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/splice.h>
  55 #include <linux/security.h>
  56 #include <linux/swapops.h>
  57 #include <linux/mempolicy.h>
  58 #include <linux/namei.h>
  59 #include <linux/ctype.h>
  60 #include <linux/migrate.h>
  61 #include <linux/highmem.h>
  62 #include <linux/seq_file.h>
  63 #include <linux/magic.h>
  64
  65 #include <asm/uaccess.h>
  66 #include <asm/div64.h>
  67 #include <asm/pgtable.h>
  68
  69 /*
  70  * The maximum size of a shmem/tmpfs file is limited by the maximum size of
  71  * its triple-indirect swap vector - see illustration at shmem_swp_entry().
  72  *
  73  * With 4kB page size, maximum file size is just over 2TB on a 32-bit kernel,
  74  * but one eighth of that on a 64-bit kernel.  With 8kB page size, maximum
  75  * file size is just over 4TB on a 64-bit kernel, but 16TB on a 32-bit kernel,
  76  * MAX_LFS_FILESIZE being then more restrictive than swap vector layout.
  77  *
  78  * We use / and * instead of shifts in the definitions below, so that the swap
  79  * vector can be tested with small even values (e.g. 20) for ENTRIES_PER_PAGE.
  80  */
  81 #define ENTRIES_PER_PAGE (PAGE_CACHE_SIZE/sizeof(unsigned long))
  82 #define ENTRIES_PER_PAGEPAGE ((unsigned long long)ENTRIES_PER_PAGE*ENTRIES_PER_PAGE)
  83
  84 #define SHMSWP_MAX_INDEX (SHMEM_NR_DIRECT + (ENTRIES_PER_PAGEPAGE/2) * (ENTRIES_PER_PAGE+1))
  85 #define SHMSWP_MAX_BYTES (SHMSWP_MAX_INDEX << PAGE_CACHE_SHIFT)
  86
  87 #define SHMEM_MAX_BYTES  min_t(unsigned long long, SHMSWP_MAX_BYTES, MAX_LFS_FILESIZE)
  88 #define SHMEM_MAX_INDEX  ((unsigned long)((SHMEM_MAX_BYTES+1) >> PAGE_CACHE_SHIFT))
  89
  90 #define BLOCKS_PER_PAGE  (PAGE_CACHE_SIZE/512)
  91 #define VM_ACCT(size)    (PAGE_CACHE_ALIGN(size) >> PAGE_SHIFT)
  92
  93 /* info->flags needs VM_flags to handle pagein/truncate races efficiently */
  94 #define SHMEM_PAGEIN     VM_READ
  95 #define SHMEM_TRUNCATE   VM_WRITE
  96
  97 /* Definition to limit shmem_truncate's steps between cond_rescheds */
  98 #define LATENCY_LIMIT    64
  99
 100 /* Pretend that each entry is of this size in directory's i_size */
 101 #define BOGO_DIRENT_SIZE 20
 102
 103 struct shmem_xattr {
 104         struct list_head list;  /* anchored by shmem_inode_info->xattr_list */
 105         char *name;             /* xattr name */
 106         size_t size;
 107         char value[0];
 108 };
 109
 110 /* Flag allocation requirements to shmem_getpage and shmem_swp_alloc */
 111 enum sgp_type {
 112         SGP_READ,       /* don't exceed i_size, don't allocate page */
 113         SGP_CACHE,      /* don't exceed i_size, may allocate page */
 114         SGP_DIRTY,      /* like SGP_CACHE, but set new page dirty */
 115         SGP_WRITE,      /* may exceed i_size, may allocate page */
 116 };
 117
 118 #ifdef CONFIG_TMPFS
 119 static unsigned long shmem_default_max_blocks(void)
 120 {
 121         return totalram_pages / 2;
 122 }
 123
 124 static unsigned long shmem_default_max_inodes(void)
 125 {
 126         return min(totalram_pages - totalhigh_pages, totalram_pages / 2);
 127 }
 128 #endif
 129
 130 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
 131         struct page **pagep, enum sgp_type sgp, gfp_t gfp, int *fault_type);
 132
 133 static inline int shmem_getpage(struct inode *inode, pgoff_t index,
 134         struct page **pagep, enum sgp_type sgp, int *fault_type)
 135 {
 136         return shmem_getpage_gfp(inode, index, pagep, sgp,
 137                         mapping_gfp_mask(inode->i_mapping), fault_type);
 138 }
 139
 140 static inline struct page *shmem_dir_alloc(gfp_t gfp_mask)
 141 {
 142         /*
 143          * The above definition of ENTRIES_PER_PAGE, and the use of
 144          * BLOCKS_PER_PAGE on indirect pages, assume PAGE_CACHE_SIZE:
 145          * might be reconsidered if it ever diverges from PAGE_SIZE.
 146          *
 147          * Mobility flags are masked out as swap vectors cannot move
 148          */
 149         return alloc_pages((gfp_mask & ~GFP_MOVABLE_MASK) | __GFP_ZERO,
 150                                 PAGE_CACHE_SHIFT-PAGE_SHIFT);
 151 }
 152
 153 static inline void shmem_dir_free(struct page *page)
 154 {
 155         __free_pages(page, PAGE_CACHE_SHIFT-PAGE_SHIFT);
 156 }
 157
 158 static struct page **shmem_dir_map(struct page *page)
 159 {
 160         return (struct page **)kmap_atomic(page, KM_USER0);
 161 }
 162
 163 static inline void shmem_dir_unmap(struct page **dir)
 164 {
 165         kunmap_atomic(dir, KM_USER0);
 166 }
 167
 168 static swp_entry_t *shmem_swp_map(struct page *page)
 169 {
 170         return (swp_entry_t *)kmap_atomic(page, KM_USER1);
 171 }
 172
 173 static inline void shmem_swp_balance_unmap(void)
 174 {
 175         /*
 176          * When passing a pointer to an i_direct entry, to code which
 177          * also handles indirect entries and so will shmem_swp_unmap,
 178          * we must arrange for the preempt count to remain in balance.
 179          * What kmap_atomic of a lowmem page does depends on config
 180          * and architecture, so pretend to kmap_atomic some lowmem page.
 181          */
 182         (void) kmap_atomic(ZERO_PAGE(0), KM_USER1);
 183 }
 184
 185 static inline void shmem_swp_unmap(swp_entry_t *entry)
 186 {
 187         kunmap_atomic(entry, KM_USER1);
 188 }
 189
 190 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
 191 {
 192         return sb->s_fs_info;
 193 }
 194
 195 /*
 196  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
 197  * for shared memory and for shared anonymous (/dev/zero) mappings
 198  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
 199  * consistent with the pre-accounting of private mappings ...
 200  */
 201 static inline int shmem_acct_size(unsigned long flags, loff_t size)
 202 {
 203         return (flags & VM_NORESERVE) ?
 204                 0 : security_vm_enough_memory_kern(VM_ACCT(size));
 205 }
 206
 207 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
 208 {
 209         if (!(flags & VM_NORESERVE))
 210                 vm_unacct_memory(VM_ACCT(size));
 211 }
 212
 213 /*
 214  * ... whereas tmpfs objects are accounted incrementally as
 215  * pages are allocated, in order to allow huge sparse files.
 216  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
 217  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
 218  */
 219 static inline int shmem_acct_block(unsigned long flags)
 220 {
 221         return (flags & VM_NORESERVE) ?
 222                 security_vm_enough_memory_kern(VM_ACCT(PAGE_CACHE_SIZE)) : 0;
 223 }
 224
 225 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
 226 {
 227         if (flags & VM_NORESERVE)
 228                 vm_unacct_memory(pages * VM_ACCT(PAGE_CACHE_SIZE));
 229 }
 230
 231 static const struct super_operations shmem_ops;
 232 static const struct address_space_operations shmem_aops;
 233 static const struct file_operations shmem_file_operations;
 234 static const struct inode_operations shmem_inode_operations;
 235 static const struct inode_operations shmem_dir_inode_operations;
 236 static const struct inode_operations shmem_special_inode_operations;
 237 static const struct vm_operations_struct shmem_vm_ops;
 238
 239 static struct backing_dev_info shmem_backing_dev_info  __read_mostly = {
 240         .ra_pages       = 0,    /* No readahead */
 241         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_SWAP_BACKED,
 242 };
 243
 244 static LIST_HEAD(shmem_swaplist);
 245 static DEFINE_MUTEX(shmem_swaplist_mutex);
 246
 247 static void shmem_free_blocks(struct inode *inode, long pages)
 248 {
 249         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 250         if (sbinfo->max_blocks) {
 251                 percpu_counter_add(&sbinfo->used_blocks, -pages);
 252                 inode->i_blocks -= pages*BLOCKS_PER_PAGE;
 253         }
 254 }
 255
 256 static int shmem_reserve_inode(struct super_block *sb)
 257 {
 258         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 259         if (sbinfo->max_inodes) {
 260                 spin_lock(&sbinfo->stat_lock);
 261                 if (!sbinfo->free_inodes) {
 262                         spin_unlock(&sbinfo->stat_lock);
 263                         return -ENOSPC;
 264                 }
 265                 sbinfo->free_inodes--;
 266                 spin_unlock(&sbinfo->stat_lock);
 267         }
 268         return 0;
 269 }
 270
 271 static void shmem_free_inode(struct super_block *sb)
 272 {
 273         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
 274         if (sbinfo->max_inodes) {
 275                 spin_lock(&sbinfo->stat_lock);
 276                 sbinfo->free_inodes++;
 277                 spin_unlock(&sbinfo->stat_lock);
 278         }
 279 }
 280
 281 /**
 282  * shmem_recalc_inode - recalculate the size of an inode
 283  * @inode: inode to recalc
 284  *
 285  * We have to calculate the free blocks since the mm can drop
 286  * undirtied hole pages behind our back.
 287  *
 288  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
 289  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
 290  *
 291  * It has to be called with the spinlock held.
 292  */
 293 static void shmem_recalc_inode(struct inode *inode)
 294 {
 295         struct shmem_inode_info *info = SHMEM_I(inode);
 296         long freed;
 297
 298         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
 299         if (freed > 0) {
 300                 info->alloced -= freed;
 301                 shmem_unacct_blocks(info->flags, freed);
 302                 shmem_free_blocks(inode, freed);
 303         }
 304 }
 305
 306 /**
 307  * shmem_swp_entry - find the swap vector position in the info structure
 308  * @info:  info structure for the inode
 309  * @index: index of the page to find
 310  * @page:  optional page to add to the structure. Has to be preset to
 311  *         all zeros
 312  *
 313  * If there is no space allocated yet it will return NULL when
 314  * page is NULL, else it will use the page for the needed block,
 315  * setting it to NULL on return to indicate that it has been used.
 316  *
 317  * The swap vector is organized the following way:
 318  *
 319  * There are SHMEM_NR_DIRECT entries directly stored in the
 320  * shmem_inode_info structure. So small files do not need an addional
 321  * allocation.
 322  *
 323  * For pages with index > SHMEM_NR_DIRECT there is the pointer
 324  * i_indirect which points to a page which holds in the first half
 325  * doubly indirect blocks, in the second half triple indirect blocks:
 326  *
 327  * For an artificial ENTRIES_PER_PAGE = 4 this would lead to the
 328  * following layout (for SHMEM_NR_DIRECT == 16):
 329  *
 330  * i_indirect -> dir --> 16-19
 331  *            |      +-> 20-23
 332  *            |
 333  *            +-->dir2 --> 24-27
 334  *            |        +-> 28-31
 335  *            |        +-> 32-35
 336  *            |        +-> 36-39
 337  *            |
 338  *            +-->dir3 --> 40-43
 339  *                     +-> 44-47
 340  *                     +-> 48-51
 341  *                     +-> 52-55
 342  */
 343 static swp_entry_t *shmem_swp_entry(struct shmem_inode_info *info, unsigned long index, struct page **page)
 344 {
 345         unsigned long offset;
 346         struct page **dir;
 347         struct page *subdir;
 348
 349         if (index < SHMEM_NR_DIRECT) {
 350                 shmem_swp_balance_unmap();
 351                 return info->i_direct+index;
 352         }
 353         if (!info->i_indirect) {
 354                 if (page) {
 355                         info->i_indirect = *page;
 356                         *page = NULL;
 357                 }
 358                 return NULL;                    /* need another page */
 359         }
 360
 361         index -= SHMEM_NR_DIRECT;
 362         offset = index % ENTRIES_PER_PAGE;
 363         index /= ENTRIES_PER_PAGE;
 364         dir = shmem_dir_map(info->i_indirect);
 365
 366         if (index >= ENTRIES_PER_PAGE/2) {
 367                 index -= ENTRIES_PER_PAGE/2;
 368                 dir += ENTRIES_PER_PAGE/2 + index/ENTRIES_PER_PAGE;
 369                 index %= ENTRIES_PER_PAGE;
 370                 subdir = *dir;
 371                 if (!subdir) {
 372                         if (page) {
 373                                 *dir = *page;
 374                                 *page = NULL;
 375                         }
 376                         shmem_dir_unmap(dir);
 377                         return NULL;            /* need another page */
 378                 }
 379                 shmem_dir_unmap(dir);
 380                 dir = shmem_dir_map(subdir);
 381         }
 382
 383         dir += index;
 384         subdir = *dir;
 385         if (!subdir) {
 386                 if (!page || !(subdir = *page)) {
 387                         shmem_dir_unmap(dir);
 388                         return NULL;            /* need a page */
 389                 }
 390                 *dir = subdir;
 391                 *page = NULL;
 392         }
 393         shmem_dir_unmap(dir);
 394         return shmem_swp_map(subdir) + offset;
 395 }
 396
 397 static void shmem_swp_set(struct shmem_inode_info *info, swp_entry_t *entry, unsigned long value)
 398 {
 399         long incdec = value? 1: -1;
 400
 401         entry->val = value;
 402         info->swapped += incdec;
 403         if ((unsigned long)(entry - info->i_direct) >= SHMEM_NR_DIRECT) {
 404                 struct page *page = kmap_atomic_to_page(entry);
 405                 set_page_private(page, page_private(page) + incdec);
 406         }
 407 }
 408
 409 /**
 410  * shmem_swp_alloc - get the position of the swap entry for the page.
