 356                 index %= ENTRIES_PER_PAGE;
 357                 subdir = *dir;
 358                 if (!subdir) {
 359                         if (page) {
 360                                 *dir = *page;
 361                                 *page = NULL;
 362                         }
 363                         shmem_dir_unmap(dir);
 364                         return NULL;            /* need another page */
 365                 }
 366                 shmem_dir_unmap(dir);
 367                 dir = shmem_dir_map(subdir);
 368         }
 369
 370         dir += index;
 371         subdir = *dir;
 372         if (!subdir) {
 373                 if (!page || !(subdir = *page)) {
 374                         shmem_dir_unmap(dir);
 375                         return NULL;            /* need a page */
 376                 }
 377                 *dir = subdir;
 378                 *page = NULL;
 379         }
 380         shmem_dir_unmap(dir);
 381         return shmem_swp_map(subdir) + offset;
 382 }
 383
 384 static void shmem_swp_set(struct shmem_inode_info *info, swp_entry_t *entry, unsigned long value)
 385 {
 386         long incdec = value? 1: -1;
 387
 388         entry->val = value;
 389         info->swapped += incdec;
 390         if ((unsigned long)(entry - info->i_direct) >= SHMEM_NR_DIRECT) {
 391                 struct page *page = kmap_atomic_to_page(entry);
 392                 set_page_private(page, page_private(page) + incdec);
 393         }
 394 }
 395
 396 /**
 397  * shmem_swp_alloc - get the position of the swap entry for the page.
 398  * @info:       info structure for the inode
 399  * @index:      index of the page to find
 400  * @sgp:        check and recheck i_size? skip allocation?
 401  *
 402  * If the entry does not exist, allocate it.
 403  */
 404 static swp_entry_t *shmem_swp_alloc(struct shmem_inode_info *info, unsigned long index, enum sgp_type sgp)
 405 {
 406         struct inode *inode = &info->vfs_inode;
 407         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 408         struct page *page = NULL;
 409         swp_entry_t *entry;
 410
 411         if (sgp != SGP_WRITE &&
 412             ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode))
 413                 return ERR_PTR(-EINVAL);
 414
 415         while (!(entry = shmem_swp_entry(info, index, &page))) {
 416                 if (sgp == SGP_READ)
 417                         return shmem_swp_map(ZERO_PAGE(0));
 418                 /*
 419                  * Test used_blocks against 1 less max_blocks, since we have 1 data
 420                  * page (and perhaps indirect index pages) yet to allocate:
 421                  * a waste to allocate index if we cannot allocate data.
 422                  */
 423                 if (sbinfo->max_blocks) {
 424                         if (percpu_counter_compare(&sbinfo->used_blocks,
 425                                                 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 cannot 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 address_space *mapping;
 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                 shmem_swp_balance_unmap();
 880                 goto found;
 881         }
 882         if (!info->i_indirect)
 883                 goto lost2;
 884
 885         dir = shmem_dir_map(info->i_indirect);
 886         stage = SHMEM_NR_DIRECT + ENTRIES_PER_PAGEPAGE/2;
 887
 888         for (idx = SHMEM_NR_DIRECT; idx < limit; idx += ENTRIES_PER_PAGE, dir++) {
 889                 if (unlikely(idx == stage)) {
 890                         shmem_dir_unmap(dir-1);
 891                         if (cond_resched_lock(&info->lock)) {
 892                                 /* check it has not been truncated */
 893                                 if (limit > info->next_index) {
 894                                         limit = info->next_index;
 895                                         if (idx >= limit)
 896                                                 goto lost2;
 897                                 }
 898                         }
 899                         dir = shmem_dir_map(info->i_indirect) +
 900                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
 901                         while (!*dir) {
 902                                 dir++;
 903                                 idx += ENTRIES_PER_PAGEPAGE;
 904                                 if (idx >= limit)
 905                                         goto lost1;
 906                         }
 907                         stage = idx + ENTRIES_PER_PAGEPAGE;
 908                         subdir = *dir;
 909                         shmem_dir_unmap(dir);
 910                         dir = shmem_dir_map(subdir);
 911                 }
 912                 subdir = *dir;
 913                 if (subdir && page_private(subdir)) {
 914                         ptr = shmem_swp_map(subdir);
 915                         size = limit - idx;
 916                         if (size > ENTRIES_PER_PAGE)
 917                                 size = ENTRIES_PER_PAGE;
 918                         offset = shmem_find_swp(entry, ptr, ptr+size);
 919                         if (offset >= 0) {
 920                                 shmem_dir_unmap(dir);
 921                                 goto found;
 922                         }
 923                         shmem_swp_unmap(ptr);
 924                 }
 925         }
 926 lost1:
 927         shmem_dir_unmap(dir-1);
 928 lost2:
 929         spin_unlock(&info->lock);
 930         return 0;
 931 found:
 932         idx += offset;
 933         ptr += offset;
 934
 935         /*
 936          * Move _head_ to start search for next from here.
 937          * But be careful: shmem_evict_inode checks list_empty without taking
 938          * mutex, and there's an instant in list_move_tail when info->swaplist
 939          * would appear empty, if it were the only one on shmem_swaplist.  We
 940          * could avoid doing it if inode NULL; or use this minor optimization.
 941          */
 942         if (shmem_swaplist.next != &info->swaplist)
 943                 list_move_tail(&shmem_swaplist, &info->swaplist);
 944
 945         /*
 946          * We rely on shmem_swaplist_mutex, not only to protect the swaplist,
 947          * but also to hold up shmem_evict_inode(): so inode cannot be freed
 948          * beneath us (pagelock doesn't help until the page is in pagecache).
 949          */
 950         mapping = info->vfs_inode.i_mapping;
 951         error = add_to_page_cache_locked(page, mapping, idx, GFP_NOWAIT);
 952         /* which does mem_cgroup_uncharge_cache_page on error */
 953
 954         if (error == -EEXIST) {
 955                 struct page *filepage = find_get_page(mapping, idx);
 956                 error = 1;
 957                 if (filepage) {
 958                         /*
 959                          * There might be a more uptodate page coming down
 960                          * from a stacked writepage: forget our swappage if so.
 961                          */
 962                         if (PageUptodate(filepage))
 963                                 error = 0;
 964                         page_cache_release(filepage);
 965                 }
 966         }
 967         if (!error) {
 968                 delete_from_swap_cache(page);
 969                 set_page_dirty(page);
 970                 info->flags |= SHMEM_PAGEIN;
 971                 shmem_swp_set(info, ptr, 0);
 972                 swap_free(entry);
 973                 error = 1;      /* not an error, but entry was found */
 974         }
 975         shmem_swp_unmap(ptr);
 976         spin_unlock(&info->lock);
 977         return error;
 978 }
 979
 980 /*
 981  * shmem_unuse() search for an eventually swapped out shmem page.
 982  */
 983 int shmem_unuse(swp_entry_t entry, struct page *page)
 984 {
 985         struct list_head *p, *next;
 986         struct shmem_inode_info *info;
 987         int found = 0;
 988         int error;
 989
 990         /*
 991          * Charge page using GFP_KERNEL while we can wait, before taking
 992          * the shmem_swaplist_mutex which might hold up shmem_writepage().
 993          * Charged back to the user (not to caller) when swap account is used.
 994          * add_to_page_cache() will be called with GFP_NOWAIT.
 995          */
 996         error = mem_cgroup_cache_charge(page, current->mm, GFP_KERNEL);
 997         if (error)
 998                 goto out;
 999         /*
1000          * Try to preload while we can wait, to not make a habit of
1001          * draining atomic reserves; but don't latch on to this cpu,
1002          * it's okay if sometimes we get rescheduled after this.
1003          */
1004         error = radix_tree_preload(GFP_KERNEL);
1005         if (error)
1006                 goto uncharge;
1007         radix_tree_preload_end();
1008
1009         mutex_lock(&shmem_swaplist_mutex);
1010         list_for_each_safe(p, next, &shmem_swaplist) {
1011                 info = list_entry(p, struct shmem_inode_info, swaplist);
1012                 found = shmem_unuse_inode(info, entry, page);
1013                 cond_resched();
1014                 if (found)
1015                         break;
1016         }
1017         mutex_unlock(&shmem_swaplist_mutex);
1018
1019 uncharge:
1020         if (!found)
1021                 mem_cgroup_uncharge_cache_page(page);
1022         if (found < 0)
1023                 error = found;
1024 out:
1025         unlock_page(page);
1026         page_cache_release(page);
1027         return error;
1028 }
1029
1030 /*
1031  * Move the page from the page cache to the swap cache.
1032  */
1033 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1034 {
1035         struct shmem_inode_info *info;
1036         swp_entry_t *entry, swap;
1037         struct address_space *mapping;
1038         unsigned long index;
1039         struct inode *inode;
1040         bool unlock_mutex = false;
1041
1042         BUG_ON(!PageLocked(page));
1043         mapping = page->mapping;
1044         index = page->index;
1045         inode = mapping->host;
1046         info = SHMEM_I(inode);
1047         if (info->flags & VM_LOCKED)
1048                 goto redirty;
1049         if (!total_swap_pages)
1050                 goto redirty;
1051
1052         /*
1053          * shmem_backing_dev_info's capabilities prevent regular writeback or
1054          * sync from ever calling shmem_writepage; but a stacking filesystem
1055          * may use the ->writepage of its underlying filesystem, in which case
1056          * tmpfs should write out to swap only in response to memory pressure,
1057          * and not for the writeback threads or sync.  However, in those cases,
1058          * we do still want to check if there's a redundant swappage to be
1059          * discarded.
1060          */
1061         if (wbc->for_reclaim)
1062                 swap = get_swap_page();
1063         else
1064                 swap.val = 0;
1065
1066         /*
1067          * Add inode to shmem_unuse()'s list of swapped-out inodes,
1068          * if it's not already there.  Do it now because we cannot take
1069          * mutex while holding spinlock, and must do so before the page
1070          * is moved to swap cache, when its pagelock no longer protects
1071          * the inode from eviction.  But don't unlock the mutex until
1072          * we've taken the spinlock, because shmem_unuse_inode() will
1073          * prune a !swapped inode from the swaplist under both locks.
1074          */
1075         if (swap.val && list_empty(&info->swaplist)) {
1076                 mutex_lock(&shmem_swaplist_mutex);
1077                 /* move instead of add in case we're racing */
1078                 list_move_tail(&info->swaplist, &shmem_swaplist);
1079                 unlock_mutex = true;
1080         }
1081
1082         spin_lock(&info->lock);
1083         if (unlock_mutex)
1084                 mutex_unlock(&shmem_swaplist_mutex);
1085
1086         if (index >= info->next_index) {
1087                 BUG_ON(!(info->flags & SHMEM_TRUNCATE));
1088                 goto unlock;
1089         }
1090         entry = shmem_swp_entry(info, index, NULL);
1091         if (entry->val) {
1092                 /*
1093                  * The more uptodate page coming down from a stacked
1094                  * writepage should replace our old swappage.
1095                  */
1096                 free_swap_and_cache(*entry);
1097                 shmem_swp_set(info, entry, 0);
1098         }
1099         shmem_recalc_inode(inode);
1100
1101         if (swap.val && add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
1102                 delete_from_page_cache(page);
1103                 shmem_swp_set(info, entry, swap.val);
1104                 shmem_swp_unmap(entry);
1105                 spin_unlock(&info->lock);
1106                 swap_shmem_alloc(swap);
1107                 BUG_ON(page_mapped(page));
1108                 swap_writepage(page, wbc);
1109                 return 0;
1110         }
1111
1112         shmem_swp_unmap(entry);
1113 unlock:
1114         spin_unlock(&info->lock);
1115         /*
1116          * add_to_swap_cache() doesn't return -EEXIST, so we can safely
1117          * clear SWAP_HAS_CACHE flag.
1118          */
1119         swapcache_free(swap, NULL);
1120 redirty:
1121         set_page_dirty(page);
1122         if (wbc->for_reclaim)
1123                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
1124         unlock_page(page);
1125         return 0;
1126 }
1127
1128 #ifdef CONFIG_NUMA
1129 #ifdef CONFIG_TMPFS
1130 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1131 {
1132         char buffer[64];
1133
1134         if (!mpol || mpol->mode == MPOL_DEFAULT)
1135                 return;         /* show nothing */
1136
1137         mpol_to_str(buffer, sizeof(buffer), mpol, 1);
1138
1139         seq_printf(seq, ",mpol=%s", buffer);
1140 }
1141
1142 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1143 {
1144         struct mempolicy *mpol = NULL;
1145         if (sbinfo->mpol) {
1146                 spin_lock(&sbinfo->stat_lock);  /* prevent replace/use races */
1147                 mpol = sbinfo->mpol;
1148                 mpol_get(mpol);
1149                 spin_unlock(&sbinfo->stat_lock);
1150         }
1151         return mpol;
1152 }
1153 #endif /* CONFIG_TMPFS */
1154
1155 static struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1156                         struct shmem_inode_info *info, unsigned long idx)
1157 {
1158         struct mempolicy mpol, *spol;
1159         struct vm_area_struct pvma;
1160         struct page *page;
1161
1162         spol = mpol_cond_copy(&mpol,
1163                                 mpol_shared_policy_lookup(&info->policy, idx));
1164
1165         /* Create a pseudo vma that just contains the policy */
1166         pvma.vm_start = 0;
1167         pvma.vm_pgoff = idx;
1168         pvma.vm_ops = NULL;
1169         pvma.vm_policy = spol;
1170         page = swapin_readahead(entry, gfp, &pvma, 0);
1171         return page;
1172 }
1173
1174 static struct page *shmem_alloc_page(gfp_t gfp,
1175                         struct shmem_inode_info *info, unsigned long idx)
1176 {
1177         struct vm_area_struct pvma;
1178
1179         /* Create a pseudo vma that just contains the policy */
1180         pvma.vm_start = 0;
1181         pvma.vm_pgoff = idx;
1182         pvma.vm_ops = NULL;
1183         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, idx);
1184
1185         /*
1186          * alloc_page_vma() will drop the shared policy reference
1187          */
1188         return alloc_page_vma(gfp, &pvma, 0);
1189 }
1190 #else /* !CONFIG_NUMA */
1191 #ifdef CONFIG_TMPFS
1192 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *p)
1193 {
1194 }
1195 #endif /* CONFIG_TMPFS */
1196
1197 static inline struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1198                         struct shmem_inode_info *info, unsigned long idx)
1199 {
1200         return swapin_readahead(entry, gfp, NULL, 0);
1201 }
1202
1203 static inline struct page *shmem_alloc_page(gfp_t gfp,
1204                         struct shmem_inode_info *info, unsigned long idx)
1205 {
1206         return alloc_page(gfp);
1207 }
1208 #endif /* CONFIG_NUMA */
1209
1210 #if !defined(CONFIG_NUMA) || !defined(CONFIG_TMPFS)
1211 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1212 {
1213         return NULL;
1214 }
1215 #endif
1216
1217 /*
1218  * shmem_getpage - either get the page from swap or allocate a new one
1219  *
1220  * If we allocate a new one we do not mark it dirty. That's up to the
1221  * vm. If we swap it in we mark it dirty since we also free the swap
1222  * entry since a page cannot live in both the swap and page cache
1223  */
1224 static int shmem_getpage(struct inode *inode, unsigned long idx,
1225                         struct page **pagep, enum sgp_type sgp, int *type)
1226 {
1227         struct address_space *mapping = inode->i_mapping;
1228         struct shmem_inode_info *info = SHMEM_I(inode);
1229         struct shmem_sb_info *sbinfo;
1230         struct page *filepage = *pagep;
1231         struct page *swappage;
1232         struct page *prealloc_page = NULL;
1233         swp_entry_t *entry;
1234         swp_entry_t swap;
1235         gfp_t gfp;
1236         int error;
1237
1238         if (idx >= SHMEM_MAX_INDEX)
1239                 return -EFBIG;
1240
1241         if (type)
1242                 *type = 0;
1243
1244         /*
1245          * Normally, filepage is NULL on entry, and either found
1246          * uptodate immediately, or allocated and zeroed, or read
1247          * in under swappage, which is then assigned to filepage.
1248          * But shmem_readpage (required for splice) passes in a locked
1249          * filepage, which may be found not uptodate by other callers
1250          * too, and may need to be copied from the swappage read in.
1251          */
1252 repeat:
1253         if (!filepage)
1254                 filepage = find_lock_page(mapping, idx);
1255         if (filepage && PageUptodate(filepage))
1256                 goto done;
1257         gfp = mapping_gfp_mask(mapping);
1258         if (!filepage) {
1259                 /*
1260                  * Try to preload while we can wait, to not make a habit of
1261                  * draining atomic reserves; but don't latch on to this cpu.