 411  * @info:       info structure for the inode
 412  * @index:      index of the page to find
 413  * @sgp:        check and recheck i_size? skip allocation?
 414  * @gfp:        gfp mask to use for any page allocation
 415  *
 416  * If the entry does not exist, allocate it.
 417  */
 418 static swp_entry_t *shmem_swp_alloc(struct shmem_inode_info *info,
 419                         unsigned long index, enum sgp_type sgp, gfp_t gfp)
 420 {
 421         struct inode *inode = &info->vfs_inode;
 422         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 423         struct page *page = NULL;
 424         swp_entry_t *entry;
 425
 426         if (sgp != SGP_WRITE &&
 427             ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode))
 428                 return ERR_PTR(-EINVAL);
 429
 430         while (!(entry = shmem_swp_entry(info, index, &page))) {
 431                 if (sgp == SGP_READ)
 432                         return shmem_swp_map(ZERO_PAGE(0));
 433                 /*
 434                  * Test used_blocks against 1 less max_blocks, since we have 1 data
 435                  * page (and perhaps indirect index pages) yet to allocate:
 436                  * a waste to allocate index if we cannot allocate data.
 437                  */
 438                 if (sbinfo->max_blocks) {
 439                         if (percpu_counter_compare(&sbinfo->used_blocks,
 440                                                 sbinfo->max_blocks - 1) >= 0)
 441                                 return ERR_PTR(-ENOSPC);
 442                         percpu_counter_inc(&sbinfo->used_blocks);
 443                         inode->i_blocks += BLOCKS_PER_PAGE;
 444                 }
 445
 446                 spin_unlock(&info->lock);
 447                 page = shmem_dir_alloc(gfp);
 448                 spin_lock(&info->lock);
 449
 450                 if (!page) {
 451                         shmem_free_blocks(inode, 1);
 452                         return ERR_PTR(-ENOMEM);
 453                 }
 454                 if (sgp != SGP_WRITE &&
 455                     ((loff_t) index << PAGE_CACHE_SHIFT) >= i_size_read(inode)) {
 456                         entry = ERR_PTR(-EINVAL);
 457                         break;
 458                 }
 459                 if (info->next_index <= index)
 460                         info->next_index = index + 1;
 461         }
 462         if (page) {
 463                 /* another task gave its page, or truncated the file */
 464                 shmem_free_blocks(inode, 1);
 465                 shmem_dir_free(page);
 466         }
 467         if (info->next_index <= index && !IS_ERR(entry))
 468                 info->next_index = index + 1;
 469         return entry;
 470 }
 471
 472 /**
 473  * shmem_free_swp - free some swap entries in a directory
 474  * @dir:        pointer to the directory
 475  * @edir:       pointer after last entry of the directory
 476  * @punch_lock: pointer to spinlock when needed for the holepunch case
 477  */
 478 static int shmem_free_swp(swp_entry_t *dir, swp_entry_t *edir,
 479                                                 spinlock_t *punch_lock)
 480 {
 481         spinlock_t *punch_unlock = NULL;
 482         swp_entry_t *ptr;
 483         int freed = 0;
 484
 485         for (ptr = dir; ptr < edir; ptr++) {
 486                 if (ptr->val) {
 487                         if (unlikely(punch_lock)) {
 488                                 punch_unlock = punch_lock;
 489                                 punch_lock = NULL;
 490                                 spin_lock(punch_unlock);
 491                                 if (!ptr->val)
 492                                         continue;
 493                         }
 494                         free_swap_and_cache(*ptr);
 495                         *ptr = (swp_entry_t){0};
 496                         freed++;
 497                 }
 498         }
 499         if (punch_unlock)
 500                 spin_unlock(punch_unlock);
 501         return freed;
 502 }
 503
 504 static int shmem_map_and_free_swp(struct page *subdir, int offset,
 505                 int limit, struct page ***dir, spinlock_t *punch_lock)
 506 {
 507         swp_entry_t *ptr;
 508         int freed = 0;
 509
 510         ptr = shmem_swp_map(subdir);
 511         for (; offset < limit; offset += LATENCY_LIMIT) {
 512                 int size = limit - offset;
 513                 if (size > LATENCY_LIMIT)
 514                         size = LATENCY_LIMIT;
 515                 freed += shmem_free_swp(ptr+offset, ptr+offset+size,
 516                                                         punch_lock);
 517                 if (need_resched()) {
 518                         shmem_swp_unmap(ptr);
 519                         if (*dir) {
 520                                 shmem_dir_unmap(*dir);
 521                                 *dir = NULL;
 522                         }
 523                         cond_resched();
 524                         ptr = shmem_swp_map(subdir);
 525                 }
 526         }
 527         shmem_swp_unmap(ptr);
 528         return freed;
 529 }
 530
 531 static void shmem_free_pages(struct list_head *next)
 532 {
 533         struct page *page;
 534         int freed = 0;
 535
 536         do {
 537                 page = container_of(next, struct page, lru);
 538                 next = next->next;
 539                 shmem_dir_free(page);
 540                 freed++;
 541                 if (freed >= LATENCY_LIMIT) {
 542                         cond_resched();
 543                         freed = 0;
 544                 }
 545         } while (next);
 546 }
 547
 548 void shmem_truncate_range(struct inode *inode, loff_t start, loff_t end)
 549 {
 550         struct shmem_inode_info *info = SHMEM_I(inode);
 551         unsigned long idx;
 552         unsigned long size;
 553         unsigned long limit;
 554         unsigned long stage;
 555         unsigned long diroff;
 556         struct page **dir;
 557         struct page *topdir;
 558         struct page *middir;
 559         struct page *subdir;
 560         swp_entry_t *ptr;
 561         LIST_HEAD(pages_to_free);
 562         long nr_pages_to_free = 0;
 563         long nr_swaps_freed = 0;
 564         int offset;
 565         int freed;
 566         int punch_hole;
 567         spinlock_t *needs_lock;
 568         spinlock_t *punch_lock;
 569         unsigned long upper_limit;
 570
 571         truncate_inode_pages_range(inode->i_mapping, start, end);
 572
 573         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 574         idx = (start + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 575         if (idx >= info->next_index)
 576                 return;
 577
 578         spin_lock(&info->lock);
 579         info->flags |= SHMEM_TRUNCATE;
 580         if (likely(end == (loff_t) -1)) {
 581                 limit = info->next_index;
 582                 upper_limit = SHMEM_MAX_INDEX;
 583                 info->next_index = idx;
 584                 needs_lock = NULL;
 585                 punch_hole = 0;
 586         } else {
 587                 if (end + 1 >= inode->i_size) { /* we may free a little more */
 588                         limit = (inode->i_size + PAGE_CACHE_SIZE - 1) >>
 589                                                         PAGE_CACHE_SHIFT;
 590                         upper_limit = SHMEM_MAX_INDEX;
 591                 } else {
 592                         limit = (end + 1) >> PAGE_CACHE_SHIFT;
 593                         upper_limit = limit;
 594                 }
 595                 needs_lock = &info->lock;
 596                 punch_hole = 1;
 597         }
 598
 599         topdir = info->i_indirect;
 600         if (topdir && idx <= SHMEM_NR_DIRECT && !punch_hole) {
 601                 info->i_indirect = NULL;
 602                 nr_pages_to_free++;
 603                 list_add(&topdir->lru, &pages_to_free);
 604         }
 605         spin_unlock(&info->lock);
 606
 607         if (info->swapped && idx < SHMEM_NR_DIRECT) {
 608                 ptr = info->i_direct;
 609                 size = limit;
 610                 if (size > SHMEM_NR_DIRECT)
 611                         size = SHMEM_NR_DIRECT;
 612                 nr_swaps_freed = shmem_free_swp(ptr+idx, ptr+size, needs_lock);
 613         }
 614
 615         /*
 616          * If there are no indirect blocks or we are punching a hole
 617          * below indirect blocks, nothing to be done.
 618          */
 619         if (!topdir || limit <= SHMEM_NR_DIRECT)
 620                 goto done2;
 621
 622         /*
 623          * The truncation case has already dropped info->lock, and we're safe
 624          * because i_size and next_index have already been lowered, preventing
 625          * access beyond.  But in the punch_hole case, we still need to take
 626          * the lock when updating the swap directory, because there might be
 627          * racing accesses by shmem_getpage(SGP_CACHE), shmem_unuse_inode or
 628          * shmem_writepage.  However, whenever we find we can remove a whole
 629          * directory page (not at the misaligned start or end of the range),
 630          * we first NULLify its pointer in the level above, and then have no
 631          * need to take the lock when updating its contents: needs_lock and
 632          * punch_lock (either pointing to info->lock or NULL) manage this.
 633          */
 634
 635         upper_limit -= SHMEM_NR_DIRECT;
 636         limit -= SHMEM_NR_DIRECT;
 637         idx = (idx > SHMEM_NR_DIRECT)? (idx - SHMEM_NR_DIRECT): 0;
 638         offset = idx % ENTRIES_PER_PAGE;
 639         idx -= offset;
 640
 641         dir = shmem_dir_map(topdir);
 642         stage = ENTRIES_PER_PAGEPAGE/2;
 643         if (idx < ENTRIES_PER_PAGEPAGE/2) {
 644                 middir = topdir;
 645                 diroff = idx/ENTRIES_PER_PAGE;
 646         } else {
 647                 dir += ENTRIES_PER_PAGE/2;
 648                 dir += (idx - ENTRIES_PER_PAGEPAGE/2)/ENTRIES_PER_PAGEPAGE;
 649                 while (stage <= idx)
 650                         stage += ENTRIES_PER_PAGEPAGE;
 651                 middir = *dir;
 652                 if (*dir) {
 653                         diroff = ((idx - ENTRIES_PER_PAGEPAGE/2) %
 654                                 ENTRIES_PER_PAGEPAGE) / ENTRIES_PER_PAGE;
 655                         if (!diroff && !offset && upper_limit >= stage) {
 656                                 if (needs_lock) {
 657                                         spin_lock(needs_lock);
 658                                         *dir = NULL;
 659                                         spin_unlock(needs_lock);
 660                                         needs_lock = NULL;
 661                                 } else
 662                                         *dir = NULL;
 663                                 nr_pages_to_free++;
 664                                 list_add(&middir->lru, &pages_to_free);
 665                         }
 666                         shmem_dir_unmap(dir);
 667                         dir = shmem_dir_map(middir);
 668                 } else {
 669                         diroff = 0;
 670                         offset = 0;
 671                         idx = stage;
 672                 }
 673         }
 674
 675         for (; idx < limit; idx += ENTRIES_PER_PAGE, diroff++) {
 676                 if (unlikely(idx == stage)) {
 677                         shmem_dir_unmap(dir);
 678                         dir = shmem_dir_map(topdir) +
 679                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
 680                         while (!*dir) {
 681                                 dir++;
 682                                 idx += ENTRIES_PER_PAGEPAGE;
 683                                 if (idx >= limit)
 684                                         goto done1;
 685                         }
 686                         stage = idx + ENTRIES_PER_PAGEPAGE;
 687                         middir = *dir;
 688                         if (punch_hole)
 689                                 needs_lock = &info->lock;
 690                         if (upper_limit >= stage) {
 691                                 if (needs_lock) {
 692                                         spin_lock(needs_lock);
 693                                         *dir = NULL;
 694                                         spin_unlock(needs_lock);
 695                                         needs_lock = NULL;
 696                                 } else
 697                                         *dir = NULL;
 698                                 nr_pages_to_free++;
 699                                 list_add(&middir->lru, &pages_to_free);
 700                         }
 701                         shmem_dir_unmap(dir);
 702                         cond_resched();
 703                         dir = shmem_dir_map(middir);
 704                         diroff = 0;
 705                 }
 706                 punch_lock = needs_lock;
 707                 subdir = dir[diroff];
 708                 if (subdir && !offset && upper_limit-idx >= ENTRIES_PER_PAGE) {
 709                         if (needs_lock) {
 710                                 spin_lock(needs_lock);
 711                                 dir[diroff] = NULL;
 712                                 spin_unlock(needs_lock);
 713                                 punch_lock = NULL;
 714                         } else
 715                                 dir[diroff] = NULL;
 716                         nr_pages_to_free++;
 717                         list_add(&subdir->lru, &pages_to_free);
 718                 }
 719                 if (subdir && page_private(subdir) /* has swap entries */) {
 720                         size = limit - idx;
 721                         if (size > ENTRIES_PER_PAGE)
 722                                 size = ENTRIES_PER_PAGE;
 723                         freed = shmem_map_and_free_swp(subdir,
 724                                         offset, size, &dir, punch_lock);
 725                         if (!dir)
 726                                 dir = shmem_dir_map(middir);
 727                         nr_swaps_freed += freed;
 728                         if (offset || punch_lock) {
 729                                 spin_lock(&info->lock);
 730                                 set_page_private(subdir,
 731                                         page_private(subdir) - freed);
 732                                 spin_unlock(&info->lock);
 733                         } else
 734                                 BUG_ON(page_private(subdir) != freed);
 735                 }
 736                 offset = 0;
 737         }
 738 done1:
 739         shmem_dir_unmap(dir);
 740 done2:
 741         if (inode->i_mapping->nrpages && (info->flags & SHMEM_PAGEIN)) {
 742                 /*
 743                  * Call truncate_inode_pages again: racing shmem_unuse_inode
 744                  * may have swizzled a page in from swap since
 745                  * truncate_pagecache or generic_delete_inode did it, before we
 746                  * lowered next_index.  Also, though shmem_getpage checks
 747                  * i_size before adding to cache, no recheck after: so fix the
 748                  * narrow window there too.