1262                  */
1263                 error = radix_tree_preload(gfp & ~__GFP_HIGHMEM);
1264                 if (error)
1265                         goto failed;
1266                 radix_tree_preload_end();
1267                 if (sgp != SGP_READ && !prealloc_page) {
1268                         /* We don't care if this fails */
1269                         prealloc_page = shmem_alloc_page(gfp, info, idx);
1270                         if (prealloc_page) {
1271                                 if (mem_cgroup_cache_charge(prealloc_page,
1272                                                 current->mm, GFP_KERNEL)) {
1273                                         page_cache_release(prealloc_page);
1274                                         prealloc_page = NULL;
1275                                 }
1276                         }
1277                 }
1278         }
1279         error = 0;
1280
1281         spin_lock(&info->lock);
1282         shmem_recalc_inode(inode);
1283         entry = shmem_swp_alloc(info, idx, sgp);
1284         if (IS_ERR(entry)) {
1285                 spin_unlock(&info->lock);
1286                 error = PTR_ERR(entry);
1287                 goto failed;
1288         }
1289         swap = *entry;
1290
1291         if (swap.val) {
1292                 /* Look it up and read it in.. */
1293                 swappage = lookup_swap_cache(swap);
1294                 if (!swappage) {
1295                         shmem_swp_unmap(entry);
1296                         /* here we actually do the io */
1297                         if (type && !(*type & VM_FAULT_MAJOR)) {
1298                                 __count_vm_event(PGMAJFAULT);
1299                                 *type |= VM_FAULT_MAJOR;
1300                         }
1301                         spin_unlock(&info->lock);
1302                         swappage = shmem_swapin(swap, gfp, info, idx);
1303                         if (!swappage) {
1304                                 spin_lock(&info->lock);
1305                                 entry = shmem_swp_alloc(info, idx, sgp);
1306                                 if (IS_ERR(entry))
1307                                         error = PTR_ERR(entry);
1308                                 else {
1309                                         if (entry->val == swap.val)
1310                                                 error = -ENOMEM;
1311                                         shmem_swp_unmap(entry);
1312                                 }
1313                                 spin_unlock(&info->lock);
1314                                 if (error)
1315                                         goto failed;
1316                                 goto repeat;
1317                         }
1318                         wait_on_page_locked(swappage);
1319                         page_cache_release(swappage);
1320                         goto repeat;
1321                 }
1322
1323                 /* We have to do this with page locked to prevent races */
1324                 if (!trylock_page(swappage)) {
1325                         shmem_swp_unmap(entry);
1326                         spin_unlock(&info->lock);
1327                         wait_on_page_locked(swappage);
1328                         page_cache_release(swappage);
1329                         goto repeat;
1330                 }
1331                 if (PageWriteback(swappage)) {
1332                         shmem_swp_unmap(entry);
1333                         spin_unlock(&info->lock);
1334                         wait_on_page_writeback(swappage);
1335                         unlock_page(swappage);
1336                         page_cache_release(swappage);
1337                         goto repeat;
1338                 }
1339                 if (!PageUptodate(swappage)) {
1340                         shmem_swp_unmap(entry);
1341                         spin_unlock(&info->lock);
1342                         unlock_page(swappage);
1343                         page_cache_release(swappage);
1344                         error = -EIO;
1345                         goto failed;
1346                 }
1347
1348                 if (filepage) {
1349                         shmem_swp_set(info, entry, 0);
1350                         shmem_swp_unmap(entry);
1351                         delete_from_swap_cache(swappage);
1352                         spin_unlock(&info->lock);
1353                         copy_highpage(filepage, swappage);
1354                         unlock_page(swappage);
1355                         page_cache_release(swappage);
1356                         flush_dcache_page(filepage);
1357                         SetPageUptodate(filepage);
1358                         set_page_dirty(filepage);
1359                         swap_free(swap);
1360                 } else if (!(error = add_to_page_cache_locked(swappage, mapping,
1361                                         idx, GFP_NOWAIT))) {
1362                         info->flags |= SHMEM_PAGEIN;
1363                         shmem_swp_set(info, entry, 0);
1364                         shmem_swp_unmap(entry);
1365                         delete_from_swap_cache(swappage);
1366                         spin_unlock(&info->lock);
1367                         filepage = swappage;
1368                         set_page_dirty(filepage);
1369                         swap_free(swap);
1370                 } else {
1371                         shmem_swp_unmap(entry);
1372                         spin_unlock(&info->lock);
1373                         if (error == -ENOMEM) {
1374                                 /*
1375                                  * reclaim from proper memory cgroup and
1376                                  * call memcg's OOM if needed.
1377                                  */
1378                                 error = mem_cgroup_shmem_charge_fallback(
1379                                                                 swappage,
1380                                                                 current->mm,
1381                                                                 gfp);
1382                                 if (error) {
1383                                         unlock_page(swappage);
1384                                         page_cache_release(swappage);
1385                                         goto failed;
1386                                 }
1387                         }
1388                         unlock_page(swappage);
1389                         page_cache_release(swappage);
1390                         goto repeat;
1391                 }
1392         } else if (sgp == SGP_READ && !filepage) {
1393                 shmem_swp_unmap(entry);
1394                 filepage = find_get_page(mapping, idx);
1395                 if (filepage &&
1396                     (!PageUptodate(filepage) || !trylock_page(filepage))) {
1397                         spin_unlock(&info->lock);
1398                         wait_on_page_locked(filepage);
1399                         page_cache_release(filepage);
1400                         filepage = NULL;
1401                         goto repeat;
1402                 }
1403                 spin_unlock(&info->lock);
1404         } else {
1405                 shmem_swp_unmap(entry);
1406                 sbinfo = SHMEM_SB(inode->i_sb);
1407                 if (sbinfo->max_blocks) {
1408                         if (percpu_counter_compare(&sbinfo->used_blocks,
1409                                                 sbinfo->max_blocks) >= 0 ||
1410                             shmem_acct_block(info->flags))
1411                                 goto nospace;
1412                         percpu_counter_inc(&sbinfo->used_blocks);
1413                         spin_lock(&inode->i_lock);
1414                         inode->i_blocks += BLOCKS_PER_PAGE;
1415                         spin_unlock(&inode->i_lock);
1416                 } else if (shmem_acct_block(info->flags))
1417                         goto nospace;
1418
1419                 if (!filepage) {
1420                         int ret;
1421
1422                         if (!prealloc_page) {
1423                                 spin_unlock(&info->lock);
1424                                 filepage = shmem_alloc_page(gfp, info, idx);
1425                                 if (!filepage) {
1426                                         shmem_unacct_blocks(info->flags, 1);
1427                                         shmem_free_blocks(inode, 1);
1428                                         error = -ENOMEM;
1429                                         goto failed;
1430                                 }
1431                                 SetPageSwapBacked(filepage);
1432
1433                                 /*
1434                                  * Precharge page while we can wait, compensate
1435                                  * after
1436                                  */
1437                                 error = mem_cgroup_cache_charge(filepage,
1438                                         current->mm, GFP_KERNEL);
1439                                 if (error) {
1440                                         page_cache_release(filepage);
1441                                         shmem_unacct_blocks(info->flags, 1);
1442                                         shmem_free_blocks(inode, 1);
1443                                         filepage = NULL;
1444                                         goto failed;
1445                                 }
1446
1447                                 spin_lock(&info->lock);
1448                         } else {
1449                                 filepage = prealloc_page;
1450                                 prealloc_page = NULL;
1451                                 SetPageSwapBacked(filepage);
1452                         }
1453
1454                         entry = shmem_swp_alloc(info, idx, sgp);
1455                         if (IS_ERR(entry))
1456                                 error = PTR_ERR(entry);
1457                         else {
1458                                 swap = *entry;
1459                                 shmem_swp_unmap(entry);
1460                         }
1461                         ret = error || swap.val;
1462                         if (ret)
1463                                 mem_cgroup_uncharge_cache_page(filepage);
1464                         else
1465                                 ret = add_to_page_cache_lru(filepage, mapping,
1466                                                 idx, GFP_NOWAIT);
1467                         /*
1468                          * At add_to_page_cache_lru() failure, uncharge will
1469                          * be done automatically.
1470                          */
1471                         if (ret) {
1472                                 spin_unlock(&info->lock);
1473                                 page_cache_release(filepage);
1474                                 shmem_unacct_blocks(info->flags, 1);
1475                                 shmem_free_blocks(inode, 1);
1476                                 filepage = NULL;
1477                                 if (error)
1478                                         goto failed;
1479                                 goto repeat;
1480                         }
1481                         info->flags |= SHMEM_PAGEIN;
1482                 }
1483
1484                 info->alloced++;
1485                 spin_unlock(&info->lock);
1486                 clear_highpage(filepage);
1487                 flush_dcache_page(filepage);
1488                 SetPageUptodate(filepage);
1489                 if (sgp == SGP_DIRTY)
1490                         set_page_dirty(filepage);
1491         }
1492 done:
1493         *pagep = filepage;
1494         error = 0;
1495         goto out;
1496
1497 nospace:
1498         /*
1499          * Perhaps the page was brought in from swap between find_lock_page
1500          * and taking info->lock?  We allow for that at add_to_page_cache_lru,
1501          * but must also avoid reporting a spurious ENOSPC while working on a
1502          * full tmpfs.  (When filepage has been passed in to shmem_getpage, it
1503          * is already in page cache, which prevents this race from occurring.)