 749                  */
 750                 truncate_inode_pages_range(inode->i_mapping, start, end);
 751         }
 752
 753         spin_lock(&info->lock);
 754         info->flags &= ~SHMEM_TRUNCATE;
 755         info->swapped -= nr_swaps_freed;
 756         if (nr_pages_to_free)
 757                 shmem_free_blocks(inode, nr_pages_to_free);
 758         shmem_recalc_inode(inode);
 759         spin_unlock(&info->lock);
 760
 761         /*
 762          * Empty swap vector directory pages to be freed?
 763          */
 764         if (!list_empty(&pages_to_free)) {
 765                 pages_to_free.prev->next = NULL;
 766                 shmem_free_pages(pages_to_free.next);
 767         }
 768 }
 769 EXPORT_SYMBOL_GPL(shmem_truncate_range);
 770
 771 static int shmem_setattr(struct dentry *dentry, struct iattr *attr)
 772 {
 773         struct inode *inode = dentry->d_inode;
 774         int error;
 775
 776         error = inode_change_ok(inode, attr);
 777         if (error)
 778                 return error;
 779
 780         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
 781                 loff_t oldsize = inode->i_size;
 782                 loff_t newsize = attr->ia_size;
 783                 struct page *page = NULL;
 784
 785                 if (newsize < oldsize) {
 786                         /*
 787                          * If truncating down to a partial page, then
 788                          * if that page is already allocated, hold it
 789                          * in memory until the truncation is over, so
 790                          * truncate_partial_page cannot miss it were
 791                          * it assigned to swap.
 792                          */
 793                         if (newsize & (PAGE_CACHE_SIZE-1)) {
 794                                 (void) shmem_getpage(inode,
 795                                         newsize >> PAGE_CACHE_SHIFT,
 796                                                 &page, SGP_READ, NULL);
 797                                 if (page)
 798                                         unlock_page(page);
 799                         }
 800                         /*
 801                          * Reset SHMEM_PAGEIN flag so that shmem_truncate can
 802                          * detect if any pages might have been added to cache
 803                          * after truncate_inode_pages.  But we needn't bother
 804                          * if it's being fully truncated to zero-length: the
 805                          * nrpages check is efficient enough in that case.
 806                          */
 807                         if (newsize) {
 808                                 struct shmem_inode_info *info = SHMEM_I(inode);
 809                                 spin_lock(&info->lock);
 810                                 info->flags &= ~SHMEM_PAGEIN;
 811                                 spin_unlock(&info->lock);
 812                         }
 813                 }
 814                 if (newsize != oldsize) {
 815                         i_size_write(inode, newsize);
 816                         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 817                 }
 818                 if (newsize < oldsize) {
 819                         loff_t holebegin = round_up(newsize, PAGE_SIZE);
 820                         unmap_mapping_range(inode->i_mapping, holebegin, 0, 1);
 821                         shmem_truncate_range(inode, newsize, (loff_t)-1);
 822                         /* unmap again to remove racily COWed private pages */
 823                         unmap_mapping_range(inode->i_mapping, holebegin, 0, 1);
 824                 }
 825                 if (page)
 826                         page_cache_release(page);
 827         }
 828
 829         setattr_copy(inode, attr);
 830 #ifdef CONFIG_TMPFS_POSIX_ACL
 831         if (attr->ia_valid & ATTR_MODE)
 832                 error = generic_acl_chmod(inode);
 833 #endif
 834         return error;
 835 }
 836
 837 static void shmem_evict_inode(struct inode *inode)
 838 {
 839         struct shmem_inode_info *info = SHMEM_I(inode);
 840         struct shmem_xattr *xattr, *nxattr;
 841
 842         if (inode->i_mapping->a_ops == &shmem_aops) {
 843                 shmem_unacct_size(info->flags, inode->i_size);
 844                 inode->i_size = 0;
 845                 shmem_truncate_range(inode, 0, (loff_t)-1);
 846                 if (!list_empty(&info->swaplist)) {
 847                         mutex_lock(&shmem_swaplist_mutex);
 848                         list_del_init(&info->swaplist);
 849                         mutex_unlock(&shmem_swaplist_mutex);
 850                 }
 851         }
 852
 853         list_for_each_entry_safe(xattr, nxattr, &info->xattr_list, list) {
 854                 kfree(xattr->name);
 855                 kfree(xattr);
 856         }
 857         BUG_ON(inode->i_blocks);
 858         shmem_free_inode(inode->i_sb);
 859         end_writeback(inode);
 860 }
 861
 862 static inline int shmem_find_swp(swp_entry_t entry, swp_entry_t *dir, swp_entry_t *edir)
 863 {
 864         swp_entry_t *ptr;
 865
 866         for (ptr = dir; ptr < edir; ptr++) {
 867                 if (ptr->val == entry.val)
 868                         return ptr - dir;
 869         }
 870         return -1;
 871 }
 872
 873 static int shmem_unuse_inode(struct shmem_inode_info *info, swp_entry_t entry, struct page *page)
 874 {
 875         struct address_space *mapping;
 876         unsigned long idx;
 877         unsigned long size;
 878         unsigned long limit;
 879         unsigned long stage;
 880         struct page **dir;
 881         struct page *subdir;
 882         swp_entry_t *ptr;
 883         int offset;
 884         int error;
 885
 886         idx = 0;
 887         ptr = info->i_direct;
 888         spin_lock(&info->lock);
 889         if (!info->swapped) {
 890                 list_del_init(&info->swaplist);
 891                 goto lost2;
 892         }
 893         limit = info->next_index;
 894         size = limit;
 895         if (size > SHMEM_NR_DIRECT)
 896                 size = SHMEM_NR_DIRECT;
 897         offset = shmem_find_swp(entry, ptr, ptr+size);
 898         if (offset >= 0) {
 899                 shmem_swp_balance_unmap();
 900                 goto found;
 901         }
 902         if (!info->i_indirect)
 903                 goto lost2;
 904
 905         dir = shmem_dir_map(info->i_indirect);
 906         stage = SHMEM_NR_DIRECT + ENTRIES_PER_PAGEPAGE/2;
 907
 908         for (idx = SHMEM_NR_DIRECT; idx < limit; idx += ENTRIES_PER_PAGE, dir++) {
 909                 if (unlikely(idx == stage)) {
 910                         shmem_dir_unmap(dir-1);
 911                         if (cond_resched_lock(&info->lock)) {
 912                                 /* check it has not been truncated */
 913                                 if (limit > info->next_index) {
 914                                         limit = info->next_index;
 915                                         if (idx >= limit)
 916                                                 goto lost2;
 917                                 }
 918                         }
 919                         dir = shmem_dir_map(info->i_indirect) +
 920                             ENTRIES_PER_PAGE/2 + idx/ENTRIES_PER_PAGEPAGE;
 921                         while (!*dir) {
 922                                 dir++;
 923                                 idx += ENTRIES_PER_PAGEPAGE;
 924                                 if (idx >= limit)
 925                                         goto lost1;
 926                         }
 927                         stage = idx + ENTRIES_PER_PAGEPAGE;
 928                         subdir = *dir;
 929                         shmem_dir_unmap(dir);
 930                         dir = shmem_dir_map(subdir);
 931                 }
 932                 subdir = *dir;
 933                 if (subdir && page_private(subdir)) {
 934                         ptr = shmem_swp_map(subdir);
 935                         size = limit - idx;
 936                         if (size > ENTRIES_PER_PAGE)
 937                                 size = ENTRIES_PER_PAGE;
 938                         offset = shmem_find_swp(entry, ptr, ptr+size);
 939                         shmem_swp_unmap(ptr);
 940                         if (offset >= 0) {
 941                                 shmem_dir_unmap(dir);
 942                                 ptr = shmem_swp_map(subdir);
 943                                 goto found;
 944                         }
 945                 }
 946         }
 947 lost1:
 948         shmem_dir_unmap(dir-1);
 949 lost2:
 950         spin_unlock(&info->lock);
 951         return 0;
 952 found:
 953         idx += offset;
 954         ptr += offset;
 955
 956         /*
 957          * Move _head_ to start search for next from here.
 958          * But be careful: shmem_evict_inode checks list_empty without taking
 959          * mutex, and there's an instant in list_move_tail when info->swaplist
 960          * would appear empty, if it were the only one on shmem_swaplist.  We
 961          * could avoid doing it if inode NULL; or use this minor optimization.
 962          */
 963         if (shmem_swaplist.next != &info->swaplist)
 964                 list_move_tail(&shmem_swaplist, &info->swaplist);
 965
 966         /*
 967          * We rely on shmem_swaplist_mutex, not only to protect the swaplist,
 968          * but also to hold up shmem_evict_inode(): so inode cannot be freed
 969          * beneath us (pagelock doesn't help until the page is in pagecache).
 970          */
 971         mapping = info->vfs_inode.i_mapping;
 972         error = add_to_page_cache_locked(page, mapping, idx, GFP_NOWAIT);
 973         /* which does mem_cgroup_uncharge_cache_page on error */
 974
 975         if (error != -ENOMEM) {
 976                 delete_from_swap_cache(page);
 977                 set_page_dirty(page);
 978                 info->flags |= SHMEM_PAGEIN;
 979                 shmem_swp_set(info, ptr, 0);
 980                 swap_free(entry);
 981                 error = 1;      /* not an error, but entry was found */
 982         }
 983         shmem_swp_unmap(ptr);
 984         spin_unlock(&info->lock);
 985         return error;
 986 }
 987
 988 /*
 989  * shmem_unuse() search for an eventually swapped out shmem page.
 990  */
 991 int shmem_unuse(swp_entry_t entry, struct page *page)
 992 {
 993         struct list_head *p, *next;
 994         struct shmem_inode_info *info;
 995         int found = 0;
 996         int error;
 997
 998         /*
 999          * Charge page using GFP_KERNEL while we can wait, before taking
1000          * the shmem_swaplist_mutex which might hold up shmem_writepage().
1001          * Charged back to the user (not to caller) when swap account is used.
1002          * add_to_page_cache() will be called with GFP_NOWAIT.
1003          */
1004         error = mem_cgroup_cache_charge(page, current->mm, GFP_KERNEL);
1005         if (error)
1006                 goto out;
1007         /*
1008          * Try to preload while we can wait, to not make a habit of
1009          * draining atomic reserves; but don't latch on to this cpu,
1010          * it's okay if sometimes we get rescheduled after this.
1011          */
1012         error = radix_tree_preload(GFP_KERNEL);
1013         if (error)
1014                 goto uncharge;
1015         radix_tree_preload_end();
1016
1017         mutex_lock(&shmem_swaplist_mutex);
1018         list_for_each_safe(p, next, &shmem_swaplist) {
1019                 info = list_entry(p, struct shmem_inode_info, swaplist);
1020                 found = shmem_unuse_inode(info, entry, page);
1021                 cond_resched();
1022                 if (found)
1023                         break;
1024         }
1025         mutex_unlock(&shmem_swaplist_mutex);
1026
1027 uncharge:
1028         if (!found)
1029                 mem_cgroup_uncharge_cache_page(page);
1030         if (found < 0)
1031                 error = found;
1032 out:
1033         unlock_page(page);
1034         page_cache_release(page);
1035         return error;
1036 }
1037
1038 /*
1039  * Move the page from the page cache to the swap cache.
1040  */
1041 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1042 {
1043         struct shmem_inode_info *info;
1044         swp_entry_t *entry, swap;
1045         struct address_space *mapping;
1046         unsigned long index;
1047         struct inode *inode;
1048
1049         BUG_ON(!PageLocked(page));
1050         mapping = page->mapping;
1051         index = page->index;
1052         inode = mapping->host;
1053         info = SHMEM_I(inode);
1054         if (info->flags & VM_LOCKED)
1055                 goto redirty;
1056         if (!total_swap_pages)
1057                 goto redirty;
1058
1059         /*
1060          * shmem_backing_dev_info's capabilities prevent regular writeback or
1061          * sync from ever calling shmem_writepage; but a stacking filesystem
1062          * might use ->writepage of its underlying filesystem, in which case
1063          * tmpfs should write out to swap only in response to memory pressure,
1064          * and not for the writeback threads or sync.
1065          */
1066         if (!wbc->for_reclaim) {
1067                 WARN_ON_ONCE(1);        /* Still happens? Tell us about it! */
1068                 goto redirty;
1069         }
1070         swap = get_swap_page();
1071         if (!swap.val)
1072                 goto redirty;
1073
1074         /*
1075          * Add inode to shmem_unuse()'s list of swapped-out inodes,
1076          * if it's not already there.  Do it now because we cannot take
1077          * mutex while holding spinlock, and must do so before the page
1078          * is moved to swap cache, when its pagelock no longer protects
1079          * the inode from eviction.  But don't unlock the mutex until
1080          * we've taken the spinlock, because shmem_unuse_inode() will
1081          * prune a !swapped inode from the swaplist under both locks.
1082          */
1083         mutex_lock(&shmem_swaplist_mutex);
1084         if (list_empty(&info->swaplist))
1085                 list_add_tail(&info->swaplist, &shmem_swaplist);
1086
1087         spin_lock(&info->lock);
1088         mutex_unlock(&shmem_swaplist_mutex);
1089
1090         if (index >= info->next_index) {
1091                 BUG_ON(!(info->flags & SHMEM_TRUNCATE));
1092                 goto unlock;
1093         }
1094         entry = shmem_swp_entry(info, index, NULL);
1095         if (entry->val) {
1096                 WARN_ON_ONCE(1);        /* Still happens? Tell us about it! */
1097                 free_swap_and_cache(*entry);
1098                 shmem_swp_set(info, entry, 0);
1099         }
1100         shmem_recalc_inode(inode);
1101
1102         if (add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
1103                 delete_from_page_cache(page);
1104                 shmem_swp_set(info, entry, swap.val);
1105                 shmem_swp_unmap(entry);
1106                 swap_shmem_alloc(swap);
1107                 spin_unlock(&info->lock);
1108                 BUG_ON(page_mapped(page));
1109                 swap_writepage(page, wbc);
1110                 return 0;
1111         }
1112
1113         shmem_swp_unmap(entry);
1114 unlock:
1115         spin_unlock(&info->lock);
1116         /*
1117          * add_to_swap_cache() doesn't return -EEXIST, so we can safely
1118          * clear SWAP_HAS_CACHE flag.
1119          */
1120         swapcache_free(swap, NULL);
1121 redirty:
1122         set_page_dirty(page);
1123         if (wbc->for_reclaim)
1124                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
1125         unlock_page(page);
1126         return 0;
1127 }
1128
1129 #ifdef CONFIG_NUMA
1130 #ifdef CONFIG_TMPFS
1131 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1132 {
1133         char buffer[64];
1134
1135         if (!mpol || mpol->mode == MPOL_DEFAULT)
1136                 return;         /* show nothing */
1137
1138         mpol_to_str(buffer, sizeof(buffer), mpol, 1);
1139
1140         seq_printf(seq, ",mpol=%s", buffer);
1141 }
1142
1143 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1144 {
1145         struct mempolicy *mpol = NULL;
1146         if (sbinfo->mpol) {
1147                 spin_lock(&sbinfo->stat_lock);  /* prevent replace/use races */
1148                 mpol = sbinfo->mpol;
1149                 mpol_get(mpol);
1150                 spin_unlock(&sbinfo->stat_lock);
1151         }
1152         return mpol;
1153 }
1154 #endif /* CONFIG_TMPFS */
1155
1156 static struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1157                         struct shmem_inode_info *info, unsigned long idx)
1158 {
1159         struct mempolicy mpol, *spol;
1160         struct vm_area_struct pvma;
1161         struct page *page;
1162
1163         spol = mpol_cond_copy(&mpol,
1164                                 mpol_shared_policy_lookup(&info->policy, idx));
1165
1166         /* Create a pseudo vma that just contains the policy */
1167         pvma.vm_start = 0;
1168         pvma.vm_pgoff = idx;
1169         pvma.vm_ops = NULL;
1170         pvma.vm_policy = spol;
1171         page = swapin_readahead(entry, gfp, &pvma, 0);
1172         return page;
1173 }
1174
1175 static struct page *shmem_alloc_page(gfp_t gfp,
1176                         struct shmem_inode_info *info, unsigned long idx)
1177 {
1178         struct vm_area_struct pvma;
1179
1180         /* Create a pseudo vma that just contains the policy */
1181         pvma.vm_start = 0;
1182         pvma.vm_pgoff = idx;
1183         pvma.vm_ops = NULL;
1184         pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, idx);
1185
1186         /*
1187          * alloc_page_vma() will drop the shared policy reference
1188          */
1189         return alloc_page_vma(gfp, &pvma, 0);
1190 }
1191 #else /* !CONFIG_NUMA */
1192 #ifdef CONFIG_TMPFS
1193 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *p)
1194 {
1195 }
1196 #endif /* CONFIG_TMPFS */
1197
1198 static inline struct page *shmem_swapin(swp_entry_t entry, gfp_t gfp,
1199                         struct shmem_inode_info *info, unsigned long idx)
1200 {
1201         return swapin_readahead(entry, gfp, NULL, 0);
1202 }
1203
1204 static inline struct page *shmem_alloc_page(gfp_t gfp,
1205                         struct shmem_inode_info *info, unsigned long idx)
1206 {
1207         return alloc_page(gfp);
1208 }
1209 #endif /* CONFIG_NUMA */
1210
1211 #if !defined(CONFIG_NUMA) || !defined(CONFIG_TMPFS)
1212 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1213 {
1214         return NULL;
1215 }
1216 #endif
1217
1218 /*
1219  * shmem_getpage_gfp - find page in cache, or get from swap, or allocate
1220  *
1221  * If we allocate a new one we do not mark it dirty. That's up to the
1222  * vm. If we swap it in we mark it dirty since we also free the swap
1223  * entry since a page cannot live in both the swap and page cache
1224  */
1225 static int shmem_getpage_gfp(struct inode *inode, pgoff_t idx,
1226         struct page **pagep, enum sgp_type sgp, gfp_t gfp, int *fault_type)
1227 {
1228         struct address_space *mapping = inode->i_mapping;
1229         struct shmem_inode_info *info = SHMEM_I(inode);
1230         struct shmem_sb_info *sbinfo;
1231         struct page *page;
1232         struct page *prealloc_page = NULL;
1233         swp_entry_t *entry;
1234         swp_entry_t swap;
1235         int error;
1236         int ret;
1237
1238         if (idx >= SHMEM_MAX_INDEX)
1239                 return -EFBIG;
1240 repeat:
1241         page = find_lock_page(mapping, idx);
1242         if (page) {
1243                 /*
1244                  * Once we can get the page lock, it must be uptodate:
1245                  * if there were an error in reading back from swap,
1246                  * the page would not be inserted into the filecache.
1247                  */
1248                 BUG_ON(!PageUptodate(page));
1249                 goto done;
1250         }
1251
1252         /*
1253          * Try to preload while we can wait, to not make a habit of
1254          * draining atomic reserves; but don't latch on to this cpu.
1255          */
1256         error = radix_tree_preload(gfp & GFP_RECLAIM_MASK);
1257         if (error)
1258                 goto out;
1259         radix_tree_preload_end();
1260
1261         if (sgp != SGP_READ && !prealloc_page) {
1262                 prealloc_page = shmem_alloc_page(gfp, info, idx);
1263                 if (prealloc_page) {
1264                         SetPageSwapBacked(prealloc_page);
1265                         if (mem_cgroup_cache_charge(prealloc_page,
1266                                         current->mm, GFP_KERNEL)) {
1267                                 page_cache_release(prealloc_page);
1268                                 prealloc_page = NULL;
1269                         }
1270                 }
1271         }
1272
1273         spin_lock(&info->lock);
1274         shmem_recalc_inode(inode);
1275         entry = shmem_swp_alloc(info, idx, sgp, gfp);
1276         if (IS_ERR(entry)) {
1277                 spin_unlock(&info->lock);
1278                 error = PTR_ERR(entry);
1279                 goto out;
1280         }
1281         swap = *entry;
1282
1283         if (swap.val) {
1284                 /* Look it up and read it in.. */
1285                 page = lookup_swap_cache(swap);
1286                 if (!page) {
1287                         shmem_swp_unmap(entry);
1288                         spin_unlock(&info->lock);
1289                         /* here we actually do the io */
1290                         if (fault_type)
1291                                 *fault_type |= VM_FAULT_MAJOR;
1292                         page = shmem_swapin(swap, gfp, info, idx);
1293                         if (!page) {
1294                                 spin_lock(&info->lock);
1295                                 entry = shmem_swp_alloc(info, idx, sgp, gfp);
1296                                 if (IS_ERR(entry))
1297                                         error = PTR_ERR(entry);
1298                                 else {
1299                                         if (entry->val == swap.val)
1300                                                 error = -ENOMEM;
1301                                         shmem_swp_unmap(entry);
1302                                 }
1303                                 spin_unlock(&info->lock);
1304                                 if (error)
1305                                         goto out;
1306                                 goto repeat;
1307                         }
1308                         wait_on_page_locked(page);
1309                         page_cache_release(page);
1310                         goto repeat;
1311                 }
1312
1313                 /* We have to do this with page locked to prevent races */
1314                 if (!trylock_page(page)) {
1315                         shmem_swp_unmap(entry);
1316                         spin_unlock(&info->lock);
1317                         wait_on_page_locked(page);
1318                         page_cache_release(page);
1319                         goto repeat;
1320                 }
1321                 if (PageWriteback(page)) {
1322                         shmem_swp_unmap(entry);
1323                         spin_unlock(&info->lock);
1324                         wait_on_page_writeback(page);
1325                         unlock_page(page);
1326                         page_cache_release(page);
1327                         goto repeat;
1328                 }
1329                 if (!PageUptodate(page)) {
1330                         shmem_swp_unmap(entry);
1331                         spin_unlock(&info->lock);
1332                         unlock_page(page);
1333                         page_cache_release(page);
1334                         error = -EIO;
1335                         goto out;
1336                 }
1337
1338                 error = add_to_page_cache_locked(page, mapping,
1339                                                  idx, GFP_NOWAIT);
1340                 if (error) {
1341                         shmem_swp_unmap(entry);
1342                         spin_unlock(&info->lock);
1343                         if (error == -ENOMEM) {
1344                                 /*
1345                                  * reclaim from proper memory cgroup and
1346                                  * call memcg's OOM if needed.
1347                                  */
1348                                 error = mem_cgroup_shmem_charge_fallback(
1349                                                 page, current->mm, gfp);
1350                                 if (error) {
1351                                         unlock_page(page);
1352                                         page_cache_release(page);
1353                                         goto out;
1354                                 }
1355                         }
1356                         unlock_page(page);
1357                         page_cache_release(page);
1358                         goto repeat;
1359                 }
1360
1361                 info->flags |= SHMEM_PAGEIN;
1362                 shmem_swp_set(info, entry, 0);
1363                 shmem_swp_unmap(entry);
1364                 delete_from_swap_cache(page);
1365                 spin_unlock(&info->lock);
1366                 set_page_dirty(page);
1367                 swap_free(swap);
1368
1369         } else if (sgp == SGP_READ) {
1370                 shmem_swp_unmap(entry);
1371                 page = find_get_page(mapping, idx);
1372                 if (page && !trylock_page(page)) {
1373                         spin_unlock(&info->lock);
1374                         wait_on_page_locked(page);
1375                         page_cache_release(page);
1376                         goto repeat;
1377                 }
1378                 spin_unlock(&info->lock);
1379
1380         } else if (prealloc_page) {
1381                 shmem_swp_unmap(entry);
1382                 sbinfo = SHMEM_SB(inode->i_sb);
1383                 if (sbinfo->max_blocks) {
1384                         if (percpu_counter_compare(&sbinfo->used_blocks,
1385                                                 sbinfo->max_blocks) >= 0 ||
1386                             shmem_acct_block(info->flags))
1387                                 goto nospace;
1388                         percpu_counter_inc(&sbinfo->used_blocks);
1389                         inode->i_blocks += BLOCKS_PER_PAGE;
1390                 } else if (shmem_acct_block(info->flags))
1391                         goto nospace;
1392
1393                 page = prealloc_page;
1394                 prealloc_page = NULL;
1395
1396                 entry = shmem_swp_alloc(info, idx, sgp, gfp);
1397                 if (IS_ERR(entry))
1398                         error = PTR_ERR(entry);
1399                 else {
1400                         swap = *entry;
1401                         shmem_swp_unmap(entry);
1402                 }
1403                 ret = error || swap.val;
1404                 if (ret)
1405                         mem_cgroup_uncharge_cache_page(page);
1406                 else
1407                         ret = add_to_page_cache_lru(page, mapping,
1408                                                 idx, GFP_NOWAIT);
1409                 /*
1410                  * At add_to_page_cache_lru() failure,
1411                  * uncharge will be done automatically.
1412                  */
1413                 if (ret) {
1414                         shmem_unacct_blocks(info->flags, 1);
1415                         shmem_free_blocks(inode, 1);
1416                         spin_unlock(&info->lock);
1417                         page_cache_release(page);
1418                         if (error)
1419                                 goto out;
1420                         goto repeat;
1421                 }
1422
1423                 info->flags |= SHMEM_PAGEIN;
1424                 info->alloced++;
1425                 spin_unlock(&info->lock);
1426                 clear_highpage(page);
1427                 flush_dcache_page(page);
1428                 SetPageUptodate(page);
1429                 if (sgp == SGP_DIRTY)
1430                         set_page_dirty(page);
1431
1432         } else {
1433                 spin_unlock(&info->lock);
1434                 error = -ENOMEM;
1435                 goto out;
1436         }
1437 done:
1438         *pagep = page;
1439         error = 0;
1440 out:
1441         if (prealloc_page) {
1442                 mem_cgroup_uncharge_cache_page(prealloc_page);
1443                 page_cache_release(prealloc_page);
1444         }
1445         return error;
1446
1447 nospace:
1448         /*
1449          * Perhaps the page was brought in from swap between find_lock_page
1450          * and taking info->lock?  We allow for that at add_to_page_cache_lru,
1451          * but must also avoid reporting a spurious ENOSPC while working on a
1452          * full tmpfs.