1504          */
1505         if (!filepage) {
1506                 struct page *page = find_get_page(mapping, idx);
1507                 if (page) {
1508                         spin_unlock(&info->lock);
1509                         page_cache_release(page);
1510                         goto repeat;
1511                 }
1512         }
1513         spin_unlock(&info->lock);
1514         error = -ENOSPC;
1515 failed:
1516         if (*pagep != filepage) {
1517                 unlock_page(filepage);
1518                 page_cache_release(filepage);
1519         }
1520 out:
1521         if (prealloc_page) {
1522                 mem_cgroup_uncharge_cache_page(prealloc_page);
1523                 page_cache_release(prealloc_page);
1524         }
1525         return error;
1526 }
1527
1528 static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1529 {
1530         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1531         int error;
1532         int ret;
1533
1534         if (((loff_t)vmf->pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
1535                 return VM_FAULT_SIGBUS;
1536
1537         error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_CACHE, &ret);
1538         if (error)
1539                 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1540
1541         return ret | VM_FAULT_LOCKED;
1542 }
1543
1544 #ifdef CONFIG_NUMA
1545 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *new)
1546 {
1547         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1548         return mpol_set_shared_policy(&SHMEM_I(i)->policy, vma, new);
1549 }
1550
1551 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
1552                                           unsigned long addr)
1553 {
1554         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1555         unsigned long idx;
1556
1557         idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1558         return mpol_shared_policy_lookup(&SHMEM_I(i)->policy, idx);
1559 }
1560 #endif
1561
1562 int shmem_lock(struct file *file, int lock, struct user_struct *user)
1563 {
1564         struct inode *inode = file->f_path.dentry->d_inode;
1565         struct shmem_inode_info *info = SHMEM_I(inode);
1566         int retval = -ENOMEM;
1567
1568         spin_lock(&info->lock);
1569         if (lock && !(info->flags & VM_LOCKED)) {
1570                 if (!user_shm_lock(inode->i_size, user))
1571                         goto out_nomem;
1572                 info->flags |= VM_LOCKED;
1573                 mapping_set_unevictable(file->f_mapping);
1574         }
1575         if (!lock && (info->flags & VM_LOCKED) && user) {
1576                 user_shm_unlock(inode->i_size, user);
1577                 info->flags &= ~VM_LOCKED;
1578                 mapping_clear_unevictable(file->f_mapping);
1579                 scan_mapping_unevictable_pages(file->f_mapping);
1580         }
1581         retval = 0;
1582
1583 out_nomem:
1584         spin_unlock(&info->lock);
1585         return retval;
1586 }
1587
1588 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1589 {
1590         file_accessed(file);
1591         vma->vm_ops = &shmem_vm_ops;
1592         vma->vm_flags |= VM_CAN_NONLINEAR;
1593         return 0;
1594 }
1595
1596 static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
1597                                      int mode, dev_t dev, unsigned long flags)
1598 {
1599         struct inode *inode;
1600         struct shmem_inode_info *info;
1601         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1602
1603         if (shmem_reserve_inode(sb))
1604                 return NULL;
1605
1606         inode = new_inode(sb);
1607         if (inode) {
1608                 inode->i_ino = get_next_ino();
1609                 inode_init_owner(inode, dir, mode);
1610                 inode->i_blocks = 0;
1611                 inode->i_mapping->backing_dev_info = &shmem_backing_dev_info;
1612                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1613                 inode->i_generation = get_seconds();
1614                 info = SHMEM_I(inode);
1615                 memset(info, 0, (char *)inode - (char *)info);
1616                 spin_lock_init(&info->lock);
1617                 info->flags = flags & VM_NORESERVE;
1618                 INIT_LIST_HEAD(&info->swaplist);
1619                 cache_no_acl(inode);
1620
1621                 switch (mode & S_IFMT) {
1622                 default:
1623                         inode->i_op = &shmem_special_inode_operations;
1624                         init_special_inode(inode, mode, dev);
1625                         break;
1626                 case S_IFREG:
1627                         inode->i_mapping->a_ops = &shmem_aops;
1628                         inode->i_op = &shmem_inode_operations;
1629                         inode->i_fop = &shmem_file_operations;
1630                         mpol_shared_policy_init(&info->policy,
1631                                                  shmem_get_sbmpol(sbinfo));
1632                         break;
1633                 case S_IFDIR:
1634                         inc_nlink(inode);
1635                         /* Some things misbehave if size == 0 on a directory */
1636                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
1637                         inode->i_op = &shmem_dir_inode_operations;
1638                         inode->i_fop = &simple_dir_operations;
1639                         break;
1640                 case S_IFLNK:
1641                         /*
1642                          * Must not load anything in the rbtree,
1643                          * mpol_free_shared_policy will not be called.
1644                          */
1645                         mpol_shared_policy_init(&info->policy, NULL);
1646                         break;
1647                 }
1648         } else
1649                 shmem_free_inode(sb);
1650         return inode;
1651 }
1652
1653 #ifdef CONFIG_TMPFS
1654 static const struct inode_operations shmem_symlink_inode_operations;
1655 static const struct inode_operations shmem_symlink_inline_operations;
1656
1657 /*
1658  * Normally tmpfs avoids the use of shmem_readpage and shmem_write_begin;
1659  * but providing them allows a tmpfs file to be used for splice, sendfile, and
1660  * below the loop driver, in the generic fashion that many filesystems support.
1661  */
1662 static int shmem_readpage(struct file *file, struct page *page)
1663 {
1664         struct inode *inode = page->mapping->host;
1665         int error = shmem_getpage(inode, page->index, &page, SGP_CACHE, NULL);
1666         unlock_page(page);
1667         return error;
1668 }
1669
1670 static int
1671 shmem_write_begin(struct file *file, struct address_space *mapping,
1672                         loff_t pos, unsigned len, unsigned flags,
1673                         struct page **pagep, void **fsdata)
1674 {
1675         struct inode *inode = mapping->host;
1676         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1677         *pagep = NULL;
1678         return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1679 }
1680
1681 static int
1682 shmem_write_end(struct file *file, struct address_space *mapping,
1683                         loff_t pos, unsigned len, unsigned copied,
1684                         struct page *page, void *fsdata)
1685 {
1686         struct inode *inode = mapping->host;
1687
1688         if (pos + copied > inode->i_size)
1689                 i_size_write(inode, pos + copied);
1690
1691         set_page_dirty(page);
1692         unlock_page(page);
1693         page_cache_release(page);
1694
1695         return copied;
1696 }
1697
1698 static void do_shmem_file_read(struct file *filp, loff_t *ppos, read_descriptor_t *desc, read_actor_t actor)
1699 {
1700         struct inode *inode = filp->f_path.dentry->d_inode;
1701         struct address_space *mapping = inode->i_mapping;
1702         unsigned long index, offset;
1703         enum sgp_type sgp = SGP_READ;
1704
1705         /*
1706          * Might this read be for a stacking filesystem?  Then when reading
1707          * holes of a sparse file, we actually need to allocate those pages,
1708          * and even mark them dirty, so it cannot exceed the max_blocks limit.
1709          */
1710         if (segment_eq(get_fs(), KERNEL_DS))
1711                 sgp = SGP_DIRTY;
1712
1713         index = *ppos >> PAGE_CACHE_SHIFT;
1714         offset = *ppos & ~PAGE_CACHE_MASK;
1715
1716         for (;;) {
1717                 struct page *page = NULL;
1718                 unsigned long end_index, nr, ret;
1719                 loff_t i_size = i_size_read(inode);
1720
1721                 end_index = i_size >> PAGE_CACHE_SHIFT;
1722                 if (index > end_index)
1723                         break;
1724                 if (index == end_index) {
1725                         nr = i_size & ~PAGE_CACHE_MASK;
1726                         if (nr <= offset)
1727                                 break;
1728                 }
1729
1730                 desc->error = shmem_getpage(inode, index, &page, sgp, NULL);
1731                 if (desc->error) {
1732                         if (desc->error == -EINVAL)
1733                                 desc->error = 0;
1734                         break;
1735                 }
1736                 if (page)
1737                         unlock_page(page);
1738
1739                 /*
1740                  * We must evaluate after, since reads (unlike writes)
1741                  * are called without i_mutex protection against truncate
1742                  */
1743                 nr = PAGE_CACHE_SIZE;
1744                 i_size = i_size_read(inode);
1745                 end_index = i_size >> PAGE_CACHE_SHIFT;
1746                 if (index == end_index) {
1747                         nr = i_size & ~PAGE_CACHE_MASK;
1748                         if (nr <= offset) {
1749                                 if (page)
1750                                         page_cache_release(page);
1751                                 break;
1752                         }
1753                 }
1754                 nr -= offset;
1755
1756                 if (page) {
1757                         /*
1758                          * If users can be writing to this page using arbitrary
1759                          * virtual addresses, take care about potential aliasing
1760                          * before reading the page on the kernel side.
1761                          */
1762                         if (mapping_writably_mapped(mapping))
1763                                 flush_dcache_page(page);
1764                         /*
1765                          * Mark the page accessed if we read the beginning.
1766                          */
1767                         if (!offset)
1768                                 mark_page_accessed(page);
1769                 } else {
1770                         page = ZERO_PAGE(0);
1771                         page_cache_get(page);
1772                 }
1773
1774                 /*
1775                  * Ok, we have the page, and it's up-to-date, so
1776                  * now we can copy it to user space...
1777                  *
1778                  * The actor routine returns how many bytes were actually used..
1779                  * NOTE! This may not be the same as how much of a user buffer
1780                  * we filled up (we may be padding etc), so we can only update
1781                  * "pos" here (the actor routine has to update the user buffer
1782                  * pointers and the remaining count).