1453          */
1454         page = find_get_page(mapping, idx);
1455         spin_unlock(&info->lock);
1456         if (page) {
1457                 page_cache_release(page);
1458                 goto repeat;
1459         }
1460         error = -ENOSPC;
1461         goto out;
1462 }
1463
1464 static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1465 {
1466         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1467         int error;
1468         int ret = VM_FAULT_LOCKED;
1469
1470         if (((loff_t)vmf->pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
1471                 return VM_FAULT_SIGBUS;
1472
1473         error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_CACHE, &ret);
1474         if (error)
1475                 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1476
1477         if (ret & VM_FAULT_MAJOR) {
1478                 count_vm_event(PGMAJFAULT);
1479                 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1480         }
1481         return ret;
1482 }
1483
1484 #ifdef CONFIG_NUMA
1485 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *new)
1486 {
1487         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1488         return mpol_set_shared_policy(&SHMEM_I(i)->policy, vma, new);
1489 }
1490
1491 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
1492                                           unsigned long addr)
1493 {
1494         struct inode *i = vma->vm_file->f_path.dentry->d_inode;
1495         unsigned long idx;
1496
1497         idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
1498         return mpol_shared_policy_lookup(&SHMEM_I(i)->policy, idx);
1499 }
1500 #endif
1501
1502 int shmem_lock(struct file *file, int lock, struct user_struct *user)
1503 {
1504         struct inode *inode = file->f_path.dentry->d_inode;
1505         struct shmem_inode_info *info = SHMEM_I(inode);
1506         int retval = -ENOMEM;
1507
1508         spin_lock(&info->lock);
1509         if (lock && !(info->flags & VM_LOCKED)) {
1510                 if (!user_shm_lock(inode->i_size, user))
1511                         goto out_nomem;
1512                 info->flags |= VM_LOCKED;
1513                 mapping_set_unevictable(file->f_mapping);
1514         }
1515         if (!lock && (info->flags & VM_LOCKED) && user) {
1516                 user_shm_unlock(inode->i_size, user);
1517                 info->flags &= ~VM_LOCKED;
1518                 mapping_clear_unevictable(file->f_mapping);
1519                 scan_mapping_unevictable_pages(file->f_mapping);
1520         }
1521         retval = 0;
1522
1523 out_nomem:
1524         spin_unlock(&info->lock);
1525         return retval;
1526 }
1527
1528 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
1529 {
1530         file_accessed(file);
1531         vma->vm_ops = &shmem_vm_ops;
1532         vma->vm_flags |= VM_CAN_NONLINEAR;
1533         return 0;
1534 }
1535
1536 static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
1537                                      int mode, dev_t dev, unsigned long flags)
1538 {
1539         struct inode *inode;
1540         struct shmem_inode_info *info;
1541         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
1542
1543         if (shmem_reserve_inode(sb))
1544                 return NULL;
1545
1546         inode = new_inode(sb);
1547         if (inode) {
1548                 inode->i_ino = get_next_ino();
1549                 inode_init_owner(inode, dir, mode);
1550                 inode->i_blocks = 0;
1551                 inode->i_mapping->backing_dev_info = &shmem_backing_dev_info;
1552                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1553                 inode->i_generation = get_seconds();
1554                 info = SHMEM_I(inode);
1555                 memset(info, 0, (char *)inode - (char *)info);
1556                 spin_lock_init(&info->lock);
1557                 info->flags = flags & VM_NORESERVE;
1558                 INIT_LIST_HEAD(&info->swaplist);
1559                 INIT_LIST_HEAD(&info->xattr_list);
1560                 cache_no_acl(inode);
1561
1562                 switch (mode & S_IFMT) {
1563                 default:
1564                         inode->i_op = &shmem_special_inode_operations;
1565                         init_special_inode(inode, mode, dev);
1566                         break;
1567                 case S_IFREG:
1568                         inode->i_mapping->a_ops = &shmem_aops;
1569                         inode->i_op = &shmem_inode_operations;
1570                         inode->i_fop = &shmem_file_operations;
1571                         mpol_shared_policy_init(&info->policy,
1572                                                  shmem_get_sbmpol(sbinfo));
1573                         break;
1574                 case S_IFDIR:
1575                         inc_nlink(inode);
1576                         /* Some things misbehave if size == 0 on a directory */
1577                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
1578                         inode->i_op = &shmem_dir_inode_operations;
1579                         inode->i_fop = &simple_dir_operations;
1580                         break;
1581                 case S_IFLNK:
1582                         /*
1583                          * Must not load anything in the rbtree,
1584                          * mpol_free_shared_policy will not be called.
1585                          */
1586                         mpol_shared_policy_init(&info->policy, NULL);
1587                         break;
1588                 }
1589         } else
1590                 shmem_free_inode(sb);
1591         return inode;
1592 }
1593
1594 #ifdef CONFIG_TMPFS
1595 static const struct inode_operations shmem_symlink_inode_operations;
1596 static const struct inode_operations shmem_symlink_inline_operations;
1597
1598 static int
1599 shmem_write_begin(struct file *file, struct address_space *mapping,
1600                         loff_t pos, unsigned len, unsigned flags,
1601                         struct page **pagep, void **fsdata)
1602 {
1603         struct inode *inode = mapping->host;
1604         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1605         return shmem_getpage(inode, index, pagep, SGP_WRITE, NULL);
1606 }
1607
1608 static int
1609 shmem_write_end(struct file *file, struct address_space *mapping,
1610                         loff_t pos, unsigned len, unsigned copied,
1611                         struct page *page, void *fsdata)
1612 {
1613         struct inode *inode = mapping->host;
1614
1615         if (pos + copied > inode->i_size)
1616                 i_size_write(inode, pos + copied);
1617
1618         set_page_dirty(page);
1619         unlock_page(page);
1620         page_cache_release(page);
1621
1622         return copied;
1623 }
1624
1625 static void do_shmem_file_read(struct file *filp, loff_t *ppos, read_descriptor_t *desc, read_actor_t actor)
1626 {
1627         struct inode *inode = filp->f_path.dentry->d_inode;
1628         struct address_space *mapping = inode->i_mapping;
1629         unsigned long index, offset;
1630         enum sgp_type sgp = SGP_READ;
1631
1632         /*
1633          * Might this read be for a stacking filesystem?  Then when reading
1634          * holes of a sparse file, we actually need to allocate those pages,
1635          * and even mark them dirty, so it cannot exceed the max_blocks limit.
1636          */
1637         if (segment_eq(get_fs(), KERNEL_DS))
1638                 sgp = SGP_DIRTY;
1639
1640         index = *ppos >> PAGE_CACHE_SHIFT;
1641         offset = *ppos & ~PAGE_CACHE_MASK;
1642
1643         for (;;) {
1644                 struct page *page = NULL;
1645                 unsigned long end_index, nr, ret;
1646                 loff_t i_size = i_size_read(inode);
1647
1648                 end_index = i_size >> PAGE_CACHE_SHIFT;
1649                 if (index > end_index)
1650                         break;
1651                 if (index == end_index) {
1652                         nr = i_size & ~PAGE_CACHE_MASK;
1653                         if (nr <= offset)
1654                                 break;
1655                 }
1656
1657                 desc->error = shmem_getpage(inode, index, &page, sgp, NULL);
1658                 if (desc->error) {
1659                         if (desc->error == -EINVAL)
1660                                 desc->error = 0;
1661                         break;
1662                 }
1663                 if (page)
1664                         unlock_page(page);
1665
1666                 /*
1667                  * We must evaluate after, since reads (unlike writes)
1668                  * are called without i_mutex protection against truncate
1669                  */
1670                 nr = PAGE_CACHE_SIZE;
1671                 i_size = i_size_read(inode);
1672                 end_index = i_size >> PAGE_CACHE_SHIFT;
1673                 if (index == end_index) {
1674                         nr = i_size & ~PAGE_CACHE_MASK;
1675                         if (nr <= offset) {
1676                                 if (page)
1677                                         page_cache_release(page);
1678                                 break;
1679                         }
1680                 }
1681                 nr -= offset;
1682
1683                 if (page) {
1684                         /*
1685                          * If users can be writing to this page using arbitrary
1686                          * virtual addresses, take care about potential aliasing
1687                          * before reading the page on the kernel side.
1688                          */
1689                         if (mapping_writably_mapped(mapping))
1690                                 flush_dcache_page(page);
1691                         /*
1692                          * Mark the page accessed if we read the beginning.
1693                          */
1694                         if (!offset)
1695                                 mark_page_accessed(page);
1696                 } else {
1697                         page = ZERO_PAGE(0);
1698                         page_cache_get(page);
1699                 }
1700
1701                 /*
1702                  * Ok, we have the page, and it's up-to-date, so
1703                  * now we can copy it to user space...
1704                  *
1705                  * The actor routine returns how many bytes were actually used..
1706                  * NOTE! This may not be the same as how much of a user buffer
1707                  * we filled up (we may be padding etc), so we can only update
1708                  * "pos" here (the actor routine has to update the user buffer
1709                  * pointers and the remaining count).
1710                  */
1711                 ret = actor(desc, page, offset, nr);
1712                 offset += ret;
1713                 index += offset >> PAGE_CACHE_SHIFT;
1714                 offset &= ~PAGE_CACHE_MASK;
1715
1716                 page_cache_release(page);
1717                 if (ret != nr || !desc->count)
1718                         break;
1719
1720                 cond_resched();
1721         }
1722
1723         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1724         file_accessed(filp);
1725 }
1726
1727 static ssize_t shmem_file_aio_read(struct kiocb *iocb,
1728                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
1729 {
1730         struct file *filp = iocb->ki_filp;
1731         ssize_t retval;
1732         unsigned long seg;
1733         size_t count;
1734         loff_t *ppos = &iocb->ki_pos;
1735
1736         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1737         if (retval)
1738                 return retval;
1739
1740         for (seg = 0; seg < nr_segs; seg++) {
1741                 read_descriptor_t desc;
1742
1743                 desc.written = 0;
1744                 desc.arg.buf = iov[seg].iov_base;
1745                 desc.count = iov[seg].iov_len;
1746                 if (desc.count == 0)
1747                         continue;
1748                 desc.error = 0;
1749                 do_shmem_file_read(filp, ppos, &desc, file_read_actor);
1750                 retval += desc.written;
1751                 if (desc.error) {
1752                         retval = retval ?: desc.error;
1753                         break;
1754                 }
1755                 if (desc.count > 0)
1756                         break;
1757         }
1758         return retval;
1759 }
1760
1761 static ssize_t shmem_file_splice_read(struct file *in, loff_t *ppos,
1762                                 struct pipe_inode_info *pipe, size_t len,
1763                                 unsigned int flags)
1764 {
1765         struct address_space *mapping = in->f_mapping;
1766         struct inode *inode = mapping->host;
1767         unsigned int loff, nr_pages, req_pages;
1768         struct page *pages[PIPE_DEF_BUFFERS];
1769         struct partial_page partial[PIPE_DEF_BUFFERS];
1770         struct page *page;
1771         pgoff_t index, end_index;
1772         loff_t isize, left;
1773         int error, page_nr;
1774         struct splice_pipe_desc spd = {
1775                 .pages = pages,
1776                 .partial = partial,
1777                 .flags = flags,
1778                 .ops = &page_cache_pipe_buf_ops,
1779                 .spd_release = spd_release_page,
1780         };
1781
1782         isize = i_size_read(inode);
1783         if (unlikely(*ppos >= isize))
1784                 return 0;
1785
1786         left = isize - *ppos;
1787         if (unlikely(left < len))
1788                 len = left;
1789
1790         if (splice_grow_spd(pipe, &spd))
1791                 return -ENOMEM;
1792
1793         index = *ppos >> PAGE_CACHE_SHIFT;
1794         loff = *ppos & ~PAGE_CACHE_MASK;
1795         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1796         nr_pages = min(req_pages, pipe->buffers);
1797
1798         spd.nr_pages = find_get_pages_contig(mapping, index,
1799                                                 nr_pages, spd.pages);
1800         index += spd.nr_pages;
1801         error = 0;
1802
1803         while (spd.nr_pages < nr_pages) {
1804                 error = shmem_getpage(inode, index, &page, SGP_CACHE, NULL);
1805                 if (error)
1806                         break;
1807                 unlock_page(page);
1808                 spd.pages[spd.nr_pages++] = page;
1809                 index++;
1810         }
1811
1812         index = *ppos >> PAGE_CACHE_SHIFT;
1813         nr_pages = spd.nr_pages;
1814         spd.nr_pages = 0;
1815
1816         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
1817                 unsigned int this_len;
1818
1819                 if (!len)
1820                         break;
1821
1822                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
1823                 page = spd.pages[page_nr];
1824
1825                 if (!PageUptodate(page) || page->mapping != mapping) {
1826                         error = shmem_getpage(inode, index, &page,
1827                                                         SGP_CACHE, NULL);
1828                         if (error)
1829                                 break;
1830                         unlock_page(page);
1831                         page_cache_release(spd.pages[page_nr]);
1832                         spd.pages[page_nr] = page;
1833                 }
1834
1835                 isize = i_size_read(inode);
1836                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1837                 if (unlikely(!isize || index > end_index))
1838                         break;
1839
1840                 if (end_index == index) {
1841                         unsigned int plen;
1842
1843                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1844                         if (plen <= loff)
1845                                 break;
1846
1847                         this_len = min(this_len, plen - loff);
1848                         len = this_len;
1849                 }
1850
1851                 spd.partial[page_nr].offset = loff;
1852                 spd.partial[page_nr].len = this_len;
1853                 len -= this_len;
1854                 loff = 0;
1855                 spd.nr_pages++;
1856                 index++;
1857         }
1858
1859         while (page_nr < nr_pages)
1860                 page_cache_release(spd.pages[page_nr++]);
1861
1862         if (spd.nr_pages)
1863                 error = splice_to_pipe(pipe, &spd);
1864
1865         splice_shrink_spd(pipe, &spd);
1866
1867         if (error > 0) {
1868                 *ppos += error;
1869                 file_accessed(in);
1870         }
1871         return error;
1872 }
1873
1874 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
1875 {
1876         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
1877
1878         buf->f_type = TMPFS_MAGIC;
1879         buf->f_bsize = PAGE_CACHE_SIZE;
1880         buf->f_namelen = NAME_MAX;
1881         if (sbinfo->max_blocks) {
1882                 buf->f_blocks = sbinfo->max_blocks;
1883                 buf->f_bavail = buf->f_bfree =
1884                                 sbinfo->max_blocks - percpu_counter_sum(&sbinfo->used_blocks);
1885         }
1886         if (sbinfo->max_inodes) {
1887                 buf->f_files = sbinfo->max_inodes;
1888                 buf->f_ffree = sbinfo->free_inodes;
1889         }
1890         /* else leave those fields 0 like simple_statfs */
1891         return 0;
1892 }
1893
1894 /*
1895  * File creation. Allocate an inode, and we're done..
1896  */
1897 static int
1898 shmem_mknod(struct inode *dir, struct dentry *dentry, int mode, dev_t dev)
1899 {
1900         struct inode *inode;
1901         int error = -ENOSPC;
1902
1903         inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
1904         if (inode) {
1905                 error = security_inode_init_security(inode, dir,
1906                                                      &dentry->d_name, NULL,
1907                                                      NULL, NULL);
1908                 if (error) {
1909                         if (error != -EOPNOTSUPP) {
1910                                 iput(inode);
1911                                 return error;
1912                         }
1913                 }
1914 #ifdef CONFIG_TMPFS_POSIX_ACL
1915                 error = generic_acl_init(inode, dir);
1916                 if (error) {
1917                         iput(inode);
1918                         return error;
1919                 }
1920 #else
1921                 error = 0;
1922 #endif
1923                 dir->i_size += BOGO_DIRENT_SIZE;
1924                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1925                 d_instantiate(dentry, inode);
1926                 dget(dentry); /* Extra count - pin the dentry in core */
1927         }
1928         return error;
1929 }
1930
1931 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, int mode)
1932 {
1933         int error;
1934
1935         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
1936                 return error;
1937         inc_nlink(dir);
1938         return 0;
1939 }
1940
1941 static int shmem_create(struct inode *dir, struct dentry *dentry, int mode,
1942                 struct nameidata *nd)
1943 {
1944         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
1945 }
1946
1947 /*
1948  * Link a file..
1949  */
1950 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
1951 {
1952         struct inode *inode = old_dentry->d_inode;
1953         int ret;
1954
1955         /*
1956          * No ordinary (disk based) filesystem counts links as inodes;
1957          * but each new link needs a new dentry, pinning lowmem, and
1958          * tmpfs dentries cannot be pruned until they are unlinked.
1959          */
1960         ret = shmem_reserve_inode(inode->i_sb);
1961         if (ret)
1962                 goto out;
1963
1964         dir->i_size += BOGO_DIRENT_SIZE;
1965         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1966         inc_nlink(inode);
1967         ihold(inode);   /* New dentry reference */
1968         dget(dentry);           /* Extra pinning count for the created dentry */
1969         d_instantiate(dentry, inode);
1970 out:
1971         return ret;
1972 }
1973
1974 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
1975 {
1976         struct inode *inode = dentry->d_inode;
1977
1978         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
1979                 shmem_free_inode(inode->i_sb);
1980
1981         dir->i_size -= BOGO_DIRENT_SIZE;
1982         inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
1983         drop_nlink(inode);
1984         dput(dentry);   /* Undo the count from "create" - this does all the work */
1985         return 0;
1986 }
1987
1988 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
1989 {
1990         if (!simple_empty(dentry))
1991                 return -ENOTEMPTY;
1992
1993         drop_nlink(dentry->d_inode);
1994         drop_nlink(dir);
1995         return shmem_unlink(dir, dentry);
1996 }
1997
1998 /*
1999  * The VFS layer already does all the dentry stuff for rename,
2000  * we just have to decrement the usage count for the target if
2001  * it exists so that the VFS layer correctly free's it when it
2002  * gets overwritten.
2003  */
2004 static int shmem_rename(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
2005 {
2006         struct inode *inode = old_dentry->d_inode;
2007         int they_are_dirs = S_ISDIR(inode->i_mode);
2008
2009         if (!simple_empty(new_dentry))
2010                 return -ENOTEMPTY;
2011
2012         if (new_dentry->d_inode) {
2013                 (void) shmem_unlink(new_dir, new_dentry);
2014                 if (they_are_dirs)
2015                         drop_nlink(old_dir);
2016         } else if (they_are_dirs) {
2017                 drop_nlink(old_dir);
2018                 inc_nlink(new_dir);
2019         }
2020
2021         old_dir->i_size -= BOGO_DIRENT_SIZE;
2022         new_dir->i_size += BOGO_DIRENT_SIZE;
2023         old_dir->i_ctime = old_dir->i_mtime =
2024         new_dir->i_ctime = new_dir->i_mtime =
2025         inode->i_ctime = CURRENT_TIME;
2026         return 0;
2027 }
2028
2029 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
2030 {
2031         int error;
2032         int len;
2033         struct inode *inode;
2034         struct page *page;
2035         char *kaddr;
2036         struct shmem_inode_info *info;
2037
2038         len = strlen(symname) + 1;
2039         if (len > PAGE_CACHE_SIZE)
2040                 return -ENAMETOOLONG;
2041
2042         inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0, VM_NORESERVE);
2043         if (!inode)
2044                 return -ENOSPC;
2045
2046         error = security_inode_init_security(inode, dir, &dentry->d_name, NULL,
2047                                              NULL, NULL);
2048         if (error) {
2049                 if (error != -EOPNOTSUPP) {
2050                         iput(inode);
2051                         return error;
2052                 }
2053                 error = 0;
2054         }
2055
2056         info = SHMEM_I(inode);
2057         inode->i_size = len-1;
2058         if (len <= SHMEM_SYMLINK_INLINE_LEN) {
2059                 /* do it inline */
2060                 memcpy(info->inline_symlink, symname, len);
2061                 inode->i_op = &shmem_symlink_inline_operations;
2062         } else {
2063                 error = shmem_getpage(inode, 0, &page, SGP_WRITE, NULL);
2064                 if (error) {
2065                         iput(inode);
2066                         return error;
2067                 }
2068                 inode->i_mapping->a_ops = &shmem_aops;
2069                 inode->i_op = &shmem_symlink_inode_operations;
2070                 kaddr = kmap_atomic(page, KM_USER0);
2071                 memcpy(kaddr, symname, len);
2072                 kunmap_atomic(kaddr, KM_USER0);
2073                 set_page_dirty(page);
2074                 unlock_page(page);
2075                 page_cache_release(page);
2076         }
2077         dir->i_size += BOGO_DIRENT_SIZE;
2078         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
2079         d_instantiate(dentry, inode);
2080         dget(dentry);
2081         return 0;
2082 }
2083
2084 static void *shmem_follow_link_inline(struct dentry *dentry, struct nameidata *nd)
2085 {
2086         nd_set_link(nd, SHMEM_I(dentry->d_inode)->inline_symlink);
2087         return NULL;
2088 }
2089
2090 static void *shmem_follow_link(struct dentry *dentry, struct nameidata *nd)
2091 {
2092         struct page *page = NULL;
2093         int res = shmem_getpage(dentry->d_inode, 0, &page, SGP_READ, NULL);
2094         nd_set_link(nd, res ? ERR_PTR(res) : kmap(page));
2095         if (page)
2096                 unlock_page(page);
2097         return page;
2098 }
2099
2100 static void shmem_put_link(struct dentry *dentry, struct nameidata *nd, void *cookie)
2101 {
2102         if (!IS_ERR(nd_get_link(nd))) {
2103                 struct page *page = cookie;
2104                 kunmap(page);
2105                 mark_page_accessed(page);
2106                 page_cache_release(page);
2107         }
2108 }
2109
2110 #ifdef CONFIG_TMPFS_XATTR
2111 /*
2112  * Superblocks without xattr inode operations may get some security.* xattr
2113  * support from the LSM "for free". As soon as we have any other xattrs
2114  * like ACLs, we also need to implement the security.* handlers at
2115  * filesystem level, though.
2116  */
2117
2118 static int shmem_xattr_get(struct dentry *dentry, const char *name,
2119                            void *buffer, size_t size)
2120 {
2121         struct shmem_inode_info *info;
2122         struct shmem_xattr *xattr;
2123         int ret = -ENODATA;
2124
2125         info = SHMEM_I(dentry->d_inode);
2126
2127         spin_lock(&info->lock);
2128         list_for_each_entry(xattr, &info->xattr_list, list) {
2129                 if (strcmp(name, xattr->name))
2130                         continue;
2131
2132                 ret = xattr->size;
2133                 if (buffer) {
2134                         if (size < xattr->size)
2135                                 ret = -ERANGE;
2136                         else
2137                                 memcpy(buffer, xattr->value, xattr->size);
2138                 }
2139                 break;
2140         }
2141         spin_unlock(&info->lock);
2142         return ret;
2143 }
2144
2145 static int shmem_xattr_set(struct dentry *dentry, const char *name,
2146                            const void *value, size_t size, int flags)
2147 {
2148         struct inode *inode = dentry->d_inode;
2149         struct shmem_inode_info *info = SHMEM_I(inode);
2150         struct shmem_xattr *xattr;
2151         struct shmem_xattr *new_xattr = NULL;
2152         size_t len;
2153         int err = 0;
2154
2155         /* value == NULL means remove */
2156         if (value) {
2157                 /* wrap around? */
2158                 len = sizeof(*new_xattr) + size;
2159                 if (len <= sizeof(*new_xattr))
2160                         return -ENOMEM;
2161
2162                 new_xattr = kmalloc(len, GFP_KERNEL);
2163                 if (!new_xattr)
2164                         return -ENOMEM;
2165
2166                 new_xattr->name = kstrdup(name, GFP_KERNEL);
2167                 if (!new_xattr->name) {
2168                         kfree(new_xattr);
2169                         return -ENOMEM;
2170                 }
2171
2172                 new_xattr->size = size;
2173                 memcpy(new_xattr->value, value, size);
2174         }
2175
2176         spin_lock(&info->lock);
2177         list_for_each_entry(xattr, &info->xattr_list, list) {
2178                 if (!strcmp(name, xattr->name)) {
2179                         if (flags & XATTR_CREATE) {
2180                                 xattr = new_xattr;
2181                                 err = -EEXIST;
2182                         } else if (new_xattr) {
2183                                 list_replace(&xattr->list, &new_xattr->list);
2184                         } else {
2185                                 list_del(&xattr->list);
2186                         }
2187                         goto out;
2188                 }
2189         }
2190         if (flags & XATTR_REPLACE) {
2191                 xattr = new_xattr;
2192                 err = -ENODATA;
2193         } else {
2194                 list_add(&new_xattr->list, &info->xattr_list);
2195                 xattr = NULL;
2196         }
2197 out:
2198         spin_unlock(&info->lock);
2199         if (xattr)
2200                 kfree(xattr->name);
2201         kfree(xattr);
2202         return err;
2203 }
2204
2205
2206 static const struct xattr_handler *shmem_xattr_handlers[] = {
2207 #ifdef CONFIG_TMPFS_POSIX_ACL
2208         &generic_acl_access_handler,
2209         &generic_acl_default_handler,
2210 #endif
2211         NULL
2212 };
2213
2214 static int shmem_xattr_validate(const char *name)
2215 {
2216         struct { const char *prefix; size_t len; } arr[] = {
2217                 { XATTR_SECURITY_PREFIX, XATTR_SECURITY_PREFIX_LEN },
2218                 { XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN }
2219         };
2220         int i;
2221
2222         for (i = 0; i < ARRAY_SIZE(arr); i++) {
2223                 size_t preflen = arr[i].len;
2224                 if (strncmp(name, arr[i].prefix, preflen) == 0) {
2225                         if (!name[preflen])
2226                                 return -EINVAL;
2227                         return 0;
2228                 }
2229         }
2230         return -EOPNOTSUPP;
2231 }
2232
2233 static ssize_t shmem_getxattr(struct dentry *dentry, const char *name,
2234                               void *buffer, size_t size)
2235 {
2236         int err;
2237
2238         /*
2239          * If this is a request for a synthetic attribute in the system.*
2240          * namespace use the generic infrastructure to resolve a handler
2241          * for it via sb->s_xattr.
2242          */
2243         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2244                 return generic_getxattr(dentry, name, buffer, size);
2245
2246         err = shmem_xattr_validate(name);
2247         if (err)
2248                 return err;
2249
2250         return shmem_xattr_get(dentry, name, buffer, size);
2251 }
2252
2253 static int shmem_setxattr(struct dentry *dentry, const char *name,
2254                           const void *value, size_t size, int flags)
2255 {
2256         int err;
2257
2258         /*
2259          * If this is a request for a synthetic attribute in the system.*
2260          * namespace use the generic infrastructure to resolve a handler
2261          * for it via sb->s_xattr.
2262          */
2263         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2264                 return generic_setxattr(dentry, name, value, size, flags);
2265
2266         err = shmem_xattr_validate(name);
2267         if (err)
2268                 return err;
2269
2270         if (size == 0)
2271                 value = "";  /* empty EA, do not remove */
2272
2273         return shmem_xattr_set(dentry, name, value, size, flags);
2274
2275 }
2276
2277 static int shmem_removexattr(struct dentry *dentry, const char *name)
2278 {
2279         int err;
2280
2281         /*
2282          * If this is a request for a synthetic attribute in the system.*
2283          * namespace use the generic infrastructure to resolve a handler
2284          * for it via sb->s_xattr.