1783                  */
1784                 ret = actor(desc, page, offset, nr);
1785                 offset += ret;
1786                 index += offset >> PAGE_CACHE_SHIFT;
1787                 offset &= ~PAGE_CACHE_MASK;
1788
1789                 page_cache_release(page);
1790                 if (ret != nr || !desc->count)
1791                         break;
1792
1793                 cond_resched();
1794         }
1795
1796         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1797         file_accessed(filp);
1798 }
1799
1800 static ssize_t shmem_file_aio_read(struct kiocb *iocb,
1801                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
1802 {
1803         struct file *filp = iocb->ki_filp;
1804         ssize_t retval;
1805         unsigned long seg;
1806         size_t count;
1807         loff_t *ppos = &iocb->ki_pos;
1808
1809         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1810         if (retval)
1811                 return retval;
1812
1813         for (seg = 0; seg < nr_segs; seg++) {
1814                 read_descriptor_t desc;
1815
1816                 desc.written = 0;
1817                 desc.arg.buf = iov[seg].iov_base;
1818                 desc.count = iov[seg].iov_len;
1819                 if (desc.count == 0)
1820                         continue;
1821                 desc.error = 0;
1822                 do_shmem_file_read(filp, ppos, &desc, file_read_actor);
1823                 retval += desc.written;
1824                 if (desc.error) {
1825                         retval = retval ?: desc.error;
1826                         break;
1827                 }
1828                 if (desc.count > 0)
1829                         break;
1830         }
1831         return retval;
1832 }
1833
1834 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
1835 {
1836         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
1837
1838         buf->f_type = TMPFS_MAGIC;
1839         buf->f_bsize = PAGE_CACHE_SIZE;
1840         buf->f_namelen = NAME_MAX;
1841         if (sbinfo->max_blocks) {
1842                 buf->f_blocks = sbinfo->max_blocks;
1843                 buf->f_bavail = buf->f_bfree =
1844                                 sbinfo->max_blocks - percpu_counter_sum(&sbinfo->used_blocks);
1845         }
1846         if (sbinfo->max_inodes) {
1847                 buf->f_files = sbinfo->max_inodes;
1848                 buf->f_ffree = sbinfo->free_inodes;
1849         }
1850         /* else leave those fields 0 like simple_statfs */
1851         return 0;
1852 }
1853
1854 /*
1855  * File creation. Allocate an inode, and we're done..
1856  */
1857 static int
1858 shmem_mknod(struct inode *dir, struct dentry *dentry, int mode, dev_t dev)
1859 {
1860         struct inode *inode;
1861         int error = -ENOSPC;
1862
1863         inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
1864         if (inode) {
1865                 error = security_inode_init_security(inode, dir,
1866                                                      &dentry->d_name, NULL,
1867                                                      NULL, NULL);
1868                 if (error) {
1869                         if (error != -EOPNOTSUPP) {
1870                                 iput(inode);
1871                                 return error;
1872                         }
1873                 }
1874 #ifdef CONFIG_TMPFS_POSIX_ACL
1875                 error = generic_acl_init(inode, dir);
1876                 if (error) {
1877                         iput(inode);
1878                         return error;
1879                 }
1880 #else
1881                 error = 0;
1882 #endif
1883                 dir->i_size += BOGO_DIRENT_SIZE;
1884                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1885                 d_instantiate(dentry, inode);
1886                 dget(dentry); /* Extra count - pin the dentry in core */
1887         }
1888         return error;
1889 }
1890
1891 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, int mode)
1892 {
1893         int error;
1894
1895         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
1896                 return error;
1897         inc_nlink(dir);
1898         return 0;
1899 }
1900
1901 static int shmem_create(struct inode *dir, struct dentry *dentry, int mode,
1902                 struct nameidata *nd)
1903 {
1904         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
1905 }
1906
1907 /*
1908  * Link a file..
1909  */
1910 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
1911 {
1912         struct inode *inode = old_dentry->d_inode;
1913         int ret;
1914
1915         /*
1916          * No ordinary (disk based) filesystem counts links as inodes;
1917          * but each new link needs a new dentry, pinning lowmem, and
1918          * tmpfs dentries cannot be pruned until they are unlinked.
1919          */
1920         ret = shmem_reserve_inode(inode->i_sb);
1921         if (ret)
1922                 goto out;
1923
1924         dir->i_size += BOGO_DIRENT_SIZE;
1925         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1926         inc_nlink(inode);
1927         ihold(inode);   /* New dentry reference */
1928         dget(dentry);           /* Extra pinning count for the created dentry */
1929         d_instantiate(dentry, inode);
1930 out:
1931         return ret;
1932 }
1933
1934 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
1935 {
1936         struct inode *inode = dentry->d_inode;
1937
1938         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
1939                 shmem_free_inode(inode->i_sb);
1940
1941         dir->i_size -= BOGO_DIRENT_SIZE;
1942         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1943         drop_nlink(inode);
1944         dput(dentry);   /* Undo the count from "create" - this does all the work */
1945         return 0;
1946 }
1947
1948 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
1949 {
1950         if (!simple_empty(dentry))
1951                 return -ENOTEMPTY;
1952
1953         drop_nlink(dentry->d_inode);
1954         drop_nlink(dir);
1955         return shmem_unlink(dir, dentry);
1956 }
1957
1958 /*
1959  * The VFS layer already does all the dentry stuff for rename,
1960  * we just have to decrement the usage count for the target if
1961  * it exists so that the VFS layer correctly free's it when it
1962  * gets overwritten.
1963  */
1964 static int shmem_rename(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
1965 {
1966         struct inode *inode = old_dentry->d_inode;
1967         int they_are_dirs = S_ISDIR(inode->i_mode);
1968
1969         if (!simple_empty(new_dentry))
1970                 return -ENOTEMPTY;
1971
1972         if (new_dentry->d_inode) {
1973                 (void) shmem_unlink(new_dir, new_dentry);
1974                 if (they_are_dirs)
1975                         drop_nlink(old_dir);
1976         } else if (they_are_dirs) {
1977                 drop_nlink(old_dir);
1978                 inc_nlink(new_dir);
1979         }
1980
1981         old_dir->i_size -= BOGO_DIRENT_SIZE;
1982         new_dir->i_size += BOGO_DIRENT_SIZE;
1983         old_dir->i_ctime = old_dir->i_mtime =
1984         new_dir->i_ctime = new_dir->i_mtime =
1985         inode->i_ctime = CURRENT_TIME;
1986         return 0;
1987 }
1988
1989 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
1990 {
1991         int error;
1992         int len;
1993         struct inode *inode;
1994         struct page *page = NULL;
1995         char *kaddr;
1996         struct shmem_inode_info *info;
1997
1998         len = strlen(symname) + 1;
1999         if (len > PAGE_CACHE_SIZE)
2000                 return -ENAMETOOLONG;
2001
2002         inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0, VM_NORESERVE);
2003         if (!inode)
2004                 return -ENOSPC;
2005
2006         error = security_inode_init_security(inode, dir, &dentry->d_name, NULL,
2007                                              NULL, NULL);
2008         if (error) {
2009                 if (error != -EOPNOTSUPP) {
2010                         iput(inode);
2011                         return error;
2012                 }
2013                 error = 0;
2014         }
2015
2016         info = SHMEM_I(inode);
2017         inode->i_size = len-1;
2018         if (len <= (char *)inode - (char *)info) {
2019                 /* do it inline */
2020                 memcpy(info, symname, len);
2021                 inode->i_op = &shmem_symlink_inline_operations;
2022         } else {
2023                 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
2024                 if (error) {
2025                         iput(inode);
2026                         return error;
2027                 }
2028                 inode->i_mapping->a_ops = &shmem_aops;
2029                 inode->i_op = &shmem_symlink_inode_operations;
2030                 kaddr = kmap_atomic(page, KM_USER0);
2031                 memcpy(kaddr, symname, len);
2032                 kunmap_atomic(kaddr, KM_USER0);
2033                 set_page_dirty(page);
2034                 unlock_page(page);
2035                 page_cache_release(page);
2036         }
2037         dir->i_size += BOGO_DIRENT_SIZE;
2038         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2039         d_instantiate(dentry, inode);
2040         dget(dentry);
2041         return 0;
2042 }
2043
2044 static void *shmem_follow_link_inline(struct dentry *dentry, struct nameidata *nd)
2045 {
2046         nd_set_link(nd, (char *)SHMEM_I(dentry->d_inode));
2047         return NULL;
2048 }
2049
2050 static void *shmem_follow_link(struct dentry *dentry, struct nameidata *nd)
2051 {
2052         struct page *page = NULL;
2053         int res = shmem_getpage(dentry->d_inode, 0, &page, SGP_READ, NULL);
2054         nd_set_link(nd, res ? ERR_PTR(res) : kmap(page));
2055         if (page)
2056                 unlock_page(page);
2057         return page;
2058 }
2059
2060 static void shmem_put_link(struct dentry *dentry, struct nameidata *nd, void *cookie)
2061 {
2062         if (!IS_ERR(nd_get_link(nd))) {
2063                 struct page *page = cookie;
2064                 kunmap(page);
2065                 mark_page_accessed(page);
2066                 page_cache_release(page);
2067         }
2068 }
2069
2070 static const struct inode_operations shmem_symlink_inline_operations = {
2071         .readlink       = generic_readlink,
2072         .follow_link    = shmem_follow_link_inline,
2073 };
2074
2075 static const struct inode_operations shmem_symlink_inode_operations = {
2076         .readlink       = generic_readlink,
2077         .follow_link    = shmem_follow_link,
2078         .put_link       = shmem_put_link,
2079 };
2080
2081 #ifdef CONFIG_TMPFS_POSIX_ACL
2082 /*
2083  * Superblocks without xattr inode operations will get security.* xattr
2084  * support from the VFS "for free". As soon as we have any other xattrs
2085  * like ACLs, we also need to implement the security.* handlers at
2086  * filesystem level, though.
2087  */
2088
2089 static size_t shmem_xattr_security_list(struct dentry *dentry, char *list,
2090                                         size_t list_len, const char *name,
2091                                         size_t name_len, int handler_flags)
2092 {
2093         return security_inode_listsecurity(dentry->d_inode, list, list_len);
2094 }
2095
2096 static int shmem_xattr_security_get(struct dentry *dentry, const char *name,
2097                 void *buffer, size_t size, int handler_flags)
2098 {
2099         if (strcmp(name, "") == 0)
2100                 return -EINVAL;
2101         return xattr_getsecurity(dentry->d_inode, name, buffer, size);
2102 }
2103
2104 static int shmem_xattr_security_set(struct dentry *dentry, const char *name,
2105                 const void *value, size_t size, int flags, int handler_flags)
2106 {
2107         if (strcmp(name, "") == 0)
2108                 return -EINVAL;
2109         return security_inode_setsecurity(dentry->d_inode, name, value,
2110                                           size, flags);
2111 }
2112
2113 static const struct xattr_handler shmem_xattr_security_handler = {
2114         .prefix = XATTR_SECURITY_PREFIX,
2115         .list   = shmem_xattr_security_list,
2116         .get    = shmem_xattr_security_get,
2117         .set    = shmem_xattr_security_set,
2118 };
2119
2120 static const struct xattr_handler *shmem_xattr_handlers[] = {
2121         &generic_acl_access_handler,
2122         &generic_acl_default_handler,
2123         &shmem_xattr_security_handler,
2124         NULL
2125 };
2126 #endif
2127
2128 static struct dentry *shmem_get_parent(struct dentry *child)
2129 {
2130         return ERR_PTR(-ESTALE);
2131 }
2132
2133 static int shmem_match(struct inode *ino, void *vfh)
2134 {
2135         __u32 *fh = vfh;
2136         __u64 inum = fh[2];
2137         inum = (inum << 32) | fh[1];
2138         return ino->i_ino == inum && fh[0] == ino->i_generation;
2139 }
2140
2141 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
2142                 struct fid *fid, int fh_len, int fh_type)
2143 {
2144         struct inode *inode;
2145         struct dentry *dentry = NULL;
2146         u64 inum = fid->raw[2];
2147         inum = (inum << 32) | fid->raw[1];
2148
2149         if (fh_len < 3)
2150                 return NULL;
2151
2152         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
2153                         shmem_match, fid->raw);
2154         if (inode) {
2155                 dentry = d_find_alias(inode);
2156                 iput(inode);
2157         }
2158
2159         return dentry;
2160 }
2161
2162 static int shmem_encode_fh(struct dentry *dentry, __u32 *fh, int *len,
2163                                 int connectable)
2164 {
2165         struct inode *inode = dentry->d_inode;
2166
2167         if (*len < 3) {
2168                 *len = 3;
2169                 return 255;
2170         }
2171
2172         if (inode_unhashed(inode)) {
2173                 /* Unfortunately insert_inode_hash is not idempotent,
2174                  * so as we hash inodes here rather than at creation
2175                  * time, we need a lock to ensure we only try
2176                  * to do it once
2177                  */
2178                 static DEFINE_SPINLOCK(lock);
2179                 spin_lock(&lock);
2180                 if (inode_unhashed(inode))
2181                         __insert_inode_hash(inode,
2182                                             inode->i_ino + inode->i_generation);
2183                 spin_unlock(&lock);
2184         }
2185
2186         fh[0] = inode->i_generation;
2187         fh[1] = inode->i_ino;
2188         fh[2] = ((__u64)inode->i_ino) >> 32;
2189
2190         *len = 3;
2191         return 1;
2192 }
2193
2194 static const struct export_operations shmem_export_ops = {
2195         .get_parent     = shmem_get_parent,
2196         .encode_fh      = shmem_encode_fh,
2197         .fh_to_dentry   = shmem_fh_to_dentry,
2198 };
2199
2200 static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
2201                                bool remount)
2202 {
2203         char *this_char, *value, *rest;
2204
2205         while (options != NULL) {
2206                 this_char = options;
2207                 for (;;) {
2208                         /*
2209                          * NUL-terminate this option: unfortunately,
2210                          * mount options form a comma-separated list,
2211                          * but mpol's nodelist may also contain commas.
2212                          */
2213                         options = strchr(options, ',');
2214                         if (options == NULL)
2215                                 break;
2216                         options++;
2217                         if (!isdigit(*options)) {
2218                                 options[-1] = '\0';
2219                                 break;
2220                         }
2221                 }
2222                 if (!*this_char)
2223                         continue;
2224                 if ((value = strchr(this_char,'=')) != NULL) {
2225                         *value++ = 0;
2226                 } else {
2227                         printk(KERN_ERR
2228                             "tmpfs: No value for mount option '%s'\n",
2229                             this_char);
2230                         return 1;
2231                 }
2232
2233                 if (!strcmp(this_char,"size")) {
2234                         unsigned long long size;
2235                         size = memparse(value,&rest);
2236                         if (*rest == '%') {
2237                                 size <<= PAGE_SHIFT;
2238                                 size *= totalram_pages;
2239                                 do_div(size, 100);
2240                                 rest++;
2241                         }
2242                         if (*rest)
2243                                 goto bad_val;
2244                         sbinfo->max_blocks =
2245                                 DIV_ROUND_UP(size, PAGE_CACHE_SIZE);
2246                 } else if (!strcmp(this_char,"nr_blocks")) {
2247                         sbinfo->max_blocks = memparse(value, &rest);
2248                         if (*rest)
2249                                 goto bad_val;
2250                 } else if (!strcmp(this_char,"nr_inodes")) {
2251                         sbinfo->max_inodes = memparse(value, &rest);
2252                         if (*rest)
2253                                 goto bad_val;
2254                 } else if (!strcmp(this_char,"mode")) {
2255                         if (remount)
2256                                 continue;
2257                         sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
2258                         if (*rest)
2259                                 goto bad_val;
2260                 } else if (!strcmp(this_char,"uid")) {
2261                         if (remount)
2262                                 continue;
2263                         sbinfo->uid = simple_strtoul(value, &rest, 0);
2264                         if (*rest)
2265                                 goto bad_val;
2266                 } else if (!strcmp(this_char,"gid")) {
2267                         if (remount)
2268                                 continue;
2269                         sbinfo->gid = simple_strtoul(value, &rest, 0);
2270                         if (*rest)
2271                                 goto bad_val;
2272                 } else if (!strcmp(this_char,"mpol")) {
2273                         if (mpol_parse_str(value, &sbinfo->mpol, 1))
2274                                 goto bad_val;
2275                 } else {
2276                         printk(KERN_ERR "tmpfs: Bad mount option %s\n",
2277                                this_char);
2278                         return 1;
2279                 }
2280         }
2281         return 0;
2282
2283 bad_val:
2284         printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
2285                value, this_char);
2286         return 1;
2287
2288 }
2289
2290 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2291 {
2292         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2293         struct shmem_sb_info config = *sbinfo;
2294         unsigned long inodes;
2295         int error = -EINVAL;
2296
2297         if (shmem_parse_options(data, &config, true))
2298                 return error;
2299
2300         spin_lock(&sbinfo->stat_lock);
2301         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2302         if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
2303                 goto out;
2304         if (config.max_inodes < inodes)
2305                 goto out;
2306         /*
2307          * Those tests also disallow limited->unlimited while any are in
2308          * use, so i_blocks will always be zero when max_blocks is zero;
2309          * but we must separately disallow unlimited->limited, because
2310          * in that case we have no record of how much is already in use.