2285          */
2286         if (!strncmp(name, XATTR_SYSTEM_PREFIX, XATTR_SYSTEM_PREFIX_LEN))
2287                 return generic_removexattr(dentry, name);
2288
2289         err = shmem_xattr_validate(name);
2290         if (err)
2291                 return err;
2292
2293         return shmem_xattr_set(dentry, name, NULL, 0, XATTR_REPLACE);
2294 }
2295
2296 static bool xattr_is_trusted(const char *name)
2297 {
2298         return !strncmp(name, XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN);
2299 }
2300
2301 static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
2302 {
2303         bool trusted = capable(CAP_SYS_ADMIN);
2304         struct shmem_xattr *xattr;
2305         struct shmem_inode_info *info;
2306         size_t used = 0;
2307
2308         info = SHMEM_I(dentry->d_inode);
2309
2310         spin_lock(&info->lock);
2311         list_for_each_entry(xattr, &info->xattr_list, list) {
2312                 size_t len;
2313
2314                 /* skip "trusted." attributes for unprivileged callers */
2315                 if (!trusted && xattr_is_trusted(xattr->name))
2316                         continue;
2317
2318                 len = strlen(xattr->name) + 1;
2319                 used += len;
2320                 if (buffer) {
2321                         if (size < used) {
2322                                 used = -ERANGE;
2323                                 break;
2324                         }
2325                         memcpy(buffer, xattr->name, len);
2326                         buffer += len;
2327                 }
2328         }
2329         spin_unlock(&info->lock);
2330
2331         return used;
2332 }
2333 #endif /* CONFIG_TMPFS_XATTR */
2334
2335 static const struct inode_operations shmem_symlink_inline_operations = {
2336         .readlink       = generic_readlink,
2337         .follow_link    = shmem_follow_link_inline,
2338 #ifdef CONFIG_TMPFS_XATTR
2339         .setxattr       = shmem_setxattr,
2340         .getxattr       = shmem_getxattr,
2341         .listxattr      = shmem_listxattr,
2342         .removexattr    = shmem_removexattr,
2343 #endif
2344 };
2345
2346 static const struct inode_operations shmem_symlink_inode_operations = {
2347         .readlink       = generic_readlink,
2348         .follow_link    = shmem_follow_link,
2349         .put_link       = shmem_put_link,
2350 #ifdef CONFIG_TMPFS_XATTR
2351         .setxattr       = shmem_setxattr,
2352         .getxattr       = shmem_getxattr,
2353         .listxattr      = shmem_listxattr,
2354         .removexattr    = shmem_removexattr,
2355 #endif
2356 };
2357
2358 static struct dentry *shmem_get_parent(struct dentry *child)
2359 {
2360         return ERR_PTR(-ESTALE);
2361 }
2362
2363 static int shmem_match(struct inode *ino, void *vfh)
2364 {
2365         __u32 *fh = vfh;
2366         __u64 inum = fh[2];
2367         inum = (inum << 32) | fh[1];
2368         return ino->i_ino == inum && fh[0] == ino->i_generation;
2369 }
2370
2371 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
2372                 struct fid *fid, int fh_len, int fh_type)
2373 {
2374         struct inode *inode;
2375         struct dentry *dentry = NULL;
2376         u64 inum = fid->raw[2];
2377         inum = (inum << 32) | fid->raw[1];
2378
2379         if (fh_len < 3)
2380                 return NULL;
2381
2382         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
2383                         shmem_match, fid->raw);
2384         if (inode) {
2385                 dentry = d_find_alias(inode);
2386                 iput(inode);
2387         }
2388
2389         return dentry;
2390 }
2391
2392 static int shmem_encode_fh(struct dentry *dentry, __u32 *fh, int *len,
2393                                 int connectable)
2394 {
2395         struct inode *inode = dentry->d_inode;
2396
2397         if (*len < 3) {
2398                 *len = 3;
2399                 return 255;
2400         }
2401
2402         if (inode_unhashed(inode)) {
2403                 /* Unfortunately insert_inode_hash is not idempotent,
2404                  * so as we hash inodes here rather than at creation
2405                  * time, we need a lock to ensure we only try
2406                  * to do it once
2407                  */
2408                 static DEFINE_SPINLOCK(lock);
2409                 spin_lock(&lock);
2410                 if (inode_unhashed(inode))
2411                         __insert_inode_hash(inode,
2412                                             inode->i_ino + inode->i_generation);
2413                 spin_unlock(&lock);
2414         }
2415
2416         fh[0] = inode->i_generation;
2417         fh[1] = inode->i_ino;
2418         fh[2] = ((__u64)inode->i_ino) >> 32;
2419
2420         *len = 3;
2421         return 1;
2422 }
2423
2424 static const struct export_operations shmem_export_ops = {
2425         .get_parent     = shmem_get_parent,
2426         .encode_fh      = shmem_encode_fh,
2427         .fh_to_dentry   = shmem_fh_to_dentry,
2428 };
2429
2430 static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
2431                                bool remount)
2432 {
2433         char *this_char, *value, *rest;
2434
2435         while (options != NULL) {
2436                 this_char = options;
2437                 for (;;) {
2438                         /*
2439                          * NUL-terminate this option: unfortunately,
2440                          * mount options form a comma-separated list,
2441                          * but mpol's nodelist may also contain commas.
2442                          */
2443                         options = strchr(options, ',');
2444                         if (options == NULL)
2445                                 break;
2446                         options++;
2447                         if (!isdigit(*options)) {
2448                                 options[-1] = '\0';
2449                                 break;
2450                         }
2451                 }
2452                 if (!*this_char)
2453                         continue;
2454                 if ((value = strchr(this_char,'=')) != NULL) {
2455                         *value++ = 0;
2456                 } else {
2457                         printk(KERN_ERR
2458                             "tmpfs: No value for mount option '%s'\n",
2459                             this_char);
2460                         return 1;
2461                 }
2462
2463                 if (!strcmp(this_char,"size")) {
2464                         unsigned long long size;
2465                         size = memparse(value,&rest);
2466                         if (*rest == '%') {
2467                                 size <<= PAGE_SHIFT;
2468                                 size *= totalram_pages;
2469                                 do_div(size, 100);
2470                                 rest++;
2471                         }
2472                         if (*rest)
2473                                 goto bad_val;
2474                         sbinfo->max_blocks =
2475                                 DIV_ROUND_UP(size, PAGE_CACHE_SIZE);
2476                 } else if (!strcmp(this_char,"nr_blocks")) {
2477                         sbinfo->max_blocks = memparse(value, &rest);
2478                         if (*rest)
2479                                 goto bad_val;
2480                 } else if (!strcmp(this_char,"nr_inodes")) {
2481                         sbinfo->max_inodes = memparse(value, &rest);
2482                         if (*rest)
2483                                 goto bad_val;
2484                 } else if (!strcmp(this_char,"mode")) {
2485                         if (remount)
2486                                 continue;
2487                         sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
2488                         if (*rest)
2489                                 goto bad_val;
2490                 } else if (!strcmp(this_char,"uid")) {
2491                         if (remount)
2492                                 continue;
2493                         sbinfo->uid = simple_strtoul(value, &rest, 0);
2494                         if (*rest)
2495                                 goto bad_val;
2496                 } else if (!strcmp(this_char,"gid")) {
2497                         if (remount)
2498                                 continue;
2499                         sbinfo->gid = simple_strtoul(value, &rest, 0);
2500                         if (*rest)
2501                                 goto bad_val;
2502                 } else if (!strcmp(this_char,"mpol")) {
2503                         if (mpol_parse_str(value, &sbinfo->mpol, 1))
2504                                 goto bad_val;
2505                 } else {
2506                         printk(KERN_ERR "tmpfs: Bad mount option %s\n",
2507                                this_char);
2508                         return 1;
2509                 }
2510         }
2511         return 0;
2512
2513 bad_val:
2514         printk(KERN_ERR "tmpfs: Bad value '%s' for mount option '%s'\n",
2515                value, this_char);
2516         return 1;
2517
2518 }
2519
2520 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
2521 {
2522         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2523         struct shmem_sb_info config = *sbinfo;
2524         unsigned long inodes;
2525         int error = -EINVAL;
2526
2527         if (shmem_parse_options(data, &config, true))
2528                 return error;
2529
2530         spin_lock(&sbinfo->stat_lock);
2531         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
2532         if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
2533                 goto out;
2534         if (config.max_inodes < inodes)
2535                 goto out;
2536         /*
2537          * Those tests also disallow limited->unlimited while any are in
2538          * use, so i_blocks will always be zero when max_blocks is zero;
2539          * but we must separately disallow unlimited->limited, because
2540          * in that case we have no record of how much is already in use.
2541          */
2542         if (config.max_blocks && !sbinfo->max_blocks)
2543                 goto out;
2544         if (config.max_inodes && !sbinfo->max_inodes)
2545                 goto out;
2546
2547         error = 0;
2548         sbinfo->max_blocks  = config.max_blocks;
2549         sbinfo->max_inodes  = config.max_inodes;
2550         sbinfo->free_inodes = config.max_inodes - inodes;
2551
2552         mpol_put(sbinfo->mpol);
2553         sbinfo->mpol        = config.mpol;      /* transfers initial ref */
2554 out:
2555         spin_unlock(&sbinfo->stat_lock);
2556         return error;
2557 }
2558
2559 static int shmem_show_options(struct seq_file *seq, struct vfsmount *vfs)
2560 {
2561         struct shmem_sb_info *sbinfo = SHMEM_SB(vfs->mnt_sb);
2562
2563         if (sbinfo->max_blocks != shmem_default_max_blocks())
2564                 seq_printf(seq, ",size=%luk",
2565                         sbinfo->max_blocks << (PAGE_CACHE_SHIFT - 10));
2566         if (sbinfo->max_inodes != shmem_default_max_inodes())
2567                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
2568         if (sbinfo->mode != (S_IRWXUGO | S_ISVTX))
2569                 seq_printf(seq, ",mode=%03o", sbinfo->mode);
2570         if (sbinfo->uid != 0)
2571                 seq_printf(seq, ",uid=%u", sbinfo->uid);
2572         if (sbinfo->gid != 0)
2573                 seq_printf(seq, ",gid=%u", sbinfo->gid);
2574         shmem_show_mpol(seq, sbinfo->mpol);
2575         return 0;
2576 }
2577 #endif /* CONFIG_TMPFS */
2578
2579 static void shmem_put_super(struct super_block *sb)
2580 {
2581         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2582
2583         percpu_counter_destroy(&sbinfo->used_blocks);
2584         kfree(sbinfo);
2585         sb->s_fs_info = NULL;
2586 }
2587
2588 int shmem_fill_super(struct super_block *sb, void *data, int silent)
2589 {
2590         struct inode *inode;
2591         struct dentry *root;
2592         struct shmem_sb_info *sbinfo;
2593         int err = -ENOMEM;
2594
2595         /* Round up to L1_CACHE_BYTES to resist false sharing */
2596         sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
2597                                 L1_CACHE_BYTES), GFP_KERNEL);
2598         if (!sbinfo)
2599                 return -ENOMEM;
2600
2601         sbinfo->mode = S_IRWXUGO | S_ISVTX;
2602         sbinfo->uid = current_fsuid();
2603         sbinfo->gid = current_fsgid();
2604         sb->s_fs_info = sbinfo;
2605
2606 #ifdef CONFIG_TMPFS
2607         /*
2608          * Per default we only allow half of the physical ram per
2609          * tmpfs instance, limiting inodes to one per page of lowmem;
2610          * but the internal instance is left unlimited.