2311          */
2312         if (config.max_blocks && !sbinfo->max_blocks)
2313                 goto out;
2314         if (config.max_inodes && !sbinfo->max_inodes)
2315                 goto out;
2316
2317         error = 0;
2318         sbinfo->max_blocks  = config.max_blocks;
2319         sbinfo->max_inodes  = config.max_inodes;
2320         sbinfo->free_inodes = config.max_inodes - inodes;
2321
2322         mpol_put(sbinfo->mpol);
2323         sbinfo->mpol        = config.mpol;      /* transfers initial ref */
2324 out:
2325         spin_unlock(&sbinfo->stat_lock);
2326         return error;
2327 }
2328
2329 static int shmem_show_options(struct seq_file *seq, struct vfsmount *vfs)
2330 {
2331         struct shmem_sb_info *sbinfo = SHMEM_SB(vfs->mnt_sb);
2332
2333         if (sbinfo->max_blocks != shmem_default_max_blocks())
2334                 seq_printf(seq, ",size=%luk",
2335                         sbinfo->max_blocks << (PAGE_CACHE_SHIFT - 10));
2336         if (sbinfo->max_inodes != shmem_default_max_inodes())
2337                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
2338         if (sbinfo->mode != (S_IRWXUGO | S_ISVTX))
2339                 seq_printf(seq, ",mode=%03o", sbinfo->mode);
2340         if (sbinfo->uid != 0)
2341                 seq_printf(seq, ",uid=%u", sbinfo->uid);
2342         if (sbinfo->gid != 0)
2343                 seq_printf(seq, ",gid=%u", sbinfo->gid);
2344         shmem_show_mpol(seq, sbinfo->mpol);
2345         return 0;
2346 }
2347 #endif /* CONFIG_TMPFS */
2348
2349 static void shmem_put_super(struct super_block *sb)
2350 {
2351         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2352
2353         percpu_counter_destroy(&sbinfo->used_blocks);
2354         kfree(sbinfo);
2355         sb->s_fs_info = NULL;
2356 }
2357
2358 int shmem_fill_super(struct super_block *sb, void *data, int silent)
2359 {
2360         struct inode *inode;
2361         struct dentry *root;
2362         struct shmem_sb_info *sbinfo;
2363         int err = -ENOMEM;
2364
2365         /* Round up to L1_CACHE_BYTES to resist false sharing */
2366         sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
2367                                 L1_CACHE_BYTES), GFP_KERNEL);
2368         if (!sbinfo)
2369                 return -ENOMEM;
2370
2371         sbinfo->mode = S_IRWXUGO | S_ISVTX;
2372         sbinfo->uid = current_fsuid();
2373         sbinfo->gid = current_fsgid();
2374         sb->s_fs_info = sbinfo;
2375
2376 #ifdef CONFIG_TMPFS
2377         /*
2378          * Per default we only allow half of the physical ram per
2379          * tmpfs instance, limiting inodes to one per page of lowmem;
2380          * but the internal instance is left unlimited.
2381          */
2382         if (!(sb->s_flags & MS_NOUSER)) {
2383                 sbinfo->max_blocks = shmem_default_max_blocks();
2384                 sbinfo->max_inodes = shmem_default_max_inodes();
2385                 if (shmem_parse_options(data, sbinfo, false)) {
2386                         err = -EINVAL;
2387                         goto failed;
2388                 }
2389         }
2390         sb->s_export_op = &shmem_export_ops;
2391 #else
2392         sb->s_flags |= MS_NOUSER;
2393 #endif
2394
2395         spin_lock_init(&sbinfo->stat_lock);
2396         if (percpu_counter_init(&sbinfo->used_blocks, 0))
2397                 goto failed;
2398         sbinfo->free_inodes = sbinfo->max_inodes;
2399
2400         sb->s_maxbytes = SHMEM_MAX_BYTES;
2401         sb->s_blocksize = PAGE_CACHE_SIZE;
2402         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
2403         sb->s_magic = TMPFS_MAGIC;
2404         sb->s_op = &shmem_ops;
2405         sb->s_time_gran = 1;
2406 #ifdef CONFIG_TMPFS_POSIX_ACL
2407         sb->s_xattr = shmem_xattr_handlers;
2408         sb->s_flags |= MS_POSIXACL;
2409 #endif
2410
2411         inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
2412         if (!inode)
2413                 goto failed;
2414         inode->i_uid = sbinfo->uid;
2415         inode->i_gid = sbinfo->gid;
2416         root = d_alloc_root(inode);
2417         if (!root)
2418                 goto failed_iput;
2419         sb->s_root = root;
2420         return 0;
2421
2422 failed_iput:
2423         iput(inode);
2424 failed:
2425         shmem_put_super(sb);
2426         return err;
2427 }
2428
2429 static struct kmem_cache *shmem_inode_cachep;
2430
2431 static struct inode *shmem_alloc_inode(struct super_block *sb)
2432 {
2433         struct shmem_inode_info *p;
2434         p = (struct shmem_inode_info *)kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
2435         if (!p)
2436                 return NULL;
2437         return &p->vfs_inode;
2438 }
2439
2440 static void shmem_i_callback(struct rcu_head *head)
2441 {
2442         struct inode *inode = container_of(head, struct inode, i_rcu);
2443         INIT_LIST_HEAD(&inode->i_dentry);
2444         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
2445 }
2446
2447 static void shmem_destroy_inode(struct inode *inode)
2448 {
2449         if ((inode->i_mode & S_IFMT) == S_IFREG) {
2450                 /* only struct inode is valid if it's an inline symlink */
2451                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
2452         }
2453         call_rcu(&inode->i_rcu, shmem_i_callback);
2454 }
2455
2456 static void init_once(void *foo)
2457 {
2458         struct shmem_inode_info *p = (struct shmem_inode_info *) foo;
2459
2460         inode_init_once(&p->vfs_inode);
2461 }
2462
2463 static int init_inodecache(void)
2464 {
2465         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
2466                                 sizeof(struct shmem_inode_info),
2467                                 0, SLAB_PANIC, init_once);
2468         return 0;
2469 }
2470
2471 static void destroy_inodecache(void)
2472 {
2473         kmem_cache_destroy(shmem_inode_cachep);
2474 }
2475
2476 static const struct address_space_operations shmem_aops = {
2477         .writepage      = shmem_writepage,
2478         .set_page_dirty = __set_page_dirty_no_writeback,
2479 #ifdef CONFIG_TMPFS
2480         .readpage       = shmem_readpage,
2481         .write_begin    = shmem_write_begin,
2482         .write_end      = shmem_write_end,
2483 #endif
2484         .migratepage    = migrate_page,
2485         .error_remove_page = generic_error_remove_page,
2486 };
2487
2488 static const struct file_operations shmem_file_operations = {
2489         .mmap           = shmem_mmap,
2490 #ifdef CONFIG_TMPFS
2491         .llseek         = generic_file_llseek,
2492         .read           = do_sync_read,
2493         .write          = do_sync_write,
2494         .aio_read       = shmem_file_aio_read,
2495         .aio_write      = generic_file_aio_write,
2496         .fsync          = noop_fsync,
2497         .splice_read    = generic_file_splice_read,
2498         .splice_write   = generic_file_splice_write,
2499 #endif
2500 };
2501
2502 static const struct inode_operations shmem_inode_operations = {
2503         .setattr        = shmem_notify_change,
2504         .truncate_range = shmem_truncate_range,
2505 #ifdef CONFIG_TMPFS_POSIX_ACL
2506         .setxattr       = generic_setxattr,
2507         .getxattr       = generic_getxattr,
2508         .listxattr      = generic_listxattr,
2509         .removexattr    = generic_removexattr,
2510         .check_acl      = generic_check_acl,
2511 #endif
2512
2513 };
2514
2515 static const struct inode_operations shmem_dir_inode_operations = {
2516 #ifdef CONFIG_TMPFS
2517         .create         = shmem_create,
2518         .lookup         = simple_lookup,
2519         .link           = shmem_link,
2520         .unlink         = shmem_unlink,
2521         .symlink        = shmem_symlink,
2522         .mkdir          = shmem_mkdir,
2523         .rmdir          = shmem_rmdir,
2524         .mknod          = shmem_mknod,
2525         .rename         = shmem_rename,
2526 #endif
2527 #ifdef CONFIG_TMPFS_POSIX_ACL
2528         .setattr        = shmem_notify_change,
2529         .setxattr       = generic_setxattr,
2530         .getxattr       = generic_getxattr,
2531         .listxattr      = generic_listxattr,
2532         .removexattr    = generic_removexattr,
2533         .check_acl      = generic_check_acl,
2534 #endif
2535 };
2536
2537 static const struct inode_operations shmem_special_inode_operations = {
2538 #ifdef CONFIG_TMPFS_POSIX_ACL
2539         .setattr        = shmem_notify_change,
2540         .setxattr       = generic_setxattr,
2541         .getxattr       = generic_getxattr,
2542         .listxattr      = generic_listxattr,
2543         .removexattr    = generic_removexattr,
2544         .check_acl      = generic_check_acl,
2545 #endif
2546 };
2547
2548 static const struct super_operations shmem_ops = {
2549         .alloc_inode    = shmem_alloc_inode,
2550         .destroy_inode  = shmem_destroy_inode,
2551 #ifdef CONFIG_TMPFS
2552         .statfs         = shmem_statfs,
2553         .remount_fs     = shmem_remount_fs,
2554         .show_options   = shmem_show_options,
2555 #endif
2556         .evict_inode    = shmem_evict_inode,
2557         .drop_inode     = generic_delete_inode,
2558         .put_super      = shmem_put_super,
2559 };
2560
2561 static const struct vm_operations_struct shmem_vm_ops = {
2562         .fault          = shmem_fault,
2563 #ifdef CONFIG_NUMA
2564         .set_policy     = shmem_set_policy,
2565         .get_policy     = shmem_get_policy,
2566 #endif
2567 };
2568
2569
2570 static struct dentry *shmem_mount(struct file_system_type *fs_type,
2571         int flags, const char *dev_name, void *data)
2572 {
2573         return mount_nodev(fs_type, flags, data, shmem_fill_super);
2574 }
2575
2576 static struct file_system_type tmpfs_fs_type = {
2577         .owner          = THIS_MODULE,
2578         .name           = "tmpfs",
2579         .mount          = shmem_mount,
2580         .kill_sb        = kill_litter_super,
2581 };
2582
2583 int __init init_tmpfs(void)
2584 {
2585         int error;
2586
2587         error = bdi_init(&shmem_backing_dev_info);
2588         if (error)
2589                 goto out4;
2590
2591         error = init_inodecache();
2592         if (error)
2593                 goto out3;
2594
2595         error = register_filesystem(&tmpfs_fs_type);
2596         if (error) {
2597                 printk(KERN_ERR "Could not register tmpfs\n");
2598                 goto out2;
2599         }
2600
2601         shm_mnt = vfs_kern_mount(&tmpfs_fs_type, MS_NOUSER,
2602                                 tmpfs_fs_type.name, NULL);
2603         if (IS_ERR(shm_mnt)) {
2604                 error = PTR_ERR(shm_mnt);
2605                 printk(KERN_ERR "Could not kern_mount tmpfs\n");
2606                 goto out1;
2607         }
2608         return 0;
2609
2610 out1:
2611         unregister_filesystem(&tmpfs_fs_type);
2612 out2:
2613         destroy_inodecache();
2614 out3:
2615         bdi_destroy(&shmem_backing_dev_info);
2616 out4:
2617         shm_mnt = ERR_PTR(error);
2618         return error;
2619 }
2620
2621 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2622 /**
2623  * mem_cgroup_get_shmem_target - find a page or entry assigned to the shmem file
2624  * @inode: the inode to be searched
2625  * @pgoff: the offset to be searched
2626  * @pagep: the pointer for the found page to be stored
2627  * @ent: the pointer for the found swap entry to be stored
2628  *
2629  * If a page is found, refcount of it is incremented. Callers should handle
2630  * these refcount.
2631  */
2632 void mem_cgroup_get_shmem_target(struct inode *inode, pgoff_t pgoff,
2633                                         struct page **pagep, swp_entry_t *ent)
2634 {
2635         swp_entry_t entry = { .val = 0 }, *ptr;
2636         struct page *page = NULL;
2637         struct shmem_inode_info *info = SHMEM_I(inode);
2638
2639         if ((pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
2640                 goto out;
2641
2642         spin_lock(&info->lock);
2643         ptr = shmem_swp_entry(info, pgoff, NULL);
2644 #ifdef CONFIG_SWAP
2645         if (ptr && ptr->val) {
2646                 entry.val = ptr->val;
2647                 page = find_get_page(&swapper_space, entry.val);
2648         } else
2649 #endif
2650                 page = find_get_page(inode->i_mapping, pgoff);
2651         if (ptr)
2652                 shmem_swp_unmap(ptr);
2653         spin_unlock(&info->lock);
2654 out:
2655         *pagep = page;
2656         *ent = entry;
2657 }
2658 #endif
2659
2660 #else /* !CONFIG_SHMEM */
2661
2662 /*
2663  * tiny-shmem: simple shmemfs and tmpfs using ramfs code
2664  *
2665  * This is intended for small system where the benefits of the full
2666  * shmem code (swap-backed and resource-limited) are outweighed by
2667  * their complexity. On systems without swap this code should be
2668  * effectively equivalent, but much lighter weight.
2669  */
2670
2671 #include <linux/ramfs.h>
2672
2673 static struct file_system_type tmpfs_fs_type = {
2674         .name           = "tmpfs",
2675         .mount          = ramfs_mount,
2676         .kill_sb        = kill_litter_super,
2677 };
2678
2679 int __init init_tmpfs(void)
2680 {
2681         BUG_ON(register_filesystem(&tmpfs_fs_type) != 0);
2682
2683         shm_mnt = kern_mount(&tmpfs_fs_type);
2684         BUG_ON(IS_ERR(shm_mnt));
2685
2686         return 0;
2687 }
2688
2689 int shmem_unuse(swp_entry_t entry, struct page *page)
2690 {
2691         return 0;
2692 }
2693
2694 int shmem_lock(struct file *file, int lock, struct user_struct *user)
2695 {
2696         return 0;
2697 }
2698
2699 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2700 /**
2701  * mem_cgroup_get_shmem_target - find a page or entry assigned to the shmem file
2702  * @inode: the inode to be searched
2703  * @pgoff: the offset to be searched
2704  * @pagep: the pointer for the found page to be stored
2705  * @ent: the pointer for the found swap entry to be stored
2706  *
2707  * If a page is found, refcount of it is incremented. Callers should handle
2708  * these refcount.
2709  */
2710 void mem_cgroup_get_shmem_target(struct inode *inode, pgoff_t pgoff,
2711                                         struct page **pagep, swp_entry_t *ent)
2712 {
2713         struct page *page = NULL;
2714
2715         if ((pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
2716                 goto out;
2717         page = find_get_page(inode->i_mapping, pgoff);
2718 out:
2719         *pagep = page;
2720         *ent = (swp_entry_t){ .val = 0 };
2721 }
2722 #endif
2723
2724 #define shmem_vm_ops                            generic_file_vm_ops
2725 #define shmem_file_operations                   ramfs_file_operations
2726 #define shmem_get_inode(sb, dir, mode, dev, flags)      ramfs_get_inode(sb, dir, mode, dev)
2727 #define shmem_acct_size(flags, size)            0
2728 #define shmem_unacct_size(flags, size)          do {} while (0)
2729 #define SHMEM_MAX_BYTES                         MAX_LFS_FILESIZE
2730
2731 #endif /* CONFIG_SHMEM */
2732
2733 /* common code */
2734
2735 /**
2736  * shmem_file_setup - get an unlinked file living in tmpfs
2737  * @name: name for dentry (to be seen in /proc/<pid>/maps
2738  * @size: size to be set for the file
2739  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
2740  */
2741 struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
2742 {
2743         int error;
2744         struct file *file;
2745         struct inode *inode;
2746         struct path path;
2747         struct dentry *root;
2748         struct qstr this;
2749
2750         if (IS_ERR(shm_mnt))
2751                 return (void *)shm_mnt;
2752
2753         if (size < 0 || size > SHMEM_MAX_BYTES)
2754                 return ERR_PTR(-EINVAL);
2755
2756         if (shmem_acct_size(flags, size))
2757                 return ERR_PTR(-ENOMEM);
2758
2759         error = -ENOMEM;
2760         this.name = name;
2761         this.len = strlen(name);
2762         this.hash = 0; /* will go */
2763         root = shm_mnt->mnt_root;
2764         path.dentry = d_alloc(root, &this);
2765         if (!path.dentry)
2766                 goto put_memory;
2767         path.mnt = mntget(shm_mnt);
2768
2769         error = -ENOSPC;
2770         inode = shmem_get_inode(root->d_sb, NULL, S_IFREG | S_IRWXUGO, 0, flags);
2771         if (!inode)
2772                 goto put_dentry;
2773
2774         d_instantiate(path.dentry, inode);
2775         inode->i_size = size;
2776         inode->i_nlink = 0;     /* It is unlinked */
2777 #ifndef CONFIG_MMU
2778         error = ramfs_nommu_expand_for_mapping(inode, size);
2779         if (error)
2780                 goto put_dentry;
2781 #endif
2782
2783         error = -ENFILE;
2784         file = alloc_file(&path, FMODE_WRITE | FMODE_READ,
2785                   &shmem_file_operations);
2786         if (!file)
2787                 goto put_dentry;
2788
2789         return file;
2790
2791 put_dentry:
2792         path_put(&path);
2793 put_memory:
2794         shmem_unacct_size(flags, size);
2795         return ERR_PTR(error);
2796 }
2797 EXPORT_SYMBOL_GPL(shmem_file_setup);
2798
2799 /**
2800  * shmem_zero_setup - setup a shared anonymous mapping
2801  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
2802  */
2803 int shmem_zero_setup(struct vm_area_struct *vma)
2804 {
2805         struct file *file;
2806         loff_t size = vma->vm_end - vma->vm_start;
2807
2808         file = shmem_file_setup("dev/zero", size, vma->vm_flags);
2809         if (IS_ERR(file))
2810                 return PTR_ERR(file);
2811
2812         if (vma->vm_file)
2813                 fput(vma->vm_file);
2814         vma->vm_file = file;
2815         vma->vm_ops = &shmem_vm_ops;
2816         vma->vm_flags |= VM_CAN_NONLINEAR;
2817         return 0;
2818 }