2611          */
2612         if (!(sb->s_flags & MS_NOUSER)) {
2613                 sbinfo->max_blocks = shmem_default_max_blocks();
2614                 sbinfo->max_inodes = shmem_default_max_inodes();
2615                 if (shmem_parse_options(data, sbinfo, false)) {
2616                         err = -EINVAL;
2617                         goto failed;
2618                 }
2619         }
2620         sb->s_export_op = &shmem_export_ops;
2621 #else
2622         sb->s_flags |= MS_NOUSER;
2623 #endif
2624
2625         spin_lock_init(&sbinfo->stat_lock);
2626         if (percpu_counter_init(&sbinfo->used_blocks, 0))
2627                 goto failed;
2628         sbinfo->free_inodes = sbinfo->max_inodes;
2629
2630         sb->s_maxbytes = SHMEM_MAX_BYTES;
2631         sb->s_blocksize = PAGE_CACHE_SIZE;
2632         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
2633         sb->s_magic = TMPFS_MAGIC;
2634         sb->s_op = &shmem_ops;
2635         sb->s_time_gran = 1;
2636 #ifdef CONFIG_TMPFS_XATTR
2637         sb->s_xattr = shmem_xattr_handlers;
2638 #endif
2639 #ifdef CONFIG_TMPFS_POSIX_ACL
2640         sb->s_flags |= MS_POSIXACL;
2641 #endif
2642
2643         inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
2644         if (!inode)
2645                 goto failed;
2646         inode->i_uid = sbinfo->uid;
2647         inode->i_gid = sbinfo->gid;
2648         root = d_alloc_root(inode);
2649         if (!root)
2650                 goto failed_iput;
2651         sb->s_root = root;
2652         return 0;
2653
2654 failed_iput:
2655         iput(inode);
2656 failed:
2657         shmem_put_super(sb);
2658         return err;
2659 }
2660
2661 static struct kmem_cache *shmem_inode_cachep;
2662
2663 static struct inode *shmem_alloc_inode(struct super_block *sb)
2664 {
2665         struct shmem_inode_info *p;
2666         p = (struct shmem_inode_info *)kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
2667         if (!p)
2668                 return NULL;
2669         return &p->vfs_inode;
2670 }
2671
2672 static void shmem_i_callback(struct rcu_head *head)
2673 {
2674         struct inode *inode = container_of(head, struct inode, i_rcu);
2675         INIT_LIST_HEAD(&inode->i_dentry);
2676         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
2677 }
2678
2679 static void shmem_destroy_inode(struct inode *inode)
2680 {
2681         if ((inode->i_mode & S_IFMT) == S_IFREG) {
2682                 /* only struct inode is valid if it's an inline symlink */
2683                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
2684         }
2685         call_rcu(&inode->i_rcu, shmem_i_callback);
2686 }
2687
2688 static void init_once(void *foo)
2689 {
2690         struct shmem_inode_info *p = (struct shmem_inode_info *) foo;
2691
2692         inode_init_once(&p->vfs_inode);
2693 }
2694
2695 static int init_inodecache(void)
2696 {
2697         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
2698                                 sizeof(struct shmem_inode_info),
2699                                 0, SLAB_PANIC, init_once);
2700         return 0;
2701 }
2702
2703 static void destroy_inodecache(void)
2704 {
2705         kmem_cache_destroy(shmem_inode_cachep);
2706 }
2707
2708 static const struct address_space_operations shmem_aops = {
2709         .writepage      = shmem_writepage,
2710         .set_page_dirty = __set_page_dirty_no_writeback,
2711 #ifdef CONFIG_TMPFS
2712         .write_begin    = shmem_write_begin,
2713         .write_end      = shmem_write_end,
2714 #endif
2715         .migratepage    = migrate_page,
2716         .error_remove_page = generic_error_remove_page,
2717 };
2718
2719 static const struct file_operations shmem_file_operations = {
2720         .mmap           = shmem_mmap,
2721 #ifdef CONFIG_TMPFS
2722         .llseek         = generic_file_llseek,
2723         .read           = do_sync_read,
2724         .write          = do_sync_write,
2725         .aio_read       = shmem_file_aio_read,
2726         .aio_write      = generic_file_aio_write,
2727         .fsync          = noop_fsync,
2728         .splice_read    = shmem_file_splice_read,
2729         .splice_write   = generic_file_splice_write,
2730 #endif
2731 };
2732
2733 static const struct inode_operations shmem_inode_operations = {
2734         .setattr        = shmem_setattr,
2735         .truncate_range = shmem_truncate_range,
2736 #ifdef CONFIG_TMPFS_XATTR
2737         .setxattr       = shmem_setxattr,
2738         .getxattr       = shmem_getxattr,
2739         .listxattr      = shmem_listxattr,
2740         .removexattr    = shmem_removexattr,
2741 #endif
2742 };
2743
2744 static const struct inode_operations shmem_dir_inode_operations = {
2745 #ifdef CONFIG_TMPFS
2746         .create         = shmem_create,
2747         .lookup         = simple_lookup,
2748         .link           = shmem_link,
2749         .unlink         = shmem_unlink,
2750         .symlink        = shmem_symlink,
2751         .mkdir          = shmem_mkdir,
2752         .rmdir          = shmem_rmdir,
2753         .mknod          = shmem_mknod,
2754         .rename         = shmem_rename,
2755 #endif
2756 #ifdef CONFIG_TMPFS_XATTR
2757         .setxattr       = shmem_setxattr,
2758         .getxattr       = shmem_getxattr,
2759         .listxattr      = shmem_listxattr,
2760         .removexattr    = shmem_removexattr,
2761 #endif
2762 #ifdef CONFIG_TMPFS_POSIX_ACL
2763         .setattr        = shmem_setattr,
2764 #endif
2765 };
2766
2767 static const struct inode_operations shmem_special_inode_operations = {
2768 #ifdef CONFIG_TMPFS_XATTR
2769         .setxattr       = shmem_setxattr,
2770         .getxattr       = shmem_getxattr,
2771         .listxattr      = shmem_listxattr,
2772         .removexattr    = shmem_removexattr,
2773 #endif
2774 #ifdef CONFIG_TMPFS_POSIX_ACL
2775         .setattr        = shmem_setattr,
2776 #endif
2777 };
2778
2779 static const struct super_operations shmem_ops = {
2780         .alloc_inode    = shmem_alloc_inode,
2781         .destroy_inode  = shmem_destroy_inode,
2782 #ifdef CONFIG_TMPFS
2783         .statfs         = shmem_statfs,
2784         .remount_fs     = shmem_remount_fs,
2785         .show_options   = shmem_show_options,
2786 #endif
2787         .evict_inode    = shmem_evict_inode,
2788         .drop_inode     = generic_delete_inode,
2789         .put_super      = shmem_put_super,
2790 };
2791
2792 static const struct vm_operations_struct shmem_vm_ops = {
2793         .fault          = shmem_fault,
2794 #ifdef CONFIG_NUMA
2795         .set_policy     = shmem_set_policy,
2796         .get_policy     = shmem_get_policy,
2797 #endif
2798 };
2799
2800
2801 static struct dentry *shmem_mount(struct file_system_type *fs_type,
2802         int flags, const char *dev_name, void *data)
2803 {
2804         return mount_nodev(fs_type, flags, data, shmem_fill_super);
2805 }
2806
2807 static struct file_system_type tmpfs_fs_type = {
2808         .owner          = THIS_MODULE,
2809         .name           = "tmpfs",
2810         .mount          = shmem_mount,
2811         .kill_sb        = kill_litter_super,
2812 };
2813
2814 int __init init_tmpfs(void)
2815 {
2816         int error;
2817
2818         error = bdi_init(&shmem_backing_dev_info);
2819         if (error)
2820                 goto out4;
2821
2822         error = init_inodecache();
2823         if (error)
2824                 goto out3;
2825
2826         error = register_filesystem(&tmpfs_fs_type);
2827         if (error) {
2828                 printk(KERN_ERR "Could not register tmpfs\n");
2829                 goto out2;
2830         }
2831
2832         shm_mnt = vfs_kern_mount(&tmpfs_fs_type, MS_NOUSER,
2833                                 tmpfs_fs_type.name, NULL);
2834         if (IS_ERR(shm_mnt)) {
2835                 error = PTR_ERR(shm_mnt);
2836                 printk(KERN_ERR "Could not kern_mount tmpfs\n");
2837                 goto out1;
2838         }
2839         return 0;
2840
2841 out1:
2842         unregister_filesystem(&tmpfs_fs_type);
2843 out2:
2844         destroy_inodecache();
2845 out3:
2846         bdi_destroy(&shmem_backing_dev_info);
2847 out4:
2848         shm_mnt = ERR_PTR(error);
2849         return error;
2850 }
2851
2852 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2853 /**
2854  * mem_cgroup_get_shmem_target - find a page or entry assigned to the shmem file
2855  * @inode: the inode to be searched
2856  * @pgoff: the offset to be searched
2857  * @pagep: the pointer for the found page to be stored
2858  * @ent: the pointer for the found swap entry to be stored
2859  *
2860  * If a page is found, refcount of it is incremented. Callers should handle
2861  * these refcount.
2862  */
2863 void mem_cgroup_get_shmem_target(struct inode *inode, pgoff_t pgoff,
2864                                         struct page **pagep, swp_entry_t *ent)
2865 {
2866         swp_entry_t entry = { .val = 0 }, *ptr;
2867         struct page *page = NULL;
2868         struct shmem_inode_info *info = SHMEM_I(inode);
2869
2870         if ((pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
2871                 goto out;
2872
2873         spin_lock(&info->lock);
2874         ptr = shmem_swp_entry(info, pgoff, NULL);
2875 #ifdef CONFIG_SWAP
2876         if (ptr && ptr->val) {
2877                 entry.val = ptr->val;
2878                 page = find_get_page(&swapper_space, entry.val);
2879         } else
2880 #endif
2881                 page = find_get_page(inode->i_mapping, pgoff);
2882         if (ptr)
2883                 shmem_swp_unmap(ptr);
2884         spin_unlock(&info->lock);
2885 out:
2886         *pagep = page;
2887         *ent = entry;
2888 }
2889 #endif
2890
2891 #else /* !CONFIG_SHMEM */
2892
2893 /*
2894  * tiny-shmem: simple shmemfs and tmpfs using ramfs code
2895  *
2896  * This is intended for small system where the benefits of the full
2897  * shmem code (swap-backed and resource-limited) are outweighed by
2898  * their complexity. On systems without swap this code should be
2899  * effectively equivalent, but much lighter weight.
2900  */
2901
2902 #include <linux/ramfs.h>
2903
2904 static struct file_system_type tmpfs_fs_type = {
2905         .name           = "tmpfs",
2906         .mount          = ramfs_mount,
2907         .kill_sb        = kill_litter_super,
2908 };
2909
2910 int __init init_tmpfs(void)
2911 {
2912         BUG_ON(register_filesystem(&tmpfs_fs_type) != 0);
2913
2914         shm_mnt = kern_mount(&tmpfs_fs_type);
2915         BUG_ON(IS_ERR(shm_mnt));
2916
2917         return 0;
2918 }
2919
2920 int shmem_unuse(swp_entry_t entry, struct page *page)
2921 {
2922         return 0;
2923 }
2924
2925 int shmem_lock(struct file *file, int lock, struct user_struct *user)
2926 {
2927         return 0;
2928 }
2929
2930 void shmem_truncate_range(struct inode *inode, loff_t start, loff_t end)
2931 {
2932         truncate_inode_pages_range(inode->i_mapping, start, end);
2933 }
2934 EXPORT_SYMBOL_GPL(shmem_truncate_range);
2935
2936 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2937 /**
2938  * mem_cgroup_get_shmem_target - find a page or entry assigned to the shmem file
2939  * @inode: the inode to be searched
2940  * @pgoff: the offset to be searched
2941  * @pagep: the pointer for the found page to be stored
2942  * @ent: the pointer for the found swap entry to be stored
2943  *
2944  * If a page is found, refcount of it is incremented. Callers should handle
2945  * these refcount.
2946  */
2947 void mem_cgroup_get_shmem_target(struct inode *inode, pgoff_t pgoff,
2948                                         struct page **pagep, swp_entry_t *ent)
2949 {
2950         struct page *page = NULL;
2951
2952         if ((pgoff << PAGE_CACHE_SHIFT) >= i_size_read(inode))
2953                 goto out;
2954         page = find_get_page(inode->i_mapping, pgoff);
2955 out:
2956         *pagep = page;
2957         *ent = (swp_entry_t){ .val = 0 };
2958 }
2959 #endif
2960
2961 #define shmem_vm_ops                            generic_file_vm_ops
2962 #define shmem_file_operations                   ramfs_file_operations
2963 #define shmem_get_inode(sb, dir, mode, dev, flags)      ramfs_get_inode(sb, dir, mode, dev)
2964 #define shmem_acct_size(flags, size)            0
2965 #define shmem_unacct_size(flags, size)          do {} while (0)
2966 #define SHMEM_MAX_BYTES                         MAX_LFS_FILESIZE
2967
2968 #endif /* CONFIG_SHMEM */
2969
2970 /* common code */
2971
2972 /**
2973  * shmem_file_setup - get an unlinked file living in tmpfs
2974  * @name: name for dentry (to be seen in /proc/<pid>/maps
2975  * @size: size to be set for the file
2976  * @flags: VM_NORESERVE suppresses pre-accounting of the entire object size
2977  */
2978 struct file *shmem_file_setup(const char *name, loff_t size, unsigned long flags)
2979 {
2980         int error;
2981         struct file *file;
2982         struct inode *inode;
2983         struct path path;
2984         struct dentry *root;
2985         struct qstr this;
2986
2987         if (IS_ERR(shm_mnt))
2988                 return (void *)shm_mnt;
2989
2990         if (size < 0 || size > SHMEM_MAX_BYTES)
2991                 return ERR_PTR(-EINVAL);
2992
2993         if (shmem_acct_size(flags, size))
2994                 return ERR_PTR(-ENOMEM);
2995
2996         error = -ENOMEM;
2997         this.name = name;
2998         this.len = strlen(name);
2999         this.hash = 0; /* will go */
3000         root = shm_mnt->mnt_root;
3001         path.dentry = d_alloc(root, &this);
3002         if (!path.dentry)
3003                 goto put_memory;
3004         path.mnt = mntget(shm_mnt);
3005
3006         error = -ENOSPC;
3007         inode = shmem_get_inode(root->d_sb, NULL, S_IFREG | S_IRWXUGO, 0, flags);
3008         if (!inode)
3009                 goto put_dentry;
3010
3011         d_instantiate(path.dentry, inode);
3012         inode->i_size = size;
3013         inode->i_nlink = 0;     /* It is unlinked */
3014 #ifndef CONFIG_MMU
3015         error = ramfs_nommu_expand_for_mapping(inode, size);
3016         if (error)
3017                 goto put_dentry;
3018 #endif
3019
3020         error = -ENFILE;
3021         file = alloc_file(&path, FMODE_WRITE | FMODE_READ,
3022                   &shmem_file_operations);
3023         if (!file)
3024                 goto put_dentry;
3025
3026         return file;
3027
3028 put_dentry:
3029         path_put(&path);
3030 put_memory:
3031         shmem_unacct_size(flags, size);
3032         return ERR_PTR(error);
3033 }
3034 EXPORT_SYMBOL_GPL(shmem_file_setup);
3035
3036 /**
3037  * shmem_zero_setup - setup a shared anonymous mapping
3038  * @vma: the vma to be mmapped is prepared by do_mmap_pgoff
3039  */
3040 int shmem_zero_setup(struct vm_area_struct *vma)
3041 {
3042         struct file *file;
3043         loff_t size = vma->vm_end - vma->vm_start;
3044
3045         file = shmem_file_setup("dev/zero", size, vma->vm_flags);
3046         if (IS_ERR(file))
3047                 return PTR_ERR(file);
3048
3049         if (vma->vm_file)
3050                 fput(vma->vm_file);
3051         vma->vm_file = file;
3052         vma->vm_ops = &shmem_vm_ops;
3053         vma->vm_flags |= VM_CAN_NONLINEAR;
3054         return 0;
3055 }
3056
3057 /**
3058  * shmem_read_mapping_page_gfp - read into page cache, using specified page allocation flags.
3059  * @mapping:    the page's address_space
3060  * @index:      the page index
3061  * @gfp:        the page allocator flags to use if allocating
3062  *
3063  * This behaves as a tmpfs "read_cache_page_gfp(mapping, index, gfp)",
3064  * with any new page allocations done using the specified allocation flags.
3065  * But read_cache_page_gfp() uses the ->readpage() method: which does not
3066  * suit tmpfs, since it may have pages in swapcache, and needs to find those
3067  * for itself; although drivers/gpu/drm i915 and ttm rely upon this support.
3068  *
3069  * i915_gem_object_get_pages_gtt() mixes __GFP_NORETRY | __GFP_NOWARN in
3070  * with the mapping_gfp_mask(), to avoid OOMing the machine unnecessarily.
3071  */
3072 struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
3073                                          pgoff_t index, gfp_t gfp)
3074 {
3075 #ifdef CONFIG_SHMEM
3076         struct inode *inode = mapping->host;
3077         struct page *page;
3078         int error;
3079
3080         BUG_ON(mapping->a_ops != &shmem_aops);
3081         error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE, gfp, NULL);
3082         if (error)
3083                 page = ERR_PTR(error);
3084         else
3085                 unlock_page(page);
3086         return page;
3087 #else
3088         /*
3089          * The tiny !SHMEM case uses ramfs without swap
3090          */
3091         return read_cache_page_gfp(mapping, index, gfp);
3092 #endif
3093 }
3094 EXPORT_SYMBOL_GPL(shmem_read_mapping_page_gfp);